id
string | text
string | source
string | added
string | created
string | metadata
dict |
|---|---|---|---|---|---|
PMC545209
|
Streptococcus pneumoniae is a common bacterium that is present in the nasopharynx of many children and some adults, where it causes no harm to its carrier but can be transmitted to others. If it moves beyond the nasopharynx, however, it can cause ear infections or invasive disease, such as pneumonia or meningitis. Invasive disease from this organism occurs especially in children, the elderly, and individuals with weakened immune systems.
The protective effect of antibodies against bacterial pneumonia has been appreciated since the 1930s, when it was shown that serum therapy---the transfer of serum from an immunized animal to a patient with acute disease caused by the same bacterial strain---could reduce mortality from pneumococcal pneumonia by half. Subsequent development of vaccines based on the bacterium\'s polysaccharide capsule, which could protect against infection, confirmed that an endogenous antibody response can provide protection against invasive disease.[](#pmed-0020020-g001){ref-type="fig"}
::: {#pmed-0020020-g001 .fig}
::: {.caption}
###### Parallel age-incidence curves for pneumococcal serotypes suggest a common mechanism of protection
:::
:::
One challenge for vaccine development has been the existence of many different serotypes (the same species of bacteria but with different composition of the polysaccharide capsule). As protection usually doesn\'t extend to different serotypes, vaccination with capsule components from different serotypes is necessary to ensure broad protection. Such vaccines have been shown to be efficient and safe. They are now recommended in many countries for infants and toddlers, and for people over 65---the two age groups in which invasive disease is most common---and for others who are at increased risk of pneumococcal disease (e.g., patients with heart, kidney, liver, or lung disease, or who have had a splenectomy).
Even without vaccination, however, most exposed individuals will never get invasive disease. Instead, they develop natural immunity against the different serotypes, though this immunity gradually declines with old age. Marc Lipsitch and colleagues wanted to understand the immunological basis of this natural immunity, and specifically whether it was due to anticapsular antibodies.
If protection from invasive disease is due to acquiring anticapsular antibodies against each of the pneumococcal serotypes, they argued, this would lead to two predictions about the age distribution of disease caused by the different serotypes in the non-vaccinated population. First, for serotypes that are more common and therefore encountered earlier in life, children should develop immunity more quickly, causing disease from these types to drop off earlier in life than disease from the less common types. Second, protection against invasive disease from a particular serotype should coincide with the acquisition of antibodies against that serotype, on both the individual and population level.
Neither prediction was borne out by the actual data the researchers analyzed, suggesting that there is more to natural immunity against pneumococcal disease than just anticapsular antibodies. The study doesn\'t demonstrate what the additional components are, but additional research might not just teach us about our immune system but also provide clues for further vaccine development. As the authors say, "A better understanding of the mechanisms that underlie natural immunity to pneumococcus could pave the way for the development of more effective, species-specific pneumococcal vaccines."
|
PubMed Central
|
2024-06-05T03:55:51.720708
|
2005-1-25
|
{
"license": "Creative Commons Zero - Public Domain - https://creativecommons.org/publicdomain/zero/1.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545209/",
"journal": "PLoS Med. 2005 Jan 25; 2(1):e20",
"authors": []
}
|
PMC545210
|
Tyrosine kinases regulate signaling pathways that control cell growth, proliferation, motility, and other critical cellular processes. Mutations in tyrosine kinase genes can lead to abnormal kinase activity, and some tumors become dependent upon this activity for growth and survival. Thus, kinases are attractive targets for anti-cancer drugs. Examples of new kinase inhibitors include gefitinib and erlotinib, which have recently shown promise in treating non-small-cell lung cancer. Unfortunately, gefitinib and erlotinib work only in a subset of patients, and they can have severe side effects, albeit infrequently. So researchers have been trying to find ways to predict who will benefit from therapy with these drugs and who won\'t.[](#pmed-0020021-g001){ref-type="fig"}
::: {#pmed-0020021-g001 .fig}
::: {.caption}
###### Assessing lung tumors for gene mutations could help guide therapy
:::
:::
Following the work of Lynch et al. (N Engl J Med 350: 2129--2139) and Paez et al. (Science 304: 1497--1500), William Pao and colleagues have previously shown that the epidermal growth factor receptor (EGFR), a tyrosine kinase, is often mutated in non-small-cell lung cancers, and that tumors that harbor such mutations are sensitive to gefitinib and erlotinib.
In this new study, they focused on a signaling protein called KRAS, which functions downstream of many tyrosine kinases, including EGFR. The KRAS gene is also often mutated in lung cancers, but very few cancers have mutations in both EGFR and the KRAS gene. To find out whether KRAS mutations could help to predict which patients would respond to gefitinib or erlotinib, the researchers looked for mutations in EGFR and KRAS genes in 60 tumors for which sensitivity to either drug was known.
They extended their earlier findings that EGFR mutations (which were found in 17 of the tumors) were associated with sensitivity to the kinase inhibitors, and found that tumors that had mutations in KRAS (a total of nine) were refractory (i.e., did not respond) to either drug.
These results need to be validated in larger and prospective trials that use standardized mutation detection techniques. If they are confirmed, knowing the mutation status of EGFR and KRAS in tumors could help physicians decide which patients should receive gefitinib and/or erlotinib. As Inoue and Nukiwa state in a Perspective that accompanies the article, "By combining all the factors that relate to response or resistance, patients who will benefit from treatment can hopefully be identified. Undoubtedly we have taken a great step forward in molecular therapy for lung cancer treatment."
|
PubMed Central
|
2024-06-05T03:55:51.721260
|
2005-1-25
|
{
"license": "Creative Commons Zero - Public Domain - https://creativecommons.org/publicdomain/zero/1.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545210/",
"journal": "PLoS Med. 2005 Jan 25; 2(1):e21",
"authors": []
}
|
PMC545211
|
Many different things combine to cause epidemics of disease. Among these factors are the characteristics of the infecting organism, the resistance of the host, and, as is increasingly realized, climatic conditions.
El Niño, the best known climatic disturbance, is caused by a warming of the Pacific Ocean, which then affects the climate globally. Previous work has suggested that this recurring phenomenon can have a profound effect on the incidence of many diseases, including dengue, malaria, and diarrheal diseases.
In a paper in this month\'s *PLoS Medicine*, Sultan and colleagues from a climate research institute and an infectious diseases center in France looked at the relation between climate and meningitis outbreaks in Mali in West Africa, a region that every year between February and May sees devastating epidemics of meningococcal meningitis affecting up to 200,000 people. The most important recurring climatic event in this region is a dry wind, known as the Harmattan, that blows throughout the winter, causing a drop in humidity and the production of vast quantities of dust.
What the authors found was that over the years 1994--2002, the week of the onset of the yearly meningitis epidemic came at around the same time as the peak of one measure of the wind---the sixth week of the year. As Pascual and Dobson say in their Perspective article on this study, "Sultan and colleagues\' study is exceptional in that it illustrates a clear relationship between an external environmental variable and the initiation of disease outbreaks."
How do climatic changes influence disease? In some cases, such as the role of flooding in spreading a waterborne disease, the causes are perhaps obvious, but why should a dry wind affect disease incidence? Previous works have suggested that the climate can work in a number of ways, by influencing the life cycle of both disease vectors and the disease-causing organism, and, as here perhaps, by affecting the resistance of the host. Sultan and colleagues speculate that the drying effects of the wind on the mucous membranes could increase the chances of the organism getting established in the human host. Whatever the causes, one very useful feature of climate is that, once the patterns are understood, they can often be predicted.
A way of predicting these meningitis epidemics could be enormously useful. Sultan and colleagues looked at only a few years, but if these findings are confirmed over a longer time period, they could make preparing for an epidemic much more efficient.
|
PubMed Central
|
2024-06-05T03:55:51.721651
|
2005-1-25
|
{
"license": "Creative Commons Zero - Public Domain - https://creativecommons.org/publicdomain/zero/1.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545211/",
"journal": "PLoS Med. 2005 Jan 25; 2(1):e24",
"authors": []
}
|
PMC545212
|
The study by Planche et al. [@pmed-0020027-b1] provides important new information addressing intracellular volume depletion in children with severe childhood malaria, but does not address the question of whether intravascular volume depletion (hypovolemic shock) is present. Using sophisticated methodology to determine total body water and extracellular water, they demonstrate a 6.7% deficit in total body water and an 11.7% deficit of intracellular water, providing an important indication of the volumes of fluid that may be required to optimize hydration. The data, however, do not address the degree of filling of the intravascular compartment, nor should they be used to answer the question about the state of tissue and organ perfusion. Indeed, we believe that these new data present no conflict with our previously reported findings. Using methods to study critical illness physiology that are widely employed within pediatric intensive care units for interpretation of circulatory status, we have demonstrated evidence for hypovolemia in 53 Kenyan children with severe malaria complicated by metabolic acidosis [@pmed-0020027-b2]. Our children were younger, had longer capillary refilling times (\>3 s), lower central venous pressures (mean 2.9 cm H~2~O) and higher creatinines (\>80 µmol/l): all features of compensated hypovolemic shock. Furthermore, hypotension (systolic BP \< 80 mm Hg) was present in 44% of children with severe acidosis (base deficit \>15). These findings also indicate important baseline differences in two cohorts of children studied. We agree that reconsideration of guidelines for acute fluid management is warranted, particularly when current recommendations await an adequate evidence base. Nevertheless, conflicting opinions on the question of volume status in children with severe malaria can be satisfactorily resolved only through prospective randomized trials that include both fluid resuscitation and control groups. While the design and conduct of such trials will involve considerable challenges, optimal fluid management will never be resolved on the basis of theoretical consideration alone.
**Citation:** Maitland K (2005) Volume status in severe malaria: No evidence provided for the degree of filling of the intravascular compartment. PLoS Med 2(1): e27.
[^1]: **Competing Interests:** The authors declare that they have no competing interests.
|
PubMed Central
|
2024-06-05T03:55:51.722068
|
2005-1-25
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545212/",
"journal": "PLoS Med. 2005 Jan 25; 2(1):e27",
"authors": [
{
"first": "Kathryn",
"last": "Maitland"
},
{
"first": "Charles",
"last": "Newton"
},
{
"first": "Kevin",
"last": "Marsh"
},
{
"first": "Mike",
"last": "Levin"
}
]
}
|
PMC545213
|
Dr. Gerberding outlines critical steps for arresting the HIV/AIDS epidemic [@pmed-0020028-b1]. She suggests moving ahead with "ABCs" and with "D" for diagnosis and "R" for responsibility. These are good suggestions---with increased HIV testing and individuals taking responsibility for their role in HIV spread, the epidemic might be slowed. We could continue to add incrementally to the alphabet soup of public health. But instead, we could choose to immediately implement the mainstays of public health---universal testing and contact tracing \[[@pmed-0020028-b2],[@pmed-0020028-b3],[@pmed-0020028-b4]\]. Every sexually active individual and every individual at risk for HIV deserves to know their HIV status. Thus, every HIV-infected individual must be called upon to be accountable for preventing HIV transmission. Contact tracing should be instituted for HIV just as it is for other infectious diseases. Those who have been exposed to HIV have a right to know how to protect themselves and if they too are infected, to be offered treatment [@pmed-0020028-b5]. HIV testing has too often focused on testing of women in a perinatal setting rather than universal testing in routine clinical care. Without universal voluntary HIV testing and contact tracing, we will see the continued tilt of the epidemic toward women, now at 55% of all HIV infections and in all likelihood at 75%--80% in another 8 to 10 years \[[@pmed-0020028-b6],[@pmed-0020028-b7]\]. For too long the debate has been that contact tracing will result in physical abuse of women. Confining our definition of abuse of women to physical abuse alone is to have too narrow an ethical focus---HIV infection itself is an abuse of women or of anyone else. Universal HIV testing and contact tracing adds an essential comprehensive public health approach to the epidemic that will be successful in reducing the ever-escalating numbers of new infections.
**Citation:** Ammann A (2005) Completing the public health HIV/AIDS alphabet. PLoS Med 2(1): e28.
[^1]: **Competing Interests:** The author declares that he has no competing interests.
|
PubMed Central
|
2024-06-05T03:55:51.722560
|
2005-1-25
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545213/",
"journal": "PLoS Med. 2005 Jan 25; 2(1):e28",
"authors": [
{
"first": "Arthur",
"last": "Ammann"
}
]
}
|
PMC545214
|
I would like to share my experience from nearly a decade of civil war between the Maoist rebels and the Royal Nepalese Army in Nepal in reference to the article by Zwi [@pmed-0020029-b1] on the expanding role of health communities in times of conflict. The current war in Nepal has led to widespread destruction of limited infrastructure and has adversely impacted access to health-care services and personnel, affecting family planning, maternal and child health programs, and immunization services throughout the country. While Nepal is flooded with non-governmental organizations, paradoxically, humanitarian assistance may have unknowingly exacerbated the conflict by perpetuating the same inequalities that led to the conflict in the first place. This has brought to the fore the need for "conflict-sensitive development" [@pmed-0020029-b2]---development sensitive to the (conflict) environments in which they operate, in order to reduce the negative impacts of their activities---and to increase their positive impacts---on the situation and its dynamics. Development projects can continue in less affected areas with a need for transitional programs in conflict areas that can adapt to the rapidly changing environment. If agencies are unable to function, they have required the help of humanitarian agencies such as Médicins Sans Frontières with experience in conflict settings. Some agencies have adopted a participatory role in development and have involved neutral local agencies, increasing community participation in their projects with good success. But there is a need for increasing coordination between organizations working in various health-related projects. Health-care workers across the world in different conflicts are in a unique position to leverage something of universal importance---the promise of good health [@pmed-0020029-b3]. Raising awareness of the issues surrounding conflicts will act as a catalyst for change.
**Citation:** Singh S (2005) Nepal\'s war and conflict-sensitive development. PLoS Med 2(1): e29.
[^1]: **Competing Interests:** I declare that I have no competing interests.
|
PubMed Central
|
2024-06-05T03:55:51.723128
|
2005-1-25
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545214/",
"journal": "PLoS Med. 2005 Jan 25; 2(1):e29",
"authors": [
{
"first": "Sonal",
"last": "Singh"
}
]
}
|
PMC545215
|
After reading the Learning Forum by Fleming and Lynn [@pmed-0020030-b1], I would like to suggest three learning points that, in my opinion, should receive more attention.
\(1) Morphology: the essential point of dermatological diagnosis is morphology, a low tech, but hard to master, skill. Dermatological diagnosis, as any other medical diagnosis, starts by collecting adequate information from the patient, and follows by its elaboration. Many doctors consider that dermatological diagnosis can be made on a quick recognition basis, but an ordered and syndromic approach is essential to get to an adequate diagnosis. I think that most dermatologists would agree that a good description of a patient by an experienced colleague is a better starting point for diagnosis than many pictures. I would describe the lesions seen in Figure 1 of \[1\] not simply as shallow ulcers, but as clearly polycyclic erosions (a finding highly suggestive of herpetic infection).
\(2) Indicated investigations: Tzanck test is the microscopic evaluation of cell morphology on a cutaneous smear. It can be done in about 15 minutes, requiring a microscope and a trained doctor. Access to this test is probably much easier than to viral cultures or polymerase chain reaction tests. In this setting, a positive Tzanck test would be enough to confirm the clinical diagnosis at a minimum cost. Considering the widespread audience of *PLoS Medicine*, with many readers in less developed countries, this test should not be forgotten.
\(3) This case, and the suspicion about systemic manifestations of skin disease, is a wonderful opportunity to disseminate an old concept, very frequently forgotten in medical literature: the skin is an organ, in fact, the biggest one in the body. Its main functions are to act as a barrier, to control temperature, to serve immunological and hormonal roles, and, physiologically less important but very important for patient well-being, to participate in personal relationships. When these functions are not adequately performed, skin failure appears, exactly as is the case with heart or renal failure. Skin failure can have many manifestations, including noninfectious fever, bacteremia, or sepsis. As is the case with renal or cardiac failure, it is easier and more practical to learn about this syndrome than to discuss the systemic manifestations of the many diseases that can cause it. I would highly recommend the following references for doctors interested in the subject: \[[@pmed-0020030-b2],[@pmed-0020030-b3].
**Citation:** Garcia-Doval I (2005) Three more learning points. PLoS Med 1(2): e30.
[^1]: **Competing Interests:** I declare that I have no competing interests.
|
PubMed Central
|
2024-06-05T03:55:51.723587
|
2005-1-25
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545215/",
"journal": "PLoS Med. 2005 Jan 25; 2(1):e30",
"authors": [
{
"first": "Ignacio",
"last": "Garcia-Doval"
}
]
}
|
PMC545216
|
We are pleased that Dr. Maitland and colleagues consider our data on volume status (intra- and extracellular) of Gabonese children to be important. We did not consider our children with severe malaria to have intravascular volume depletion for the following reasons. When we measured central venous pressures in a proportion of children on admission, there was no evidence of intravascular volume depletion (median \[interquartile range\] = 6.5 \[3--7.5\] cm water), and these values did not change significantly over 24 h, suggesting that our severely ill children had adequate filling pressures. Consistent with this observation, our severely ill children improved rapidly when markers of tissue hypoxia (blood lactate concentrations, tachycardia, and tachypnoea) were serially monitored and children were managed with a relatively conservative fluid replacement regimen. Interestingly, extracellular volume was not increased at admission or afterwards either. Capillary leakage, which commonly accompanies hypovolaemia associated with septic shock, was therefore unlikely to be a significant pathophysiological process in these children with malaria. There may be differences in the severe syndromes of malaria seen in different geographical locations, perhaps accounting for the clinical features attributable to compensated hypovolemic shock reported by Maitland and colleagues. Such differences can be assessed using simple and recently calibrated bioelectrical impedance analysis methodology as well as other techniques that monitor intravascular volumes. The design of optimal fluid management regimens for children with severe malaria can thus be informed not only by theoretical considerations, but also by appropriate physiological assessments.
**Citation:** Krishna S, Planche T (2005) Authors\' reply. PLoS Med 2(1): e32.
[^1]: **Competing Interests:** The authors declare that they have no competing interests.
|
PubMed Central
|
2024-06-05T03:55:51.724006
|
2005-1-25
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545216/",
"journal": "PLoS Med. 2005 Jan 25; 2(1):e32",
"authors": [
{
"first": "Sanjeev",
"last": "Krishna"
},
{
"first": "Timothy",
"last": "Planche"
}
]
}
|
PMC545488
|
Background
==========
Compared to other European countries, Germany still has little of a research tradition in general practice. Increasingly policy-makers have realized that the continuity and the efficacy of the healthcare system have to be improved. For this a well-developed primary medical care system is needed. In recent years a number of new chairs of general practice have been established and a national funding programme was created in order to promote General Practice as an academic discipline in Germany. The University of Heidelberg, which is the oldest university in Germany, is among the beneficiaries of these developments. Despite of a well established network of teaching practices, the research group for general practice and health services research was only created in 2002 \[[@B1]\]. This group faces the challenge to perform studies with general practitioners who have little experience with participation in scientific research. It is known, that by specifically addressing strategies significant improvements in participation rates can be achieved \[[@B2]\]. The aim of the study was to investigate the willingness of GPs to participate in research and to learn about their attitude towards research in their field in general. These data should help to create successful approaches for further projects.
Methods
=======
Study design
------------
We performed a cross sectional observational study collecting qualitative data. The Ethical Commission of the University of Heidelberg approved the study.
Study population
----------------
A random sample of 76 GPs in the area of Heidelberg was approached for the study. The GPs were selected by choosing every third of an alphabetical list of 250 practices. These GPs were associated with the university by frequently teaching students in their practices. Due to old data, in six cases the GPs did not practice any more. So finally 76 GPs were included. All of the selected GPs were in practice for more than five years. Former studies indicated that relevance of the topic has a positive predictive value for the recruitment rate. Therefore we selected a topic with a high clinical relevance in daily practice: osteoarthritis \[[@B3],[@B4]\]. Based on this information we performed a fictitious study, aiming at improving the quality of care of patients with osteoarthritis. The GPs received an official letter from the Department of General Practice and Health Services Research. This letter contained detailed information about the relevance of the topic, the aim of the study and the possible benefit for GPs, their teams and their patients. They were also informed about the time requirement for the study, which was estimated to be 30 minutes. The allowance for participating was fixed to 50 Euro to exclude financial reasons for consent. The letter concluded with the request to fax an agreement form back to the university.
Measures
--------
No letters were returned to the university because of wrong addresses. A reminder or anything similar did not follow the first letter. One week after the letter, every GP was called by the principal investigator and was asked -- after giving him again information on the study -- if she or he wanted to participate. This approach was chosen to get qualitative information of all approached GPs about their willingness to participate and their opinion in general. This way of data collecting has already been used in this field of research and enables not only a high rate of data response, it is also a feasible way of collecting qualitative data \[[@B5]\]. If the GP decided not to participate, her or his reason to do so were recorded without further discussion. Every GP, whether he denied or agreed to participate was asked about his opinions concerning research in general practice in general and the relevance of the research topic to him or her. The GPs who agreed to participate where asked to fax the sheet of agreement.
Analysis
--------
We were mainly focused on qualitative information. Therefore the statements of the GPs were grouped and coded by two separate researchers and then discussed in order to agree on the selected categorisation according to the guidelines for qualitative researchers \[[@B6]\].
Results
=======
A total of 18 GPs (23.8%) of the approached GPs was female, 58 GPs (72.2 %) were male. Only two GPs faxed their agreement-sheet within the first week, before they were phoned and interviewed by the principal investigator. During the telephone calls 25 GPs (32.8 %) agreed to participate and promised to fax the sheet. Out of this group 5 GPs (18.5%) sent their fax during the subsequent two weeks. A total of 8 (10.5 %) faxes were returned. Five female (27.7 %) and 22 male GPs agreed (37.9 %) to participate. A total of 27 GPs agreed to participate ultimately.
Table [2](#T2){ref-type="table"} shows the GPs reasons for non-participation. 24 (31.5 %) of the GPs argued they had no time, because of overwork in their practice caused by the daily routine work. The second most frequent reason named was the regular administrative workload. Seven GPs specified this argument by blaming especially the newly introduced \"disease management program, DMP\", founded by German sick funds for chronic illnesses like diabetes and hypertension. This program was perceived to increase the daily paperwork tremendously. Other important reasons for non-participation were disbelief that possible results can be implemented in daily work without financial incentives. GPs argued that changes, which are accompanied by any additional time effort, could only be implemented in daily practice if they receive adequate financial reimbursement. \"Money sets the course\", as one GP stated. Two GPs declared they had no problem in dealing with osteoarthritis and regarded also dealing patients suffering from osteoarthritis quite easy. Four GPs named participation in courses and congresses as a reason for non-participating. One GP mentioned that this kind of research is only for academic interest and helps only the career of the researcher. An other GP argued that he already feels monitored by all the data collected by health insurance and the government.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Reasons mentioned by GPs for non-participation in research
:::
n \%
----------------------------------------------------------------------------------------- -------- ---------
Overwork in practice 24 31.6
Already too much paperwork / bureaucracy 13 17.1
The results might not be implemented in practice because of financial constrains 10 13.2
Overload because of \"disease management program\" 7 9.2
No belief in results because of the degenerative progress of the illness 5 6.6
Personal time exposure for courses, etc. 4 5.3
Private reasons 2 2.6
Adherence to an other study at the same time 2 2.6
To less connection between (theoretical) university research and practical work as a GP 2 2.6
No problem in treating arthritis patients 2 2.6
No decision 1 1.3
Feeling of being monitored 1 1.3
Only the researcher takes benefit out of this research 1 1.3
**Total** **76** **100**
:::
As can be seen in table [3](#T3){ref-type="table"}, 85.6 % of the GPs had positive attitudes regarding research in their field. They consider it reasonable and eligible, but in most of these cases the answer was not substantiated with a further argumentation. Interestingly, answers, which were allocated to the category „makes sense because it improves the reputation of GPs and documents our quality of care\", were only given by GPs who agreed to participate in the study. So this aspect seemed to be the most important motivation for an GP to take part in research. In addition, this particular group of GPs regularly added further comments regarding role of the GP in the German health care system. Important reasons for scepticism were the gap between theoretical research and practical work and the domination of research by specialists. One GP argued it would be better to spend more money on treatment than on research.
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
GPs\' attitudes regarding research in General Practice in general
:::
n \%
---------------------------------------------------------------------------------- -------- ---------
Reasonable and eligible 54 71.1
Makes sense because it improves the reputation of GPs 11 14.5
Not sure if it makes sense (\"I am not convinced\"), no further explanation 3 3.9
University research and daily work in family medicine have only little in common 3 3.9
Makes no sense because research is dominated by specialists 2 2.6
Does not lead to results (without more explanation) 1 1.3
Better more money for the GPs then for research 1 1.3
Feeling of being monitored 1 1
**Total** **76** **100**
:::
Discussion
==========
There were three main conclusions that can be drawn out of our interview results. Firstly, the research topic improving the quality of care for patients suffering from osteoarthritis was considered as highly relevant by the interviewed GPs. This is concordant to our assumptions based on epidemiological data, which led to the fictitious research topic. The same reasoning causes GPs to seek support in the daily treatment of patients with osteoarthritis. Consequently this will be subject of future research projects.
Secondly, most of the GPs appreciate research in general practice, but a few were very sceptical. German GPs still don\'t realise it as a professional obligation as their colleagues in countries like e.g. the Netherlands or the United Kingdom, with a much longer tradition in research, do \[[@B7]\].
The third main result of our survey has not yet been shown in former studies. It is the fact that the willingness for participating in research emanates mainly out of the motivation to improve the reputation of family medicine in general by documenting the high quality of care with data attained in solid surveys. This may reflect the increasing self-confidence of German GPs, which are about to expend the influence in the health care system, and their awareness that an own research culture helps to enhance this. Facing decreasing financial resources in the Health care system, GPs may also be aware that a solid database documenting the quality of care will get more important for the distribution of financial resources in the near future.
The revealed barriers against participating in studies mentioned in our telephone survey are in line with results from previous studies in other countries \[[@B8]\]. According to those former results, relevance of the research topic, reimbursement and compatibility with routine general practice work are important factors. Ideally the GPs are embedded in an existing research culture \[[@B7],[@B9],[@B10]\]. Study nurses or mentors could be an important factor to enhance GPs\' preparedness to participate in General Practice research because they reduce the administrative workload for GPs and enhance the motivation to participate in research \[[@B8],[@B10],[@B11]\]. Furthermore financial incentives for participation are essential because of time constraints and overwhelming administrative work that compete with research and represent important barriers \[[@B8],[@B11]\]. An unexpected quantitative result of this study was that being involved with the training of medical students and being linked with the University is not reflected per se in a higher motivation in participating in research. Participating rates of about 30 % are usually achieved in random postal mailings to GPs without academic affiliation \[[@B5],[@B9],[@B13],[@B14]\]. Previous studies have shown that involvement in student teaching represents a positive predictive factor for participation in research, so we assumed to achieve a much higher participation rate. It appears that a well-established teaching network does not necessarily yield much benefit for research purposes \[[@B11]\].
Conclusions
===========
Previous studies were mainly focused on formal or external barriers for GPs against participating in research, or revealed approaches that cannot easily be transferred, as e.g. the enrollment of friendly GPs \[[@B15]\]. What this study adds is that there is an important target to aim at, if GPs have to be involved in research: the motivation to underline their daily work with solid data reflecting their high quality of care. With this knowledge GPs may be easier approached if they need to be motivated to participate in future projects. Aiming more on psychological targets, this approach should be transferable to other countries as well. However, researchers should be aware that beside the chance of motivating GPs, this strategy also contains a risk: GPs could be discouraged and kept away from future participation if the anticipated demonstration of their quality of care is not as obvious as expected.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
TR conceived and performed the study and draft the manuscript. JS participated in the study design. All authors read and approved the final manuscript.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Agreement for participation related to sex
:::
Sex n Agreement after letter Agreement during telephone call Total agreement
-------- ---- ------------------------ --------------------------------- -----------------
Male 58 2 20 22 (37.9%)
Female 18 0 5 5 (27.8%)
Total 76 2 (2.6 %) 25 (32.8 %) 27 (35.5 %)
:::
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2296/5/31/prepub>
|
PubMed Central
|
2024-06-05T03:55:51.724435
|
2004-12-21
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545488/",
"journal": "BMC Fam Pract. 2004 Dec 21; 5:31",
"authors": [
{
"first": "Thomas",
"last": "Rosemann"
},
{
"first": "Joachim",
"last": "Szecsenyi"
}
]
}
|
PMC545489
|
Background
==========
Anxiety and depressive disorders are some of the most common psychiatric diagnoses in children and adolescents respectively. Both are known to be associated with major impairment in childhood and adverse consequences in later life \[[@B1]-[@B4]\]. Estimates of the prevalence of any childhood anxiety disorder are in the order of 3 to 12 % \[[@B1],[@B4]\] and rise to as high as 40% or over if impairment is not required for a diagnosis \[[@B4]\]. In general, epidemiological studies show that rates of any anxiety disorder are higher in children than adolescents \[[@B5]\]. In contrast, rates of depressive disorder in young people show higher rates in adolescence (2 to 8 %) than in childhood (1 to 3 %) \[[@B6]\]. Depression and anxiety in childhood and adolescence have long-term deleterious outcomes for a significant proportion of young people. Depression and anxiety, once experienced in childhood are very likely to recur in adulthood \[[@B2],[@B7]\]. Early onset depressive and anxiety disorders are also associated with substantial social impairment \[[@B8]\]. Even sub-clinical levels of depression in children and adolescents are associated with significant morbidity in the form of psychosocial impairment and service utilisation \[[@B9],[@B10]\]. Furthermore, adolescents identified as having high levels of depressive or anxiety symptoms are significantly more likely to experience depressive disorder in adulthood than adolescents with depression levels within the normal range \[[@B11],[@B12]\]. The observations that sub-clinical symptoms of depression are associated with significant morbidity, and that high levels of depression and anxiety symptoms predict depressive and anxiety disorders add to the evidence that depression and anxiety can be regarded as continua \[[@B13]-[@B15]\].
Depression and anxiety co-occur more commonly than would be expected by chance in children and adolescents. This co-occurrence has been identified both in clinical studies of children and adolescents and general population samples that have examined sub-clinical levels of depression and anxiety symptoms \[[@B16],[@B17]\]. More specifically, anxiety symptoms or disorders most often precede depressive symptoms or disorders \[[@B18]-[@B21]\]. Moreover, although certain sub-types of anxiety, namely social phobia and panic rarely precede depression \[[@B22]\], individuals with these disorders and depression are very likely to have had a different anxiety disorder that predated the onset of depression \[[@B20]\]. Twin studies provide a means of examining the extent to which the genetic and environmental aetiological factors contributing to two disorders or symptom groups overlap and to what extent they are distinct. Longitudinal data, collected at more than one time point provide a further test, namely, that one set of symptoms or disorder is a risk factor for another. This approach is especially useful in the study of anxiety and depression. Several groups have suggested that anxiety may be a developmental precursor of depression, particularly in young people who are at increased risk of depression due to parental depression \[[@B23]-[@B25]\]. Indeed Kovacs & Devlin \[[@B17]\] suggested that children may be biologically \'prepared\' to experience symptoms of anxiety rather than depression. The results of other studies are consistent with this proposal. For example, in a sample of depressed children, in those children who had a comorbid anxiety disorder, the anxiety disorder was found to have preceded depressive disorder in two thirds of cases \[[@B18]\]. Similar evidence that anxiety disorders tend to precede depression has been reported in longitudinal epidemiological \[[@B19],[@B21],[@B26]\] and clinical studies \[[@B20]\]. Indeed, a recently convened National Institute for Mental Health (NIMH) workgroup recommended research into childhood anxiety as a known precursor of depression as a priority \[[@B27]\]. Despite clear indications that anxiety and depression in childhood and adolescence are associated, it remains unclear as to how the transition from anxiety to depression is mediated over time. Possible factors include aetiological factors in common -- these may be 1) genetic or 2) psycho-social risk factors, or 3) a direct risk effect of anxiety leading to later depression.
Cross-sectional twin studies of children \[[@B28],[@B29]\] and adults \[[@B30]\] and one longitudinal twin study of girls \[[@B31]\] have shown that to a large extent, the overlap between anxiety and depression is due to a common set of genes that influence both depression and anxiety. However, shared environmental factors have also been shown to be important sources of covariation between anxiety and depression symptoms for children but not adults \[[@B28],[@B31]\] which suggests the importance of shared psycho-social risk factors for anxiety and depression. Nevertheless, two out of these three twin studies were based on cross-sectional data and were therefore not able to determine the genetic and environmental associations between anxiety and depression over time. Furthermore, despite the importance of understanding why anxiety tends to precede depression, \[[@B27]\] no twin study of children and adolescents has yet specifically tested the hypothesis that anxiety is a phenotypic risk factor for depression. The present study also adds to the existing literature in that data on depression symptoms from different raters (mother and child) are available, thus allowing associations to be examined with data from different informants.
In the present study, we set out to examine two hypotheses that may explain the observed associations between early anxiety and later depression.
1\. Early anxiety symptoms and later depression symptoms are associated because of shared risk factors.
2\. The association between anxiety symptoms and later depression symptoms is mediated by a risk effect of the phenotype of anxiety.
Method
======
Participants
------------
Families from a systematically ascertained, population-based register of all twin births between 1980 and 1991 in South Wales, U.K. were invited to participate. This register forms a sub-sample of the Cardiff Study of All-Wales and North West of England Twins (CASTANET). Twins who had emigrated were excluded, as were cases in which one of the twins had died or had a serious illness. At the first wave of data collection in 1997, there were a total of 1109 pairs of twins aged 5--17 years although not all of these individuals were eligible to participate at both time points (see below). Data were collected by postal questionnaire. Families received three reminders and telephone reminders when numbers could be traced. The same methods were used three years later to collect longitudinal data except that families received four reminders. To be invited to participate in the follow-up study, we required that the twins were living together in the same home and were under the age of 18 years. Twins were required to live in the same home in order to minimise heterogeneity of environmental risk factors that can impact on genetic and environmental parameter estimates. The focus of the follow-up study was childhood psychopathology and for that reason young people aged 18 and over were not included. At time 1, there were 986 twin pairs who were within the age range of the study at both time points.
In the first wave of the study (1997; time 1), 670 families provided questionnaire responses giving a response rate of 61%. Comparison of responders and non-responders using Townsend Scores \[[@B32]\] which index the level of deprivation of an electoral area revealed no significant socio-demographic differences between the two groups at time 1 (t = .373, p = 0.709). Families with children aged 8--17 were re-contacted in 2000 (time 2). Of the 670 families who replied at time one, 85 had moved away, there were 8 new contraindications and there were 123 children who were out of the age range of the study and did not live in the same home. This left a total of 454 families who were eligible at time 2.
Of these, 338 families replied, giving a total response rate of 75%. There were no significant socio-demographic differences between responders and non-responders at time 2 (t = 1.71, p = 0.09). Zygosity was assigned using a twin similarity questionnaire which has been shown to be over 90% accurate in distinguishing identical (monozygotic; MZ) from fraternal (dizygotic; DZ) twins \[[@B33]\]. There were 198 MZ girls (99 pairs), 134 MZ boys, 128 DZ girls, 116 DZ boys, 270 opposite sex DZ twins.
Measures
--------
At time 1, parents were asked to complete the Children\'s Revised Manifest Anxiety Scale \[[@B34]\] which assesses symptoms over the past three months. It has previously been found to be a reliable and valid instrument \[[@B35]\] (Cronbach\'s α = .8662 twin 1, α = .8708 twin 2). Parents also completed the general functioning scale of the McMasters Family Assessment Device (FAD) \[[@B36]\]. At both time points, parents completed the short version of the Mood and Feelings Questionnaire (MFQ) \[[@B37]\]. At the second wave children aged 11 or above also completed the MFQ. The MFQ is based on DSM-III-R symptoms of depression and has been successfully used as a screening questionnaire for clinical depression in community populations \[[@B38]\] (α = .9231 twin 1, α = 9320 twin 2).
Analysis
--------
### Descriptive statistics
For descriptive statistics, (correlations and mean comparisons), the survey commands in the program STATA \[[@B39]\] were used. These commands take into account the clustering of the data from twin pairs (i.e. each twin pair provides two data points) by likening the twin data to a two-stage cluster design with the twin pairs as the primary sampling unit. Since reliability coefficients can not be calculated using these commands these were presented for first and second-born twins separately.
### Univariate
Analysing data from twins provides a means of estimating the relative contribution of genetic and environmental effects on individual variation in behaviour. In the basic (ACE) model, variation can arise from three sources: 1) additive genetic effects (A); 2) common environmental effects (C); 3) unique environmental effects (E). Common environmental effects are non-genetic factors that serve to make twins more similar to one another while unique environmental effects are non-genetic factors that uniquely influence one individual within a twin pair and tend to make the individuals in a twin pair different from each other. Model fitting was carried out using the programs Mx \[[@B41]\] and LISREL \[[@B42]\] and continuous measures of anxiety and depressive symptoms were analysed. The significance of the A, C and E parameters can be tested by dropping them from the model and comparing the fit of the reduced model to that of the full model using the χ^2^critical value for the number of degrees of freedom gained in the reduced model.
### Bivariate
Bivariate analysis allows the covariance of two traits to be partitioned into covariance that is due to additive genetic factors, common environmental factors and unique environmental factors. The covariance parameters for the Cholesky model presented (see figure [1](#F1){ref-type="fig"}) include those factors in trait 1 (anxiety) that also influence trait 2 (depression). A bivariate model in which anxiety symptoms at time 1 precede depressive symptoms at time 2 was fitted consistent with clinical and epidemiological data showing that anxiety precedes depression more often than vice versa. In addition, a 3 variable model that included anxiety and depressive symptoms at time 1 and depressive symptoms at time 2 was fitted. This model estimated the genetic and environmental associations between anxiety at time 1 and depression at time 2 when the effects of concurrent depression were included. A causal model was then fitted (see figure [2](#F2){ref-type="fig"}). Comparing the fit of this causal model to that of the general bivariate (Cholesky) model allows two competing explanations of the association between anxiety (time 1) and depression (time 2) to be tested: 1) the association of anxiety and depression is due to genetic and /or environmental risk factors common to both anxiety and depression: 2) the association is due to a risk effect of the phenotype of early anxiety on later depression. A unidirectional causal model from anxiety to depression was fitted given that the data presented are longitudinal. Although the reliabilities of the anxiety and depression scales were good and comparable, a causal model that included residual error was included in line with the suggestion of Neale & Cardon \[[@B43]\]. This was fitted since in direction of causation models it cannot be assumed that measurement error will be confounded with non-shared environmental effects \[[@B44]\]. Fitting this type of model does not constrain measurement error to be transmitted phenotypically and thus is likely to provide more realistic parameter estimates than a casual model without residual error terms.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Bivariate Cholesky decomposition of anxiety at time 1 and depression at time 2.
Aanx genetic influences on anxiety Canx common environmental influences on anxiety Eanx non shared environmental influences on anxiety Ac genetic influences on anxiety that also influence depression Cc common environmental influences on anxiety that also influence depression Ec non shared environmental influences on anxiety that also influence depression Adep genetic influences specific to depression Cdep common environmental influences specific to depression Edep non shared environmental influences specific to depression
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Unidirectional causal model from anxiety at time 1 to depression at time 2
Aanx genetic influences on anxiety Canx common environmental influences on anxiety Eanx non shared environmental influences on anxiety Adep genetic influences specific to depression Cdep common environmental influences specific to depression Edep non shared environmental influences specific to depression
:::
:::
Sex effects
-----------
Univariate analyses were performed to test for both quantitative and qualitative sex differences. Quantitative sex differences, are tested by estimating the size of parameter estimates for the genders (\'the common effects sex limitation model\'). Qualitative sex differences test whether the set of genes influencing the phenotype differs by gender i.e. different genes (\'the general effects sex limitation model\'), this is done by estimating the genetic correlation for opposite sex DZ pairs and comparing the fit of this model to one that constrained the genetic correlation to 0.5. In addition, a bivariate sex limitation model was tested \[[@B45]\]. This model estimates whether the covariation between anxiety and depression is different for boys and girls. All analyses reported were based on the five twin groups (MZ male, MZ female, DZ male, DZ female, DZ opposite sex).
Results
=======
Descriptive statistics
----------------------
There were no significant mean differences on anxiety or depressive symptoms by gender (t = -0.047, p = .962; t = -1.628, p = .104). Age was not associated with anxiety or depressive symptoms (r = .014, p = .707; r = .082, p = .078). The mean age of children at time 1 was 10.58, range 5.58--17.83, and at time 2, was, 12.64, range 8.75--17.25. Symptoms of early anxiety and later depression were strongly correlated (parent-rated symptoms r = .479. The correlation was slightly lower for symptoms across rater (parent rated anxiety and adolescent rated depression), r = .335.
Sex effects
-----------
For anxiety, univariate analysis of parent-rated data showed no significant gender differences in the magnitude of genetic parameter estimates (Δ χ^2^= 1.505, Δ df = 3), nor were there qualitative genetic differences (Δ χ^2^= 0, Δ df = 1). For depression, univariate models for parent-rated and self-rated scores indicated no significant gender effects for the magnitude of genetic effects (parent rated, Δ χ^2^= 2.679, Δ df = 3; self-rated, Δ χ^2^= 0.586, Δ df = 3) nor qualitative gender differences (parent rated, Δ χ^2^= 0, Δ df = 1; self-rated, Δ χ^2^= 0, Δ df = 1). Finally, for parent rated symptoms, results from the bivariate sex limitation Cholesky model showed no significant gender differences in the covariation between anxiety and depression in that the genetic and environmental covariation could be equated across the genders with no significant deterioration in fit (parent-rated, Δ χ^2^= 0.445, Δ df = 3). However, for self-rated symptoms, one parameter, i.e., the non-shared environmental covariation parameter, could not be equated across the genders (Δ χ^2^= 6.107, Δ df = 1). Estimates from this model for boys were; Aanx = 50, Canx = 17, Eanx = 32, Ac = 10, Cc = 47, Ec = -11, Adep = 9, Cdep = 5, Edep = 40; and for girls were; Aanx = 47, Canx = 22, Eanx = 30, Ac = 12, Cc = 13, Ec = 8, Adep = 30, Cdep = 12, Edep = 24; χ^2^= 30.93, df = 32, AIC = -33.07). These results are presented in addition to those for the combined sample (see table [1](#T1){ref-type="table"}).
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Bivariate model fitting for time 1 anxiety and time 2 depression
:::
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
**Rater** **Aanx** **Canx** **Eanx** **Ac** **Cc** **Ec** **Adep** **Cdep** **Edep** **χ^2^** **AIC**
-------------------------------------------------------------------------- ---------- -- ---------- -- ---------- -- -------- -- ---------- -- -------- -- ---------- -- ---------- -- ---------- -- ---------- -- ---------
Parent rated total sample\ 46\*\*\* 24\*\*\* 30\*\*\* 13\*\* 36\*\*\* 2 ns 24\* 0\ 26\*\*\* 15.36\ -6.64
NMZ = 138\ ns df = 11
NDZ = 210
Parent rated -- adolescents only (8--14 at time 1 and 11--17 at time 2)\ 46\*\*\* 23\*\* 31\*\*\* 11\* 34\*\* 4 ns 25\* 0\ 25\*\*\* 15.02\ -6.98
NMZ = 90\ ns df = 11
NDZ = 136
Parent rated anxiety time 1 and self rated depression time 2\ 53\*\*\* 15 ns 32\*\*\* 13\* 17 ns 2 ns 36\*\* 0\ 31\*\*\* 5.80\ -16.20
NMZ = 92\ ns df = 11
NDZ = 128
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Ns = non significant
\* p = .05
\*\* p = .01
\*\*\* p = .001
Aanx genetic influences on anxiety
Canx common environmental influences on anxiety
Eanx non shared environmental influences on anxiety
Ac genetic influences on anxiety that also influence depression
Cc common environmental influences on anxiety that also influence depression
Ec non shared environmental influences on anxiety that also influence depression
Adep genetic influences specific to depression
Cdep common environmental influences specific to depression
Edep non shared environmental influences specific to depression
NMZ = number of MZ twin pairs
NDZ = number of DZ twin pairs
Parameter estimates equated to sum 100 in direction of arrows on each trait (anxiety time 1 & depression time 2)
:::
Bivariate analysis
------------------
Table [1](#T1){ref-type="table"} shows results from bivariate analyses of anxiety and depression. Both the genetic covariation (Ac) between anxiety and depression as well as the common environmental covariation (Cc) were significant, while unique environmental covariation (Ec) was negligible. However, despite significant covariation, there were also significant genetic (Adep) and unique environmental effects (Edep) specific to later depression. Thus, although significant genetic covariation between anxiety and depression was observed, the genetic effects on depression were not entirely mediated through genetic effects that were common with anxiety. This illustrates that the genetic covariation between anxiety and depression over time is not complete. Moreover, this observed genetic covariation does not appear to derive from the association of early depression with later depression in that the genetic covariation between anxiety and later depression remained significant when early depressive symptoms were included in the model (see figure [3](#F3){ref-type="fig"}). Figure [3](#F3){ref-type="fig"} shows that there is a significant genetic path linking early anxiety and later depression (Ac ~(anx1\ dep2)~= 11, p = .001).
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Trivariate Cholesky decomposition of anxiety at time 1, depression at time 1 and depression at time 2
χ^2^= 39.74, df = 24, AIC = -8.26 Ns = non significant \* p = .05 \*\* p = .01 \*\*\* p = .001 Parameter estimates equated to sum 100 in direction of arrows on each trait (anxiety time 1, depression time 1 & depression time 2)
:::
:::
It has been shown that the aetiology of depression varies significantly according to age, with greater common environmental effects in children aged 8--10 than in adolescents aged 11--17 \[[@B28],[@B46],[@B47]\]. The three year follow-up period of the study meant that we were unable to analyse children\'s and adolescent\'s symptoms separately. However, given previous findings of age effects, we carried out analyses including only those twins who were \'adolescents\', that is, aged 11 or over at the second time point with the expectation that the common environmental covariation path (Cc) might decrease. Nonetheless, a significant common environmental component of variance remained.
Family functioning and rater effects
------------------------------------
Following this finding of significant common environmental covariation the nature of this latent factor was further examined. Information on a potential common environmental variable, general family functioning (all twins lived in the parental home) was available. Family functioning was correlated with anxiety symptoms at time 1 (r = .210, p = .001). Bivariate analyses controlling for family functioning are shown in table [2](#T2){ref-type="table"} and it can be seen that this exerted only a slight effect on the estimate of common environmental covariation (drop of Cc from .36 to .31). However, associations between anxiety and depression symptoms could be due to the fact that a single informant (parents) rated their children\'s anxiety and depression symptoms at both time points. In bivariate analyses, similarities between variables that are due to shared rater effects will be partitioned into the common environmental component of covariance. A distinction should be made here between shared rater effects and rater bias. Rater biases that result in common environmental effects arise when a proxy informant, usually a parent rates a pair of twins as more similar than more objective measures would find them. This sort of rater bias results in deflated MZ phenotypic variance compared to DZ variance \[[@B43]\]. This pattern of variance has not been observed in this sample (anxiety DZ standard error = .035, MZ standard error = .048, t = 1.89, p = .06; depression DZ standard error = .056, MZ standard error = .067, t = .514, p =.608), in fact the MZ variance is greater than the DZ variance. Thus, rater bias can not account for the observed common environmental effects. On the other hand, shared rater effects come about simply as an effect of the same informant rating two sets of symptoms or risk variable and outcome and are therefore not exclusive to parental ratings. In order to test any potential shared rater effects, a bivariate model with data from different informants was tested (parent-rated anxiety symptoms at time 1 and adolescent-rated depression symptoms at time 2). It can be seen from Table [1](#T1){ref-type="table"} that the common environmental covariance influencing anxiety and depression (Cc) is no longer significant when cross-informant information is used. This suggests that at least a proportion of the Cc estimate observed in analyses that used only parent-rated data is likely to be due to shared rater effects, i.e. that part of the shared environmental covariation is due to the fact that the same informant rated both phenotypes. The observation that when family functioning was included as a measured environmental variable in the analyses of parent-rated information, the common environmental covariance estimate was only slightly reduced is consistent with the possibility that these Cc effects may be due to shared rater effects. However, given the small effect sizes of most single measured risk factors (genetic or environmental) \[[@B48]\], one might not expect single environmental risk factors to account for large proportions of variance. Indeed, risk variables for symptoms of depression and anxiety are likely to have multiplicative effects \[[@B48],[@B49]\]. With this in mind, the same family functioning was also included in a cross-informant model. As can be seen from table [2](#T2){ref-type="table"}, including family functioning resulted in a small decrease in the Cc estimate (17 to 14).
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Bivariate model fitting for time 1 anxiety and time 2 depression controlling for measured environmental variables
:::
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
**Rater** **Aanx** **Canx** **Eanx** **Ac** **Cc** **Ec** **Adep** **Cdep** **Edep** **χ^2^** **AIC**
------------------------------------------------- ---------- -- ---------- -- ---------- -- -------- -- -------- -- -------- -- ---------- -- ---------- -- ---------- -- ---------- -- ---------
Parent rated total sample\ 47\*\*\* 21\*\* 32\*\*\* 12\*\* 31\*\* 1 ns 24\* 0\ 26\*\*\* 12.59\ -9.41
-**family functioning at time 1 regressed out** ns df = 11
Parent rated anxiety and self rated depression\ 56\*\*\* 11 ns 33\*\*\* 14\* 14 ns 2 ns 38\*\* 0\ 33\*\*\* 6.75\ -15.25
-**family functioning at time 1 regressed out** ns df = 11
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Ns = non significant
\* p = .05
\*\* p = .01
\*\*\* p = .001
Aanx genetic influences on anxiety
Canx common environmental influences on anxiety
Eanx non shared environmental influences on anxiety
Ac genetic influences on anxiety that also influence depression
Cc common environmental influences on anxiety that also influence depression
Ec non shared environmental influences on anxiety that also influence depression
Adep genetic influences specific to depression
Cdep common environmental influences specific to depression
Edep non shared environmental influences specific to depression
NMZ = number of MZ twin pairs
NDZ = number of DZ twin pairs
Parameter estimates equated to sum 100 in direction of arrows on each trait (anxiety time 1 & depression time 2)
:::
Phenotypic \"causal\" model
---------------------------
Finally, a model that included a direct causal path from anxiety to depression was fitted (see figure [2](#F2){ref-type="fig"}). This model allows for testing whether the phenotype of anxiety (rather than shared genetic and environmental aetiological factors) was responsible for the observed covariance between anxiety and later depression symptoms. That is, testing whether anxiety (independent of shared genetic and environmental risk factors) is itself an early risk factor for depression. The full causal model provided a significantly poorer fit than the general bivariate model (Δ χ^2^= 48.82, Δ df = 1, p \< .001). Thus, it does not appear that anxiety leads to depression through direct phenotypic effects, but that, anxiety and depression symptoms are associated over time because they share aetiological factors in common.
Discussion
==========
This investigation has used a longitudinal, epidemiological and genetically sensitive design to examine two possible explanations of the association between early anxiety and later depression symptoms in children and adolescents: 1) a common genetic/environmental aetiology or 2) a phenotypic risk effect of early anxiety. There was significant genetic covariation between anxiety and later depression. (Moreover, the genetic covariation was not explained by the effects of early depression). This result is consistent with the association between early anxiety and later depression being due to a common genetic aetiology. However, the genetic overlap between early anxiety and later depression was far from complete in that there were significant, separate genetic effects on anxiety and genetic influences specific to later depression. Thus, in this sample, symptoms of anxiety and depression in children and adolescents share only a partly common genetic aetiology.
In addition to genetic covariation, significant shared environmental covariation between anxiety and depression was also observed. It appeared that some of the shared environmental covariation between anxiety and depression observed in parent ratings of anxiety and depression was due to shared rater effects as the common environmental covariation (Cc) was no longer significant when analyses were performed with data across different informants. However, including family adversity as a measured environmental risk factor into a model with data from different informants also resulted in a small decrease in the Cc parameter estimate.
The model that included a direct causal path from early anxiety symptoms to later depression symptoms provided a significantly poorer fit than the general bivariate model. These results suggest that anxiety is not an aetiologically distinct phenotype that is in itself a risk factor for future depression symptoms, but rather that the covariation over time arises from the common genetic and environmental architecture. It should be noted though, that some of the common environmental covariation is likely due to shared rater effects, because we found that this path became non significant in cross-informant analyses.
The present findings are consistent with those of several other twin studies that have reported strong genetic correlations between symptoms of anxiety and depression in children and adolescents \[[@B28],[@B29],[@B31]\] and adults \[[@B30]\] and with a study that found significant genetic effects specific to depression \[[@B29]\]. However, only one of these studies was longitudinal \[[@B31]\] and this included girls only, and none of these studies included information from more than one informant.
Sex effects
-----------
The lack of significant univariate and bivariate gender differences in genetic and environmental parameters estimates for parent reports in the present sample is of interest. The results for self reports of depression are less clear, previous analysis of a larger sample, from which the present sample was drawn, did find sex differences for self rated depressive symptoms as measured by the long version of the Mood and Feelings Questionnaire (MFQ) \[[@B47]\], which were not detected in the present sample. However, a previous cross-sectional analysis of the full time 1 self-rated depression data did not find significant gender differences \[[@B54]\]. The only significant gender difference in the present analysis was for the non-shared environmental covariation between anxiety and self-rated depression. The lack of significant effects for univariate analysis for self-rated depression may be due to the smaller sample size and thus lower power to detect effects in the present sample, or it could be due to the fact that different versions of the MFQ that were used in the present (short version) and a previous analysis (long version) \[[@B47]\]. Moreover, the bivariate Cholesky sex limitation analysis for self-reported depression was likely under-powered as few of the parameter estimates reached statistical significance. The sample size is small for those who provided self-reports (NMZ = 92 and NDZ = 128, see table [1](#T1){ref-type="table"}) and it is therefore uncertain how reliable these results are. The non-shared environmental covariation estimate for boys also yields a negative parameter estimate (-.11) (albeit non-significant) which indicates the non-shared environment for anxiety is negatively correlated with the non-shared environment for depression. This finding is difficult to interpret, further suggesting caution in conferring too much confidence to the gender-specific findings in this model. Given the sample size for cross informant models in this study, it is not safe to draw firm conclusions about gender differences in the covariance of anxiety and depression when depression is self rated. This needs to be examined in a larger sample. However, although the prevalence of depression shows gender differences in adolescence, this does not necessarily suggest gender differences in aetiology.
How do the present findings fit with results from family studies? Several family studies of the offspring of depressed parents have found increased rates of anxiety rather than depressive disorders \[[@B23],[@B25]\] and Rende and colleagues \[[@B24]\] found that sibling resemblance for anxiety disorders was increased in the offspring of depressed parents. This familial aggregation of anxiety disorders could be due to common environmental or genetic factors. There is now consistent evidence from cross-sectional and longitudinal twin studies of children and adolescents that this observed familial association between anxiety and depression symptoms has a partly common genetic aetiology.
Limitations
===========
As mentioned previously, several groups have shown that the aetiology of depressive symptoms differs between children (8--10) and adolescents (11--17) \[[@B28],[@B46],[@B47]\]. The majority, though not all, of the present sample were \'children\' aged under 10 (range 5--14) at time 1 and \'adolescents\' aged 11 and above (range 8--17) at time 2. Thus, as there is age heterogeneity in aetiology, high levels of effects specific to each time point might be expected such as shown in the present study. Nonetheless, we might not expect to find complete genetic overlap between anxiety and depression. For instance, the genetic liability to depression and antisocial behaviour in children and adolescents has also been shown to overlap \[[@B50]\]. Thus, there may be different developmental pathways to depressive symptoms in adolescence. In addition, previous studies have suggested that gene-environment correlation \[[@B51],[@B52]\] and gene-environment interaction \[[@B49],[@B53]\] involving life events also contribute to genetic variance in adolescent depression and such effects would also be subsumed within the estimate of genetic variance.
Clinical implications
=====================
Anxiety and depressive symptoms are strongly associated over time. This association does not appear to be due to a phenotypic risk effect of early anxiety. Rather, early anxiety and later depression are associated due to a common aetiology. This was primarily a common genetic aetiology although family functioning and a single informant rating on both sets of symptoms also contributed to this association. Some of the common genetic aetiology may act as indirect genetic effects via behaviour (gene-environment correlation and gene-environment interaction).
Competing interests
===================
The author(s) declare that they have no competing interests.
Author contributions
====================
AT and FR conceived the paper. FR carried out statistical analysis and wrote the paper. AT and MBM wrote and edited the paper. MBM provided statistical support. All authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-244X/4/43/prepub>
Acknowledgements
================
We thank Peter McGuffin and Jane Scourfield who were involved in the first wave of the study. The twin registers were set up with funding from the Medical Research Council U.K.
|
PubMed Central
|
2024-06-05T03:55:51.726084
|
2004-12-13
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545489/",
"journal": "BMC Psychiatry. 2004 Dec 13; 4:43",
"authors": [
{
"first": "Frances",
"last": "Rice"
},
{
"first": "Marianne BM",
"last": "van den Bree"
},
{
"first": "Anita",
"last": "Thapar"
}
]
}
|
PMC545490
|
Background
==========
Three conserved DNA polymerase enzymes whose activities are essential for complete chromosomal DNA replication have been identified through biochemical studies in mammalian systems \[[@B1]\] and combined genetic and biochemical studies in yeast \[[@B2]\]. During S-phase, the DNA polymerase α-primase complex synthesises the short RNA-DNA segment that is used to prime synthesis of the leading strand at the chromosomal replication origin and synthesis of each Okazaki fragment on the lagging strand. The short RNA segment is synthesised by the primase and then extended by 10--20 nucleotides by Pol α. The 3\' end of the RNA-DNA primer is recognised by replication factor C (RFC), which displaces the Pol α-primase complex and catalyses the loading of the sliding clamp PCNA at the primer-template junction. PCNA then acts as a processivity factor for the Pol δ and/or Pol ε enzymes. The exact roles played by Pol δ and Pol ε remain unclear (for a recent perspective, see ref. \[[@B3]\] and references therein) but Pol δ is most likely responsible for lagging strand replication and may also play a role on the leading strand. Yeast lacking Pol ε catalytic activity are viable but are slow growing and somewhat impaired in chromosome replication \[[@B4]-[@B7]\]. In such cells, Pol δ is thought to perform the bulk of nascent DNA chain elongation, raising the possibility that this enzyme performs a similar function in wild-type cells. If this is the case, Pol ε could have a specialised role, at replication origins for example, or in the replication of sites of sister chromosome cohesion \[[@B3]\].
Each of the three essential polymerases is a multi-subunit entity, comprising a large catalytic subunit and a number of smaller subunits that are presumed to play either regulatory or structural roles \[[@B2]\]. Little is known of the biochemical functions of the smaller subunits but in yeast most are, like the three catalytic subunits, essential proteins. In the fission yeast *Schizosaccharomyces pombe*, the Pol α-primase and Pol δ complexes are both heterotetrameric in structure \[[@B2]\]. The catalytic subunits of the two complexes, Pol1 and Pol3 respectively, are members of the B family of DNA polymerases typified by bacteriophage T4 polymerase or Pfu \[[@B8]\]. Both complexes also contain related B-subunits. These proteins are members of a larger family of phosphotransferase and nuclease enzymes \[[@B9],[@B10]\], although no enzymatic activity has been detected for either of the DNA polymerase subunits. In fission yeast, the B-subunit of Pol δ is the Cdc1 protein. Fission yeast Pol δ also contains two further subunits: the C-subunit Cdc27, which functions in part to link the polymerase to PCNA \[[@B11],[@B12]\], and the D-subunit Cdm1. Of the four Pol δ subunits, only Cdm1 is non-essential \[[@B13]\]. Orthologues of all four of these proteins (Pol3, Cdc1, Cdc27 and Cdm1) make up mammalian Pol δ \[[@B14],[@B15]\].
Perhaps surprisingly, there have been few reports of physical interactions between the various polymerase enzymes believed to be present at the eukaryotic replication fork. One exception to this comes from large-scale functional analysis of the budding yeast proteome where an interaction was uncovered between the catalytic subunit of the Pol α-primase complex, Pol1, and the C-subunit of the Pol δ complex in budding yeast, Pol32 \[[@B16]\]. Interaction between these proteins was detected using the two-hybrid system, raising the possibility that the interaction was not a direct one, but was instead mediated via a third protein factor or complex. Recently, the two-hybrid result was independently confirmed and a direct interaction between the budding yeast Pol α-primase and Pol δ demonstrated by mixing of the purified enzyme complexes followed by co-immunoprecipitation \[[@B17]\]. By this method, Pol α-primase and Pol δ were shown interact in a Pol32-dependent manner \[[@B17]\]. A form of Pol δ containing a mutant Pol32 protein lacking a 40 amino acid region of the C-terminal domain (Pol32-8, lacking amino acids 270--309) was also shown to be unable to co-immunoprecipitate with Pol α-primase \[[@B17]\], providing the first indication of the location of the Pol α binding site on Pol32.
In this paper, it is shown that the orthologues of the budding yeast Pol1 and Pol32 proteins in fission yeast, Pol1 and Cdc27 respectively, also interact in the two-hybrid system. It is also shown that these proteins are capable of interacting directly with one another, *in vitro*, using purified recombinant proteins and peptides. Interaction is also seen between the human Cdc27 and Pol1 homologues, p66 and p180. The binding site for Pol1 maps to the extended C-terminal domain of the Cdc27 protein and requires the presence of a short protein sequence motif, which we have designated the DPIM, for DNA polymerase interaction motif. This short sequence is conserved in Cdc27 orthologues from a wide variety of eukaryotic species. Despite this evolutionary conservation, mutational inactivation of the Pol1 binding motif does not affect Cdc27 function *in vivo*. The implications of these results are discussed.
Results
=======
Interaction between fission yeast Pol1 and Cdc27
------------------------------------------------
To test for interaction between fission yeast Pol1 and Cdc27, the two-hybrid system was used. Full-length Cdc27 fused to the activation domain (AD) of the yeast Gal4 protein was tested for its ability to interact with amino acids 278--527 of fission yeast Pol1 fused to the DNA binding domain of the bacterial transcription factor LexA (see Materials and methods for details). Amino acids 278--527 correspond to the smallest region of budding yeast Pol1 shown to interact with Pol32 \[[@B16]\]. Reporter gene induction, measured by β-galactosidase activity assay, was observed in the presence of the Gal4 AD-Cdc27 and LexA-Pol1-278-527 (LexA-Pol1) proteins, but not when either protein alone was present (Figure [1](#F1){ref-type="fig"} and Table [1A](#T1){ref-type="table"}, upper part). Thus, as in budding yeast, the catalytic subunit of Pol α is able to interact with the C-subunit of Pol δ.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Mapping of the Pol1 interaction site on Cdc27 using the two-hybrid system.**Full-length and thirteen truncated Cdc27 proteins, expressed as Gal4 transcription activation domain fusions, were tested for their ability to interact with Cdc1-LexA, Pcn1-LexA and Pol1(278--512)-LexA baits in *S. cerevisiae*CTY10-5d. Interactions were monitored by β-galactosidase activity, as described in Material and methods, and are indicated as +++ (strong interaction, typically 80--100 Miller units of β-galactosidase activity), + (weak interaction, \<20 Miller units), -- (no interaction detectable above background, \<1 Miller unit). The area between the broken vertical lines indicates the minimal Pol1 binding region, corresponding to amino acids 293--332 of Cdc27. The grey box indicates the minimal Cdc1 binding domain, amino acids 1--160; the black box represents the Pcn1 (PCNA) binding motif (amino acids 362--372).
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Two-hybrid analysis. Prey and bait proteins were expressed as Gal4 AD and LexA BD fusions from pACT2 and pBTM116 respectively. Positive interactions corresponded to 50--100 Miller units (++) or 10--20 Miller units (+) of β-galactosidase activity. Negative interactions corresponded to \< 2 units. See text for details.
:::
Prey Bait Bridge Interaction
------------------------------- --------------------- -------- -------------
A. Fission yeast proteins
Cdc27 \- \- \-
\- Pol1(278--527) \- \-
Cdc27 Pol1(278--527) \- ++
Pcn1 Pol1(278--527) \- \-
Pcn1 \- Cdc27 \-
\- Pol1(278--527) Cdc27 \-
Pcn1 Pol1(278--527) Cdc27 \+
B. Human proteins
\- Pol α (291--540) \- \-
p66 (253--466) \- \- \-
p66 (356--466) \- \- \-
p66 (253--466) Pol α (291--540) \- ++
p66 (356--466) Pol α (291--540) \- ++
C. Cross-species interactions
\- Pol1(278--527) \- \-
p66 (253--466) Pol1(278--527) \- ++
p66 (356--466) Pol1(278--527) \- ++
Cdc27 Pol α (291--540) \- \-
D. Truncated Pol1 proteins
Cdc27 Pol1(278--527) \- ++
Cdc27 Pol1(328--527) \- \-
Cdc27 Pol1(278--477) \- \-
Cdc27 Pol1(278--487) \- \-
Cdc27 Pol1(278--497) \- \-
Cdc27 Pol1(278--507) \- \-
Cdc27 Pol1(328--477) \- \-
Cdc27 Pol1-TS13(278--527) \- \-
:::
Mapping the Pol1 binding site on Cdc27
--------------------------------------
Previously, minimal Cdc1 and PCNA binding regions on Cdc27 have been mapped \[[@B11],[@B12],[@B18]\]. Cdc1 binds within the globular 160 amino acid N-terminal domain of the 372 amino acid Cdc27 protein, whereas PCNA binds at the extreme C-terminus of Cdc27 at a conserved p21^Cip1^-like PCNA binding sequence, the PIP box. To map the Pol1 binding site, a series of thirteen truncated Cdc27 proteins fused to the Gal4 AD were tested against LexA-Pol1 in the two-hybrid system. The results of this analysis are shown in Figure [1](#F1){ref-type="fig"}. Pol1 binds within the extended C-terminal domain of Cdc27, with the smallest construct capable of binding spanning a forty amino acid region, from amino acids 293--332. Therefore, the Pol1 binding region is distinct from both the globular N-terminal domain and C-terminal PIP box. In support of this, two-hybrid analysis showed that Cdc27 could bind to both Pol1 and Pcn1 (PCNA) proteins simultaneously: Gal4 AD-Pcn1 and LexA-Pol1 fusion proteins (which do not interact in the two-hybrid system) could be brought together by co-expression of Cdc27 (Table [1A](#T1){ref-type="table"}, lower part).
Cdc27-Pol1 interactions with recombinant proteins
-------------------------------------------------
In order to test whether the interaction between Pol1 and Cdc27 was a direct one, purified recombinant Cdc27 and Pol1 proteins were assayed for their ability to interact *in vitro*. Purified GST-Cdc27-273-352 fusion protein \[[@B11]\] was tested for its ability to pull-down purified recombinant hexahistidine-tagged Pol1 (amino acids 278--527). The results (Figure [2](#F2){ref-type="fig"}) mirror those seen with the two-hybrid assays and provide the first evidence that the interaction between the Cdc27 and Pol1 proteins is a direct one, rather than being mediated via a third protein or set of proteins, such as one of the other subunits of the Pol α-primase or Pol δ complexes.
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Direct interaction between recombinant Pol1 (278--512) and Cdc27 proteins.A.**Expression of H6-Pol1 (278--512) in *E. coli*detected by immunoblotting using anti-MRGS antibodies. Total protein extracts prepared from cells carrying the control plasmid pQE32-Pol1 (278--512)-REV (lanes 1, 2), in which the relevant *pol1*^+^sequence is present in the wrong orientation, or pQE32-Pol1 (278--512) (lanes 3, 4), either before (lanes 1, 3) or after (lanes 2, 4) induction of protein expression by addition of IPTG. The position of the 25 kDa marker is shown. **B.**Purified GST and GST -- Cdc27 (273--352) proteins detected by Coomassie staining following SDS-PAGE. **C.**Binding assays using GST -- Cdc27 fusion proteins. Bound H6-Pol1 was detected by Western blotting as in part A. Lanes 1, 3, 5: binding of H6-Pol1 to GST. Lanes 2, 4, 6: binding of binding of H6-Pol1 to GST -- Cdc27 (273--352). In each case the bound fraction corresponds to approximately 15--20% of the input. Lanes 1 and 2: assay performed in PBS containing 0.1% Triton X100; lanes 3 and 4: PBS containing 0.25% Triton X100; lanes 5 and 6: PBS containing 0.5% Triton X100. In the absence of detergent Pol1 binds equally well to both GST and GST-Cdc27 (273--352) proteins.
:::
:::
Pol1 binding site sequence
--------------------------
Previously, Cdc27 homologues from fission yeast \[[@B11],[@B18]\], budding yeast \[[@B19],[@B20]\], human, mouse \[[@B21]\], and *Xenopus*\[[@B15]\] have been identified and characterised. To extend this family, BLAST \[[@B22]\] and ψ-BLAST searches \[[@B23],[@B24]\] were performed to identify putative Cdc27 family members from other eukaryotic species. In all, 24 protein sequences were identified, from organisms as diverse as vertebrates, worms, fungi and plants. Protein sequence alignments (Figure [3](#F3){ref-type="fig"}) of the C-terminal regions of these proteins identified a set of highly conserved amino acids that form a putative [D]{.underline}NA [p]{.underline}olymerase [i]{.underline}nteraction [m]{.underline}otif (or DPIM) with the consensus sequence D(D/E)-G \-- (V/I)(T/S). The sequences flanking the DPIM are generally highly charged in character and some sequence conservation (particularly of charged amino acids) is apparent in these regions. Mutagenesis and peptide binding studies (described below) suggest that, in addition to the DPIM, some of these sequences may also play a role in binding Pol1 (see Discussion).
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Sequence alignment of Pol1 interacting region from fission yeast Cdc27 and homologues in other eukaryotic species.**Conserved residues are boxed. Abbreviations and NCBI sequence accession numbers: Sp (*S. pombe*, P30261), Sc (*S. cerevisiae*, P47110), Nc (*Neurospora crassa*, XP\_328704), Ca (*Candida albicans*, EAK99562), Gz (*Gibberella zeae*, XP\_384433), Um (*Ustilago maydis*, XP\_401381), Cn (*Cryptococcus neoformans*, EAL21672), Dh (*Debaryomyces hansenii*, CAG87841), Dm (*Drosophila melanogaster*, AAD38629), Ss (*Sus scrofa*, BF078337), Mm (*Mus musculus*, Q9EQ28), Rn (*Rattus norvegicus*, XP\_215011), Cf (*Canis familiaris*, CF411342), Hs (*Homo sapiens*, Q15054), Ci (*Ciona intestinalis*, AK114729), Fr (*Fugu rubripes*, protein sequence derived by translation of clone M001240 at <http://fugu.hgmp.mrc.ac.uk/>), Tn (*Tetraodon nigroviridis*, CAF97746), Dr (*Danio rerio*, AAH76031), Xl (*Xenopus laevis*, BAC82197), Att (*Ambystoma tigrinum tigrinum*, CN059104), Gg (*Gallus gallus*, BU121824 -- note the additional glutamate residue within the DPIM), Ce (*C. elegans*, Q21730), Os (*Oryza sativa*, NP\_913217), and At (*Arabadopsis thaliana*, C96815). Note that in only a few cases (Sp, Sc, Hs, Mm, Xl) have the identities of these putative Cdc27 proteins been confirmed via purification and characterisation of Pol δ. However, all the sequences shown possess a canonical PCNA binding motif at or near their C-terminal ends (Q \-- I \-- FF), in common with the *bone fide*Cdc27 proteins.
:::
:::
Interaction between human Pol1 and Cdc27 orthologues
----------------------------------------------------
To examine whether the Pol1-Cdc27 (Pol1-Pol32) interaction observed in the yeasts was also conserved in higher eukaryotes, the catalytic subunit of human Pol α and the human Cdc27 orthologue p66/KIA00039 were assayed for interaction using the two-hybrid system. Amino acids 291--540 of the human Pol α catalytic subunit, corresponding to the minimum Cdc27 binding region (amino acids 278--527) in fission yeast Pol1, were expressed as a LexA fusion alongside Gal4 activation domain fusions of either the entire C-terminal domain of human p66 (amino acids 253--466) or the C-terminal 111 amino acids only (356--466). Interactions were tested by β-galactosidase assay. Both p66 constructs bound to Pol α (Table [1B](#T1){ref-type="table"}), indicating that the DNA polymerase interaction is a conserved feature. In addition, it was observed that human p66 was able to interact with fission yeast Pol1 (Table [1C](#T1){ref-type="table"}), suggesting that conserved sequences (or structural features) within the Cdc27/p66 C-terminal region, such as the DPIM, were important for the interaction.
Mutational analysis of Pol1 binding site on Cdc27
-------------------------------------------------
Next, the importance for Pol1 binding of the conserved amino acids in the DPIM was examined. Eight mutant Cdc27 proteins (Cdc27-P1 through Cdc27-P7, and Cdc27-Q1, see Figure [4](#F4){ref-type="fig"}) were expressed as Gal4-Cdc27-273-352 fusion proteins and tested for their ability to bind to LexA-Pol1 (278--527) in the two-hybrid system. In each mutant protein one or more conserved amino acids is replaced with alanine. For example, the Cdc27-P1 mutant sees the conserved triplet DEE (residues 309--311) substituted with AAA. The results of this analysis are shown in Figure [4](#F4){ref-type="fig"}. Mutating the conserved amino acids of the DPIM (mutants Cdc27-P1, P2 and Q1) completely abolished Pol1 binding *in vivo*by Cdc27. Similar reductions were seen with the Cdc27-P4 and Cdc27-P6 mutants, where the mutated residues are located N-terminal and C-terminal to the DPIM respectively. The conserved amino acids of the DPIM are therefore essential for Pol1 binding by Cdc27, although sequences flanking the conserved motif also play a role. Only three of the eight mutant proteins (Cdc27-P3, P5 and P7) were able to interact with LexA-Pol1 (278--527), though the strength of the interaction was reduced to 30--40% of the wild-type value for Cdc27-P3 and Cdc27-P5 (both of which affect residues N-terminal to the conserved motif), and to \~ 10% of wild-type for Cdc27-P7 (located C-terminal to the conserved motif). Immunoblotting showed that all the mutant proteins were present in yeast protein extracts at the same level as the wild-type Gal4-Cdc27 fusion protein (data not shown).
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Mutational analysis of Pol1 binding domain on Cdc27.A.**Summary schematic of minimal Pol1 binding region on Cdc27, showing conserved amino acids (boxed) and the eight mutant alleles P1 through P7 and Q1. Each mutant sees two or three adjacent residues being replaced with the same number of alanines. In the case of P1 and P3 through P7, adjacent basic (P3, P4 and P5) or acidic (P1, P6 and P7) amino acids are mutated. **B.**Quantitation of β-galactosidase activity in liquid cultures. Key: C (pACT2 vector); W (pACT-Cdc27-273-352), mutants P1 -- P7, plus Q1 (mutated forms of pACT-Cdc27-273-352); F (pACT-Cdc27, i.e. full-length Cdc27 fused to Gal4).
:::
:::
The observation that sequences flanking the DPIM play a role in Pol1 binding was supported by the results of studies performed with a nested set of overlapping 20 mer peptides derived from the *S. pombe*Cdc27 sequence (see Figure [5A](#F5){ref-type="fig"}). The peptides were tested for their ability to bind to an epitope-tagged form of the Pol1 protein (Pol1-13Myc, see Materials and methods) in fission yeast protein extracts. As can be seen in Figure [5B](#F5){ref-type="fig"}, Pol1-13myc is precipitated by peptides SpB, SpC and SpD, but not SpA and SpE. Peptide SpC spans the groups of basic amino acids N-terminal to the conserved motif, including those shown to be required for Pol1 binding as defined by the Cdc27-P3 and -P5 mutants above, but does not include the conserved sequence DEEGFLVT indicating that, in this *in vitro*situation, the conserved amino acids are not absolutely required for Pol1 binding, in contrast to what is observed *in vivo*. Similar results were observed with peptides derived for the human p66 protein sequence (Figure [5C](#F5){ref-type="fig"},[5D](#F5){ref-type="fig"}), again illustrating the potential for cross-species interaction between human p66 and fission yeast Pol1 (see Table [1C](#T1){ref-type="table"}).
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Peptide binding studies.A.**Schematic of the structure of the Cdc27 protein (upper part) with the sequences of overlapping peptides SpA-SpE shown beneath. **B.**Immunoblot using anti-myc mAb showing pull-down of Pol1-13Myc protein from fission yeast protein extracts by Sp series peptides. **C.**Sequences of human peptide set HsA-HsE corresponding exactly to SpA-SpE above. **D.**Immunoblot using anti-myc mAb showing pull-down of Pol1-13Myc protein from fission yeast protein extracts by Hs series peptides.
:::
:::
Mapping the Cdc27 binding site on Pol1
--------------------------------------
In an effort to map more precisely where within the Pol1 protein Cdc27 binds, several truncated forms of Pol1 (278--527) were constructed and tested as LexA fusions in the two-hybrid system (see Materials and methods for details). Removal of fifty amino acids from the N-terminus, or twenty amino acids from the C-terminus, of the Pol1 (278--527) protein was found to abolish the interaction with Cdc27 altogether (Table [1D](#T1){ref-type="table"}).
Previously, the isolation of a temperature-sensitive mutant *pol1*allele, *pol1-ts13*, had been reported \[[@B25]\]. Sequence analysis revealed that this allele differed from the wild-type *pol1*^+^by deletion of 9 bp from the ORF, resulting in loss of three amino acids (L470, S471, R472) from within the minimal Cdc27 binding domain defined above. In this study, the ability of a Pol1-TS13(278--527) bait to bind to Cdc27 was tested using the two-hybrid system. No interaction could be detected at a range of growth temperatures (Table [1D](#T1){ref-type="table"}, and data not shown), indicating that amino acids 470--472 are required for Cdc27 binding by Pol1(278--527). To test whether overproduction of Cdc27 might rescue the temperature-sensitive phenotype of *pol1-ts13*, these cells were transformed with plasmid pREP3X-Cdc27 \[[@B18]\] and transformants plated at restrictive and semi-restrictive temperatures on thiamine-free medium, to induce high-level expression of *cdc27*^+^from the thiamine-repressible nmt promoter. No suppression of the *pol1-ts13*phenotype was observed, however (data not shown). Indeed, no suppression was observed of any of three *pol1*alleles that were analysed in this way, the others being *pol1-1*\[[@B26]\] and *pol1-H4*\[[@B27]\].
Expression of DPIM mutants *in vivo*
------------------------------------
To assay the *in vivo*role of the DPIM in Cdc27, four of the eight DPIM mutant alleles (*cdc27-P1*through *cdc27-P4*) were cloned into plasmid pREP3X, 3\' to the repressible nmt1 promoter \[[@B28],[@B29]\], and transformed into a *cdc27*^+^/*cdc27::his7*^+^diploid strain. Transformant colonies were then induced to sporulate and the spores plated on media containing thiamine, to repress the nmt1 promoter. Under these conditions, residual low level expression from the nmt1 promoter ensures that the level of Cdc27 protein present in the cell is comparable to that seen in wild-type cells \[[@B12]\]. Analysis of the meiotic products showed that the mutant Cdc27 proteins were able to support growth of *cdc27Δ*haploid cells; indeed, no phenotypic defects were apparent (data not shown).
To confirm this, a fission yeast strain was constructed in which the endogenous *cdc27*^+^gene was precisely replaced with the *cdc27-Q1*mutant allele. In the Cdc27-Q1 mutant protein, the central five amino acids of the DPIM are replaced with alanine, resulting in loss of Cdc27-Pol1 interaction in the two-hybrid system (Figure [4](#F4){ref-type="fig"}). Construction of the *cdc27-Q1*strain was achieved by first replacing one copy of *cdc27*^+^with *ura4*^+^in a diploid strain, before then replacing the *cdc27::ura4*^+^allele with *cdc27-Q1*via 5-FOA counterselection in the *cdc27::ura4*^+^haploid carrying *cdc27-Q1*on a plasmid (see Materials and methods for details). PCR analysis of genomic DNA using primers specific for wild-type *cdc27*^+^and *cdc27-Q1*sequences (see Figure [6](#F6){ref-type="fig"} and legend) allowed the identification of haploid strains in which *cdc27-Q1*was correctly integrated at the endogenous *cdc27*^+^locus, and also allowed the *cdc27-Q1*mutant to be conveniently followed through genetic crosses.
::: {#F6 .fig}
Figure 6
::: {.caption}
######
**Construction and analysis of *cdc27-Q1*mutant yeast.A.**Upper part: Schematic of the *cdc27*^+^gene region (3.1 kb HindIII-BamHI region) showing location of oligonucleotides used for PCR amplification. (Key to oligonucleotides: A~w~= CDC27-Q1W-DIAG2, A~m~= CDC27-Q1M-DIAG2, B = CDC27-B, C = CDC27-SEQ2005, D = CDC27-H, X = CDC1-AB, and Y = CDC1-XY -- see Material and methods for sequences). The boxes indicate approximate positions of *cdc27*^+^exons. The NotI site shown is found in the *cdc27-Q1*allele only. Lower part: Genomic DNA prepared from wild-type (WT) and *cdc27-Q1*(Q1) strains was amplified using the primer pairs shown. The D+B PCR product from *cdc27-Q1*alone can be digested with NotI (data not shown). The primer pair X and Y amplify an unrelated region of genome, and were included as a control. Molecular weight markers (kb) are shown to the right of the gel. **B.**Wild-type (*cdc27*^+^, left) and *cdc27-Q1*(right) cells plated on YE medium and incubated for 3 days at 32°C. Mutation of the DPIM did not affect the efficiency of colony formation or growth rate. See text for details.
:::
:::
Phenotypic analysis of *cdc27-Q1*cells, in comparison to an otherwise isogenic *cdc27*^+^control, revealed the following: cells carrying the *cdc27-Q1*allele grew normally at a range of temperatures (18 -- 36.5°C), with a generation time indistinguishable from wild-type (110 minutes at 32°C in YE medium). At 32°C, *cdc27-Q1*cells underwent division at \~ 14.1 μm (compared to wild-type at \~ 14.4 μm). Further analysis showed that the *cdc27-Q1*cells were indistinguishable from wild-type in all respects examined, including responses to the DNA replication inhibitor hydroxyurea (HU) and the DNA damaging agents methylmethane sulphonate (MMS), camptothecin (CPT), bleomycin sulphate (BMS) and UV light (see Materials and methods for details). The kinetics of cell division arrest in response to HU were also examined, but again no difference was detectable between *cdc27-Q1*and wild-type strains (data not shown). That cell number increase is arrested in *cdc27-Q1*cultures following treatment with HU, and that both wild-type and *cdc27-Q1*cells become highly elongated under these conditions, is indicative of the presence of a functional DNA replication checkpoint in these cells. Taken together, these results strongly suggest that the Pol1-Cdc27 interaction does not play an essential role either during S-phase or in various DNA repair pathways in fission yeast.
The consequences of introducing additional mutations into the *cdc27-Q1*background were also investigated. *cdc27-Q1*was crossed to strains carrying mutations in the other three subunits of Pol δ (*pol3-ts3*, *cdc1-P13*, *cdm1Δ*) and in the helicase-endonuclease Dna2 and its associated protein Cdc24 (*dna2-C2*and *cdc24-M38*). \[[@B13],[@B18],[@B30]-[@B33]\]. The double mutant strains were then analysed as described in Materials and methods. In every case examined, the properties of the double mutant with *cdc27-Q1*were indistinguishable from the single mutant. The *cdc27-Q1*mutation was also combined with *rad3Δ*\[[@B34]\] and *cds1Δ*\[[@B27]\] alleles, to create *cdc27-Q1 rad3Δ*and *cdc27-Q1 cds1Δ*double mutants, both of which were viable. The Rad3 and Cds1 proteins are key components of various DNA structure checkpoints in fission yeast \[[@B35]\]. That the *cdc27-Q1*double mutants were viable indicates that *cdc27-Q1*cells do not require the presence of a functional checkpoint for viability. When wild-type cells are treated with hydroxyurea, activation of Cds1 results in cell cycle arrest and replication fork stabilisation. In the absence of Cds1, however, replication forks are believed to collapse, resulting in loss of viability \[[@B35]\]. The *rad3Δ*and *cds1Δ*double mutants were tested for their sensitivity to HU and CPT, but as before, no differences were observed between single and double mutants with *cdc27-Q1*(data not shown).
Discussion
==========
In fission yeast, DNA polymerase δ is a multisubunit complex comprising a large catalytic subunit that is required for chromosomal replication as well as three smaller subunits, two of which are also essential for cell viability \[[@B13],[@B18],[@B36]\]. Understanding how the four subunits of the complex interact with one another, and how they interact with other components of the replication machinery, is an important goal. Once interactions are identified and the sites of interaction precisely mapped, reverse genetic analysis allows determination of the effects of disrupting individual protein-protein interactions on replisome function.
In this paper we show that the C-subunit of fission yeast Pol δ, Cdc27, is able to interact both *in vivo*and *in vitro*with Pol1, the catalytic subunit of Pol α. The Pol1 binding site on Cdc27 has been mapped by *in vivo*and *in vitro*approaches and a region of 40 amino acids, from amino acids 293 -- 332, has been shown to be sufficient for binding. Protein sequence alignments of Cdc27 homologues across this 40 amino acid region identify a short protein sequence motif (D \-- G \--VT) that is highly conserved and which is essential for Pol1 binding. This DNA polymerase interaction motif (DPIM) is flanked by relatively highly charged sequences. Ten basic amino acids are found flanking the DPIM in the fission yeast Cdc27 protein sequence, nine of which are located N-terminal to the central conserved motif. Similarly, there are ten acidic amino acids, all of which lie either within the conserved DPIM or C-terminal to it. Our mutagenesis data clearly implicates several of these charged groups in the binding to Pol1 (Figure [4B](#F4){ref-type="fig"}). Secondary structure predictions suggest that the conserved DPIM is likely to form a turn or loop, raising the possibility that the sequences on either side of the conserved motif interact with one another. The distribution of positively and negatively charged amino acids in most, though not all, of the Cdc27 homologues in other species (Figure [3](#F3){ref-type="fig"}) is similar to that in fission yeast. The most obvious exception to this is in the *C. albicans*Cdc27 protein, where six acidic amino acids are found N-terminal, and seven basic amino acids C-terminal, to the DPIM. This organisation of charged residues appears again to be compatible with the notion the sequences on either side of the DPIM may interact with one another.
The Cdc27 protein has an elongated shape, with a frictional ratio of 1.85 \[[@B12]\]. The same is true of its budding yeast orthologue Pol32 which has a frictional ratio of 2.22 \[[@B37]\]. The elongated shape of Cdc27 is due to the C-terminal region of the protein. The N-terminal Cdc1 binding domain, comprising amino acids 1--160 \[[@B11]\], behaves in solution as a globular protein \[[@B12]\]. The protein-protein interaction motif described in this study is the second to be mapped to the extended C-terminal domain. Previously we showed that Cdc27 interacts with PCNA via a conserved sequence at the extreme C-terminus of the protein \[[@B11]\]. Our results indicated that this interaction was essential for cell viability. Consistent with this, Cdc27-PCNA contact is vital for maximal polymerase processivity *in vitro*. However, recently we obtained evidence that the Cdc27-PCNA interaction is a non-essential one (H. Tanaka, G. Ryu, Y.-S. Seo and S.M., submitted).
Recently, the results of a deletion analysis of Pol32, the budding yeast orthologue of Cdc27, were reported \[[@B17]\], including analysis of Pol32-Pol1 interaction. The results of these studies showed that amino acids 270--309 of Pol32 were required for Pol1 binding in the two-hybrid system. This region corresponds to amino acids 286--325 in fission yeast Cdc27 and includes the conserved DPIM sequence (Figure [3](#F3){ref-type="fig"}). Deletion of amino acids 250--289 (266--305 in Cdc27) greatly reduced Pol1 binding, consistent with the results reported here with the Cdc27-P3, -P4 and -P5 mutants (Figure [4](#F4){ref-type="fig"}), as did deleting amino acids 310--343 (326--363), consistent with the Cdc27-P7 mutant result. (As expected, the latter Pol32 deletion also disrupted binding to PCNA.) In addition, it was shown that that Pol α-primase and Pol δ could be co-immunoprecipitated following mixing of the purified complexes, but that this interaction was effectively abolished when Pol δ contained the 270--309 deletion of Pol32 rather than the full-length protein \[[@B17]\].
In this paper, cells expressing the DPIM mutant protein Cdc27-Q1 were shown to be no more sensitive then wild-type to HU, MMS, CPT, UV and BMS (see Results), while analysis of budding yeast cells expressing Pol32 lacking the DPIM (270--309 deletion) showed them to be no more sensitive than wild-type to HU and UV, and to show normal rates of mutagenesis following UV exposure \[[@B17]\]. Indeed, in all situations examined, *S. pombe cdc27-Q1*cells were indistinguishable from wild-type. No genetic interactions were observed between *cdc27-Q1*and various other DNA replication or checkpoint mutants, including the key checkpoint kinase Rad3.
In conclusion, while the interaction between Pol α-primase and Pol δ mediated via Cdc27 could play an important role in coordinating the events of lagging strand synthesis, we have yet to obtain any evidence that this is the case. This raises two possibilities. First, that the observed interaction does not play an important role in chromosomal replication. We believe that this is unlikely to be the case, given the high degree of conservation of the DPIM sequence across evolution, in a region of the Cdc27 protein that is very poorly conserved at the primary sequence level. The second possibility is that there are multiple redundant interactions within the lagging strand machinery. In this case, an important role for the Cdc27-Pol1 interaction might only be uncovered when another protein-protein interaction is perturbed, in which case, the *cdc27-Q1*mutation might be expected to be synthetically lethal or sick with a mutant that disrupted the overlapping redundant function. With the genetic tools available in yeast, this is a hypothesis that is readily testable.
Conclusions
===========
In fission yeast, interaction between Pol α and Pol δ is mediated, at least in part, by direct binding of the Pol δ C-subunit Cdc27 to the Pol α catalytic subunit Pol1, and requires the presence of a short sequence motif (DPIM) in the C-terminal region of Cdc27. The DPIM is conserved in all known Cdc27 orthologues. Despite this, it has not been possible to identify any phenotypic consequences associated with deletion of the DPIM sequence, raising the possibility that the observed interaction does not play a crucial role *in vivo*.
Methods
=======
Yeast strains, media and methods
--------------------------------
All fission yeast strains were as described previously, except for *pol1-H4*\[[@B27]\] and *pol1-1*\[[@B26]\], which were obtained from Dr J. Hayles (CR-UK, London, U.K.), and *pol1-ts13*\[[@B25]\], which was obtained from Dr T.S.F. Wang (Stanford, U.S.A.). Note that this allele was originally designated *polα-ts13*but is correctly renamed here to bring it in line with accepted nomenclature standards (see <http://www.sanger.ac.uk/Projects/S_pombe/SP_Name_FAQ.shtml> for further details). *S. pombe*media and techniques were essentially as described \[[@B38]\], with the following exceptions. Routine transformation of *S. pombe*was carried out by electroporation \[[@B39]\], whereas transformation for PCR-based gene targeting was accomplished using a modified lithium acetate method \[[@B40]\]. For two-hybrid analysis. *S. cerevisiae*CTY10-5d (*MATa ade2 met- trp1-901 leu2-3-112 his3-Δ200 gal4-gal80-URA3::lexA-LacZ*) was used \[[@B18]\]. *S. cerevisiae*was cultured in YPDA and SD medium.
Plasmids for two-hybrid assay
-----------------------------
Two-hybrid interactions were monitored using the Gal4 transcription activation domain (prey) plasmid pACT2 (Clontech), the LexA DNA binding domain (bait) plasmid pBTM116, and *S. cerevisiae lexA op-lacZ*strain CTY10-5d, as described previously \[[@B18]\].
The Pol1 bait plasmid pBTM116-Pol1-(278--527) was constructed by amplifying sequences encoding amino acids 278--527 from plasmid pTZ19R-Pol1 (prepared by subcloning a 5925 bp SmaI-PstI fragment encompassing the entire *pol1*^+^gene from *S. pombe*cosmid SPAC3H5, see <http://www.sanger.ac.uk/Projects/S_pombe>, into plasmid pTZ19R) using oligonucleotides POL15 (5\'-GTGTGGTTTG[GGATCC]{.underline}CCCTATCACCAATGACACCTTTA-3\') and POL13-2 (5\'-GTGTGGTTTG[GGATCC]{.underline}TACATCACCGTCATTGGAGGCGT-3\'), restricting the PCR product with BamHI (sites underlined), cloning to pTZ19R (Fermentas) and sequencing to confirm the absence of errors. The 763 bp BamHI fragment was then transferred to pBTM116, to generate pBTM116-Pol1(278--527). Plasmid pBTM116-Pol1-TS13(278--527) was produced in a similar manner, except that the starting PCR template was genomic DNA prepared from *pol1-ts13*cells \[[@B25]\], the final product thereby containing a 9 bp deletion in the *pol1*^+^ORF, resulting in the loss of coding capacity for amino acids L470, S471 and R472. Plasmids expressing C-terminally truncated Pol1(278--527) proteins were constructed by PCR amplification using oligo POL15 in conjunction with POL25-3 (5\'-GTGTGGTTTG[GGATCC]{.underline}TAAGGACCCATAACTCTTCTACT-3\'), POL13-487 (5\'-GTGTGGTTTG[GGATCC]{.underline}TAAAAATTTGGTTGTTGTATTTT-3\'), POL1-497 (5\'-GTGTGGTTTG[GGATCC]{.underline}TACCGGCACCAACTAGCATTTTT-3\') or POL1-507 (5\'-GTGTGGTTTG[GGATCCT]{.underline}AGTTCTGAGGTGACGAACATCC-3\'), cloning directly to pBTM116 following BamHI cleavage of the PCR product, and sequencing. The resulting constructs encoded the following proteins as Lex A fusions: Pol1(278--477), Pol1(278--487), Pol1(278--497) and Pol1(278--507). A construct expressing an N-terminally truncated Pol1(278-527) protein, designated Pol1(328--527), was generated by amplification using POL15-2 (5\'-GTGTGGTTTG[GGATCC]{.underline}CCGGCTCATTGTGTCTATTTGGC-3\') with POL13-2 (above). Amplification with POL15-2 and POL25-3 generated a construct with the potential to express a LexA-Pol1 protein truncated at both ends: LexA-Pol1(328--477).
Sequences encoding the C-terminal region from human p66/KIAA0039 were amplified by PCR from plasmid pET19b-p66 (a gift of Dr P. Hughes, Villejuif, France) using either oligo P66-51 (5\'-GTGTGGTTTG[GGATCC]{.underline}CCTCAGAACAAGCAGTGAAAGAA-3\') or P66-52 (5\'-GTGTGGTTTG[GGATCC]{.underline}CGTCTCCACCTCTTGAACCAGTG-3\') with P66-3 (5\'-GTGTGGTTTGGGATCCTTGGTCTTCACCCTTGACCACTC-3\'). The PCR products were then restricted with BamHI, cloned to pACT2 and sequenced. The resulting plasmids encode, as Gal4 AD fusions, either the entire C-terminal domain of p66/KIAA0039 (amino acids 253--466, oligos P66-51 and P66-3) or a shorter region (amino acids 356--466, oligos P66-51 and P66-3). Sequences encoding the interacting domain of the catalytic domain of human Pol α were amplified from plasmid pBR322-Pol α (a gift of Dr T. Wang, Stanford, USA) using oligos HPOL1-5 (5\'-GTGTGGTTTG[GGATCC]{.underline}CCAAAGGGACCGTGTCCTACTTA-3\') and HPOL1-3 (5\'-GTGTGGTTTG[GGATCC]{.underline}TACATCACGACAAGCGGTGGTGG-3\'), restricted with BamHI, cloned to pBTM116 and sequenced. The resulting plasmid encodes amino acids 291--540 of the human protein as a LexA fusion.
The Cdc27 prey plasmids were constructed in pACT2 (Clontech) and have been previously described, with the exception of mutants Cdc27-273-352-P1 through -P7, and Cdc27-273-352-Q1. The first four of these (P1-P4) were generated by PCR amplification from the pREP3X-Cdc27-P1 to -Cdc27-P4 plasmids described below. Oligonucleotides for PCR, with BamHI sites underlined: PMUT5 (5\'-GTTTGTTGGT[GGATCC]{.underline}CCACCGAAGCAAAATCTGCTGCA-3\') and PMUT3 (5\'-GTTGTGGGTG[GGATCC]{.underline}TACTTCTTAGTTGCAATGTTTAC-3\'). Mutant Q1 was generated by PCR amplification with same primers but using pHBLA-Cdc27-Q1 (below) as template. Plasmids expressing mutants P5-P7 were generated by overlap extension PCR using PMUT5 and PMUT3 together with the following mutagenic oligonucleotides (shown with mutated sequence underlined in top strand oligo): CDC27-P5F (5\'-AAAGAAAAGTT[GCAGCG]{.underline}TACGCGACAACGAAAG-3\'), CDC27-P5R (5\'-CTTTCGTTGTCGCGTACGCTGCAACTTTTCTTT-3\'), CDC27-P6F (5\'-TTGGTTACTAAG[GCAGCAGCA]{.underline}GTCTGGGAATCA-3\'), CDC27-P6R (5\'-TGATTCCCAGACTGCTGCTGCCTTAGTAACCAA-3\'), CDC27-P7F (5\'-GAATCATTTTCT[GCAGCTGCA]{.underline}AACATCTCAACT-3\'), CDC27-P7R (5\'-AGTTGAGATGTTTGCAGCTGCAGAAAATGATTC-3\').
Two-hybrid assays
-----------------
Quantitative data for β-galactosidase activity was obtained as described previously \[[@B18]\] and is expressed in Miller units \[[@B41]\]. To analyse prey protein levels, yeast total protein extracts, prepared from the same cultures used for β-galactosidase assay, were subjected to SDS-PAGE and immunoblotted using antibodies against the HA epitope (12CA5, Roche Applied Science) present in proteins expressed from pACT2.
Expression and purification of recombinant H6-Pol1 (278--527)
-------------------------------------------------------------
The Pol1 (278--527) domain was expressed in *E. coli*with an N-terminal MRGSH~6~-tag to facilitate purification and detection. To achieve this, the 763 bp Pol1 (278--527)-encoding BamHI fragment described above was cloned into pQE32 (Qiagen). The resulting plasmid, pQE32-Pol1 (278--527), was transformed into *E. coli*M15 (pREP4) and recombinant protein expression induced in 500 ml cultures (OD~600\ nm~of 0.6) by addition of IPTG to a final concentration of 1 mM. Following incubation for 4 hours at 37°C, the cells were pelleted and the pellet resuspended in 40 ml of ice-cold buffer A (150 mM NaCl, 50 mM Tris-HCl pH 8.0, 20 mM imidazole) containing EDTA-free Complete™ inhibitors (Roche Applied Science) and 1 mM PMSF. The cells were lysed by sonication, then centrifuged at 25000 g for 15 minute at 4°C. The soluble supernatant was added to 4 ml of 50% (v/v) Ni-NTA agarose (Qiagen) in buffer A, mixed on a wheel at 4°C for one hour, then packed into a disposable chromatography column at 4°C. The column was drained of the flow-through and subsequently washed with 100 ml of buffer B (1 M NaCl, 50 mM Tris-HCl pH 8.0, 20 mM imidazole) containing Complete™ inhibitors and PMSF, then 100 ml of buffer A containing Complete™ inhibitors and PMSF, before the bound H6-Pol1 protein was eluted using 4 ml of buffer C (250 mM imidazole, 50 mM Tris-HCl pH 8.0). Following elution, samples were analysed by SDS-PAGE and the peak fractions pooled and dialysed overnight in PBS at 4°C. Protein concentration was determined by BCA assay (Pierce) versus BSA standards.
Binding assays
--------------
GST and GST-Cdc27-273-352 fusion proteins were prepared as described previously \[[@B11]\] from plasmid pGEX6P-1B, a modified version of pGEX6P-1 (Amersham Pharmacia) in which the reading frame of the polylinker is altered (S.M., unpublished). Sequences encoding wild-type and mutant forms of Cdc27-273-352 were subcloned into this vector from the equivalent pACT2 two-hybrid constructs described above. Binding assays were performed by mixing \~ 30 pmol of GST or GST-Cdc27-273-352 with 90 pmol of H6-Pol1 in PBS containing 1.0% Triton X100 for 1 hour at 4°C on a rotating wheel. The GST proteins were then precipitated using Glutathione Sepharose™ 4 Fast Flow resin (Amersham Biosciences) for 30 minutes at 4°C on a rotating wheel. Following extensive washing with PBS containing 0.1 -- 1.0% Triton X100, SDS-PAGE sample buffer was added and the samples heated at 95°C for 5 minutes. Bound H6-Pol1 was visualised following electrophoresis of 15% SDS-PAGE gels by PAGE Blue G90 (Fluka) staining or by immunoblotting with anti-MRGS antibodies (Qiagen).
Construction of Pol1-13Myc strain
---------------------------------
PCR-based gene targeting was used to tag the chromosomal *pol1*^+^locus in an otherwise wild-type *leu1-32 ura4-D18*h^-^strain with sequences encoding thirteen copies of the 9E10 (c-myc) mAb epitope, such that the encoded Pol1 protein (termed Pol1-13Myc) carries the 9E10 epitopes at its C-terminus. Oligonucleotides for amplification from pFA6a-13Myc-kanMX6 \[[@B40]\] were as follows: POL1-13MYC-5 (5\'-TGCCATCAACAAAAATATCTCTCGAATAATGAACAAAAATGCGCGTGAATTTGTAGATATGGGACTGATATTTTCATCG[CGGATCCCCGGGTTAATTAA]{.underline}-3\', with plasmid-specific sequence underlined) and POL1-13MYC-3 (5\'-GGCAATTCCCAAGTCTTTGAAACAGGTATTCCCATCAACATTTCTTGTACTGCATGAGCAAATATCTGTTCGAGGTGTC[GAATTCGAGCTCGTTTAAAC]{.underline}-3\'). Following transformation of \~ 10 μg of PCR product, correct G418-resistant integrants were identified by PCR amplification of genomic DNA using primer FCGPOL1-5 (5\'-ATGTCGTGGAAGCGTTCATT-3\') within the *pol1*^+^ORF and KAN269R (5\'-GATCGCAGTGGTGAGTAACCATGCATCATC-3\') within the *kanMX6*cassette, and by Western blotting using the mouse anti-Myc mAb 9E10 (Roche).
Peptide binding assays
----------------------
Peptides were synthesized by Mimotopes (Australia). All peptides were synthesized to contain the sequence biotin-SGSG at the N-terminus. Peptide sequences: SpA AAPDEPQEIIKSVSGGKRRG; SpB PQEIIKSVSGGKRRGKRKVK; SpC KSVSGGKRRGKRKVKKYATT; SpD GKRRGKRKVKKYATTKDEEG: SpE KRKVKKYATTKDEEGFLVTK; HsA PKTEPEPPSVKSSSGENKRK; HsB EPPSVKSSSGENKRKRKRVL; HsC KSSSGENKRKRKRVLKSKTY; HsD ENKRKRKRVLKSKTYLDGEG; HsE RKRVLKSKTYLDGEGCIVTE. For binding assays, the peptides were initially bound to 10 μl Streptavidin-agarose beads in PBS in a final volume of 100 μl at room temperature for 1 hour with gentle agitation. After binding, the beads were washed 3 times with 100 μl PBS. Protein extracts (100 μg of fission yeast extract) were added to the beads, and incubated at 4°C for 1 hr with gentle agitation. The beads were pelleted by centrifugation at 2000 rpm for 3 minutes, and subsequently washed extensively with NP40 buffer (50 mM Tris-HCl pH 8.0, 150 mM NaCl, 1% (v/v) NP-40), prior to final resuspension with 20 μl of SDS loading buffer. Following SDS-PAGE, Pol1-13Myc was visualised by immunoblotting with the mouse anti-Myc mAb 9E10 (Roche). Fission yeast protein extracts were prepared from 8 × 10^7^cells in mid-exponential growth in EMM medium. Cells were harvested, washed once in STOP buffer (150 mM NaCl, 50 mM NaF, 10 mM EDTA, 1 mM NaN~3~) and disrupted in buffer A (10 mM sodium phosphate buffer pH 7.0, 1% Triton, 0.1% SDS, 1 mM EDTA, 150 mM NaCl, 1 mM PMSF) supplemented with Complete™ protease inhibitors (Roche), using a bead beater (Hybaid Ribolyser). Protein concentrations were determined at 595 nm using the BioRad protein assay reagent according to the manufacturer\'s instructions, with BSA as standard.
Plasmids expressing DPIM mutants
--------------------------------
Plasmids pREP3X-Cdc27-P1 to -Cdc27-P4 were constructed by cloning mutagenised *cdc27*^+^cDNAs (SalI-BamHI) from pTZ19R-Cdc27-cDNA \[[@B18]\] into pREP3X. Mutagenesis of pTZ19R-Cdc27-cDNA was accomplished using the MutaGene II mutagenesis kit (BioRad) according to the manufacturer\'s instructions. Oligonucleotides with mutated residues underlined: CDC27-P1 (5\'-AAAAGTACGCGACAACGAAA[GCCGCAGCA]{.underline}GGATTCTTGGTTACTAAGG-3\'), CDC27-P2 (5\'-CGACAACGAAAGATGAAGAA[GCCGCC]{.underline}TTGGTTACTAAGGAAGAAG-3\'), CDC27-P3 (5\'-TCAAATCCGTATCCGGTGGA[GCCGCAGCA]{.underline}GGGAAAAGAAAAGTTAAAAAG-3\'), CDC27-P4 (5\'-CCGGTGGAAAGAGACGTGGG[GCCGCAGCA]{.underline}GTTAAAAAGTACGCGACAAC-3\'). The resulting mutant alleles were tested for function by transforming a *cdc27*^+^*/cdc27::his7*^+^*leu1-32/leu1-32 ura4-D18/ura4-D18 his7-366/his7-366 ade6-M210/ade6-M216 h*^-^/*h*^+^diploid \[[@B11]\], transferring the transformants onto malt extract medium to induce sporulation, before finally plating the helicase-treated spores onto EMM plates supplemented with 5 μg/ml thiamine with/without histidine, at 32°C.
Construction of DPIM mutant strain
----------------------------------
A 3.1 kb HindIII-BamHI genomic DNA fragment carrying the *cdc27*^+^gene was first cloned into pTZ19R to make pHB-Cdc27. This vector was then modified by addition of the *S. cerevisiae LEU2*gene and *S. pombe ars1*, to make pHBLA-Cdc27. The *cdc27*^+^gene was then subjected to oligonucleotide-directed *in vitro*mutagenesis using the QuikChange method (Stratagene) with oligonucleotides CDC27-QC1 (5\'-GTACGCGACAACGAAA[GCTGCAgcggccgcC]{.underline}TTGGTTACTAAGGAAGAAG-3\', with mutated sequence underlined and NotI site in lower case) and CDC27-QC2 (5\'-CTTCTTCCTTAGTAACCAA[GgcggccgcTGCAGC]{.underline}TTTCGTTGTCGCGTAC-3\'), to create plasmid pHBLA-Cdc27-Q1. This was then was transformed into a *cdc27*^+^/*cdc27::ura4*^+^diploid, and *cdc27::ura4*^+^(pHBLA-Cdc27-Q1) haploids obtained following sporulation and regrowth. These were then plated on YE plates containing 1 mg/ml 5-FOA. 5-FOA resistant colonies were identified, purified and characterised by PCR amplification of genomic DNA (prepared using the method of Bähler and coworkers \[[@B40]\]), to ensure loss of the chromosomal *ura4*^+^marker, its replacement with *cdc27-Q1*, and loss of the pHBLA-Cdc27-Q1. Oligonucleotides for diagnostic PCR shown in Figure [6](#F6){ref-type="fig"}: CDC27-Q1W-DIAG2 (5\'-GCGACAACGAAA[GATGAAGAAGGATTC]{.underline}-3\'; note that the underlined region anneals only to the wild-type *cdc27*^+^and not *cdc27-Q1*); CDC27-Q1M-DIAG2 (5\'-GCGACAACGAAA[GCTGCAGgcggccgcC]{.underline}-3\'; underlined region anneals only to *cdc27-Q1*and not to *cdc27*^+^; NotI site in lower case); CDC27-H (5\'-ACTGGTAGAATTGCGTTCGCGCTC-3\'); CDC27-B (5\'-TCTAGGATCAGAGTGAACTGATTG-3\'); CDC27-SEQ2005 (5\'-AGGTTGTACTAACATTAACAG-3\'). The resulting strain, *cdc27-Q1 leu1-32 ura4-D18 his7-366 ade6-M216*h^-^was then analysed alongside the wild-type *leu1-32 ura4-D18 his7-366 ade6-M216*h^-^, as described below.
Phenotypic analysis
-------------------
Cells were grown to mid-exponential phase (\~ 5 × 10^6^cells/ml) in YE medium and \~ 2000 cells plated on YE medium supplemented with varying concentrations of hydroxyurea (2.5, 5, 7.5, 10, 12.5 mM), methylmethanesulphonate (0.01, 0.005, 0.0025, 0.001, 0.0005, 0.00025, 0.0001%), bleomycin sulphate (2.5, 5, 7.5, 10, 12.5 mU/ml), or camptothecin (4, 4.5, 5, 5, 5.5, 6, 6.5, 7, 7.5, 8 μM). Cells were also plated on plates containing sub-lethal doses of both HU and CPT, specifically 7.5 mM HU with either 5.5 or 6 μM CPT, or 10 mM HU with either 5.5 or 6 μM CPT. To analyse the effects of HU treatment in liquid culture, HU was added to EMM medium to a final concentration of 12 mM. Cell number per ml of culture was monitored using a Coulter Z1 electronic particle counterTo test sensitivity to UV, 1000 cells were plated on EMM plates, allowed to dry for 20 minutes, then irradiated using either a Stratalinker UV source (Stratagene) over the range 0 -- 250 J/m^2^. Following UV treatment, plates were placed immediately in the dark to avoid photoreversal. For all treatments, the efficiency of colony formation was determined after 4 days growth at 32°C. Growth rate (YE medium, 32°C) was determined by cell counting using a particle counter. Cell length at cell division was determined using a graduated eyepiece.
DPIM double mutants
-------------------
Double mutants were constructed by standard methods, with the *cdc27-Q1*allele being identified by PCR analysis of genomic DNA using CDC27-Q1W-DIAG2 and CDC27-Q1M-DIAG2 oligonucleotides described above. Double mutants were created with the following: *pol3-ts3*\[[@B30]\], *cdc1-P13*\[[@B42]\], *cdm1::ura4*^+^\[[@B13]\], *cdc24-M38*\[[@B43]\], *dna2-C2*\[[@B31]\], *rad3::ura4*^+^\[[@B34]\] and *cds1::ura4*^+^\[[@B27]\].
List of abbreviations used
==========================
DPIM (DNA polymerase interaction motif); PCNA (proliferating cell nuclear antigen).
Authors\' contributions
=======================
In Edinburgh, SM conceived of the study, performed some of the experimental work and prepared and revised the final manuscript, while FG performed the remainder of the experimental work. In Dundee, EW and JRGP designed and carried out the peptide binding studies. All four authors read and approved the manuscript.
Acknowledgements
================
We would like to thank our friends and colleagues for their assistance, in particular those who supplied reagents for use in this study. SM and FG are supported by a Wellcome Trust Senior Research Fellowship in Basic Biomedical Science awarded to SM. JRGP was supported by a grant from the Association of International Cancer Research (AICR) awarded to EW who was herself supported by Cancer Research UK.
|
PubMed Central
|
2024-06-05T03:55:51.729144
|
2004-12-3
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545490/",
"journal": "BMC Mol Biol. 2004 Dec 3; 5:21",
"authors": [
{
"first": "Fiona C",
"last": "Gray"
},
{
"first": "J Richard G",
"last": "Pohler"
},
{
"first": "Emma",
"last": "Warbrick"
},
{
"first": "Stuart A",
"last": "MacNeill"
}
]
}
|
PMC545593
|
Background
==========
The completion of the sequencing of the human, mouse and other genomes has enabled efforts to extensively annotate these genomes using a combination of computational and experimental approaches. Generating a comprehensive list of transcripts coupled with basic information on where the different transcripts are expressed is an important first step towards annotating a genome once it has been fully sequenced. The task of identifying the transcribed regions of a sequenced genome is complicated by the fact that transcripts are composed of multiple short exons that are distributed over much larger regions of genomic DNA. This challenge is underscored by the widely divergent predictions of the number of genes in the human genome. For example, direct clustering of human expressed sequence tag (EST) sequences has predicted as many as 120,000 genes \[[@B1]\], whereas sampling and sequence-similarity-based methods have predicted far lower numbers, ranging from 28,000 to 35,000 genes \[[@B2]-[@B5]\], and a hybrid approach has suggested an intermediate number \[[@B6]\]. Furthermore, the availability of a completed draft sequence of the human genome has yielded neither a proven method for gene identification nor a definitive count of human genes. Two initial analyses of the human genome sequence that used strikingly different methods both suggested the human genome contains 30,000 to 40,000 genes \[[@B2],[@B3]\]. However, a direct comparison of the predicted genes revealed agreement in the identification of well-characterized genes but little overlap of the novel predictions. Specifically, 84% of the RefSeq transcripts agreed with fewer than 20% of the predicted transcripts matching between the two analyses. This result suggests that, individually, these datasets are incomplete and that the human genome potentially contains substantially more unidentified genes \[[@B7]\].
Several recent studies have highlighted the limitations of relying solely on computational approaches to identify genes in the draft of the human genome \[[@B8]-[@B13]\]. Furthermore, substantial experimental data from direct assays of gene expression provide evidence for many genes that would not have been recognized in the analyses just mentioned. Saha and colleagues used a new LongSAGE technology to provide strong evidence that there are thousands of genes left to be discovered in the human genome \[[@B9]\]. Specifically, they sequenced over 27,000 tags from a human colorectal cell line that collapsed down to 5,641 unique groups. Interestingly, only 61% (3,419) of the tags matched known or predicted genes, whereas 10% (575) matched novel internal exons and 14% (803) appear to represent completely novel genes \[[@B9]\]. They extrapolate from these data to predict as many as 7,500 exons from previously unrecognized genes. A recent analysis by Camargo *et al*. \[[@B8]\] also indicates that we are far from defining a complete catalog of human genes based on the analysis of 700,000 ORESTES (Open Reading Frame ESTs) that were recently released into GenBank. Finally, Kapranov and colleagues recently constructed genome-tiling arrays for human chromosomes 21 and 22 to comprehensively query transcription activity over 11 human tissues and cell lines \[[@B10]\]. They detected significant, widespread expression activity over a substantial proportion of these chromosomes outside of all known and predicted gene regions.
Most current methods in widespread use for identifying novel genes in genomic sequence depend on sequence similarity to expressed sequence and protein data. For example, *ab initio*prediction programs operate by recognizing coding potential in stretches of genomic sequence, where the recognition capability of these programs depends on a training set of known coding regions \[[@B14]\]. Therefore, genes identified by *ab initio*prediction programs or assembled from EST data are also inaccurate or incomplete much of the time \[[@B10]-[@B12]\]. While *ab initio*prediction programs perform well at identifying known genes, predictions that do not use existing expressed sequence and protein data often miss exons, incorrectly identify exon boundaries, and fail to accurately detect the 3\' and 5\' untranslated regions UTRs \[[@B14]\]. Similarly, EST data may be biased towards the 3\' or 5\' UTR \[[@B13]\]. These deficiencies are addressed in full-length gene cloning strategies \[[@B13]\], but cloning is still a laborious process which could be accelerated if we were able to start from a more accurate view of a putative gene \[[@B13]\].
Recently, several groups have used microarrays to test computational gene predictions experimentally and to tile across genomic sequence to discover the transcribed regions in the human and other genomes \[[@B10]-[@B12],[@B15]-[@B17]\]. These array-based approaches detected widespread transcriptional activity outside of the annotated gene regions in the human, *Arabidopsis thaliana*and *Escherichia coli*genomes. The recent sequencing and analysis of the mouse genome indicates extensive homology between intergenic regions of the human and mouse genomes, further highlighting the potential for other classes of transcribed regions \[[@B18]\]. Interestingly, recent tiling data suggests that many of these conserved intergenic regions are transcribed \[[@B15],[@B16]\].
In the study reported here, we describe hybridization results generated from two large microarray-based gene-expression experiments involving predicted transcript arrays spanning the entire human genome and a comprehensive set of genomic tiling arrays for human chromosomes 20 and 22. mRNA samples collected from a diversity of conditions were amplified using a strand-specific labeling protocol that was optimized to generate full-length copies of the transcripts. Analyses of the resulting hybridization data from both sets of arrays revealed widespread transcriptional activity in both known or high-confidence predicted genes, as well as regions outside current annotations. The results from this analysis are summarized with respect to published genes on chromosomes 20 and 22 in addition to our own extensive set of genome alignments and gene predictions. Combining computational and experimental approaches has allowed us to generate a comprehensive transcript index for the human genome, which has been a valuable resource for guiding our array design and full-length cloning efforts. In addition, the expression data from the 60 conditions provides a comprehensive atlas of human gene expression over a unique set of gene predictions \[[@B19]\].
Results
=======
Generating a comprehensive transcript index of the human genome
---------------------------------------------------------------
Figure [1](#F1){ref-type="fig"} illustrates the process we used to generate a comprehensive transcript index (CTI) for the human genome that represents just over 28,000 known and predicted transcripts with some level of experimental validation. The first step in this process was to generate a \'primary transcript index\' (PTI) by mapping a comprehensive set of computationally and experimentally derived annotations onto the genomic sequence. The computational predictions include the output of gene-finding algorithms and protein similarities, while the experimentally derived alignments are based on ESTs, serial analysis of gene expression (SAGE), and full-length cDNAs. The resulting list of transcripts in the PTI can be loosely ranked or classified into different categories, ranging from high confidence to low confidence, on the basis of the level of underlying experimental support. The advantages of a PTI are that the computations can be performed on a genome-wide scale and it incorporates the massive amounts of publicly available EST, SAGE and cDNA sequence data. However, the resulting transcript index has two significant limitations. First, the *ab initio*gene-finding algorithms tend to have a high false-positive rate when applied at a low-stringency setting to cast as broad a discovery net as possible. Second, gene-finding algorithms are trained on known protein-coding genes, which may limit their ability to detect truly novel classes of transcribed sequences.
The second step towards the CTI is the use of two different types of microarrays to address these limitations (Figure [1](#F1){ref-type="fig"}). First, predicted transcript arrays (PTA) were used to determine experimentally which of the lower-confidence predictions in the PTI were likely to represent real transcripts. Second, genomic tiling arrays were used to survey transcriptional activity in a completely unbiased and comprehensive fashion. As shown in Figure [1](#F1){ref-type="fig"}, the CTI plays a central part in the subsequent design of screening arrays. These are used to monitor RNA levels for all the transcripts across a large number of diverse conditions to begin the process of assigning biological functions to novel genes based on co-regulation with known genes \[[@B20]\]. The CTI is also used to design exon/junction arrays that can be used to discover and monitor alternative splicing across different tissues and stages of development \[[@B21]\].
Generating a PTI
----------------
To generate the PTI, three distinct computational analysis steps were executed in parallel: predictions based on similarity to expressed sequences from human and mouse; predictions based on similarity to all known proteins; and *ab initio*gene predictions. The process resulted in mapping 91% of the well characterized genes found in the RefSeq database \[[@B22]\], a percentage consistent with initial genome annotation results \[[@B2],[@B3]\]. The mapping results were generated by collapsing overlapping gene models and regions of similarity to define locus projections, which comprise the distinct transcribed regions making up our PTI. While the reliance on gene predictions and protein alignments biases the PTI towards protein-coding genes, the alignment of all expressed sequences should represent many of the non-coding genes reported to date. A comprehensive index of non-coding genes would require tiling arrays, as described later.
All locus projections were classified into one of eight categories on the basis of the level of underlying evidence from expressed sequence similarity, protein similarity and *ab initio*predictions. The categories, in decreasing order of support, are as follows: (1) known genes, taken as the set of 11,214 human genes represented in the RefSeq database when the arrays were designed; (2) *ab initio*gene models with expressed sequence and protein support; (3) *ab initio*gene models with expressed sequence support; (4) *ab initio*gene models with protein support; (5) alignments of expressed sequence and protein data; (6) alignments of expressed sequence data, requiring at least two overlapping expressed sequences; (7) *ab initio*gene models with no expressed sequence or protein support; and (8) alignments of protein data. Because of the limitations discussed in the previous section, we considered predictions with a single line of evidence (categories 6-8) as low confidence.
Table [1](#T1){ref-type="table"} provides summaries resulting from a comparison between our PTI and the published Sanger Institute data for chromosomes 20 and 22 \[[@B23],[@B24]\]. Our locus projections overlap 1,177 of 1,297 (91%) Sanger genes on chromosome 20 and 854 of 936 (91%) Sanger genes on chromosome 22, and our predicted exons overlap 7,306 of 7,556 (97%) and 4,819 of 5,014 (96%) total Sanger chromosome 20 and 22 exons, respectively. This comparison highlights the fact that our annotations result in the detection of both genes and exons in genomic sequence with high sensitivity.
Predicted transcript arrays
---------------------------
We previously described a high-throughput, experimental procedure to validate predicted exons and assemble exons into genes by using co-regulated expression over a diversity of conditions \[[@B11]\]. Here we employ a similar strategy over the entire genome by hybridizing RNA from 60 diverse tissue and cell-line samples to a set of arrays designed from the PTI. For a complete list of the transcripts represented on the predicted transcript arrays and 60 tissues and cell lines hybridized to these arrays (see Additional data files 1 and 2). We designed two probes per exon, where possible, for exons containing the highest-scoring probes as described in the methods from each transcript in our PTI set (on average, a total of four probes per transcript). This was done to balance the poor specificity of *ab initio*gene-finding algorithms \[[@B14],[@B25],[@B26]\] against the significant microarray costs associated with large-scale gene-expression experiments. The resulting hybridization data provides experimental validation of those low-confidence predicted genes that are either unsupported or minimally supported by existing EST data, thereby providing a means of determining which transcripts are included in the CTI.
Summary of predicted transcript validation on chromosomes 20 and 22
-------------------------------------------------------------------
We used an enhanced version of a previously described gene-detection algorithm to analyze the predicted transcript array dataset \[[@B11]\]. Basically, the hybridization data from probes each transcript from the PTI were examined to identify those transcripts with probes that appear to be more highly correlated over the 60 diverse conditions. Transcripts with probes that behaved similarly over the different conditions tested were considered to be expression-validated genes (EVGs). Unlike our original algorithm that used Pearson correlations to group similarly behaving probes, our enhanced algorithm incorporated a probe-specific model to assess the most likely set of probes making up a transcriptional unit \[[@B27]\] (see Materials and methods for details). We used the extensive publicly available annotations on chromosomes 20 and 22 to assess the sensitivity and specificity of our array-based detection procedure.
The sensitivity of our procedure was assessed by computing the EVG detection rate for those Sanger genes that overlap predictions (locus projections) represented in our PTI (Table [2](#T2){ref-type="table"}). The average detection rate for our locus projections on chromosomes 20 and 22 is approximately 70% for those overlapping Sanger genes and just over 80% for those locus projections derived from RefSeq alignments (locus category = known) that represent Sanger genes. A true positive in this instance was defined as an expression-verified gene containing at least two probes, where at least one of the probes was contained within the exon of a Sanger or RefSeq gene.
This 20% false-negative rate is the result of a complex mixture of issues, including limitations in our EVG-detection algorithm, limitations in the probe design step, lack of expression in the conditions profiled, and/or alternative splicing events. While the EVG-detection algorithm provides an efficient method to assemble probes into transcript units, the detection capabilities of this model could be expected to improve as the number of samples and the number of probes targeting any given transcript increases. The use of four probes per predicted transcript was determined to be sufficient for detection of most transcripts, as supported by the overall detection rate of known genes, although in many cases the probe design step was limited by our ability to find four high-quality probes per transcript. For many transcripts, there were not four nonoverlapping probes predicted to have good hybridization characteristics for the microarray experiment carried out here. The 60 samples were chosen to represent a broad array of tissue types, as an exhaustive list of human tissues is impossible to obtain. Because no replicate tissues/cell lines were run for any of the 60 chosen samples, we relied on the replication inherent in monitoring the same transcripts over 60 different conditions. In this case, genes expressed in multiple samples provide the replication necessary to increase our confidence in the detections. However, there are clear limitations in not replicating tissues/cell lines, as genes may be expressed in only a single condition or may be switched on only under certain physiological conditions or only during a certain stages of development. In such cases, we would have reduced power to detect these genes.
Genes in the lower-confidence categories of our PTI annotations, which are not typically considered genes by Sanger, were detected at a significantly reduced rate. Interestingly, of the 337 (188 +149) higher-confidence transcripts on chromosomes 20 and 22 that did not intersect with Sanger genes, 47 (or 14%) were detected as EVGs (Table [2](#T2){ref-type="table"}). These transcripts represent potential novel transcripts on these two highly characterized chromosomes.
However, before we can make claims to the discovery potential for this method over the entire genome, we need to assess the false-positive detection rates. To this end, we defined as false positives all detections made in regions with support by only a single gene model that fell outside Sanger-annotated genes on chromosomes 20 and 22. Applying this definition over all transcripts in our PTI leads to a false-positive rate of 3% (11 out of 406). Because we cannot exclude the possibility that some of the transcripts supported by a single gene model represent real genes, we consider this false-detection rate as an upper bound on the actual false-positive rate. Accepting that the Sanger annotations represent the gold standard for chromosome 22, we detected 70% of all Sanger-annotated genes, while only 4% of the chromosome 22 locus projections that did not intersect Sanger genes were detected by our procedure, highlighting the sensitivity and specificity of this approach. In addition, the enrichment for EVG detections in Sanger genes versus the non-Sanger PTI on chromosomes 20 and 22 was extremely significant with a *p*-value effectively equal to 0 when using the chi-square test for independence (*χ*^2^= 3,093, with 1 degree of freedom (df)).
Summarizing EVG data over the entire genome and assessing the discovery potential. The last column of Table [2](#T2){ref-type="table"} provides the number of expression verified genes detected over the entire genome for locus projections in our PTI. This represents the most comprehensive direct experimental screening of *ab initio*gene predictions ever undertaken. We can use the false-positive and negative rates derived above to assess the discovery potential on that part of the genome that has not been as extensively characterized as chromosomes 20 and 22. First, we note that our detection rates over the genome were similar to that given for chromosomes 20 and 22. That is, 75% of the category 1 genes (RefSeq genes) were detected over the entire genome, compared to 80% for chromosomes 20 and 22. In total, 15,642 genes in the PTI were experimentally validated using this array-based approach. Assuming the false-positive rate of 3% defined above and a conservative false-negative rate of 30%, defined as the percentage of Sanger genes we failed to detect on chromosomes 20 and 22, these data suggest there are close to 21,675 potential coding genes represented in our PTI set. Because our PTI misses close to 10% of the Sanger genes, we corrected this number for those genes not represented in this set and provide an estimate of the total number of protein-coding genes in the human genome supported by our data to be approximately 25,000. This number is consistent with estimates given in the current release (22.34d.1) of the Ensembl database \[[@B28],[@B29]\].
However, we caution that the estimate provided is based solely on the data described here, and that orthogonal sources of data \[[@B30]\] continue to suggest that the actual number of genes will be known only after the transcriptome has been completely characterized.
From Table [2](#T2){ref-type="table"} we note that 2,093 (1,428 + 555 + 110) of the transcripts that were detected as EVGs had only one line of evidence (EST alignment, protein alignment or *ab initio*prediction). These 2,093 transcripts represent a rich source of potential discoveries in our PTI. To assess the potential biological functions of this novel gene set, we annotated translations of this set by searching the domains represented in the Protein Families database (Pfam) \[[@B31]\]. The search results were used to assign each of the translations to Gene Ontology (GO) \[[@B32]\] codes as described in the methods. Figure [2](#F2){ref-type="fig"} graphically depicts the breakdown of the most common GO codes for two of the three major GO categories. These data suggest there may still be a significant number of protein-coding genes with important biological functions, given that domains/motifs represented in these predicted genes are similar to those found in known genes. The 339 predictions that were validated as EVGs and that had protein domains of biological interest would be natural candidates for full-length cloning, over the 24,532 (7,170 + 16,822 + 540 from Table [2](#T2){ref-type="table"}) other lower-confidence predictions in our set.
EVG data as an expression index
-------------------------------
Because multiple probes in each of the approximate 50,000 predicted genes in the human genome have been monitored over 60 different tissues and cell lines, the EVG data represent a significant atlas of human gene expression that is now publicly available \[[@B19]\]. For each transcript, the intensity information from the corresponding probes was optimally combined as described by Johnson *et al.*\[[@B21]\] to provide a quantitative measure of the relative abundance across the panel of 60 conditions, as shown in Figure [3](#F3){ref-type="fig"}.
Tiling arrays for chromosomes 20 and 22
---------------------------------------
To complement the use of PTI arrays, we constructed a set of genome tiling arrays comprised of 60 mer oligonucleotide probes tiled in 30 base-pair steps through both strands of human chromosomes 20 and 22. Repetitive sequences identified by RepeatMasker were ignored for probe design. These genome tiling arrays allow for an unbiased view of the transcriptional activity outside of known and predicted genes on these two chromosomes. mRNA from six (chromosome 20) or eight (chromosome 22) conditions was amplified and hybridized to the tiling arrays (see \[[@B19]\] and Additional data files 3 and 4). As with the PTI arrays, the amplification protocol generated strand-specific cDNA copies of the transcripts, which were full-length. Using a two-step procedure, the resulting data were analyzed to detect sequences expressed in at least one condition \[[@B33]\]. First, we examined probe behavior over conditions in overlapping windows of size 15,000 bp to identify windows that probably contained transcribed sequences, using a robust principal component analysis (PCA) method \[[@B33]\]. Second, for regions identified as likely to contain transcribed sequences, we attempted to discriminate between probes corresponding to expressed sequences (expressed \'exons\') and probes corresponding to untranscribed sequences (\'introns\' or intergenic sequence) using a clustering procedure on variables derived from the PCA procedure \[[@B33]\]. All analysis results derived from this procedure were interpreted in the light of the Sanger annotations and our custom PTI set described above.
Figure [4](#F4){ref-type="fig"} provides two representative examples of tiling data for two known Sanger genes, *KDELR3*and *EWRS1*. In the first case (Figure [4a](#F4){ref-type="fig"}), the tiling data almost perfectly correspond to the RefSeq annotation of *KDELR3*, with just two potential false positives out of the 178 intron probes. The *KDELR3*gene is annotated as having two alternative transcripts in the RefSeq database, given by the RefSeq accession numbers NM\_006855 and NM\_016657. The NCBI Acembly alternative splicing predictions further suggest the presence of additional isoforms of this gene (see Figure [4](#F4){ref-type="fig"}). One of the alternative forms, *KDELR3*.e, depicted in Figure [4a](#F4){ref-type="fig"}, includes a novel 5\' exon. The presence of this exon is supported by the EST with GenBank accession number BM921831. The tiling data for the *KDELR3*gene in two conditions clearly show expression of NM\_006855 but not NM\_016657, thereby reliably detecting distinct splice forms. Further, there is a significant signal 5\' to exon 2 in both transcripts that seems to suggest a novel exon, as opposed to a true false positive. This putative exon exactly matches the location of the first exon given in the Acembly prediction track noted in Figure [4a](#F4){ref-type="fig"} (*KDELR3*.e).
Figure [4b](#F4){ref-type="fig"} shows the tiling data for the *EWSR1*gene. In contrast to the first example, this gene has intense transcriptional activity outside of the annotated exons. Specifically, the *EWSR1*gene has 43 potentially false-positive calls out of 203 intron probes. However, the EST data and alternative splicing predictions strongly suggest that these probes represent biologically relevant transcriptional activity. As with the *KDELR3*gene, *EWRS1*is annotated by RefSeq as having two transcripts: NM\_005243 and NM\_013986. The Acembly predictions identify four additional alternative splice forms; most noteworthy among these are *EWSR1*.b and *EWSR*.g, shown in Figure [4b](#F4){ref-type="fig"}. These predictions indicate that alternative transcripts may exist for the *EWSR1*gene that essentially divide the largest transcript into two transcripts, suggesting that multiple promoter and transcription-stop signals are present in this gene. The tiling data depicted in Figure [4b](#F4){ref-type="fig"} shows that all exons from both RefSeq splice forms were detected. In addition, there is a region to the right of probe position 400 in Figure [4b](#F4){ref-type="fig"} that indicates significant transcription activity but where there are no RefSeq exons annotated. However, the green bars indicate exons that are supported by EST data as well as the *EWSR*.b and *EWSR*.g predicted alternative splice forms, providing experimental support that these predictions represent actual isoforms of this gene. In fact, these data may provide a more accurate representation of the putative structure of this gene, as they support multiple alternatively spliced transcripts in this gene, beyond what has already been annotated in the RefSeq database. In all, 5% of the probes detected as expressed in intronic sequence mapped to predicted alternative splice forms. Given the extent of alternative splicing that is yet to be characterized \[[@B21]\], we believe a significant proportion of the \'intron\' transcriptional activity in our data may represent alternative splicing.
Summarizing the tiling results
------------------------------
Our genome tiling arrays consisted of 2,119,794 and 1,201,632 probes for chromosomes 20 and 22, respectively. Of these, 1,615,034 probes fell into Sanger gene regions, with 239,542 probes actually overlapping Sanger exons. Under stringent criteria 64,241 probes were detected as expressed, with 34,245 of these falling within Sanger exons, 18,551 falling within Sanger introns, and 15,835 probes falling completely outside all Sanger annotations. This widespread transcriptional activity outside annotated regions of the human genome is consistent with other reports from multiple species \[[@B10],[@B12],[@B15],[@B16]\]. Overall, at least one exon in each of 876 Sanger genes was detected as expressed out of 1,703 total genes covered by probes (excluding annotated pseudogenes), leading to an overall gene detection rate of 52%. The bias of probes identified as exon probes that actually fall in exons is striking, given that exons comprise roughly 2% of the genomic sequence (the *p*-value for this enrichment using the Fisher exact test is less than 10^-15^). To estimate the upper bound of false-positive calls, we counted as false-positive events each probe identified as expressed by the detection process, but falling within an annotated intron of the RefSeq genes we detected as expressed. This resulted in an estimated false-positive rate of 1.3%.
As indicated in Figure [4](#F4){ref-type="fig"}, a percentage of these false-positive calls will be due to unannotated isoforms of genes. Others still will be due to cross-hybridization of the intron probes to genes in other parts of the genome. We consider cross-hybridization as made up of two components: specific cross-hybridization resulting from transcripts with similar, usually homologous, sequences; and nonspecific cross-hybridization resulting from the base composition of the probe sequence (J.C. and G.C., unpublished work). Of the intron probes detected as expressed, 23% had sequence similarities to known transcripts considered to render them susceptible to specific cross-hybridization, and 17% contained sequence features associated with nonspecific cross-hybridization. Accounting for probes that were positive for both specific and nonspecific cross-hybridization, we are left with 55% of the probes detected as expressed in the introns of Sanger genes that cannot easily be explained as alternative splicing or cross-hybridization. These data support recent observations that significant levels of transcription exist within the introns of known genes \[[@B15],[@B16]\].
For those probes falling outside all Sanger genes, we again made use of our custom genome annotations to help interpret the extent of transcriptional activity in these regions. Table [3](#T3){ref-type="table"} summarizes the detections made for each of the categories described above. Filtering probes using the same cross-hybridization predictors described above suggests that 65% of those probes falling outside all annotations are not likely to be the result of cross-hybridization. Furthermore, for those detections that overlap low-confidence locus projections in our PTI, we used the classification procedure discussed above to assign GO codes to these transcripts. Only seven of the 297 transcribed regions detected outside of all Sanger genes via the tiling results (see Table [3](#T3){ref-type="table"}) contained GO protein domains. This suggests that a large fraction of the transcriptional activity detected using tiling arrays is non-coding and of unknown biological function \[[@B15],[@B34]\].
Tiling data help classify conserved sequences between species
-------------------------------------------------------------
One further advantage of the tiling data is that they can be used to discriminate between transcribed and non-transcribed sequences conserved between human and mouse, or between any other pair of species. Figure [4c](#F4){ref-type="fig"} highlights tiling data under one condition for the beta-actin gene, a gene that is constitutively expressed in all tissues and often serves as a positive control in mRNA and protein expression experiments. The genomic region containing the complete beta-actin mRNA and 10 kilobases (kb) of genomic sequence upstream of the transcription start, was obtained from the mouse and human genomes, aligned using the AVID program \[[@B35]\] and then fed into the rVista program \[[@B36]\]. From this, we identified the conserved regions in this gene between mouse and human, including several relevant transcription factor binding domains that are key to the transcriptional regulation of this gene \[[@B37]-[@B39]\]. As can be seen directly from the figure, the exons are all detected as highly expressed, but none of the conserved transcription factor regions shows activity. This combination of expressed sequence in close proximity to conserved regions that are not expressed (as determined by the tiling data), offers another powerful advantage of the tiling data in discriminating among the possible roles of conserved sequences.
Discussion
==========
A complete understanding of the human genome will only come after all genes have been identified and the functions of those genes have been determined. There has been much recent progress in defining the human transcriptome with *ab initio*methods, sequencing of EST libraries, full-length gene cloning projects, and comparative analyses between fully sequenced genomes of different species. However, we are still a long way from having a comprehensive set of annotations for the human and other genomes. There is need for new high-throughput experimental approaches to accelerate the process of annotating sequenced genomes in a comprehensive and accurate fashion. Toward this goal, we have used two microarray-based experimental approaches to provide evidence of widespread transcription activity outside of any known or predicted genes in the human genome. We have also provided experimental support for many *ab initio*predicted genes that have no other or minimal experimental sequence support, suggesting a small but significant class of genes that have evaded all other forms of experimental detection. Similar identifications have been made recently in the first extensive comparative analysis between mouse and human genomes \[[@B18]\]. Despite the extent of novel discovery, our data suggest there are only 25,000-30,000 protein-coding genes in the human genome, with perhaps an equal number of non-coding transcripts that may serve important regulatory roles \[[@B34],[@B40]\]. Finally, our data indicate widespread alternative splicing across known genes, providing a glimpse into the extent of transcript complexity that may exist in mammalian genomes.
We have used the expression data for the approximate 50,000 predicted transcripts hybridized to 60 diverse conditions in combination with genomic tiling data to generate a CTI containing 28,456 experimentally supported transcripts. The transcripts represented in the CTI include all computational predictions with two or more lines of evidence from our PTI (independent of microarray validation), in addition to the overlapping set of 15,642 transcripts detected as EVGs. This resulting comprehensive list of known and predicted transcripts provides the starting point for large-scale systematic studies to determine the biological function of genes in both normal and disease states. The primary goal of the CTI is to allow researchers to focus experimental efforts on a comprehensive set of genes that are likely to be real.
It is of note that between the time the predicted transcript arrays were designed and annotated using the custom genome annotations described above, and the time this work was published, more than 6,000 genes were added to the RefSeq collection. These newer RefSeq genes were represented by 5,100 locus projections in our original PTI that were not classified in the RefSeq category. Interestingly, 4,212 were detected as EVGs in the present analysis and had already been included in our CTI, a validation rate slightly greater than 82%. Only 19% of the non-RefSeq genes in our PTI had been detected as EVGs (see Table [2](#T2){ref-type="table"}), yet more than 82% of the new RefSeq genes coming from this set were detected as EVGs. This result speaks to the utility of the microarray-based approach to gene validation described here (see Additional data file 5 for a complete breakdown of validation rates by category).
While using microarrays to test computational gene predictions experimentally has the advantage of being economically feasible across the whole genome, the tiling data represent a more comprehensive and unbiased view of transcription. Our data indicate widespread transcriptional activity in the introns of annotated genes and in intergenic regions, where a significant proportion of this activity can be explained by nonspecific and specific cross-hybridization. The transcriptional activity noted for our low-onfidence transcripts in the PTI indicates that 25% of the activity we observe may be coding for proteins that are at least somewhat related to existing protein data. Much of the transcription activity we have noted in the introns of genes may also represent uncharacterized alternative splicing, and potentially novel genes, in addition to specific and nonspecific cross-hybridization.
Conclusions
===========
At present, predicted transcript arrays allow for the discovery of most protein-coding genes genome wide when many different conditions are considered. Until the discovery and characterization of these protein-coding genes is completed, this method will continue to be a cost-effective solution to drive such discovery. In contrast, genomic tiling represents a completely unbiased method for monitoring transcriptional activity in genomes, but due to cost will probably be limited to screening a smaller number of conditions. However, as novel transcription regions are identified from the tiling data, these regions can be represented on predicted transcript arrays that are hybridized over many more conditions, as described in Figure [1](#F1){ref-type="fig"}. As the microarray technologies have evolved, tiling the entire human genome is now possible, with such efforts presently being supported by the ENCODE (Encyclopedia of DNA Elements) project of the National Human Genome Research Institute (NHGRI) \[[@B41]\].
We believe the steps taken here are necessary for querying all potential transcription activity in the genome, for the purpose of identifying novel genes, more completely characterizing existing genes, and identifying a more comprehensive set of probes for these genes that can be used to monitor transcription abundances in more standard gene expression studies. Not all uses of microarrays demand an exhaustive representation of probes to all genes in the genome under study. However, experiments that seek to identify key drivers of pathways \[[@B42]\] or that seek to discriminate between alternative splice forms of genes within a given tissue \[[@B21]\] require a more comprehensive set of arrays to ensure success. These data provide an essential first step to generating a comprehensive set of arrays that are based on experimental support combined with computational annotation, instead of relying solely on the latter. These comprehensive arrays will be invaluable as we seek to better understand mechanisms of action for existing and novel drug targets and elucidate pathways underlying complex diseases. In addition, further study of the extensive noncoding RNA identified via the methods described here and elsewhere \[[@B10],[@B12],[@B15],[@B16]\] is likely to open new fields of biology as the functional roles for these entities are determined.
Materials and methods
=====================
Data preparation
----------------
The NCBI 8/2001 assembly of the human genome was the input data for this analysis. The 4/21/1999 release of RepeatMasker \[[@B43]\] was used to mask for human repeats. An internal database of RNA genes and bacterial and vector sequences was aligned to the genome with BLASTN. Genomic sequences with 95% or higher identity over at least 50 bases were masked. No probes were designed from masked regions.
Gene index production
---------------------
To predict genes on the basis of expressed sequence similarity, we first clustered and aligned all expressed human and mouse sequences in GenBank to create a human gene index (HGI) and a mouse gene index (MGI). Clustering and alignment were performed with the DoubleTwist Clustering and Alignment Tools (CAT) \[[@B44]\]. Input data included all mouse and human RefSeq mRNA sequences, and EST and mRNA sequences from GenBank release 124, masked as described above for repeats and contaminating sequences. For each species, the CAT software first clustered sequences and then defined subclusters on the basis of a multiple sequence alignment. The subclusters represent candidate alternatively spliced gene transcripts. The 644,168 human and 291,656 mouse sequences that were singleton ESTs or completely masked were excluded from the HGI and MGI.
Expressed sequence mapping
--------------------------
Human and mouse UniGene and RefSeq, MGI, and HGI sequences were aligned with the genome first by BLASTN 2.2.1 \[[@B45]\], followed by refinement of intron/exon boundaries by the sim4 algorithm (12/17/2000 release) \[[@B46]\]. Only the representative sequences (Hs.seq.uniq) for each UniGene cluster designated in the 3 August 2001/build 138 version of the UniGene database were used in this analysis. We only refined BLAST alignments with an E-value of less than 10^-20^for human sequences and 10^-8^for mouse sequences. For human UniGene and HGI, we refined only those BLAST hits where the target sequence showed greater than or equal to 92% identity to the genomic sequence over 75 bp. For human RefSeq, we refined hits with greater than or equal to 95% identity, and for MGI, RefSeq, and UniGene, we refined hits with greater than or equal to 80% identity. These thresholds were empirically determined to provide good sensitivity in aligning most sequences to the genome while limiting multiple alignments past those expected from paralogs present in the human genome. In all cases percent identity was measured over 75 bp. Individual sim4 exons of questionable confidence were then removed on the basis of percent identity and length thresholds. All sequence databases were downloaded from GenBank August, 2001.
Protein sequence mapping
------------------------
The GenBank nonredundant protein database (downloaded 25 August 2001) was aligned to the genomic sequence with BLASTX 2.2.1 \[[@B45]\] using an E-score threshold of 10^-5^. Adjacent protein alignments from a single protein were grouped together as a prediction whenever the protein sequence coordinates of the alignments were consistent in direction and did not significantly overlap.
*Ab initio* gene prediction
---------------------------
GrailEXP 4.0 \[[@B47]\], GENSCAN 1.0 \[[@B48]\], FGENESH \[[@B49]\], and FGENESH+ \[[@B49]\]*ab initio*gene-prediction algorithms were run independently across the entire genome assembly to augment alignment-based gene identification methods. GrailEXP 4.0, GENSCAN 1.0, and FGENESH version 1.c were run with default parameters for human sequence. GrailEXP used expressed sequence evidence from RefSeq, UniGene and DoubleTwist HGI to refine gene predictions. FGENESH+ was run with protein sequences from BLASTX with E-score lower than 10^-5^. When multiple protein alignments overlapped, all overlapping protein sequences were clustered with BLASTClust \[[@B50]\] and the lowest E-score hit was used by FGENESH+.
Synthesis and analysis
----------------------
Locus projections contained the union of all exons from all overlapping predictions in a contiguous region of the chromosome that were derived from sequence alignments or gene-finding algorithms. Predictions to a given strand of the genomic sequence that overlapped by even a single nucleotide were grouped into a single locus projection (antisense transcripts were not considered in defining the locus projections). The criteria for grouping predictions were intentionally kept loose, given that the intent was to include as many potential exons as possible in a given genomic region, and then use the experimental microarray-based approach to elucidate the actual gene structure. These merged overlapping predictions defined the 5\' and 3\' ends of the locus projections. Overlapping predicted exons were merged to form an exon prediction of maximal extent. Low-quality predicted exons from sim4 alignments that contained a high percentage of A or T were removed. We also removed sim4-predicted exons that overlapped two or more predicted exons from another sim4 alignment. Additionally, 3\' sim4 and 3\' or 5\' FGENESH+ predicted exons that were short and/or distant from internal predicted exons were removed. Finally, locus projections that contained mRNAs from RefSeq were split at the 5\' end of the RefSeq sequence.
Locus projections supported by expressed sequences alone could be portions of 3\' or 5\' UTRs of genes included in the other gene-prediction categories described in the text. To minimize the consequences of this potential artifact, we used a UTR filter to exclude locus projections from the expressed sequence alone category that were within 20 kb of a locus projection supported by an *ab initio*gene model.
All data were loaded into a relational database to count and categorize locus projections. At least one type of evidence was assigned to each predicted exon for each locus projection. Multiple types of evidence were assigned to a merged predicted exon if there was overlap between predicted exons of different types for at least 1% of the length of the merged exon prediction. One of the eight evidence categories discussed in the text was assigned to each exon on the basis of the combination of types of evidence. Locus projections inherited the highest-ranking evidence category of their constituent exons. Evidence categories were ranked in the following order: Refseq (highest); expressed sequence + protein + *ab initio*; expressed sequence + *ab initio*; protein + *ab initio*; expressed sequence + protein; *ab initio*alone; protein alone; expressed sequence alone. FGENESH+ predictions were counted as protein + *ab initio*. For the *ab initio*category, predictions from at least two of FGENESH, GENSCAN and GrailEXP were required to overlap in at least one exon to be merged.
Probe selection for the genome tiling and predicted transcript arrays
---------------------------------------------------------------------
Input sequences for probe selection were masked for vector, interspersed repeats, simple repeats, poly(A) tails, *Escherichia coli*contamination and human non-coding RNA and mitochrondrial DNA contamination using Scylla (Paracel). For genomic tiling arrays, 60 mer probes were then selected from unmasked regions of both forward and reverse complement strands at uniform 30-base intervals. For predicted transcript arrays, up to four oligonucleotide probes were selected from the unmasked regions of each transcript using a multistep process.
The first step in the probe-selection process was the generation of a pool of candidate probes 60 nucleotides long (60 mers), where each probe was required to fall entirely within an exon from the set of exons under consideration. If there were fewer than four 60 mers then all 50 mers were considered as well. If there were fewer than four 50 mers or 60 mers then all 40 mers were considered, and so on. Stilts composed of sequence from *Saccharomyces cerevisiae*were added to the 3\' ends of probes shorter than 60 nucleotides so that they had a total length of 60 bases when printed onto the arrays.
The second step in the probe-selection process was the classification and reduction of the probe pool on the basis of base composition and related filters. Probes were sorted into four classes on the basis of several criteria, including A, G, C and T content, GC content, the length of the longest homopolymeric run and the number of A residues at the 5\' end. For example, a probe had to have GC content between 35 and 45% to be in class 1, between 15 and 55% to be in class 2, and between 10 and 60% to be in class 3. After all classifications were made, probes from lower-quality classes were discarded, keeping the number of probes per gene greater than 15. In cases where a pair of probes was overlapping by more than 50 bases, only a single probe was chosen.
The final step in the probe-selection process identified probes with minimal overlap, and predicted cross-hybridization and desirable positions in the transcript sequence. Cross-hybridization prediction was based on BLAST searching of the full collection of transcript sequences \[[@B51]\]. Probes with perfect matches to transcript sequences for genes other than the one undergoing design were discarded unless they were the only probes available. Otherwise the probes with the weakest predicted cross-hybridization interactions were preferred. Probes were also selected to have as little overlap as possible, and probes located in the last 500 bp of each transcript were discarded where possible to reduce the effects of impaired amplification in this region \[[@B52]\].
All arrays included a set of standard control probes which were used for image processing and quality control. Each array also included 30 randomly distributed copies of each of 51 negative-control probes. These probes were selected for their low intensities in previous human hybridizations. The negative controls local to each experimental probe were used for background correction. Non-control probes were added to each array such that all probes for a given input sequence were grouped together and ordered by their position on the sequence.
Preparation of labeled cDNA and array hybridization
---------------------------------------------------
Hybridization material was generated through a random-priming amplification procedure using primers with a random sequence at the 3\' end and fixed motif at the 5\' end. This amplification procedure has been fully described \[[@B52]\] and has been optimized to generate strand-specific cDNA copies of the mRNA transcripts that are full-length. The 60 mRNA samples from the human tissues described in Additional data files 2 and 3 were purchased from Clontech. The 60 mRNA samples hybridized to the predicted transcript set of arrays were done in duplicate with fluor reversal to systematically correct for dye bias. For tiling hybridizations, six samples were used for chromosome 20 arrays and eight samples for chromosome 22. The mRNA samples hybridized to the set of tiling arrays were not done in duplicate as the analysis carried out on these data was intensity based, and our preliminary data demonstrated reasonable results without performing the tiling experiments in fluor-reverse pairs (data not shown). Additional data files 2-4 contain the full list of samples used for each set of arrays.
Array images were processed as described \[[@B53]\] to obtain background noise, single channel intensity and associated measurement error estimates. Expression changes between two samples were quantified as log~10~(expression ratio) where the expression ratio was taken to be the ratio between normalized, background-corrected intensity values for the two channels (red and green) for each spot on the predicted transcript arrays. An independent normalization routine was carried out on the tiling data as described \[[@B33]\] to correct for dye biases, given the lack of technical replicates for these data.
Analysis of predicted transcript array data
-------------------------------------------
Probes from each computationally determined locus were analyzed for coordinated expression over 60 tissues by adapting an additive, probe-specific model initially developed to estimate gene expression indices \[[@B27]\]. The model for a single probe in a single sample pair is given by
*y*~*ij*~= *μ*+ *φ*~*j*~+ *θ*~*i*~+ *ε*~*j*~,
where the *y*~*ij*~represent the mlratio measurements for sample pair *i*and probe *j*in the current transcriptional model, *μ*is the grand mean term, *φ*~*i*~is the probe-specific term for probe *j*in the model, *θ*~*i*~is the sample-specific term for sample *i*, and *ε*~*j*~is the probe-specific error term, which is taken to be normally distributed with mean 0 and variance . Given the above representation for an observed mlratio value, the likelihood for a single probe over *N*condition pairs is simply
From this, the likelihood for a given transcriptional model, where a transcriptional model in this context is defined as a set of probes that are adjacent to one another in the genomic sequence and that co-regulate over a number of conditions, is easily seen to be the product of the individual probe likelihoods defined above over the *M*probes comprising the current model:
The maximum likelihood estimates for the parameters of this model are obtained using standard optimization techniques.
With the likelihood model described above, probe groups making up a transcriptional model were formed by iteratively considering whether neighboring probes (within a PTI member based on genomic location) of a given probe improved the fit of the model just described. This was determined by examining the likelihood ratio statistics between the current, best transcriptional model with or without an additional probe included in the model. Thresholds for the likelihood ratio test statistic and the different model parameters were empirically determined to minimize false-positive and false-negative rates. False positives were estimated by the detection of PTI members supported by only a single *ab initio*prediction that fell outside annotated Sanger genes on chromosomes 20 and 22. False negatives were defined as Sanger genes on chromosome 20 and 22 that were not detected. Probe sets with a maximum likelihood statistic greater than 100 and an *r*^2^value for fit of data to the model greater than 0.8 were considered high-confidence candidates for EVGs.
For each high-confidence EVG candidate, probes were further assessed by considering the number of conditions in which the absolute intensity of the probe was seen to be significantly above background, and the number of times the probe was seen significantly differentially expressed. Candidate EVGs with at least one probe that was: significantly above background (*p*-value \< 0.01) in at least 10% of the samples; or significantly differentially expressed (*p*-value \< 0.01) in at least 10% of the condition pairs, were considered validated.
Analysis of tiling array data
-----------------------------
The analysis of the tiling data has been described in detail by Ying *et al.*\[[@B33]\]. Briefly, probes were separated into 15 kb windows along the genome with 7.5 kb overlap between the windows. For each window, a robust principal component analysis was applied to the between-sample correlation matrix for probes in the window. Windows containing transcriptional activity were characterized by comparing the distribution of the Mahalanobis distances for the probes in the window (the Mahalanobis distance for each probe was calculated from the probe location to the center of the data in the first dimensions of the principal component score (PCS)) space with the reference distribution calculated based on known intron probes. Individual probes were then classified as belonging to the transcribed unit or not on the basis of the log of the Mahalanobis distance and an approximation of the diagonal distance (slope) of the probe from the minimum first PCS and median second PCS. Using these measures for distance, the probes were clustered using standard clustering techniques as described \[[@B33]\].
The procedure for estimating cross-hybridization of the probes is the subject of a separate manuscript. For the analyses described in this paper, the nonspecific cross-hybridization was estimated by the presence of motifs within the probe sequence that were enriched in probes observed to have a high level of nonspecific cross-hybridization. These probes were observed to have significant intensity when hybridized to human mRNA samples despite having no EST support and falling in introns of well characterized genes on chromosomes 20 and 22. Specific cross-hybridization was estimated by the minimum predicted ΔG value for hybridization of the probe to all genes other than the intended target in the UniGene database (build 157).
Annotation of EVG and tiling data
---------------------------------
Hidden Markov model Pfam (HMMPfam) domain predictions were run on six-frame translations of the PTIs using the HFRAME software from Paracel with an E-value cutoff of 0.01 and frameshift penalty of -12. Information on Pfam \[[@B31]\] domains is available \[[@B54],[@B55]\]. GO terms \[[@B32]\] were then assigned to each locus projection using the full set of Pfam to GO mappings available from EBI FTP site \[[@B56]\]. The domain-to-ontology mapping is a part of the larger InterPro effort on annotating the proteome \[[@B57],[@B58]\]. Multiple GO categories can be assigned to a single element of the PTI.
Additional data files
=====================
The following additional data is available with the online version of this paper and at \[[@B19]\]. Additional data file [1](#s1){ref-type="supplementary-material"} gives a complete list of 48,614 transcripts in the PTI that were represented on the set of predicted transcript arrays. Additional data file [2](#s2){ref-type="supplementary-material"} gives a complete list of 60 tissues and cell lines hybridized to the predicted transcript arrays. Additional data file [3](#s3){ref-type="supplementary-material"} gives a list of six tissues and cell lines hybridized to the chromosome 20 genomic tiling arrays. Additional data file [4](#s4){ref-type="supplementary-material"} lists the eight tissues and cell lines hybridized to the chromosome 22 genomic tiling arrays. Additional data file [5](#s5){ref-type="supplementary-material"} contains a comparison of EVG predictions with RefSeq sequences (March 2004). Also available on our website \[[@B19]\] are: ratio data and body atlas data along with the EVG status, and full sequences for the locus projections in fasta format. All probe sequences and expression data are available from the GEO database \[[@B59]\]. The series accession numbers for the tiling and predicted transcript arrays are GSE1097 and GSE918 respectively.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
A complete list of 48,614 transcripts in the PTI that were represented on the set of predicted transcript arrays
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
A complete list of 60 tissues and cell lines hybridized to the predicted transcript arrays
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 3
A list of six tissues and cell lines hybridized to the chromosome 20 genomic tiling arrays
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 4
The eight tissues and cell lines hybridized to the chromosome 22 genomic tiling arrays
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 5
A comparison of EVG predictions with RefSeq sequencesP
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank D. Kessler, M. Marton and the rest of the Rosetta Gene Expression Laboratory for sample preparation and hybridization, S. Dow for reagent and primer handling, and E. Coffey and the Array Production Team for array synthesis. We also thank M. Krolewski and S. Ezekiel for database and programming support. Finally, we would like to thank B. Bush and J. Sachs for critical evaluation of the manuscript. The authors thank J. Burchard for mapping the PTI probes to the current RefSeq sequences. Rosetta Inpharmatics LLC is a wholly owned subsidiary of Merck & Co, Inc.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
A process to generate a comprehensive transcript index (CTI) for the human genome. The first step is the assembly of a comprehensive set of annotations to generate a predicted transcript index (PTI). Sets of microarrays capable of monitoring the transcription activity over the entire genome can then be designed on the basis of the PTI. The different microarray types that can be used in this process include predicted transcript arrays (PTA), exon junction arrays (EJA) \[21\] and genome tiling arrays (GTA). After hybridizing a diversity of conditions onto these arrays, the transcription data are processed to identify a comprehensive set of transcripts (the CTI) and associated probes that are capable of querying all forms of transcripts that may exist in the genome. This set of probes comprises a focused set of microarrays that can be used in more standard microarray-based experiments.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Gene Ontology (GO) classification of novel expression-validated genes (EVGs). EVGs not supported by the expressed sequence data (2,093) were submitted to a search against the Pfam database. Those with significant alignments (339) were assigned GO codes based on Pfam. The pie charts show the distribution of GO terms within this set of EVGs. Note that the total number of GO terms in each category is greater than the number of EVGs because of assignment of multiple GO terms to some EVGs. **(a)**Distribution of the different \'biological process\' GO codes assigned to the EVGs with significant hits to the Pfam database: a total of 526 GO terms. **(b)**Distribution of the different \'molecular function\' GO codes assigned to the EVGs with significant hits to the Pfam database: a total of 374 GO terms.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Utilizing PTA data as an expression index. Absolute transcript abundance over the 60 conditions described in \[19\] for two expression-supported transcripts. RLP09885002 represents a known gene (*ATP1A1*, ATPase, Na^+^/K^+^transporting, alpha 1 polypeptide) whereas RLP10406004 was supported solely by gene model predictions before microarray validation.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Examples of tiling results for known genes. The colored bars across the bottom of the data window are color matched with the corresponding exon annotations shown in the genome viewer. **(a)**The *KDELR3*gene shows strong agreement between the public transcript annotations and the tiling results. The top panel represents a screen shot from the UCSC genome browser \[60\] highlighting *KDLER3*. The bottom panel represents transcription activity as raw intensities (*y*-axis) for each probe used to tile through *KDLER3*(*x*-axis), in one of the eight conditions monitored by the genomic tiling arrays. **(b)**The *EWRS1*gene potentially contains a larger number of false-positive predictions, but more probably lends additional experimental support to previously predicted alternative splice forms (*EWSR*.b and *EWSR*.g), giving a more accurate representation of the putative structure of this gene. The top panel represents a screen shot from the UCSC genome browser \[60\] highlighting EWRS1. The bottom panel represents transcription activity as raw intensities (y-axis) for each probe used to tile through EWSR1 (x-axis), in one of the eight conditions monitored by the genomic tiling arrays. **(c)**Conserved regions between mouse and human upstream of the beta-actin gene. The tiling data readily detect all of the transcribed parts of the gene, but not the conserved regulatory regions. The green bars in the probe-intensity plot represent the annotated transcribed regions for the beta-actin gene, while the blue bars indicate regions that are not known to be transcribed. The lower section shows the sequence conservation between human and mouse as obtained through the program rVISTA \[36,61\]. Conserved coding (blue peaks) and non-coding regions (red peaks) are shown where the two genomic sequences align with 75% identity over 100-bp windows. The rows marked ELK, ETF, and SRF show binding sites for these transcription factors predicted using TRANSFAC matrix models and the MATCHTM program, which are part of the rVISTA suite. The exons for the gene are shown in blue.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Comparison of locus projections in the PTI on chromosomes 20 and 22 to Sanger-annotated genes
:::
Sanger chromosome 20, genes Non-Sanger chromosome 20, genes Sanger chromosome 22, genes Non-Sanger chromosome 22, genes
-------------------------------------- ------------------------------------------- ----------------------------- --------------------------------- ----------------------------- ---------------------------------
Sanger genes (including pseudogenes) 1,297 936
Locus projection categories
High-confidence categories RefSeq 676 (30) 8 375 (47) 12
*Ab initio*+ expressed sequence + protein 336 (63) 10 285 (127) 10
*Ab initio*+ expressed sequence 38 (2) 96 28 (7) 74
*Ab initio*+ protein 28 (11) 37 31 (18) 29
Expressed sequence + protein 38 (30) 37 36 (30) 24
Low-confidence categories *Ab initio* 22 (4) 674 50 (21) 362
Protein 17 (14) 157 18 (13) 121
Expressed sequence 22 (2) 1,591 31 (7) 1,127
Higher-confidence categories 1,116 (136) 188 755 (229) 149
All categories 1,177 (156) 2,610 854 (270) 1,759
Columns 1 and 3 provide the number of locus projections in the PTI set that overlap Sanger genes for chromosomes 20 and 22, respectively. The numbers given in parentheses indicate the number of Sanger-annotated pseudogenes; these pseudogenes were not used when summarizing the results. Columns 2 and 4 give the number of genes in the PTI set that were not overlapping Sanger genes.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Summary of expression-validated genes (EVGs) from predicted transcripts over the entire human genome
:::
Gene categories Sanger/PTI chromosome 20 Non-Sanger PTI chromosome 20 Sanger/PTI chromosome 22 Non-Sanger PTI chromosome 22 PTI genome-wide
------------------------------------------- -------------------------- ------------------------------ -------------------------- ------------------------------ -----------------
Total Sanger genes represented 1,177 (826) 854 (575)
RefSeq 676 (552) 8 (2) 375 (290) 12 (5) 10,720 (7992)
*Ab initio*+ expressed sequence + protein 336 (229) 10 (2) 285 (202) 10 (5) 8,801 (4269)
*Ab initio*+ expressed sequence 38 (17) 96 (8) 28 (15) 74 (8) 3,733 (784)
*Ab initio*+ protein 28 (9) 37 (7) 31 (16) 29 (4) 1,983 (233)
Expressed sequence + protein 38 (2) 37 (2) 36 (10) 24 (4) 1,126 (271)
Expressed sequence 22 (3) 1,591 (44) 31 (3) 1,127 (33) 7,170 (1428)
*Ab initio* 22 (12) 674 (39) 50 (35) 362 (17) 16,822 (555)
Protein 17 (2) 157 (7) 18 (4) 121 (4) 540 (110)
High-confidence categories 1,116 (809) 188 (21) 755 (533) 149 (26) 26,363 (13,549)
All categories 1,177 (826) 2,610 (111) 854 (575) 1,759 (80) 50,895 (15,642)
Columns 1 and 3 provide the total number of Sanger genes for each category for chromosomes 20 and 22, respectively, with the number of EVGs detected given in parentheses. Columns 2 and 4 provide the total number of LPs that did not overlap Sanger genes, with the number of EVGs detected given in parentheses. The last column provides the total number of LPs in the PTI represented on the PTA microarrays, with the number of EVGs detected over the entire genome given in parentheses.
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Summary of transcription activity detected from the chromosome 20 and 22 genome tiling data
:::
Locus projection categories Sanger tiling chromosome 20 Non-Sanger tiling chromosome 20 Sanger tiling chromosome 22 Non-Sanger tiling chromosome 22
-------------------------------------------- ----------------------------- --------------------------------- ----------------------------- --------------------------------- --
Total Sanger genes 1,278 933
Sanger category 1 577 (398) 368 (184)
Sanger category 2 155 (32) 121 (60)
Sanger category 3 338 (150) 144 (52)
Sanger category 4 161 (117) 294 (138)
RefSeq 3 0
*Ab initio* + expressed sequence + protein 1 0
*Ab initio* + expressed sequence 15 8
*Ab initio* + protein 6 4
Expressed sequence + protein 4 1
*Ab initio* 71 26
protein 11 21
Expressed sequence 80 46
Outside all annotations\* 1,936 1,058
High-confidence categories NA 25 NA 12
All annotation categories 1,231 (697) 191 927 (434) 106
\*Number of probes detected as components of EVGs. Columns 1 and 3 provide the number of Sanger genes represented on the genome tiling arrays for chromosomes 20 and 22, respectively, with the number of genes detected given in parentheses. Columns 2 and 4 provide the number of LPs not overlapping Sanger genes that were detected on chromosomes 20 and 22, respectively. NA, not applicable.
:::
|
PubMed Central
|
2024-06-05T03:55:51.733581
|
2004-9-23
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545593/",
"journal": "Genome Biol. 2004 Sep 23; 5(10):R73",
"authors": [
{
"first": "Eric E",
"last": "Schadt"
},
{
"first": "Stephen W",
"last": "Edwards"
},
{
"first": "Debraj",
"last": "GuhaThakurta"
},
{
"first": "Dan",
"last": "Holder"
},
{
"first": "Lisa",
"last": "Ying"
},
{
"first": "Vladimir",
"last": "Svetnik"
},
{
"first": "Amy",
"last": "Leonardson"
},
{
"first": "Kyle W",
"last": "Hart"
},
{
"first": "Archie",
"last": "Russell"
},
{
"first": "Guoya",
"last": "Li"
},
{
"first": "Guy",
"last": "Cavet"
},
{
"first": "John",
"last": "Castle"
},
{
"first": "Paul",
"last": "McDonagh"
},
{
"first": "Zhengyan",
"last": "Kan"
},
{
"first": "Ronghua",
"last": "Chen"
},
{
"first": "Andrew",
"last": "Kasarskis"
},
{
"first": "Mihai",
"last": "Margarint"
},
{
"first": "Ramon M",
"last": "Caceres"
},
{
"first": "Jason M",
"last": "Johnson"
},
{
"first": "Christopher D",
"last": "Armour"
},
{
"first": "Philip W",
"last": "Garrett-Engele"
},
{
"first": "Nicholas F",
"last": "Tsinoremas"
},
{
"first": "Daniel D",
"last": "Shoemaker"
}
]
}
|
PMC545594
|
Background
==========
The differentiation of a small number of cells in the developing embryo into the hundreds of cell and tissue types present in a human adult is associated with a multitude of changes in gene expression. In addition to many differences between tissues in transcriptional and translational regulation of genes, alternative pre-mRNA splicing (AS) is also frequently used to regulate gene expression and to generate tissue-specific mRNA and protein isoforms \[[@B1]-[@B5]\]. Between one-third and two-thirds of human genes are estimated to undergo AS \[[@B6]-[@B11]\] and the disruption of specific AS events has been implicated in several human genetic diseases \[[@B12]\]. The diverse and important biological roles of alternative splicing have led to significant interest in understanding its regulation.
Insights into the regulation of AS have come predominantly from the molecular dissection of individual genes (reviewed in \[[@B1],[@B12]\]). Prominent examples include the tissue-specific splicing of the c-*src*N1 exon \[[@B13]\], cancer-associated splicing of the *CD44*gene \[[@B14]\] and the alternative splicing cascade involved in *Drosophila melanogaster*sex determination \[[@B15]\]. Biochemical studies of these and other genes have described important classes of *trans*-acting splicing-regulatory factors, implicating members of the ubiquitously expressed serine/arginine-rich protein (SR protein) and heterogeneous nuclear ribonucleoprotein (hnRNP) families, and tissue-specific factors including members of the CELF \[[@B16]\] and NOVA \[[@B17]\] families of proteins, as well as other proteins and protein families, in control of specific splicing events. A number of *cis*-regulatory elements in exons or introns that play key regulatory roles have also been identified, using a variety of methods including site-directed mutagenesis, systematic evolution of ligands by exponential enrichment (SELEX) and computational approaches \[[@B18]-[@B22]\]. In addition, DNA microarrays and polymerase colony approaches have been developed for higher-throughput analysis of alternative mRNA isoforms \[[@B23]-[@B26]\] and a cross-linking/immunoprecipitation strategy (CLIP) has been developed for systematic detection of the RNAs bound by a given splicing factor \[[@B27]\]. These new methods suggest a path towards increasingly parallel experimental analysis of splicing regulation.
From another direction, the accumulation of large databases of cDNA and expressed sequence tag (EST) sequences has enabled large-scale computational studies, which have assessed the scope of AS in the mammalian transcriptome \[[@B3],[@B8],[@B10],[@B28]\]. Other computational studies have analyzed the tissue specificity of AS events and identified sets of exons and genes that exhibit tissue-biased expression \[[@B29],[@B30]\]. However, a number of significant questions about tissue-specific alternative splicing have not yet been comprehensively addressed. Which tissues have the highest and lowest proportions of alternative splicing? Do tissues differ in their usage of different AS types, such as exon skipping, alternative 5\' splice site choice or alternative 3\' splice site choice? Which tissues are most distinct from other tissues in the spectrum of alternative mRNA isoforms they express? And to what extent do expression levels of known splicing factors explain AS patterns in different tissues?
Here, we describe an initial effort to answer these questions using a large-scale computational analysis of ESTs derived from about two dozen human tissues, which were aligned to the assembled human genome sequence to infer patterns of AS occurring in thousands of human genes. Our results distinguish specific tissues as having high levels and distinctive patterns of AS, identify pronounced differences between the proportions of alternative 5\' splice site and alternative 3\' splice site usage between tissues, and predict candidate *cis*-regulatory elements and *trans*-acting factors involved in tissue-specific AS.
Results and discussion
======================
Variation in the levels of alternative splicing in different human tissues
--------------------------------------------------------------------------
Alternative splicing events are commonly distinguished in terms of whether mRNA isoforms differ by inclusion or exclusion of an exon, in which case the exon involved is referred to as a \'skipped exon\' (SE) or \'cassette exon\', or whether isoforms differ in the usage of a 5\' splice site or 3\' splice site, giving rise to alternative 5\' splice site exons (A5Es) or alternative 3\' splice site exons (A3Es), respectively (depicted in Figure [1](#F1){ref-type="fig"}). These descriptions are not necessarily mutually exclusive; for example, an exon can have both an alternative 5\' splice site and an alternative 3\' splice site, or have an alternative 5\' splice site or 3\' splice site but be skipped in other isoforms. A fourth type of alternative splicing, \'intron retention\', in which two isoforms differ by the presence of an unspliced intron in one transcript that is absent in the other, was not considered in this analysis because of the difficulty in distinguishing true intron retention events from contamination of the EST databases by pre-mRNA or genomic sequences. The presence of these and other artifacts in EST databases are important caveats to any analysis of EST sequence data. Therefore, we imposed stringent filters on the quality of EST to genomic alignments used in this analysis, accepting only about one-fifth of all EST alignments obtained (see Materials and methods).
To determine whether differences occur in the proportions of these three types of AS events across human tissues, we assessed the frequencies of genes containing skipped exons, alternative 3\' splice site exons or alternative 5\' splice site exons for 16 human tissues (see Figure [1](#F1){ref-type="fig"} for the list of tissues) for which sufficiently large numbers of EST sequences were available. Because the availability of a larger number of ESTs derived from a gene increases the chance of observing alternative isoforms of that gene, the proportion of AS genes observed in a tissue will tend to increase with increasing EST coverage of genes \[[@B10],[@B31]\]. Since the number of EST sequences available differs quite substantially among human tissues (for example, the dbEST database contains about eight times more brain-derived ESTs than heart-derived ESTs), in order to compare the proportion of AS in different tissues in an unbiased way, we used a sampling strategy that ensured that all genes/tissues studied were represented by equal numbers of ESTs.
It is important to point out that our analysis does not make use of the concept of a canonical transcript for each gene because it is not clear that such a transcript could be chosen objectively or that this concept is biologically meaningful. Instead, AS events are defined only through pairwise comparison of ESTs.
Our objective was to control for differences in EST abundance across tissues while retaining sufficient power to detect a reasonable fraction of AS events. For each tissue we considered genes that had at least 20 aligned EST sequences derived from human cDNA libraries specific to that tissue (\'tissue-derived\' ESTs). For each such gene, a random sample of 20 of these ESTs was chosen (without replacement) to represent the splicing of the given gene in the given human tissue. For the gene and tissue combinations included in this analysis, the median number of EST sequences per gene was not dramatically different between tissues, ranging from 25 to 35 (see Additional data file 1). The sampled ESTs for each gene were then compared to each other to identify AS events occurring within the given tissue (see Materials and methods). The random sampling was repeated 20 times and the mean fraction of AS genes observed in these 20 trials was used to assess the fraction of AS genes for each tissue (Figure [1a](#F1){ref-type="fig"}). Different random subsets of a relatively large pool will have less overlap in the specific ESTs chosen (and therefore in the specific AS events detected) than for random subsets of a smaller pool of ESTs, and increased numbers of ESTs give greater coverage of exons. However, there is no reason that the expected number of AS events detected per randomly sampled subset should depend on the size of the pool the subset was chosen from. While the error (standard deviation) of the measured AS frequency per gene should be lower when restricting to genes with larger minimum pools of ESTs, such a restriction would not change the expected value. Unfortunately, the reduction in error of the estimated AS frequency per gene is offset by an increase in the expected error of the tissue-level AS frequency resulting from the use of fewer genes. The inclusion of all genes with at least 20 tissue-derived ESTs represents a reasonable trade-off between these factors.
The human brain had the highest fraction of AS genes in this analysis (Figure [1a](#F1){ref-type="fig"}), with more than 40% of genes exhibiting one or more AS events, followed by the liver and testis. Previous EST-based analyses have identified high proportions of splicing in human brain and testis tissues \[[@B29],[@B30],[@B32]\]. These studies did not specifically control for the highly unequal representation of ESTs from different human tissues. As larger numbers of ESTs increase the chance of observing a larger fraction of the expressed isoforms of a gene, the number of available ESTs has a direct impact on estimated proportions of AS, as seen previously in analyses comparing the levels of AS in different organisms \[[@B31]\]. Thus, the results obtained in this study confirm that the human brain and testis possess an unusually high level of AS, even in the absence of EST-abundance advantages over other tissues. We also observe a high level of AS in the human liver, a tissue with much lower EST coverage, where higher levels of AS have been previously reported in cancerous cells \[[@B33],[@B34]\]. Human muscle, uterus, breast, stomach and pancreas had the lowest levels of AS genes in this analysis (less than 25% of genes). Lowering the minimum EST count for inclusion in this analysis from 20 to 10 ESTs, and sampling 10 (out of 10 or more) ESTs to represent each gene in each tissue, did not alter the results qualitatively (data not shown).
Differences in the levels of exon skipping in different tissues
---------------------------------------------------------------
Alternatively spliced genes in this analysis exhibited on average between one and two distinct AS exons. Analyzing the different types of AS events separately, we found that the human brain and testis had the highest levels of skipped exons, with more than 20% of genes containing SEs (Figure [1b](#F1){ref-type="fig"}). The high level of skipped exons observed in the brain is consistent with previous analyses \[[@B29],[@B30],[@B32]\]. At the other extreme, the human ovary, muscle, uterus and liver had the lowest levels of skipped exons (about 10% of genes).
An example of a conserved exon-skipping event observed in human and mouse brain tissue is shown in Figure [2a](#F2){ref-type="fig"} for the human fragile X mental retardation syndrome-related (*FXR1*) gene \[[@B35],[@B36]\]. In this event, skipping of the exon alters the reading frame of the downstream exon, presumably leading to production of a protein with an altered and truncated carboxy terminus. The exon sequence is perfectly conserved between the human and mouse genomes, as are the 5\' splice site and 3\' splice site sequences (Figure [2a](#F2){ref-type="fig"}), suggesting that this AS event may have an important regulatory role \[[@B37]-[@B39]\].
Differences in the levels of alternative splice site usage in different tissues
-------------------------------------------------------------------------------
Analyzing the proportions of AS events involving the usage of A5Es and A3Es revealed a very different pattern (Figure [1c,d](#F1){ref-type="fig"}). Notably, the fraction of genes containing A3Es was more than twice as high in the liver as in any other human tissue studied (Figure [1d](#F1){ref-type="fig"}), and the level of A5Es was also about 40-50% higher in the liver than in any other tissue (Figure [1c](#F1){ref-type="fig"}). The tissue with the second highest level of alternative usage for both 5\' splice sites and 3\' splice sites was the brain. Another group of human tissues including muscle, uterus, breast, pancreas and stomach - similar to the low SE frequency group above - had the lowest level of A5Es and A3Es (less than 5% of genes in each category). Thus, a picture emerges in which certain human tissues such as muscle, uterus, breast, pancreas and stomach, have low levels of AS of all types, whereas other tissues, such as the brain and testis, have relatively high levels of AS of all types and the liver has very high levels of A3Es and A5Es, but exhibits only a modest level of exon skipping. To our knowledge, this study represents the first systematic analysis of the proportions of different types of AS events occurring in different tissues. Repeating the analyses by removing ESTs from disease-associated tissue libraries, using available library classifications \[[@B40]\], gave qualitatively similar results (see Additional data files 2, 3, and 4). These data show that ESTs derived from diseased tissues show modestly higher frequencies of exon skipping, but the relative rankings of tissues remain similar. The fractions of genes containing A5Es and A3Es were not changed substantially when diseased-tissue ESTs were excluded.
From the set of genes with at least 20 human liver-derived ESTs, this analysis identified a total of 114 genes with alternative 5\' splice site and/or 3\' splice site usage in the liver. Those genes in this set that were named, annotated and for which the consensus sequences of the alternative splice sites were conserved in the orthologous mouse gene (see Materials and methods) are listed in Table [1](#T1){ref-type="table"}. Of course, conservation of splice sites alone is necessary, but not sufficient by itself, to imply conservation of the AS event in the mouse. Many essential liver metabolic and detoxifying enzyme-coding genes appear on this list, including enzymes involved in sugar metabolism (for example, *ALDOB*, *IDH1*), protein and amino acid metabolism (for example, *BHMT, CBP2*, *TDO2*, *PAH*, *GATM*), detoxification or breakdown of drugs and toxins (for example, *GSTA3*, *CYP3A4*, *CYP2C8*).
Sequences and splicing patterns for two of these genes for which orthologous mouse exons/genes and transcripts could be identified - the genes *BHMT*and *CYP2C8*- are shown in detail in Figure [2b,c](#F2){ref-type="fig"}. In the event depicted for *BHMT*, the exons involved are highly conserved between the human and mouse orthologs (Figure [2b](#F2){ref-type="fig"}), consistent with the possibility that the splicing event may have a (conserved) regulatory role. This AS event preserves the reading frame of downstream exons, so the two isoforms are both likely to produce functional proteins, differing by the insertion/deletion of 23 amino acids. In the event depicted for *CYP2C8*, usage of an alternative 3\' splice site removes 71 nucleotides, shifting the reading frame and leading to a premature termination codon in the exon (Figure [2c](#F2){ref-type="fig"}). In this case, the shorter alternative transcript is a potential substrate for nonsense-mediated decay \[[@B41],[@B42]\] and the AS event may be used to regulate the level of functional mRNA/protein produced.
Differences in splicing factor expression between tissues
---------------------------------------------------------
To explore the differences in splicing factor expression in different tissues, available mRNA expression data was obtained from two different DNA microarray studies \[[@B43]-[@B45]\]. For this *trans*-factor analysis, we obtained a list of 20 splicing factors of the SR, SR-related and hnRNP protein families from proteomic analyses of the human spliceosome \[[@B46]-[@B48]\] (see Materials and methods for the list of genes). The variation in splicing-factor expression between pairs of tissues was studied by computing the Pearson (product-moment) correlation coefficient (*r*) between the 20-dimensional vectors of splicing-factor expression values between all pairs of 26 human tissues. The DNA microarray studies analyzed 10 tissues in addition to the 16 previously studied (Figure [3](#F3){ref-type="fig"}). A low value of *r*between a pair of tissues indicates a low degree of concordance in the relative mRNA expression levels across this set of splicing factors, whereas a high value of *r*indicates strong concordance.
While most of the tissues examined showed a very high degree of correlation in the expression levels of the 20 splicing factors studied (typically with *r*\> 0.75; Figure [3](#F3){ref-type="fig"}), the human adult liver was clearly an outlier, with low concordance in splicing-factor expression to most other tissues (typically *r*\< 0.6, and often much lower). The unusual splicing-factor expression in the human liver was seen consistently in data from two independent DNA microarray studies using different probe sets (compare the two halves of Figure [3](#F3){ref-type="fig"}). The low correlation observed between liver and other tissues in splicing factor expression is statistically significant even relative to arbitrary collections of 20 genes (see Additional data file 8). Examining the relative levels of specific splicing factors in the human adult liver versus other tissues, the relative level of SRp30c message was consistently higher in the liver and the relative levels of SRp40, hnRNP A2/B2 and Srp54 messages were consistently lower. A well established paradigm in the field of RNA splicing is that usage of alternative splice sites is often controlled by the relative concentrations of specific SR proteins and hnRNP proteins \[[@B49]-[@B52]\]. This functional antagonism between particular SR and hnRNP proteins is often due to competition for binding of nearby sites on pre-mRNAs \[[@B49],[@B53],[@B54]\]. Therefore, it seems likely that the unusual patterns of expression seen in the human adult liver for these families of splicing factors may contribute to the high level of alternative splice site usage seen in this tissue. It is also interesting that splicing-factor expression in the human fetal liver is highly concordant with most other tissues, but has low concordance with the adult liver (Figure [3](#F3){ref-type="fig"}). This observation suggests that substantial changes in splicing-factor expression may occur during human liver development, presumably leading to a host of changes in the splicing patterns of genes expressed in human liver. Currently available EST data were insufficient to allow systematic analysis of the patterns of AS in fetal relative to adult liver.
An important caveat to these results is that the DNA microarray data used in this analysis measure mRNA expression levels rather than protein levels or activities. The relation between the amount of mRNA expressed from a gene and the concentration of the corresponding protein has been examined previously in several studies in yeast as well as in human and mouse liver tissues \[[@B55]-[@B58]\]. These studies have generally found that mRNA expression levels correlate positively with protein concentrations, but with fairly wide divergences for a significant fraction of genes.
Over-represented motifs in alternative exons in the human brain, testis and liver
---------------------------------------------------------------------------------
The unusually high levels of alternative splicing seen in the human brain, testis and liver prompted us to search for candidate tissue-specific splicing regulatory motifs in AS exons in genes expressed in each of these tissues. Using a procedure similar to Brudno *et al*. \[[@B59]\], sequence motifs four to six bases long that were significantly enriched in exons skipped in AS genes expressed in the human brain relative to constitutive exons in genes expressed in the brain were identified. These sequences were then compared to each other and grouped into seven clusters, each of which shared one or two four-base motifs (Table [2](#T2){ref-type="table"}). The motifs in cluster BR1 (CUCC, CCUC) resemble the consensus binding site for the polypyrimidine tract-binding protein (PTB), which acts as a repressor of splicing in many contexts \[[@B60]-[@B63]\]. A similar motif (CNCUCCUC) has been identified in exons expressed specifically in the human brain \[[@B29]\]. The motifs in cluster BR7 (containing UAGG) are similar to the high-affinity binding site UAGGG \[A/U\], identified for the splicing repressor protein hnRNP A1 by SELEX experiments \[[@B64]\]. The consensus sequences for the remaining clusters BR2 to BR6 (GGGU, UGGG, GGGA, CUCA, UAGC, respectively), as well as BR7, all resembled motifs identified in a screen for exonic splicing silencers (ESSs) in cultured human cells (Z. Wang and C.B.B., unpublished results), suggesting that most or all of the motifs BR1 to BR7 represent sequences directly involved in mediating exon skipping. In particular, G-rich elements, which are known to act as intronic splicing enhancers \[[@B65],[@B66]\], may function as silencers of splicing when present in an exonic context.
A comparison of human testis-derived skipped exons to exons constitutively included in genes expressed in the testis identified only a single cluster of sequences, TE1, which share the tetramer UAGG. Enrichment of this motif, common to the brain-specific cluster BR7, suggests a role for regulation of exon skipping by hnRNP A1 - or a *trans-*acting factor with similar binding preferences - in the testis.
Alternative splice site usage gives rise to two types of exon segments - the \'core\' portion common to both splice forms and the \'extended\' portion that is present only in the longer isoform. Two clusters of sequence motifs enriched in the core sequences of A5Es in genes expressed in the liver relative to the core segments of A5Es resulting from alignments of non-liver-derived ESTs were identified - LI1 and LI2. Both are adenosine-rich, with consensus tetramers AAAC and UAAA, respectively. The former motif matches a candidate ESE motif identified previously using the computational/experimental RESCUE-ESE approach (motif 3F with consensus \[AG\]AA \[AG\]C) \[[@B19]\]. The enrichment of a probable ESE motif in exons exhibiting alternative splice site usage in the liver is consistent with the model that such splicing events are often controlled by the relative levels of SR proteins (which bind many ESEs) and hnRNP proteins. Insufficient data were available for the analysis of motifs in the extended portions of liver A5Es (which tend to be significantly shorter than the core regions) or for the analysis of liver A3Es.
A measure of dissimilarity between mRNA isoforms
------------------------------------------------
To quantify the differences in splicing patterns between mRNAs or ESTs derived from a gene locus, a new measure called the splice junction difference ratio (SJD) was developed. For any pair of mRNAs/ESTs that align to overlapping portions of the same genomic locus, the SJD is defined as the proportion of splice junctions present in both transcripts that differ between them, including only those splice junctions that occur in regions of overlap between the transcripts (Figure [4](#F4){ref-type="fig"}). The SJD varies between zero and one, with a value of zero for any pair of transcripts that have identical splice junctions in the overlapping region (for example, transcripts 2 and 5 in Figure [4](#F4){ref-type="fig"}, or for two identical transcripts), and has a value of 1.0 for two transcripts whose splice junctions are completely different in the regions where they overlap (for example, transcripts 1 and 2 in Figure [4](#F4){ref-type="fig"}). For instance, transcripts 2 and 3 in Figure [4](#F4){ref-type="fig"} differ in the 3\' splice site used in the second intron, yielding an SJD value of 2/4 = 0.5, whereas transcripts 2 and 4 differ by skipping/inclusion of an alternative exon, which affects a larger fraction of the introns in the two transcripts and therefore yields a higher SJD value of 3/5 = 0.6.
The SJD value can be generalized to compare splicing patterns between two sets of transcripts from a gene - for example, to compare the splicing patterns of the sets of ESTs derived from two different tissues. In this case, the SJD is defined by counting the number of splice junctions that differ between all pairs of transcripts (*i*, *j*), with transcript *i*coming from set 1 (for example, heart-derived ESTs), and transcript *j*coming from set 2 (for example, lung-derived ESTs), and dividing this number by the total number of splice junctions in all pairs of transcripts compared, again considering only those splice junctions that occur in regions of overlap between the transcript pairs considered. Note that this definition has the desirable property that pairs of transcripts that have larger numbers of overlapping splice junctions contribute more to the total than transcript pairs that overlap less. As an example of the splice junction difference between two sets of transcripts, consider the set *S*1, consisting of transcripts (1,2) from Figure [4](#F4){ref-type="fig"}, and set *S*2, consisting of transcripts (3,4) from Figure [4](#F4){ref-type="fig"}. Using the notation introduced in Figure [4](#F4){ref-type="fig"}, SJD(*S*1,*S*2) = *d*(*S*1,*S*2) / *t*(*S*1,*S*2) = \[*d*(1,3) + *d*(1,4) + *d*(2,3) + *d*(2,4)\]/ \[*t*(1,3) +*t*(1,4) + *t*(2,3) + *t*(2,4)\] = \[3 + 4 + 2 + 3\]/ \[3 + 4 + 4 + 5\] = 12/16 = 0.75, reflecting a high level of dissimilarity between the isoforms in these sets, whereas the SJD falls to 0.57 for the more similar sets *S*1 = transcripts (1,2) versus *S*3 = transcripts (2,3). Note that in cases where multiple similar/identical transcripts occur in a given set, the SJD measure effectively weights the isoforms by their abundance, reflecting an average dissimilarity when comparing randomly chosen pairs of transcripts from the two tissues. For example, the SJD computed for the set *S*4 = (1,2,2,2,2), that is, one transcript aligning as transcript 1 in Figure [4](#F4){ref-type="fig"} and four transcripts aligning as transcript 2, and the set *S*5 = (2,2,2,2,3) is 23/95 = 0.24, substantially lower than the SJD value for sets *S*1 versus *S*3 above, reflecting the higher fraction of identically spliced transcripts between sets *S*4 and *S*5.
Global comparison of splicing patterns between tissues
------------------------------------------------------
To make a global comparison of patterns of splicing between two different human tissues, a tissue-level SJD value was computed by comparing the splicing patterns of ESTs from all genes for which at least one EST was available from cDNA libraries representing both tissues. The \'inter-tissue\' SJD value is then defined as the ratio of the sum of *d*(*S*~A~,*S*~B~) values for all such genes, divided by the sum of *t*(*S*~A~,*S*~B~) values for all of these genes, where *S*~A~and *S*~B~refer to the set of ESTs for a gene derived from tissues A and B, respectively, and *d*(*S*~A~,*S*~B~) and *t*(*S*~A~,*S*~B~) are defined in terms of comparison of all pairs of ESTs from the two sets as described above. This analysis uses all available ESTs for each gene in each tissue (rather than samples of a fixed size). A large SJD value between a pair of tissues indicates that mRNA isoforms of genes expressed in the two tissues tend to be more dissimilar in their splicing patterns than is the case for two tissues with a smaller inter-tissue SJD value. This definition puts greater weight on those genes for which more ESTs are available.
The SJD values were then used to globally assess tissue-level differences in alternative splicing. A set of 25 human tissues for which at least 20,000 genomically aligned ESTs were available was compiled for this comparison (see Materials and methods) and the SJD values were then computed between all pairs of tissues in this set (Figure [5a](#F5){ref-type="fig"}). A clustering of human tissues on the basis of their inter-tissue SJD values (Figure [5b](#F5){ref-type="fig"}) identified groups of tissues that cluster together very closely (for example, the ovary/thyroid/breast cluster, the heart/lymph cluster and the bone/B-cell cluster), while other tissues including the brain, pancreas, liver, peripheral nervous system (PNS) and placenta occur as outgroups. These results complement a previous clustering analysis based on data from microarrays designed to detect exon skipping \[[@B24]\]. Calculating the mean SJD value for a given tissue when compared to the remaining 24 tissues (Figure [5c](#F5){ref-type="fig"}) identified a set of human tissues including the ovary, thyroid, breast, heart, bone, B-cell, uterus, lymph and colon that have \'generic\' splicing patterns which are more similar to most other tissues. As expected, many of these tissues with generic splicing patterns overlap with the set of tissues that have low levels of AS (Figure [1](#F1){ref-type="fig"}). On the other hand, another group of tissues including the human brain, pancreas, liver and peripheral nervous system, have highly \'distinctive\' splicing patterns that differ from most other tissues (Figure [5c](#F5){ref-type="fig"}). Many of these tissues were identified as having high proportions of AS in Figure [1](#F1){ref-type="fig"}. Taken together, these observations suggest that specific human tissues such as the brain, testis and liver, make more extensive use of AS in gene regulation and that these tissues have also diverged most from other tissues in the set of spliced isoforms they express. Although we are not aware of reliable, quantitative data on the relative abundance of different cell types in these tissues, a greater diversity of cell types is likely to contribute to higher SJD values for many of these tissues.
Conclusions
===========
The systematic analysis of transcripts generated from the human genome is just beginning, but promises to deepen our understanding of how changes in the program of gene expression contribute to development and differentiation. Here, we have observed pronounced differences between human tissues in the set of alternative mRNA isoforms that they express. Because our approach normalizes the EST coverage per gene in each tissue, there is higher confidence that these differences accurately reflect differences in splicing patterns between tissues. As human tissues are generally made up of a mixture of cell types, each of which may have its own unique pattern of gene expression and splicing, it will be important in the future to develop methods for systematic analysis of transcripts in different human cell types.
Understanding the mechanisms and regulatory consequences of AS will require experimental and computational analyses at many levels. At its core, AS involves the generation of alternative transcripts mediated by interactions between *cis*-regulatory elements in exons or introns and *trans*-acting splicing factors. The current study has integrated these three elements, inferring alternative transcripts from EST-genomic alignments, identifying candidate regulatory sequence motifs enriched in alternative exons from different tissues, and analyzing patterns of splicing-factor expression in different tissues. Our results emphasize differences in the frequencies of exon skipping versus alternative splice site usage in different tissues and highlight the liver, brain and testis as having particularly high levels of AS, supporting the idea that tissue-regulated AS plays important roles in the differentiation of these tissues. The high levels of alternative splice site usage in the liver may relate to the unusual patterns of splicing-factor expression observed in the adult liver, suggesting aspects of developmental regulation of AS at the tissue level. Obtaining a more comprehensive picture of AS will require the integration of additional types of data upstream and downstream of these core interactions. Upstream, splicing factors themselves may be differentially regulated in different tissues or in response to different stimuli at the level of transcription, splicing, or translation, and are frequently regulated by post-translational modifications such as phosphorylation, so systematic measurements of splicing factor levels and activities will be required. Downstream, AS may affect the stability of alternative transcripts (for example, in cases of messages subject to nonsense-mediated mRNA decay), and frequently alters functional properties of the encoded proteins, so systematic measurements of AS transcript and protein isoforms and functional assays will also be needed to fully understand the regulatory consequences of AS events. Ultimately, it will be important to place regulatory events involving AS into the context of regulatory networks involving control at the levels of transcription, translation and post-translational modifications.
Materials and methods
=====================
Data and resources
------------------
Chromosome assemblies of the human genome (hg13) were obtained from public databases \[[@B67]\]. Transcript databases included approximately 94,000 human cDNA sequences obtained from GenBank (release 134.0, gbpri and gbhtc categories), and approximately 5 million human expressed sequence tags (ESTs) from dbEST (repository 02202003). Human ESTs were designated according to their cDNA library source (in total about 800) into different tissue types. Pertinent information about cDNA libraries and the corresponding human tissue or cell line was extracted from dbEST and subsequently integrated with library information retrieved from the Mammalian Gene Collection Initiative (MGC) \[[@B68]\], the Integrated Molecular Analysis of Gene Expression Consortium (IMAGE) \[[@B69]\] and the Cancer Genome Anatomy Project (CGAP) \[[@B70]\]. Library information obtained from MGC, IMAGE and CGAP is provided in Additional data file 5.
Genome annotation by alignment of spliced transcripts
-----------------------------------------------------
The GENOA genome annotation script \[[@B71]\] was used to align spliced cDNA and EST sequences to the human genome. GENOA uses BLASTN to detect significant blocks of identity between repeat-masked cDNA sequences and genomic DNA, and then aligns cDNAs to the genomic loci identified by BLASTN using the spliced-alignment algorithm MRNAVSGEN \[[@B71]\]. This algorithm is similar in concept to SIM4 \[[@B72]\] but was developed specifically to align high-quality cDNAs rather than ESTs and thus requires higher alignment quality (at least 93% identity) and consensus terminal dinucleotides at the ends of all introns (that is, GT..AG, GC..AG or AT..AC). EST sequences were aligned using SIM4 to those genomic regions that had aligned cDNAs. Stringent alignment criteria were imposed: ESTs were required to overlap cDNAs (so that all the genes studied were supported by at least one cDNA-genomic alignment); the first and last aligned segments of ESTs were required to be at least 30 nucleotides in length, with at least 90% sequence identity; and the entire EST sequence alignment was required to extend over at least 90% of the length of the EST with at least 90% sequence identity.
In total, GENOA aligned about 85,900 human cDNAs and about 890,300 ESTs to the human genome. The relatively low fraction of aligned ESTs (about 18%), and average aligned length of about 550 bases (the average lengths were not significantly different between different tissues, see Additional data file 6), reflect the stringent alignment-quality criteria that were imposed so as to be as confident as possible in the inferred splicing patterns. The aligned sequences yielded about 17,800 gene regions with more than one transcript aligned that exhibited a multi-exon structure. Of these, about 60% exhibited evidence of alternative splicing of internal exons. Our analysis did not examine differences in 3\'-terminal and 5\'-terminal exons, inclusion of which is frequently dictated by alternative polyadenylation or alternative transcription start sites and therefore does not represent \'pure\' AS \[[@B73],[@B74]\]. The EST alignments were then used to categorize all internal exons as: constitutive exons; A3Es, A5Es, skipped exons, multiply alternatively spliced exons (for example, exons that exhibited both skipping and alternative 5\' splice site usage), and exons that contained retained introns. An internal exon present in at least one transcript was identified as a skipped exon if it was precisely excluded in one or more other transcripts, such that the boundaries of both the 5\' and 3\' flanking exons were the same in the transcripts that included and skipped the exon (for example, exon E3 in Figure [1](#F1){ref-type="fig"}). Similarly, an internal exon present in at least one transcript was identified as an A3E or A5E if at least one other transcript contained an exon differing in length by the use of an alternative 3\' splice site or 5\' splice site. The \'core\' of an A3E or A5E is defined as the exon portion that is common to all transcripts used to infer the AS event. The extension of an alternatively spliced exon is the exon portion added to the core region by the use of an alternative 3\' splice site or 5\' splice site) that is present in some, but not all transcripts used to infer the AS event. Pairs of inferred A3Es or A5Es differing by fewer than six nucleotides were excluded from further analysis, as in \[[@B8]\], because of the possibility that such small differences might sometimes result from EST sequencing or alignment errors. As the frequency of insertion-deletions errors greater than three bases using modern sequencing techniques is vanishingly small (P. Green, personal communication), a six-base cutoff should exclude the vast majority of such errors. Alternatively spliced exons/genes identified in specific tissues are available for download from the GENOA website \[[@B71]\].
Quantifying splice junction differences between alternative mRNA isoforms
-------------------------------------------------------------------------
To quantify the difference in splicing patterns between mRNAs or ESTs derived from a gene locus, the splice junction difference ratio (SJD) was calculated. For any pair of mRNAs/ESTs that have been aligned to overlapping portions of a genomic locus, the SJD is defined as the fraction of the splice junctions that occur in overlapping portions of the two transcripts that differ in one or both splice sites. A sample calculation is given in Figure [4](#F4){ref-type="fig"}. The SJD measure was calculated by taking the ratio of the number of \'valid\' splice junctions that differ between two sequences over the total number of splice junctions, when comparing a pair of ESTs across all splice junctions present in overlapping portions of the two transcripts. A splice junction was considered valid if: the 5\' splice site and the 3\' splice site satisfied either the GT..AG or the GC..AG dinucleotide sequences at exon-intron junctions; and if the splice junction was observed at least twice in different transcripts.
Identification of candidate splicing regulatory motifs
------------------------------------------------------
Over-represented sequence motifs (*k*-mers) were identified by comparing the number of occurrences of *k*-mers (for *k*in the range of 4 to 6 bases) in a test set of alternative exons versus a control set. In this analysis, monomeric tandem repeats (for example, poly(A) sequences) were excluded. The enrichment score of candidate *k*-mers in the test set versus the control set was evaluated by computing χ^2^(chi-squared) values with a Yates correction term \[[@B75]\], using an approach similar in spirit to that described by Brudno *et al*. \[[@B59]\]. We randomly sampled 500 subsets of the same size as the test set from the control set. The enrichment scores for *k*-mers over-represented in the sampled subset versus the remainder of the control set were computed as above. The estimated *p*-value for observing the given enrichment score (χ^2^-value) associated with an over-represented sequence motif of length *k*was defined as the fraction of subsets that contained any *k*-mer with enrichment score (χ^2^-value) higher than the tested motif. Correcting for multiple testing is not required as the *p*-value was defined relative to the most enriched *k*-mer for each sampled set. For sets of skipped exons from human brain- and testis-derived EST sequences, the test sets comprised 1,265 and 517 exons skipped in brain- and testis-derived ESTs, respectively, and the control sets comprised 12,527 and 8,634 exons constitutively included in human brain- and testis-derived ESTs, respectively. Candidate sequence motifs in skipped exons from brain and testis-derived ESTs with associated *p*-values less than 0.002 were retained. For the set of A5E and A3E events from human liver-derived EST sequences, the test set comprised 44 A3Es and 45 A5Es, and the control set comprised 1,619 A3Es and 1,481 A5Es identified using ESTs from all tissues excluding liver. In this analysis, A3Es and A5Es with extension sequences of less than 25 bases were excluded and sequences longer than 150 bases were truncated to 150 bases, by retaining the exon sequence segment closest to the internal alternative splice junction. Over-represented sequence motifs in A3Es and A5Es from liver-derived EST sequences with associated *p*-values less than 0.01 were retained.
Gene-expression analysis of *trans*-acting splicing factors
-----------------------------------------------------------
SR proteins, SR-related proteins, and hnRNPs were derived from published proteomic analyses of the spliceosome \[[@B46]-[@B48]\]. Expression values for these genes were obtained from the \'gene expression atlas\' using the HG-U95A DNA microarray \[[@B43]\] and from a similar set of expression data using the HG-U133A DNA microarray \[[@B45]\]. Altogether, 20 splicing factors - ASF/SF2, SRm300, SC35, SRp40, SRp55, SRp30c, 9G8, SRp54, SFRS10, SRp20, hnRNPs A1, A2/B2, C, D, G, H1, K, L, M, and RALY - were studied in 26 different tissues present in both microarray experiments (Figure [5](#F5){ref-type="fig"}). The data from each gene chip - HG-U95A and HG-U133A - were analyzed separately. The average difference (AD) value of each probe was used as the indicator of expression level. In analyzing these microarray data, AD values smaller than 20 were standardized to 20, as in \[[@B43]\]. When two or more probes mapped to a single gene, the values from those probes were averaged. Pearson (product-moment) correlation coefficients between 20-dimensional vectors for all tissue pairs were calculated, using data from each of the two DNA microarray studies separately.
Additional data files
=====================
Additional data files containing the following supplementary data, tables and figures are available with the online version of this paper and from the GENOA genome annotation website \[[@B71]\]. The lists of GenBank accession numbers of human cDNAs and ESTs that were mapped to the human genome by the GENOA pipeline, GENOA gene locus identifiers, and gene loci with spliced alignments for the 22 human autosomes and two sex chromosomes are provided at our website \[[@B76]\]. Sets of constitutive and alternative exons in genes expressed in the human brain, testis and liver, and control sets used are also provided \[[@B77]\]. Additional data file [1](#s1){ref-type="supplementary-material"} lists the average and median number of ESTs per gene and tissue, and the total number of genes per tissue using different minimum numbers of ESTs. Additional data file [2](#s2){ref-type="supplementary-material"} lists the average total number of AS genes and AS genes containing SEs, A3Es and A5Es using ESTs derived from normal, non-diseased tissues. Additional data file [3](#s3){ref-type="supplementary-material"} lists the number of constitutively spliced and AS genes, and AS genes containing SEs, A3Es and A5Es. Additional data file [4](#s4){ref-type="supplementary-material"} shows the average fractions of AS genes and average fractions of AS genes containing SEs, A3Es and A5Es using ESTs derived from normal, non-disease-derived tissues. Additional data file [5](#s5){ref-type="supplementary-material"} lists categories of cDNA libraries and designated tissues derived from the MGC, IMAGE and CGAP. Additional data file [6](#s6){ref-type="supplementary-material"} shows the average lengths of ESTs that aligned to gene loci expressed in different tissues. Additional data file [7](#s7){ref-type="supplementary-material"} lists human splicing factors of SR, SR-related and hnRNP genes, corresponding Ensembl gene numbers and Affymetrix microarray probe identification numbers. Additional data file [8](#s8){ref-type="supplementary-material"} shows the distribution of the average Pearson correlation coefficient values across different tissues for expression levels of random sets of genes obtained from the Affymetrix microarray data.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
The average and median number of ESTs per gene and tissue, and the total number of genes per tissue using different minimum numbers of ESTs
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
The average total number of AS genes and AS genes containing SEs, A3Es and A5Es using ESTs derived from normal, non-diseased tissues
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 3
The number of constitutively spliced and AS genes, and AS genes containing SEs, A3Es and A5Es
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 4
The average fractions of AS genes and average fractions of AS genes containing SEs, A3Es and A5Es using ESTs derived from normal, non-disease-derived tissues
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 5
Categories of cDNA libraries and designated tissues derived from the MGC, IMAGE and CGAP
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 6
The average lengths of ESTs that aligned to gene loci expressed in different tissues
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 7
Human splicing factors of SR, SR-related and hnRNP genes, corresponding Ensembl gene numbers and Affymetrix microarray probe identification numbers
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 8
The distribution of the average Pearson correlation coefficient values across different tissues for expression levels of random sets of genes obtained from the Affymetrix microarray datat
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank T. Poggio and P. Sharp for stimulating discussions and the anonymous reviewers for constructive suggestions. This work was supported by grants from the National Science Foundation and the National Institutes of Health (C.B.B.), and by a Lee Kuan Yew fellowship (G.Y.) and a Whiteman fellowship (G.K.).
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Levels of alternative splicing in 16 human tissues with moderate or high EST sequence coverage. Horizontal bars show the average fraction of alternatively spliced (AS) genes of each splicing type (and estimated standard deviation) for random samplings of 20 ESTs per gene from each gene with ≥ 20 aligned EST sequences derived from a given human tissue. The different splicing types are schematically illustrated in each subplot. **(a)**Fraction of AS genes containing skipped exons, alternative 3\' splice site exons (A3Es) or 5\' splice site exons (A5Es), **(b)**fraction of AS genes containing skipped exons, **(c)**fraction of AS genes containing A3Es, **(d)**fraction of AS genes containing A5Es.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Examples of tissue-specific AS events in human genes with evidence of splice conservation in orthologous mouse genes. **(a)**Human fragile X mental retardation syndrome-related (*FXR1*) gene splicing detected in brain-derived EST sequences. *FXR1*exhibited two alternative mRNA isoforms differing by skipping/inclusion of exons E15 and E16. Exclusion of E16 creates a shift in the reading-frame, which is predicted to result in an altered and shorter carboxy terminus. The exon-skipping event is conserved in the mouse ortholog of the human *FXR1*gene, and both isoforms were detected in mouse brain-derived ESTs. **(b)**Human betaine-homocysteine *S*-methyltransferase (*BHMT*) gene splicing detected in liver-derived ESTs. *BHMT*exhibited two alternative isoforms differing by alternative 5\' splice site usage in exon E4. Sequence comparisons indicate that the exon and splice site sequences involved in both alternative 5\' splice site exon events are conserved in the mouse ortholog of the human *BHMT*gene. **(c)**Human cytochrome P450 2C8 (*CYP2C8*) gene splicing. *CYP2C8*exhibited two alternative mRNA isoforms differing in the 3\' splice site usage for exon E4 (detected in ESTs derived from several tissues), where the exclusion of a 71-base sequence creates a premature termination codon in exon E4b. Exons and splice sites involved in the AS event are conserved in the mouse ortholog of *CYP2C8*.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Correlation of mRNA expression levels of 20 known splicing factors (see Materials and methods) across 26 human tissues (lower diagonal: data from Affymetrix HU-133A DNA microarray experiment \[45\]; upper diagonal: data from Affymetrix HU-95A DNA microarray experiment \[43\]). Small squares are colored to represent the extent of the correlation between the mRNA expression patterns of the 20 splicing factor genes in each pair of tissues (see scale at top of figure).
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Computation of splice junction difference ratio (SJD). The SJD value for a pair of transcripts is computed as the number of splice junctions in each transcript that are not represented in the other transcript, divided by the total number of splice junctions in the two transcripts, in both cases considering only those splice junctions that occur in portions of the two transcripts that overlap (see Materials and methods for details). SJD value calculations for different combinations of the transcripts shown in the upper part of the figure are also shown.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Comparison of alternative mRNA isoforms across 25 human tissues. **(a)**Color-coded representation of *SJD*values between pairs of tissues (see Figure 4 and Materials and methods for definition of SJD). **(b)**Hierarchical clustering of SJD values using average-linkage clustering. Groups of tissues in clusters with short branch lengths (for example, thyroid/ovary, B-cell/bone) have highly similar patterns of AS. **(c)**Mean SJD values (versus other 24 tissues) for each tissue.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Human genes expressed in the liver with alternative 3\' splice site exons (A3Es) or alternative 5\' splice site exons (A5Es)
:::
Splicing type Ensembl gene ID Gene name Exon numbers Fold-change above median expression, HG-U95A Fold-change above median expression, MG-U74A
--------------- ----------------- ----------------------------------------------------------------- -------------- ---------------------------------------------- ----------------------------------------------
A5E;A3E 091513 Serotransferrin precursor, *TF* 8, 9; 4 100 100
A5E;A3E 115414 Fibronectin precursor, *FN1* 36; 31 10 \-
A5E;A3E 117601 Antithrombin-III precursor, *SERPINC1* 5; 4 100 100
A5E;A3E 136872 Fructose-bisphosphate aldolase, *ALDOB* 3, 8; 4 100 10
A5E;A3E 140833 Haptoglobin-related protein precursor, *HPR* 3 100 10
A5E;A3E 151790 Tryptophan 2,3-dioxygenase, TDO2 3, 5; 4 10 100
A5E;A3E 171759 Phenylalanine-4-hydroxylase, *PAH* 6; 4,10 \- 100
A5E 047457 Ceruloplasmin precursor, *CP* 14, 16 3 \-
A5E 055957 Inter-alpha-trypsin inhibitor heavy chain H1 precursor, *ITIH1* 21 100 10
A5E 111275 Aldehyde dehydrogenase, *ALDH2* 12 3 3
A5E 132386 Pigment epithelium-derived factor precursor, *SERPINF1* 4 10 10
A5E 138356 Aldehyde oxidase, *AOX1* 27, 29 3 3
A5E 138413 Isocitrate dehydrogenase, *IDH1* 3 1 \-
A5E 145692 Betaine-homocysteine *S*-methyltransferase, *BHMT* 4 10 100
A5E 160868 Cytochrome P450, *CYP3A4* 5 10 10
A5E 171766 Glycine amidinotransferase, *GATM* 8 3 3
A3E 080618 Carboxypeptidase, *CBP2* 10 \- \-
A3E 080824 Heat shock protein HSP 90-alpha, *HSPCA* 8 \- \-
A3E 096087 Glutathione *S*-transferase, *GSTA2* 4, 6 10 10
A3E 106927 Protein precursor, *AMBP* 5, 9 100 100
A3E 110958 Telomerase-binding protein P23, *TEBP* 5 \<1 1
A3E 134240 Hydroxymethylglutaryl-CoA synthase, *HMGCS2* 8 10 \-
A3E 138115 Cytochrome P450, *CYP2C8* 4 100 10
A3E 145192 Alpha-2-HS-glycoprotein precursor, *AHSG* 6 100 100
A3E 163631 Serum albumin precursor, *ALB* 9 100 100
A3E 171557 Fibrinogen gamma chain precursor, *FGG* 4 100 100
A3E 174156 Glutathione S-transferase, *GSTA3* 4, 6 10 10
Examples of human AS genes found to exhibit A3E and/or A5E splicing with both isoforms detected in liver-derived ESTs. AS types are listed in the first column, followed by the last six digits of the Ensembl gene number, the gene name and alternative exon numbers. The last two columns list expression levels in human liver and mouse liver tissues, respectively, expressed in terms of the fold-change relative to the median expression level in other tissues (from the DNA microarray data of \[43\] and \[45\], respectively).
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Sequence motifs enriched in skipped exons (SEs) and alternative 5\' splice site exons (A5Es)
:::
AS type /tissue (motif name) Oligonucleotides Occurrences Consensus (% of exons containing)
------------------------------ ------------------ ------------- -----------------------------------
SE/brain (BR1) ` CUCCUG` 169 `CUCC` (45.3)
` CUCCU` 323
` CUCCC` 264
` CUCC` 945
` CCUCCC` 137 `CCUC` (41.0)
` CCUCC` 363
` CCUC` 1021
` GCCUCC` 136
` GCCUC` 375
` GCCUCA` 122
`GGCCUC` 118
`UGCCUC` 108
SE/brain (BR2) ` GGGUU` 97 `GGGU` (25.6)
` GGGU` 411
` AGGGU` 116
SE/brain (BR3) ` UGGGA` 324 `UGGG` (47.2)
` UGGG` 948
` CUGGG` 426
`CCUGGG` 171
SE/brain (BR4) ` GGGAUU` 58 `GGGA` (45.5)
` GGGAU` 176
` GGGA` 840
SE/brain (BR5) ` CUCA` 925 `CUCA` (46.5)
` CUCAC` 206
`GCCUCA` 122
`GGCUCA` 102
` GCUCAC` 79
` CUCAGC` 126
SE/brain (BR6) ` UAGC` 269 `UAGC` (18.0)
` UAGCU` 106
` GUAGC` 96
` GUAGCU` 51
`AGUAGC` 47
` UAGCUG` 54
SE/brain (BR7) ` UAGG` 186 `UAGG` (13.8)
` UUAGG` 63
` UUAGGG` 24
SE/testis (TE1) ` UAGG` 99 `UAGG` (16.6)
` UUAGG` 33
Core A5E/liver (LI1) ` AAAC` 42 `AAAC` (53.3)
` AAAAC` 18
Core A5E/liver (LI2) ` UAAA` 29 `UAAA` (40.0)
` UAAACC` 5
Sequence motifs of length four to six bases that are significantly over-represented (*p*\< 0.002) in SEs relative to constitutively spliced exons from brain- or testis-derived ESTs are shown, followed by the number of occurrences in SEs in these tissues. Sequence motifs are grouped/aligned by similarity, and shared tetramers are shown in bold and listed in the last column, followed by the fraction of SEs that contain the given tetramer. Sequence motifs significantly over-represented (*p*\< 0.01) in the core of A5Es from human liver-derived ESTs are shown at the bottom, followed by the number of A5E occurrences and the fraction of A5Es that contain the given tetramer. Statistical significance was evaluated as described in Materials and methods.
:::
|
PubMed Central
|
2024-06-05T03:55:51.740680
|
2004-9-13
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545594/",
"journal": "Genome Biol. 2004 Sep 13; 5(10):R74",
"authors": [
{
"first": "Gene",
"last": "Yeo"
},
{
"first": "Dirk",
"last": "Holste"
},
{
"first": "Gabriel",
"last": "Kreiman"
},
{
"first": "Christopher B",
"last": "Burge"
}
]
}
|
PMC545595
|
Background
==========
Alternative splicing is a widespread mechanism involved in regulation of gene expression, which enables production of many structurally and functionally different forms of proteins from a single gene, adding to the complexity of the genomes \[[@B1]-[@B3]\]. Different mRNA transcripts of a gene can be expressed in different tissues or developmental stages or physiological conditions \[[@B4],[@B5]\].
An expanding body of expressed sequence data from the human and mouse genomes indicates that alternative splicing is an important mechanism in creating protein diversity, and adds to functional complexity encoded in eukaryotic genomes. Earlier studies indicate that at least 50% of the genes in the human genome are alternatively spliced \[[@B6]\]. Examples include the vast majority of immune system and nervous system genes \[[@B7]\].
Comprehensive analysis of alternative splicing is essential to understand fully the proteomes of organisms \[[@B8]\]. Several reports have indicated that variant splice forms result in proteins with different functions. These can range from minimal changes in function to absolutely opposite functions. For example, the cAMP-response element modulator has three different isoforms with entirely different DNA-binding domains, which are all transcription activators. On the other hand, isoforms of the human transcription factor AML1 function both as positive and as negative regulators of transcription \[[@B9]\]. However, for the majority of genes, the functional significance of alternative splicing is still not known \[[@B8]\].
Transcription is a critical process that specifies the mRNAs and the proteins expressed within a cell. Expression of a given gene is dependent on the interactions of different transcription factors and their cofactors with the regulatory regions of that gene. These transcription factors are in turn regulated by processes that include interaction with other proteins and signaling cascades \[[@B9]\].
Alternative splicing is a mechanism that regulates transcription factor (TF) activity by generating a variety of protein isoforms from a single gene. As noted by Lopez, alternative splicing can affect TF structure in two primary ways \[[@B9]\]: alterations can be in the DNA-binding domains affecting their affinity or specificity; or alterations can modulate interactions of transcription factors with their cofactors. Such changes have been observed experimentally to alter specificity or binding strength or to switch between activator and repressor isoforms of the same TF \[[@B10]\]. TF isoforms can have stage-specific and tissue-specific expression patterns throughout the development of an organism \[[@B9]\]. Little is known about the tissue specificity of alternative splicing \[[@B11]\].
In this paper, we use an integrated approach to analyze DNA and protein sequence data jointly to determine the potential effect of alternative splicing on protein structure and function. We perform a detailed analysis of tissue-specific distribution of alternatively spliced mouse TFs to gain biologically meaningful insights into regulation of gene expression by alternative splicing.
Results
=======
Definitions
-----------
For our joint DNA-protein analysis described here, we developed MouSDB3 \[[@B12]\], which identifies, classifies, computes, stores and answers queries about splice variants within the mouse genome. As described in Materials and methods, MouSDB3 uses the mouse genome and expressed sequences in GenBank \[[@B13]\] and dbEST \[[@B14]\] to compute splice variants of mouse transcripts organized by genomic loci. This section provides definitions of terms used in MouSDB3 and in the joint DNA-protein analysis method described here. A \'transcript\' is a sequence transcribed from the genomic DNA sequence. MouSDB3 is restricted to transcripts with at least one splice junction. A \'locus\' is a genomic region that includes a set of overlapping transcripts mapped to the genome such that a transcript appears in only one locus and all transcripts whose genome coordinates overlap by at least one nucleotide are included in the locus. Within a locus, a \'cassette exon\' is completely included in some transcripts and completely excluded in others. A \'length variant exon\' has alternative 5\' or 3\' splice sites, or both, in different transcripts. An exon can be both length variant and cassette. A \'variant exon\' is either cassette or length variant or both. We consider an exon whose number of nucleotides is a multiple of three and which starts at the first base of a codon to be an \'in-frame exon\'. Such exons do not introduce an amino-acid substitution or a stop codon when skipped, unless they are terminal exons within the coding sequence. A \'genomic exon\' is an uninterrupted series of nucleotides, each of which is mapped to a transcript. By this definition the genomic exon for a length variant exon reflects the outermost splice sites. A \'cluster\' is the set of transcripts that map to a locus. A \'variant cluster\' contains one or more variant exons. An \'invariant cluster\' has no variant exons.
MouSDB3 cluster analysis
------------------------
Our cluster analysis revealed that out of the 461 TF clusters, 62% are variant, compared to 29% of all genes in MouSDB3 (Table [1](#T1){ref-type="table"}). The majority (62%) of the variation in TFs is due to cassette exons, which is comparable to cassette-exon distribution in the entire transcriptome (68% of the variant exons in all loci are cassette). As the majority of alternative splicing is due to cassette exons, we focus on these exons for our analyses.
Cassette exon analysis
----------------------
We screened the 287 variant TF clusters for the presence of cassette exons within coding sequences. We categorized MouSDB3 transcripts into three categories with respect to each cassette exon within a cluster. Category 1 transcripts contain the exon and are referred to as \'long transcripts\'. Category 2 transcripts skip the exon and are referred to as \'short transcripts\'. Category 3 transcripts do not overlap with the cassette exon due to 5\' or 3\' truncations. In our structural analysis, we computationally delete in-frame cassette exons from Category 1 transcripts to produce an \'altered transcript\'. Figure [1](#F1){ref-type="fig"} displays a MouSDB3 cluster and illustrates these categories.
The 287 variant TF loci contain 324 cassette exons of which 23% (76 exons) are in-frame. Only 11% of cassette exons are expected to be multiples of three and in codon position 1 randomly. The twofold difference between expected and observed numbers indicates a bias towards in-frame cassette exons. The exons which are a multiple of three and in codon position 2 and 3 comprise 10% and 7%, respectively. When deleted, these exons introduce an amino-acid substitution to the sequence. As exons which are a multiple of three starting at codon position 1 are enriched and do not introduce an amino-acid substitution when deleted, our study focuses on these exons only.
As shown in Figure [2](#F2){ref-type="fig"}, of the 76 in-frame cassette exons, 66 have domain architectures predicted by SMART. The remaining 10 exons are either from transcripts with too short sequences or these transcripts do not have any of the domains annotated in SMART. Of the 66 in-frame cassette exons, 80% (53) induce a domain-structure alteration to the protein when skipped. Of these 53 structure-altering exons, 68% are within coding regions for the domains that are important for TF activity, such as DNA-binding or activation domains. The remaining 32% (17) of exons are proximal to the computed domain boundaries; that is, the domain is coded by the upstream or the downstream neighboring exon of the cassette exon. When the cassette exon is removed, the sequence no longer meets the computational criteria for the domain (Figure [2](#F2){ref-type="fig"}).
Assessing domain architecture alterations
-----------------------------------------
SMART \[[@B15],[@B16]\] and Pfam \[[@B17],[@B18]\] entries for the altered domains revealed that 75% of the domains affected by alternative splicing with known functions are DNA-binding domains. The names of all altered domains and links to their annotated biological functions are provided on our web page \[[@B19]\]. There we provide the 53 in-frame cassette exons (shown in Figure [2](#F2){ref-type="fig"}), which alter the domain architecture of their transcripts when skipped. Links to MouSDB3 clusters containing these transcripts and links to their GenBank entries are provided. In addition, we provide the names of the domains altered by these 53 exons as active links to their SMART and Pfam annotations. All sequences for long transcripts, altered transcripts and in-frame cassette exons are provided as links to fasta files on the same web page. Our domain-alteration results correlate with recent findings of Resch *et al*. \[[@B20]\], who show that alternative splicing preferentially removes certain domains more frequently.
Tissue-distribution analysis
----------------------------
Part two of our analysis assessed the tissue distribution of alternatively spliced transcription factors. We chose 18 tissues from the existing libraries in MouSDB3 on the basis of the fact that they contain both variant and invariant transcripts annotated as TFs. There are a total of 1,413 library names in MouSDB3 imported from expressed sequence records in GenBank and dbEST. Of these, 328 are ambiguous in that they list several different tissues or cell types for a single library, such as \'mixture of brain and testis\' or no tissues at all, such as \'embryo or carcinoma\'. For the work described here we did not include tissue information from such ambiguous libraries. There are a total of 95 libraries in MouSDB3 for which there are TF transcripts. In addition, to account for library ambiguities within these 95 libraries, we pooled different parts of a tissue into one library. For example, the term \'brain\' corresponds to all parts of the brain found in MouSDB3, including cerebellum, thalamus, hippocampus and 16 other libraries. When analyzing the tissue distribution of all genes, only the libraries that contain TF transcripts have been used.
Transcript counts within variant loci for 18 pooled libraries indicated that in 14 of the 18 analyzed tissues, the proportion of TFs that are variant is higher than the proportion of all genes that are variant (Figure [3a](#F3){ref-type="fig"}). This finding, together with the observation that 62% of TF loci are variant, indicates the widespread impact of alternative splicing on regulation of gene expression via TFs.
For each of the 18 tissues in Figure [3a](#F3){ref-type="fig"}, we compared the proportion of TFs that vary to the proportion of all genes that vary. As shown in Figure [3b](#F3){ref-type="fig"}, eight tissues exhibited more than twofold difference in variant TFs versus variant genes in total. (Note that values in Figure [3b](#F3){ref-type="fig"} are base 2 logarithms of the ratios. Tissues with twofold differences have log~2~values above 1 on the graph). In salivary gland, skeletal muscle, urinary bladder and testis, the fold-differences are 8.7, 5.6, 3.8 and 3.0-fold respectively. Spinal cord, liver, adipose tissue and eye also exhibit more than twofold differences. These values are independent of the sampling depth of the transcripts from these tissues, as illustrated in Figures [4a](#F4){ref-type="fig"} and [4b](#F4){ref-type="fig"}. Sampling depth is the number of transcripts sequenced per tissue (either a single library or a pooled library as in the case of \'brain\'). Figure [4a](#F4){ref-type="fig"} displays absolute numbers of variant TF transcripts and Figure [4b](#F4){ref-type="fig"} displays absolute numbers of the entire variant transcripts of the transcriptome. In Figures [4a,b](#F4){ref-type="fig"}, tissues are presented along the *x*-axis as in Figure [3b](#F3){ref-type="fig"} for the reader\'s convenience. The correlation coefficient of the absolute numbers of TFs and the fold-differences between variant TFs and all genes is -0.13, indicating that they do not correlate. Likewise, the correlation coefficient of the absolute numbers of all genes and the fold-differences between variant TFs and all genes is -0.46. Additionally, the scatter-plots in Figures [4c,d](#F4){ref-type="fig"} show that there is no correlation between the fold-differences and sampling depth. The datasets used in calculating the correlation coefficients can be found on our web page \[[@B19]\].
Isoform heterogeneity
=====================
We analyzed the presence of different isoforms of transcription factors within and across these 18 tissues. For this analysis we consider transcripts with coding sequence information only. We ignore variation due to 5\' and 3\' truncation of transcripts. We consider only cassette exons within coding sequences when assessing the differences between isoforms. Within a cluster we compute homogeneity and heterogeneity within a single tissue by checking for the transcripts from the same library and comparing the cassette exons within their coding sequences. If all transcripts from the same tissue contain the same cassette exons with same splice sites they are termed \'homogeneous within\'. If the cassette exon distribution within the coding sequences of these transcripts differ, they are termed \'heterogeneous within\'. We compute \'homogeneity across\' and \'heterogeneity across\' tissues in the same way by taking into account transcripts within the same clusters but from different libraries. As shown in Figure [5](#F5){ref-type="fig"}, when heterogeneity to homogeneity ratios are compared within and across tissues, there is significantly more heterogeneity of isoforms across tissues than within a single tissue (*p*-value = 0.04). This is true for both transcription factors and the rest of the genes in the mouse transcriptome.
When single tissues are taken into account, TFs are more homogenous within each tissue analyzed. As shown in Figure [6](#F6){ref-type="fig"}, heterogeneity to homogeneity ratios in all tissues are lower than 1, indicating that these tissues are more homogeneous in terms of TF isoforms. In fact, except for brain and thymus, all values for TFs are zero, hence the absence of blue bars from Figure [6](#F6){ref-type="fig"}. When all genes are considered, heterogeneity to homogeneity ratios are also below 1, indicating homogeneity of isoforms of all genes within these tissues. However, there is still a significant difference in heterogeneity to homogeneity ratios between TF isoforms and isoforms of all genes: TFs are significantly more homogeneous within single tissues when compared to all genes (*p*-value = 0.02). (The data used in calculating the homogeneity and heterogeneity values can be found on our web page \[[@B19]\].)
Figures [5](#F5){ref-type="fig"} and [6](#F6){ref-type="fig"} show that the majority of TF isoforms and the isoforms of all alternatively spliced genes differ across tissues: within a given single tissue there generally is only one isoform. These data indicate the presence of tissue-specific alternative splicing throughout the mouse transcriptome. In addition, our findings indicate expression of different TF isoforms in different tissues. This implies contribution of alternative splicing to regulation of gene expression in a tissue-specific manner by controlling activation or repression of different sets of genes in different tissues via variant TF isoforms. These data have significant implications in further understanding the regulation of tissue-specific gene expression and control of transcription.
Discussion
==========
Through integrated analyses of DNA and protein sequences for TF genes, we show that alternative splicing of TFs are more prevalent in the entire mouse transcriptome and in specific tissues when compared to alternatively spliced forms of all the genes. In 78% of the tissues analyzed, higher proportions of TFs exhibit alternative splicing compared to all the genes in the mouse transcriptome. This result, along with the finding that 62% of TF loci are variant, indicates the widespread impact of alternative splicing on regulation of TF function.
We also show that alternative splicing changes TF structure by adding or deleting domains. This study reveals that 80% of alternatively spliced TFs have different domain architectures due to introduction of an in-frame cassette exon by alternative splicing. Of the altered domains, 75% have a role in DNA binding. These findings provide quantitative evidence for the role of alternative splicing in controlling the presence of domains in the proteins. They also suggest that alternative splicing might regulate TF activity by changing the architecture of the DNA-binding domains of these proteins.
Our analyses revealed that within a single tissue there generally is only one TF isoform, and that across tissues, isoforms differ. This finding indicates tissue specificity of alternatively spliced TFs and suggests that TFs might regulate gene expression in a tissue-specific manner by having different isoforms in different tissues. These findings further indicate the role of alternative splicing in regulation of tissue-specific gene expression. Activation and repression of different sets of genes within different tissues can be regulated through variant TF isoforms created by alternative splicing. These findings will significantly aid further understanding of control of transcription and tissue-specific gene expression. In addition, our study shows that all variant loci in the mouse transcriptome display isoform homogeneity within single tissues and heterogeneity across tissues. This finding demonstrates the presence of tissue-specific alternative splicing across the mouse transcriptome and greatly expands the knowledge on the tissue specificity of alternatively spliced genes.
Conclusions
===========
Overall, our study provides quantitative evidence for the effect of alternative splicing on protein structure and sheds light on how alternative splicing might regulate transcription factor function in a tissue-specific manner. This, in turn, reveals the contribution of alternative splicing to regulation of gene expression via tissue-specific TF isoforms. The work described here implies that future high-throughput screens of gene expression analyses should be sensitive to multiple alternatively spliced forms of TFs. Because gene-expression arrays are intended to measure transcription, the next generation of arrays should contain probes specific to all known isoforms of genes represented on the arrays. Given that alternatively spliced exons are highly conserved across species \[[@B21],[@B22]\], it would be of further interest to extend this study to other organisms. Strong sequence homology between mouse, human and rat exons suggests that a comparative analysis of human, mouse and rat TF variations will be a natural extension of the studies described here.
Materials and methods
=====================
Development of the alternative splicing database MouSDB3
--------------------------------------------------------
For this analysis, we constructed a database of alternatively spliced mouse transcripts called MouSDB3 \[[@B12]\], using the methods described in \[[@B23]\]. Briefly, full-length transcript nucleotide sequences were obtained by an Entrez query on 5 August 2003 from GenBank \[[@B24]\] with molecule selected as mRNA and limits used to exclude expressed sequence tags (ESTs), sequence-tagged sites (STSs), genome sequence survey (GSS), third-party annotation (TPA), working draft and patents. EST sequences were downloaded on 31 July 2003 from dbEST \[[@B25]\] by extracting only *Mus musculus*entries. All expressed sequences were mapped to a region of the University of California Santa Cruz (UCSC) February 2003 version mm3 of the mouse genome assembly using BLAT \[[@B26]\]. BLAT tools gfServer and gfClient were installed from jksrc444 dated 15 July 2002 \[[@B27]\]. This was followed by a careful alignment by SIM4 \[[@B28]\] version 3/3/2002 to establish splice sites of exons. A post-processing analysis computed genomic exons and determined types of variation for each exon, transcript and locus.
Cassette exon analysis
----------------------
We identified in-frame cassette exons and extracted from MouSDB3 nucleotide and amino-acid sequences for transcripts containing these exons. The selected amino-acid sequences were then analyzed with SMART \[[@B29],[@B30]\] to compute protein-domain architecture for each transcript within a cluster.
Tissue distribution of alternatively spliced TFs
------------------------------------------------
From MouSDB3, we then extracted library information for the transcripts within clusters and their annotations. We used these data to compute the tissue distribution of variant transcripts as reported in Results. All scripts and README files used to carry out this data-gathering process are available upon request from the Laboratory of Computational Genomics of The Rockefeller University.
Acknowledgements
================
We acknowledge support from Mathers Foundation and Hirschl Foundation. This work has been partially funded by NSF grant DBI9984882 and NIH grant GM62529 to T.G. We thank Joseph A. Sorge for suggestions regarding the tissue-distribution analyses and members of Laboratory of Computational Genomics for their support. Corresponding author T.G. can be reached at <gaasterland@ucsd.edu> as well as at <gaasterl@genomes.rockefeller.edu>.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Transcripts of a MouSDB3 cluster. **(a)**Partial image of MouSDB3 cluster number scl24819 \[31\] displaying alternatively spliced transcripts. **(b)**Categorization of transcripts with respect to the cassette exon indicated by the arrow. This figure shows an example transcript for each of the three categories from the scl24819 cluster. Category 1, long transcript with cassette exon indicated by the arrow. Category 2, short transcript skips the cassette exon. Category 3, cassette exon is missing owing to a 5\' truncation. Pink bars represent in-frame cassette exons. Green and blue bars represent exons with other types of splice variation. (Green are invariant, blue are length-variant exons). The red line represents intronic regions of the genome sequence.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Transcription factor cassette exon analysis. This figure illustrates the distribution of 324 cassette exons within variant TF transcripts. These 324 exons are from 287 different variant MouSDB3 clusters. When 76 of the 324 cassette exons are skipped, the altered transcripts are in-frame; exclusion of remaining exons either introduces an amino-acid substitution or causes frameshifting. Of the in-frame exons, 53 alter domain architecture and 13 do not. Of the exons that cause domain alteration, 36 are in coding regions for domains and 17 are proximal to these coding regions. In-frame cassette exon sequences, sequences of their transcripts and annotations of the domains they alter are provided on our web page \[19\].
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
TF variation is higher in the majority of tissues compared to all genes. **(a)**Tissue distribution of alternatively spliced TFs versus tissue distribution of all alternatively spliced genes. For each tissue, the number of variant TF transcripts in tissue normalized by the total number of variant TF transcripts in MouSDB3 is represented as a blue bar. This number is computed as follows: *t*= number of variant TF transcripts in tissue; *T*= total number of variant TF transcripts; bar value = (*t*/*T*× 100). Red bars represent the number of variant transcripts of all genes in the tissue normalized by the total number of variant transcripts in MouSDB3. This value is computed as follows: *a*= total number of variant transcripts in tissue; *A*= total number of all variant transcripts in MouSDB3; bar value = (*a*/*A*× 100). **(b)**Fold differences in variant number of transcripts between TFs and all genes. This value is computed as follows: bar value = log~2~((*t*/*T*)/(*a*/*A*)). Tissues are in descending order from highest to lowest fold difference of variation in TF versus variation in all genes. Tissue abbreviations: SG, salivary gland; SM, skeletal muscle; UB, urinary bladder; SC, spinal cord; AT, adipose tissue; MG, mammary gland.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Higher variation in TFs is independent of sampling depth from each tissue. **(a)**Absolute number of variant TF transcripts per tissue. **(b)**Absolute number of all variant transcripts per tissue. **(c)**For each tissue (labeled to the right of each data point), *x*-axis: ratio of variant TF transcripts to all variant transcripts (*x*= (*t*/*T*)/(*a*/*A*)); *y*-axis: absolute numbers of variant TF transcripts. See Figure 3 legend or definitions of *t*, *T*, *a*and *A*. **(d)**For each tissue (labeled to the right of each data point), *x*-axis: ratio of variant TF transcripts to all variant transcripts (*x*= (*t*/*T*)/(*a*/*A*)); *y*-axis: absolute numbers of all variant transcripts. Tissue abbreviations: SG, salivary gland; SM, skeletal muscle; UB, urinary bladder; SC, spinal cord; AT, adipose tissue; MG, mammary gland.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Isoforms of alternatively spliced genes are more heterogeneous across different tissues than within single tissues. The blue bars represent the ratio of all TF clusters with multiple isoforms within a tissue to all TF clusters with only one isoform within each tissue. The red bars represent the ratio of all variant clusters with multiple isoforms within a tissue to all variant clusters with only one isoform within each tissue.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
Heterogeneity versus homogeneity of isoforms in single tissues. The blue bars represent the ratio of TF clusters with multiple isoforms within the given tissue to TF clusters with only one isoform within that tissue. The red bars represent the ratio of variant clusters with multiple isoforms within the given tissue to variant clusters with only one isoform within that tissue. Tissue abbreviations: MG, mammary gland; SM, skeletal muscle; SC, spinal cord.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Cluster analyses of transcription factors and entire MouSDB3
:::
Transcription factors Entire MouSDB3
------------------------------ ----------------------- ----------------
Total number of clusters 461 55,087
Number of invariant clusters 174 (38%) 39,273 (71%)
Number of variant clusters 287 (62%) 15,814 (29%)
:::
|
PubMed Central
|
2024-06-05T03:55:51.746344
|
2004-9-30
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545595/",
"journal": "Genome Biol. 2004 Sep 30; 5(10):R75",
"authors": [
{
"first": "Bahar",
"last": "Taneri"
},
{
"first": "Ben",
"last": "Snyder"
},
{
"first": "Alexey",
"last": "Novoradovsky"
},
{
"first": "Terry",
"last": "Gaasterland"
}
]
}
|
PMC545596
|
Background
==========
Complete genome analysis showed the tremendous extent to which gene and genome duplication events have shaped genomes over time. Remarkably, 30% of the *Saccharomyces cerevisae*genome, 40% that of *Drosophila melanogaster*, 50% that of *Caenorhabditis elegans*, and 38% of the human genome are composed of duplicated genes \[[@B1],[@B2]\]. According to Ohno\'s theory \[[@B3]\], such duplication events should have provided genetic raw material, a source of evolutionary novelties, that could have led to the emergence of new genes and functions through mutations followed by natural selection. But despite the recent increase in genomic knowledge, the patterns by which gene duplications might give rise to new gene functions over the course of evolution remain poorly understood. This is mainly explained by the fact that there are very few ways of experimentally investigating the evolution of function of duplicated genes. Studying the function of duplicated genes usually means estimating the extent of the conservation/divergence between duplicates from comparison of actual sequences. For this purpose, the sequence divergence, the divergence time and the selective constraints on gene pairs are usually calculated (as in \[[@B4]\]). Given that these calculations are only valid on a relatively short timescale \[[@B4],[@B5]\], they exclude *de facto*the study of ancient duplication events (such as the complete duplication of the yeast genome \[[@B6]-[@B8]\]), even though remnants of such events are still present in the genomes \[[@B9]\]. Enlarging the timescale on which we are able to work is thus a desirable goal, which may be reached by using other means to evaluate the functional conservation/divergence between duplicates.
In addition, sequence analysis generally only reveal the possible molecular (biochemical) function(s) of proteins and even this only applies when domains of known function are identified in the sequences. As discussed previously \[[@B10]\], the function of a gene or protein can be defined at several integrated levels of complexity (molecular, cellular, tissue, organismal) As far as genome evolution is concerned, consideration of the functional evolution of genes and proteins not only at the basal molecular level, but also at upper, more integrated, levels is particularly important. In this respect, it is essential to consider the cellular function of genes/proteins - that is, the biological processes they are involved in. One can easily imagine, for instance, that the evolution of a duplicated pair of protein kinases, having the same molecular function, could potentially result in the emergence of a new signaling pathway involved in a different cellular function. Being able to study the evolutionary fate of duplicated genes at the level of cellular function using bioinformatics methods, something that was quite difficult until now, may thus provide new insights into the field. To do so, one needs to be able to easily compare the functions of many proteins at once and to estimate their functional similarities at the cellular level.
Function comparison was one of our aims while developing PRODISTIN, a computational method that we recently proposed \[[@B11]\]. This method permits the functional classification of proteins solely on the basis of protein-protein interaction data, independently of sequence data. It clusters proteins with respect to their common interactors and defines classes of proteins found to be involved in the same cellular functions.
In the work presented here, we addressed the question of the cellular functional fate of duplicated genes in the yeast *S. cerevisiae*, focusing on the 899 duplicated genes which represent remnants of an ancient whole-genome duplication (WGD) \[[@B6]-[@B8]\]. This event took place 100-150 million years ago in the *Saccharomyces*lineage, after the divergence from *Kluyveromyces waltii*, and was probably followed by a gene-loss event leading to the current *S. cerevisiae*genome \[[@B8]\]. Overall, these duplicated genes form 460 pairs of paralogs, accounting for 16% of the current genome \[[@B6]\].
After applying the PRODISTIN method to the yeast interactome, we established and analyzed the functional classification of the duplicated yeast genes originating from the WGD. This analysis allowed us to compare the cellular function(s) of 41 paralog pairs for which enough interaction data was available. Three different behaviors of the pairs of paralogs in respect of the PRODISTIN classification were identified from this analysis, allowing us to establish a scale of functional divergence for the duplicated genes based on the protein-protein network analysis. This work validates the use of interaction data and the analysis of interaction networks as a new means of investigating evolutionary processes at the level of the cellular function.
Results
=======
GO annotations do not functionally distinguish between duplicated pairs from the ancient genome duplication
-----------------------------------------------------------------------------------------------------------
To obtain a first estimation of the functional conservation/divergence of the yeast duplicated genes, we analyzed available textual information relative to the actual functions of the 460 pairs of paralogs from the WGD. For this purpose, we used the Gene Ontology (GO) annotations. The Gene Ontology consortium \[[@B12]\] develops structured controlled vocabularies describing three aspects of gene function: \'Molecular Function\' describes the biochemical function of proteins (their molecular activity); \'Biological Process\' describes their cellular function (the \"broad biological goals that are accomplished by ordered assemblies of molecular functions\"); and \'Cellular Component\' describes their subcellular localization. These structured vocabularies, or ontologies, are not organized as hierarchies but as directed acyclic graphs (DAGs), in which child terms (the more specialized terms) can have several parent terms (less specialized terms). These functional annotations thus provide a means of comparing gene functions as long as one is able to take into account the structure of the ontology in the comparison process. We performed a pairwise comparison of the functions of the 460 pairs of duplicates by processing their functional GO annotations with GOproxy \[[@B13]\]. This tool calculates a functional distance between genes based on the shared and specific GO annotations. The calculation is made separately for the three ontologies, and for each gene the complete hierarchy of GO terms, from the root term to the leaf term of the DAG, is considered in the comparison process without differentiating the two parent-child relationships existing in GO (the \'is-a\' and the \'has-a\' relations) (for details see Materials and methods). Two genes that do not share any GO terms would have a maximum distance value (equal to 1), whereas two genes sharing exactly the same set of GO terms would have a minimum distance value (equal to 0).
The distributions of the calculated distance values are showed in Figure [1](#F1){ref-type="fig"}. First, as expected, paralog pairs are globally closer in term of functional distance based on the annotations (Figure [1a](#F1){ref-type="fig"}) than pairs of proteins chosen randomly from the proteome (Figure [1a](#F1){ref-type="fig"}, inset). Indeed, the distribution of the distances peaks at the minimum distance value for the paralogs while it peaks at the maximum distance value for the randomly selected pairs.
Second, the vast majority of the duplicated pairs do not differ significantly when Molecular Function terms are compared: 74.5% of the pairs have a zero distance based on annotations (Figure [1a](#F1){ref-type="fig"}, purple bars). This could be explained by the fact that on one hand, a tight relationship exists between protein sequence similarity and molecular function(s) similarity, and on the other the majority of the paralogs share a percentage sequence identity above the \'twilight zone\' (20-35%) \[[@B14]\], usually considered as a threshold for molecular function similarity.
Given that paralogs with the same molecular function may potentially be involved in different cellular functions, we also considered the Biological Process annotations of gene products. Interestingly, the majority of the paralogs also display a zero distance value, suggesting that a majority of duplicated genes from the ancient duplication do not significantly differ when considering the cellular function annotations. However, although the distribution of the distances between the duplicates for the Biological Process annotations displays the same overall shape, only 56.5% of the pairs show a zero value (Figure [1a](#F1){ref-type="fig"}, blue bars) as compared to 74.5% for the Molecular Function annotations. The fact that, on average, the molecular functions of duplicated pairs are more conserved than their corresponding cellular functions may reflect the fact that changes in function that occurred during evolution are more measurable and discernible at the cellular level than at the molecular level at the present time. This is corroborated by the fact that paralog pairs are found to be globally closer according to the Molecular Function annotation compared to the Biological Process annotation when the expectation values are calculated for each distribution, whereas the converse is encountered for randomly selected pairs (see Additional data file 1). Similarly, changes in subcellular localization (Cellular Component annotations, Figure [1a](#F1){ref-type="fig"}, yellow bars) also appear to be more apparent than changes in Molecular Function (see Additional data file 1).
PRODISTIN interaction network analysis: three classification behaviors
----------------------------------------------------------------------
Immediately after a genome-duplication event, the two duplicated proteins will have the same interactors. As time goes by and mutations occur, these proteins may gain or lose interactors; that is, the number of interactors for each protein of the pair may change as well as their identity. Taking account of the fact that protein action is seldom isolated but rather is exerted in concert with other proteins, studying duplicates according to the interactors they still share and the ones they have lost or acquired since the duplication event may give a hint about how their cellular functions have evolved.
We thus applied the PRODISTIN method \[[@B11]\] to 4,143 selected binary protein-protein interactions involving 2,643 yeast proteins. Briefly, the PRODISTIN method consists of three different steps: first, a functional distance is calculated between all possible pairs of proteins in the interaction network with regard to the number of interactors they share (proteins must have at least three interactors to be considered further); second, all distance values are clustered, leading to a classification tree; third, the tree is visualized and subdivided into formal classes. A PRODISTIN class is defined as the largest possible sub-tree composed of at least three proteins sharing the same functional annotation and representing at least 50% of the individual class members for which a functional annotation is available. Classes of proteins are then analyzed for their biological relevance and tested for their statistical robustness (see Materials and methods and \[[@B11]\] for a detailed explanation). The relevance of the method has been assessed biologically and statistically in a previous study (its first application to a smaller interaction dataset led to the prediction of the cellular function of 42 uncharacterized yeast proteins with a success rate of 67% \[[@B11]\]). In the present work, 890 proteins were classified (Figure [2](#F2){ref-type="fig"}). Among them, 154 correspond to products of duplicated genes from the ancient duplication and 82/154 form 41 pairs of paralogs. These 41 pairs thus correspond to the only pairs from the ancient duplication for which more than three interaction partners per protein are presently known. Then, following the PRODISTIN procedure, the clustering of the proteins was analyzed, defining classes of proteins involved in the same cellular function(s) according to the GO Biological Process ontology (for details, see Materials and methods). In total, 123 classes corresponding to 53 different cellular functions were identified in the tree (see Additional data file 2) and evaluated statistically (data not shown), allowing the classification of 38/41 pairs of duplicated genes (Table [1](#T1){ref-type="table"}).
We then investigated the details of the distribution of the duplicates in the tree by analyzing the PRODISTIN classes. Interestingly enough, three different situations were encountered (Figure [2](#F2){ref-type="fig"}, Table [1](#T1){ref-type="table"}). First, for 26 pairs both gene products were found in the same class. This means that their list of interactors is very similar and that these proteins should thus be involved in the same biological process. This is illustrated by Tif4631 and Tif4632 (Figure [2](#F2){ref-type="fig"}), which are subunits of the translation initiation complex that binds the cap on the 5\' end of mRNAs \[[@B15]\]. In our analysis they both belong to a class devoted to \'Protein biosynthesis\'. Interestingly, they are clustered with other actors of the initiation of translation (Cdc33, Pab1), as well as with proteins involved in cell-wall biogenesis (Kre6, Pkc1, Stt3), thus reinforcing the recent proposal of the existence of a functional link between these two biological processes \[[@B16]\].
Second, three other pairs of duplicates were recovered in different PRODISTIN classes, relatively far away when considering the tree topology (they therefore no longer share the majority of their interactors), but interestingly, the classes containing the duplicates were dedicated to the same biological process. This is reminiscent of a previous observation we made while studying in detail the rationale sustaining the PRODISTIN clustering \[[@B11]\]: classes distant in the tree but corresponding to the grouping of proteins involved in the same biological process often correspond to different aspects of the same biological process. This is the case for the pair composed of Tub1 and Tub4 (Figure [2](#F2){ref-type="fig"}), which are classified in different PRODISTIN classes both annotated \'cytoplasm organization and biogenesis\' and \'cell cycle\' (PRODISTIN classes may be annotated with several cellular functions \[[@B11]\]). These two proteins are structural components of the cytoskeleton that are implicated in microtubule organization. But strikingly, these two paralogous genes have different roles relative to microtubules. Tub1 is an alpha-tubulin and thus a component of the microtubule itself, whereas Tub4 is a gamma-tubulin involved in the nucleation of the microtubules on both the nuclear and the cytoplasmic sides of the spindle-pole body \[[@B17]\]. Consequently, the class containing Tub1 is more structural and mainly composed of proteins implicated in microtubule formation, orientation and catabolism (Kar9, Bim1, Pre4), whereas the class containing Tub4 includes actors of the nuclear processes in which the microtubules are involved: chromosome segregation, spindle orientation and nuclear migration (Spc72, Spc97, Spc98, Spc110, Mcm16, Yfr008w, Far3, Vps64, Ylr238w, Ynl127w). Thus, it appears that the PRODISTIN classification of these two paralogous proteins reflects their functions in two different aspects of the same biological process.
Finally, nine pairs of duplicated genes were found in different classes devoted to different biological processes. This is exemplified by the case of Ace2 and Swi5 (Figure [2](#F2){ref-type="fig"}), which are two transcription factors regulating the expression of cell-cycle-specific genes. Although they regulate a shared set of genes *in vivo*, they display different specificities in some cases. Swi5 specifically promotes transcription of the *HO*gene whereas Ace2 localizes to daughter cell nuclei after cytokinesis, regulates the expression of daughter-specific genes and delays the G1 progression in daughters \[[@B18]-[@B20]\]. The PRODISTIN classification was successful in pointing towards these differences as Swi5 and Ace2 localize in different classes annotated for \'transcription\' and \'cell cycle\', respectively. Indeed, Swi5 is found with Pho2, a transcription factor acting in a combinatorial manner, with which it interacts to regulate *HO*transcription \[[@B21]\]. Other Pho2 partners populate the rest of the class. On the other hand, Ace2 partitioned with Mob2 and Cbk1, which form a kinase complex regulating the localization of Ace2 in the daughter cell \[[@B20]\].
Overall, this analysis shows that the duplicated gene pairs from the ancient duplication present in the tree display three different behaviors in respect of the PRODISTIN classification (Table [2](#T2){ref-type="table"}). The three groups are populated differently: 63% of the protein pairs are located in the same class, and are therefore involved in the same biological process (behavior I); 7.5% of the duplicated pairs are located in different classes with the same function, therefore suggesting that they are involved in different aspects of the same biological process (behavior II); and, finally, the remaining 22% are implicated in different cellular functions because they are located in different classes devoted to different biological processes (behavior III).
We propose considering the three behaviors identified by the PRODISTIN classification as a scale of functional divergence for duplicated pairs. First, the duplicated pairs found in the same class and which essentially have identical interactors would compose the basic level of the scale. This level represents paralogous genes for which cellular function is identical or highly conserved. Higher in the functional scale of divergence are found the duplicates that have different interactors. They are found either in different classes of the same cellular function, thus defining the intermediate level of the functional scale of divergence, or in different classes of different function. This latter case populates the higher level of the scale and represents paralogs for which the cellular function has diverged.
The relationship between the functional distance based on annotation and the classification behavior based on protein-protein interactions
------------------------------------------------------------------------------------------------------------------------------------------
As noted above, most of the 460 duplicated gene pairs from the ancient duplication were not distinguishable when considering either the functional annotations for Molecular Function or Biological Process as their functional distances based on annotations were mainly equal or close to zero. We have also shown (Figure [1b](#F1){ref-type="fig"}) that the subset of 41 paralogous pairs characterized in the PRODISTIN analysis exhibits the same distribution of distance values based on annotations as the 460 pairs. Because the PRODISTIN method allowed us to distinguish three categories of duplicated gene pairs with different types of functional similarities, we wondered if and how the results of the annotation and interaction clustering were correlated. To investigate this, we reported the PRODISTIN behaviours of the paralogs on the distribution of their functional distance based on the Biological Process annotations (Figure [3](#F3){ref-type="fig"}). Among the duplicated pairs that are similarly annotated, we were able not only to distinguish gene pairs found in the same class, as expected for a correlation between the results of the two approaches (behavior I, blue), but also gene pairs involved in different aspects of the same biological process (behavior II, pink) as well as gene pairs not implicated in the same biological processes (behavior III, gray). The last two cases reveal that whereas annotations do not allow us to differentiate certain paralogs from each other functionally, interactions do unveil subtle functional differences. Conversely, paralogous genes may be grouped in the same PRODISTIN class even though their annotations are not completely similar (up to an annotation-based functional distance equal to 0.6). Interestingly, pairs of duplicated genes partitioning into different classes with different functions are encountered independently of the functional distance based on annotation range. This again underlines the fact that the classification based on interactions identifies functional details that are not discernible at the level of annotation only. Therefore, the protein-protein interactions processed by PRODISTIN bring supplementary functional information about the function of the duplicated genes.
Sequence evolution versus functional evolution of duplicated genes
------------------------------------------------------------------
The availability of 41 yeast paralog pairs for which a pairwise functional comparison can be proposed, offers for the first time the possibility of studying the relationship (if any) between sequence conservation/divergence and evolution of cellular function. Because we have proposed here a three-level scale of possible functional divergence between paralog pairs, what can be said about the sequence-identity patterns shown by protein pairs within and between these three groups? To answer this question, 41 binary sequence comparison analyses were performed (one for each paralogue pair) and the results are displayed according to the classification behavior of the pair identified in the PRODISTIN analysis (Figure [4](#F4){ref-type="fig"}). If paralogs displaying behaviors I, II and III are compared, three observations can be made: first, all gene pairs that show more than 55% sequence identity display behavior I, with one noticeable exception. It is clear, however, that despite the fact that all the protein pairs of this class have been classified by the PRODISTIN analysis as essentially having a conserved function, their degree of sequence identity covers, in a nearly uniform manner, a wide range comprising 16 to 95% sequence identity. Second, and conversely, gene pairs with between 15 and 55% sequence identity are found in all three classes, clearly indicating that neither cellular functional similarity nor divergence can confidently be deduced for paralog pairs with sequence identity falling in this range. Third and strikingly, no clear distinction can be made on the basis of sequence identity between paralogs found in different classes with (behavior II) or without (behavior III) identical functions. In summary, as suggested by a preliminary study \[[@B22]\], a simple relationship cannot be established between sequence identity and the cellular functional similarity revealed by the interaction-network analysis. So, as previously shown for the annotations, the functional classification based on interactions is able to underline properties of the duplicates that are not discernible when only sequences are compared.
Discussion
==========
Bioinformatic study of the interaction network as a tool to investigate the function of the duplicated genes
------------------------------------------------------------------------------------------------------------
We have shown here that studying the cellular interactome using bioinformatics methods leads to a proposal of a functional scale of divergence for yeast duplicated genes. As our work makes use of functional gene annotations and interaction lists, it is important to examine how the quality of these two types of data could potentially affect the conclusions that can be drawn from our studies.
Gene annotations provided by the GO consortium \[[@B12]\] are the result of collaborative work by experts, and all annotations are supported by at least one type of experimental evidence. This, together with the use of a controlled vocabulary consistently applied for all annotations, is in principle a good guarantee of annotation quality. However, several potential problems should be taken into account when using annotations. First, all gene products are not annotated. This is the case for 30% of the pairs of duplicated genes, for which at least one gene is not annotated. Second, annotation errors can propagate in the databases, due to the transfer of annotations from gene to gene based only on sequence or structural similarities. In GO, some functional annotations are \"inferred from sequence or structural similarity\" (ISS), meaning that the annotation assignment is not supported by experimental evidence *per se*. It can then can be argued that paralog pairs may be more prone to such annotation transfers than other genes because of their sequence identity. In such a case, our measure of functional distance according to annotations would be largely meaningless. We thus estimated the amount of genes for which GO annotations are solely \'inferred from sequence or structural similarity\'. Interestingly enough, they account, at the level of the complete genome, for only 10.3% and 4.95% of the Molecular Function and the Biological Process annotations, respectively. Similar low values are encountered for the 460 pairs of paralogs (11.2% and 4.5%), allowing us to neglect the weight of such inferred annotations in our distance calculation.
As far as the quality of interactions is concerned, two main problems result from erroneous (false-positive) interactions and missing (false-negative) interactions. Taking into account that the PRODISTIN method was largely statistically assessed for robustness against the presence of false interactions in our previous study \[[@B11]\], we can anticipate that the classification behaviors found in the present analysis will be confirmed, or only slightly modified, in the near future when new interactions are discovered.
The ancestral yeast genome duplication as a case study for functional evolution of paralogs
-------------------------------------------------------------------------------------------
In the present analysis, we worked solely on pairs of paralogs that supposedly originated from the ancient WGD \[[@B6],[@B7]\]. This choice was made for several reasons. First, after the yeast WGD hypothesis, we can consider that all genes, remnants from this event, have duplicated simultaneously. This sets a \'time 0\' for the duplication event and therefore enables us to avoid the problem of determining the age of the duplication events, a problem inherent in all genome-wide analyses of paralogs. Second, after a WGD, polyploidization preserves the necessary stoechiometric relationships between gene products, while the duplication of a single gene does not: duplicates are then out of balance with their interacting partners. This is an important parameter to consider when one wants to study the evolution of the duplicated genes through the analysis of interactions, as we did in this work. Third, studying the remnants of a WGD after more than 100 million years \[[@B7],[@B23]\] allows one to estimate how the sequence, function and interactors of the paralog gene products have evolved since their origin, when their sequence, function(s) and interactor(s) were identical.
An important issue for the interpretation of our results is the validity of the hypothesis of the existence of a WGD in *S. cerevisiae*. Initially proposed by Wolfe and Shields \[[@B7]\], the WGD model has been controversial and alternative models of local duplications have been proposed \[[@B24]-[@B27]\]. Very recently, a novel proof of WGD was provided \[[@B8]\]. Among the 460 paralog pairs we studied, 362 were shown by this new analysis to arise from the WGD. Revisiting our results to take into account the new dataset of duplicated genes did not change them drastically. The distribution of the duplicated pairs becomes 68, 4.5 and 18% for the three different categories of classification behaviors (I, II, III), respectively, compared to 63, 7.5 and 22% for the dataset we used (Table [2](#T2){ref-type="table"}).
The evolution of cellular function: from the scale of functional divergence to the evolutionary fates of the duplicated genes
-----------------------------------------------------------------------------------------------------------------------------
Our study was driven by the idea that investigating the cellular rather than the molecular function of the duplicated genes might provide new information about the extent of their actual divergence and, consequently, might help us to envisage how their cellular function has evolved since the duplication event. Indeed, the first important outcome of our study, based on the comparison of annotations for duplicated pairs, is that although both the molecular and cellular functions of the majority of protein pairs have been conserved since the date of the WGD, cellular functions have evolved more rapidly than molecular functions. Although this finding could seem rather intuitive, it is, to the best of our knowledge, the first time that evidence has been proposed in its favor. Conservation of the same molecular function for two duplicated proteins while allowing the diversification of their cellular functions may represent a simple and economical way of introducing functional diversity and complexity in a controlled manner during evolution. This may be the result of a change in interaction partners and/or subcellular localization.
The second important result of our study is that since the date of the ancient WGD, cellular functions have evolved at variable rates, since a scale of functional divergence can be detected. In this respect, we propose to interpret this functional scale of divergence in the light of different theoritical evolutionary scenarios for cellular function.
First, the first level of the functional scale (behavior I) may contain duplicates which have been conserved as such, because keeping two copies may confer an evolutionary benefit on the cell (for instance, Rps26A/Rps26B; Table [1](#T1){ref-type="table"}).
Second, we propose that the majority of the paralog pairs populating the two first levels of the functional scale of divergence based on interactions (behaviors I and II) evolved functionally according to the duplication-degeneration-complementation (DDC) or subfunctionalization model proposed by Force *et al*. \[[@B28]\]. This predicts that duplicated genes are preserved by the partitioning of the function(s) of the ancestral gene between the two duplicates. This may happen, for instance, by the complementary loss of regulatory elements or the modification of the coding regions. Even though our analysis does not pretend to reveal the molecular mechanisms by which the subfunctionalization of the duplicated pairs has occurred, several lines of evidence sustain our proposal. First, the first level of the functional scale is populated by paralog pairs, which have kept their interactors identical or still share common interactors. This is in good agreement with a situation in which duplicates have slightly diverged by subfunctionalization to form two subunits of a same complex (for example, Tif4631/Tif4632, Rfc3/Rfc4, Yck1/Yck2; Table [1](#T1){ref-type="table"}) or to increase the complexity of a signaling pathway (for instance, Mkk1/Mkk2; Table [1](#T1){ref-type="table"}). Second, the intermediate level of the functional scale of divergence (behavior II) contains paralog pairs that do not have the same interactors but have still conserved their cellular function(s) since the duplication event. They may represent paralog pairs involved in different aspects of the same biological process (see Results and \[[@B11]\]) and/or pairs for which the spatio-temporal regulation has evolved by subfunctionalization, therefore implying a new cast of interactors.
Finally, the third level of the functional scale (behavior III) may correspond to duplicates that have evolved by neofunctionalization, as not only their interactors are different but they are also involved in different cellular processes (for instance, Swi5/Ace2). These genes may illustrate Ohno\'s theory \[[@B3]\] of the emergence of new functions from gene duplication events. Even though we have shown here that there is no simple relationship between sequence identity and cellular function, it is interesting to note that data newly generated by Kellis *et al*. \[[@B8]\] strengthen our proposal. Indeed, the frequency of pairs showing accelerated protein evolution is almost twice as high among the paralog pairs displaying behavior III (37.5% (3/8) of the pairs common to both studies) than among pairs with the same function (20% (5/25) of the pairs common to both studies with behaviors I and II). Overall, these results corroborated our proposal.
Conclusions
===========
Most network analyses carried out up to now either emphasized the prediction of function for uncharacterized proteins \[[@B29],[@B30]\] or, in the frame of evolutionary studies, estimated the rate of evolution of proteins according to their number of interactors \[[@B31]\] and addressed the issue of the link between protein dispensability and rate of protein evolution \[[@B32],[@B33]\]. As far as we know, this work constitutes the first attempt to address the functional evolutionary fate of duplicated genes using a bioinformatic analysis of the protein-protein interaction network in which the products of these genes are involved, and to provide detailed protein function comparisons based on interaction data. Our approach might thus provide a new way to analyze the evolution of the function of duplicated genes in different organisms.
A limitation of this type of analysis is the present knowledge of interaction networks. Even in a well-studied organism such as *Saccharomyces cerevisiae*, less than 10% of the gene pairs, remnants of the WGD, are amenable to such a detailed analysis. As our knowledge on interaction networks is increasing and as more interactions become available, we can expect to improve both the coverage of duplicated pairs of interactors and the relevance of the functional clusters found by the PRODISTIN method.
Finally, it should be emphasized that the study of evolutionary processes greatly benefits by being approached using different tools not only at the sequence level, as is usual, but also directly at the functional level. In the case of the study of the 41 paralog pairs reported here, functional conclusions inferred from the sequence level would have been incomplete and even erroneous in several instances.
Materials and methods
=====================
Functional distance based on GO annotations
-------------------------------------------
GOproxy \[[@B13]\], a tool that calculates the Czekanowski-Dice distance between gene annotations was used to compare the GO annotations \[[@B12]\] of the duplicated gene products as well as that of five datasets of 460 pairs of proteins randomly selected from the yeast genome. The Molecular Function, Biological Process and Cellular Component ontologies were processed separately. The Czekanowski-Dice distance formula used in the algorithm is:
Dist(*i,j*) = number of (Terms(*i*) ΔTerms(*j*))/ \[number of (Terms(*i*) ∪ Terms(*j*)) + number of (Terms(*i*) ∩ Terms(*j*))\],
in which, *i*and *j*denote two genes, Terms(*i*) and Terms(*j*) are the lists of their GO terms and Δ is the symmetrical difference between the two sets. This distance formula increases the weight of the shared GO terms by giving more weight to similarities than to differences. The GOToolBox website can be accessed at \[[@B13]\].
Protein-protein interaction dataset
-----------------------------------
The protein-protein interaction dataset we investigated contains a total of 4,143 selected interactions involving 2,643 proteins. We updated our former dataset \[[@B11]\] with 1,244 new interactions taken from the Munich Information Center for Protein Sequences (MIPS) \[[@B34]\] and from the literature. As previously, only direct binary interactions were selected according to the method used for their identification (two-hybrid experiments, *in vitro*binding, far western, gel retardation and biochemical experiments).
PRODISTIN analysis
------------------
PRODISTIN, a computational method we recently proposed \[[@B11]\], was used to analyze the protein-protein interaction dataset. Starting with a binary list of interactions, only proteins involved in at least three binary interactions were selected for further classification (because poorly connected proteins have a higher chance of being involved in false-positive interactions). A graph in which vertices are proteins and edges correspond to the relation \'interact with and/or share at least one common interactor\' was computed and the Czekanowski-Dice distance was calculated between all possible pairs of proteins belonging to the connected component of this graph (using the formula above and applying it to the list of protein interactors instead of the list of GO terms). The distance matrix was then clustered using BioNJ \[[@B35]\] and the tree was visualized using TreeDyn \[[@B36]\]. PRODISTIN classes corresponding to the largest possible subtree composed of at least three proteins sharing the same functional annotation and representing at least 50% of the individual class members for which a functional annotation is available were detected in the tree. GO annotations corresponding to the Biological Process ontology were used for this purpose. Given that GO is organized as a DAG, proteins may be annotated at different levels of the ontology. Our goal was to analyze subtrees regarding to the proteins commonly annotated as participating in them, so we considered annotations for all proteins at a specific level of the ontology. We chose to work at level 4 because we estimated, on previous experience using the Yeast Proteome Database \[[@B37]\] system of annotation, that this particular level provides a good representation of the complexity of cellular functions. For this, we used GODiet, a tool enabling us to restrict the list of GO terms to a given depth in the ontology \[[@B13]\].
Sequence analysis
-----------------
Pairwise sequence alignments were carried out on the set of 460 pairs of duplicated protein sequences using the Needleman-Wunsch (global alignment) algorithm. The program used is available at \[[@B38]\]. The chosen alignment matrix was BLOSUM50, and the gap-opening and gap-extension penalties were set to 12 and 2, respectively. The resulting 460 alignments have been processed to calculate the percent identity for each protein pair.
Additional data files
=====================
The following additional data are available with the online version of this paper. Additional data file [1](#s1){ref-type="supplementary-material"} contains the expectation values for the distribution of functional distances based on the GO annotations. Additional data file [2](#s2){ref-type="supplementary-material"} contains details of the 123 PRODISTIN classes contained in the classification tree.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
The expectation values for the distribution of functional distances based on the GO annotations
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
Details of the 123 PRODISTIN classes contained in the classification tree
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank Didier Casane for helpful discussions and David Martin for help in processing GO annotations. This project is supported by an Action Bioinformatique inter-EPST grant and an ACI IMPBio (EIDIPP project) to B.J. A.B. and C.B. respectively thank the Ministère de la Recherche et de la Technologie and the Fondation pour la Recherche Médicale for financial support.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Distribution of functional distances between duplicated pairs based on Gene Ontology annotations. The annnotations are for \'Biological Process\' (blue), \'Molecular Function\' (purple) and \'Cellular Component\' (light yellow). Distributions of distances (ranging from 0 to 1) based on annotations for **(a)**the 460 duplicated pairs, (a, inset) randomly selected pairs and **(b)**the 41 duplicated pairs present in the PRODISTIN tree.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
PRODISTIN classification tree for 890 yeast proteins. PRODISTIN classes have been colored according to their corresponding Biological Process annotations. Protein names have been omitted for clarity. The tree contains 41 out of 460 duplicated pairs, the remnant of the ancient whole-genome duplication. Examples of PRODISTIN classes illustrating the three different behaviors of duplicated pairs have been extracted and enlarged from the tree. Their original position in the tree is shown by dashed lines.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Repartition of the 3 different PRODISTIN behaviors in respect to the distribution of the GO-based functional distances (ranging from 0 to 1) between the 41 duplicated pairs. Behaviors are classified as: same class, same function (behavior I, blue); different classes, same function (behavior II, pink); different classes, different functions (behavior III, gray); not classified (green). Results are shown for the Biological Process annotations only.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Percent of sequence identity between the 41 duplicated protein pairs. Proteins were classified as belonging to the same class (blue diamonds), different classes with the same function (pink diamonds), different classes with different functions (gray diamonds), or not classified (green triangles).
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Details of the behaviors of the 41 duplicated pairs present in the PRODISTIN classification tree
:::
Behavior class Gene 1 Gene 2 Localization in same PRODSTIN class Same cellular function Annotation of the PRODISTIN classes by cellular function
---------------- --------- --------- ------------------------------------- ------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
I ARF1 ARF2 \+ \+ Vesicle-mediated transport, secretory pathway, intracellular transport (50)
ASM4 NUP53 \+ \+ Nuclear organization and biogenesis (22), nucleobase nucleoside nucleotide and nucleic acid transport, protein targeting, RNA localization (32), nucleobase nucleoside nucleotide and nucleic acid metabolism, intracellular transport (48)
BMH2 BMH1 \+ \+ Energy derivation by oxidation of organic compounds, polysaccharide metabolism, carbohydrate metabolism (6)
BOI1 BOI2 \+ \+ Nuclear organization and biogenesis (22), nucleobase nucleoside nucleotide and nucleic acid transport, protein targeting, RNA localization (32), nucleobase nucleoside nucleotide and nucleic acid metabolism, intracellular transport (48)
ECI1 DCI1 \+ \+ Cytoplasm organization and biogenesis, protein targeting (7)
GIC2 GIC1 \+ \+ Bud growth (6), intracellular signaling cascade (26), signal transduction (58), cytoplasm organization and biogenesis (94)
GZF3 DAL80 \+ \+ Transcription, nitrogen utilization (5), nucleobase nucleoside nucleotide and nucleic acid metabolism (66)
KCC4 GIN4 \+ \+ Cell cycle(16), nucleobase nucleoside nucleotide and nucleic acid metabolism, intracellular transport (48)
MKK1 MKK2 \+ \+ Phosphate metabolism, protein modification (6), conjugation with cellular fusion, sensory perception, perception of abiotic stimulus (20), signal transduction (58), cytoplasm organization and biogenesis (94)
MYO3 MYO5 \+ \+ Polar budding, vesicle-mediated transport, response to osmotic stress (5), cytoplasm organization and biogenesis (10), nucleobase nucleoside nucleotide and nucleic acid metabolism (55)
NUP100 NUP116 \+ \+ Nuclear organization and biogenesis (22), nucleobase nucleoside nucleotide and nucleic acid transport, protein targeting, RNA localization(32), nucleobase nucleoside nucleotide and nucleic acid metabolism, intracellular transport (48)
PCL6 PCL7 \+ \+ Energy derivation by oxidation of organic compounds, polysaccharide metabolism, carbohydrate metabolism (5), transcription (17)
RAS2 RAS1 \+ \+ Intracellular signaling cascade(4), cell proliferation (20)
RFC3 RFC4 \+ \+ DNA repair, response to DNA damage stimulus, cell cycle(18), nucleobase nucleoside nucleotide and nucleic acid metabolism (23)
SEC4 YPT7 \+ \+ Vesicle-mediated transport, secretory pathway, intracellular transport (50)
SIZ1 NFI1 \+ \+ External encapsulating structure organization and biogenesis, cell proliferation, cellular morphogenesis (8), signal transduction (58), cytoplasm organization and biogenesis (94)
SSK22 SSK2 \+ \+ Phosphate metabolism, intracellular signaling cascade, protein modification (5), cell surface receptor linked signal transduction nucleobase nucleoside, nucleotide and nucleic acid metabolism (7)
SSO2 SSO1 \+ \+ Vesicle-mediated transport (14)
TIF4632 TIF4631 \+ \+ Protein biosynthesis (7), macromolecule biosynthesis (12), nucleobase nucleoside nucleotide and nucleic acid metabolism (55)
VPS64 YLR238W \+ \+ Response to pheromone during conjugation with cellular fusion, sensory perception, perception of abiotic stimulus (6), cell cycle, cytoplasm organization and biogenesis (16)
YIL105C YNL047C \+ \+ Unknown (4)
YPT31 YPT32 \+ \+ Vesicle-mediated transport, secretory pathway, intracellular transport (50)
YPT53 VPS21 \+ \+ Cytoplasm organization and biogenesis (6), vesicle-mediated transport, secretory pathway, intracellular transport (50)
ZDS2 ZDS1 \+ \+ Cell aging, response to DNA damage stimulus, chromatin silencing(5), intracellular signaling cascade (26), cytoplasm organization and biogenesis (94), signal transduction (58)
RPS26B RPS26A \+ \+ Nucleobase, nucleoside, nucleotide and nucleic acid metabolism (29)
YCK1 YCK2 \+ \+ Transport (6), nucleobase, nucleoside, nucleotide and nucleic acid metabolism (202)
II BUB1 MAD3 \- \+ Cell cycle, cell proliferation (40), nucleobase, nucleoside, nucleotide and nucleic acid metabolism (66)
TUB4 TUB1 \- \+ Cell cycle, cytoplasm organization and biogenesis (16)
Cell cycle, cytoplasm organization and biogenesis (7)
ENT1 ENT2 \- \+ Cytokinesis, vesicle-mediated transport, cytoplasm organization and biogenesis (4), cell proliferation (20)
Vesicle-mediated transport (14)
III YAP1802 YAP1801 \- \- Cell proliferation (20)
Vesicle-mediated transport (14)
YMR181C YPL229W \- \- Cell proliferation (20)
Transcription (8), nucleobase, nucleoside, nucleotide and nucleic acid metabolism (202)
NUP170 NUP157 \- \- Nuclear organization and biogenesis (22), nucleobase nucleoside nucleotide and nucleic acid transport, protein targeting, RNA localization (32), nucleobase nucleoside nucleotide and nucleic acid metabolism, intracellular transport (48)
Cell cycle, cytoplasm organization and biogenesis (7)
APP2 GYP5 \- \- Vesicle-mediated transport (18), transport (21), cytoplasm organization and biogenesis (94)
RNA metabolism (29), nucleobase, nucleoside, nucleotide and nucleic acid metabolism (202)
SIR2 HST1 \- \- Cell cycle, chromatin silencing(6), nucleobase, nucleoside, nucleotide and nucleic acid metabolism (14)
RNA metabolism (9), nucleobase, nucleoside, nucleotide and nucleic acid metabolism (202)
GSP1 GSP2 \- \- Nuclear organization and biogenesis (22), nucleobase, nucleoside, nucleotide and nucleic acid transport, protein targeting, RNA localization(32), nucleobase, nucleoside, nucleotide and nucleic acid metabolism, intracellular transport (48)
Cell cycle (4)
SWI5 ACE2 \- \- Transcription (6), macromolecule biosynthesis (11), nucleobase, nucleoside, nucleotide and nucleic acid metabolism (55)
Cell cycle (4)
LSB1 PIN3 \- \- Unknown (5), nucleobase, nucleoside, nucleotide and nucleic acid metabolism (23)
RNA metabolism (29), nucleobase, nucleoside, nucleotide and nucleic acid metabolism (202)
YBR270C BIT61 \- \- Unknown (4)
Transport (21), cytoplasm organization and biogenesis (94)
NC EBS1 EST1
MTH1 STD1
NMA2 NMA1
\+ and - indicate the status of the duplicates in respect of their localization in the same PRODISTIN class and whether they have the same cellular functions. NC, not classified, indicating the pairs for which at least one of the genes does not belong to a PRODISTIN class. The last column shows the annotation of the PRODISTIN classes containing the duplicated genes and the number of class members (in parentheses). When the 2 genes of the pair belong to different classes (behavior II and III), the first list of annotations corresponds to the class containing gene 1 and the second list to the one containing gene 2.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Summary of the behaviors of the 41 duplicated genes
:::
Classification behaviors Number of duplicated pairs
----------------------------------------------------- ----------------------------
I Same class, same biological process 26 (63%)
II Different classes, same biological process 3 (7.5%)
III Different classes, different biological process 9 (22%)
Not classified 3 (7.5%)
Total 41
:::
|
PubMed Central
|
2024-06-05T03:55:51.748330
|
2004-9-15
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545596/",
"journal": "Genome Biol. 2004 Sep 15; 5(10):R76",
"authors": [
{
"first": "Anaïs",
"last": "Baudot"
},
{
"first": "Bernard",
"last": "Jacq"
},
{
"first": "Christine",
"last": "Brun"
}
]
}
|
PMC545597
|
Background
==========
*Bacillus licheniformis*is a Gram-positive, spore-forming bacterium widely distributed as a saprophytic organism in the environment. This species is a close relative of *Bacillus subtilis*, an organism that is second only to *Escherichia coli*in the level of detail at which it has been studied. Unlike most other bacilli, which are predominantly aerobic, *B. licheniformis*is a facultative anaerobe, which may allow it to grow in additional ecological niches. Certain *B. licheniformis*isolates are capable of denitrification; the relevance of this characteristic to environmental denitrification may be small, however, as the species generally persists in soil as endospores \[[@B1]\].
There are numerous commercial and agricultural uses for *B. licheniformis*and its extracellular products. The species has been used for decades in the manufacture of industrial enzymes including several proteases, α-amylase, penicillinase, pentosanase, cycloglucosyltransferase, β-mannanase and several pectinolytic enzymes. The proteases from *B. licheniformis*are used in the detergent industry as well as for dehairing and bating of leather \[[@B2],[@B3]\]. Amylases from *B. licheniformis*are deployed for the hydrolysis of starch, desizing of textiles and sizing of paper \[[@B3]\]. Specific *B. licheniformis*strains are also used to produce peptide antibiotics such as bacitracin and proticin in addition to a number of specialty chemicals such as citric acid, inosine, inosinic acid and poly-γ-glutamic acid \[[@B4]\]. Some *B. licheniformis*isolates can mitigate the affects of fungal pathogens on maize, grasses and vegetable crops \[[@B5]\]. As an endospore-forming bacterium, the ability of the organism to survive under unfavorable environmental conditions may enhance its potential as a natural biocontrol agent.
*B. licheniformis*can be differentiated from other bacilli on the basis of metabolic and physiological tests \[[@B6],[@B7]\]; however, biochemical and phenotypic characteristics may be ambiguous among closely related species. Recent taxonomic studies indicate that *B. licheniformis*is closely related to *B. subtilis*and *Bacillus amyloliquefaciens*on the basis of comparisons of 16S rDNA and 16S-23S internal transcribed spacer (ITS) nucleotide sequences \[[@B8]\]. Lapidus *et al.*\[[@B9]\] recently constructed a physical map of the *B. licheniformis*chromosome using a PCR approach, and established a number of regions of colinearity where gene content and organization were conserved with the *B. subtilis*genome.
Given that *B. licheniformis*is an industrial organism used for the manufacture of enzymes, antibiotics, and chemicals, important in nutrient cycling in the environment, and a species that is taxonomically related to *B. subtilis*, perhaps the best studied of all Gram-positive bacteria, we derived the complete nucleotide sequence of the *B. licheniformis*type strain (ATCC 14580) genome. With this data in hand, functional and comparative genomics studies can be initiated that may ultimately lead to new strategies for improving industrial strains as well as better understanding of genome evolution among the species within the *subtilis-licheniformis*group.
Results and discussion
======================
General features of the *B. licheniformis*genome
------------------------------------------------
The genome of *B. licheniformis*ATCC 14580 consists of a circular chromosome of 4,222,336 base-pairs (bp) with an average G+C content of 46.2% (Table [1](#T1){ref-type="table"}). No plasmids were found during the genome analysis, and none were found by agarose gel electrophoresis (data not shown). Using a combination of several gene-finding programs and manual inspection, 4,208 protein-coding sequences (CDSs) were predicted. These CDSs constitute 87% of the genome and have an average length of 873 bp (ranging from 78 to 10,767 bp). They are oriented on the chromosome primarily in the direction of replication (Figure [1](#F1){ref-type="fig"}) with 74.4% of the genes on the leading strand and 25.6% on the lagging strand. Among the 4,208 protein coding genes, 3,948 (94%) had significant similarity to proteins in PIR, 3,187 (76%) of these gene models contain Interpro motifs, and 2,895 (69%) contain protein motifs found in PFAM. The number of hypothetical and conserved hypothetical proteins in the *B. licheniformis*genome with hits in the PIR database was 1,318 (212 conserved hypothetical proteins). Among the list of hypothetical and conserved hypothetical gene products, 683 (52%) have protein motifs contained in PFAM (148 conserved hypothetical proteins). There are 72 tRNA genes representing all 20 amino acids and seven rRNA operons.
The likely origin of replication (Figure [1](#F1){ref-type="fig"}) was identified by similarities to several features of the corresponding regions in *B. subtilis*and other bacteria. These included co-localization of four genes (*rpmH*, *dnaA*, *dnaN*, and *recF*) near the origin, GC nucleotide skew ((G-C)/(G+C)) analysis, and the presence of multiple *dnaA*-boxes and AT-rich sequences immediately upstream of the *dnaA*gene \[[@B10]-[@B12]\]. On the basis of these observations we assigned a cytosine residue of the *Bst*BI restriction site between the *rpmH*and *dnaA*genes to be the first nucleotide of the *B. licheniformis*genome. The replication termination site was localized near 2.02 megabases (Mb) by GC skew analysis. This region lies roughly opposite the origin of replication (Figure [1](#F1){ref-type="fig"}). Unlike *B. subtilis*, there was no apparent gene encoding a replication terminator protein (*rtp*) in *B. licheniformis*. The *Bacillus halodurans*genome also lacks an obvious *rtp*function \[[@B13]\]; therefore, it seems likely that *B. subtilis*acquired the *rtp*gene following its divergence from *B. halodurans*and *B. licheniformis*.
Transposable elements, prophages and atypical regions
-----------------------------------------------------
The genome of *B. licheniformis*ATCC 14580 contains nine identical copies of a 1,285 bp insertion sequence element termed *IS3Bli1*\[[@B9]\]. This sequence shares a number of features with other *IS3*family elements \[[@B9]\] including direct repeats of 3-5 bp, a 10-bp left inverted repeat, and a 9 bp right inverted repeat (Figure [2](#F2){ref-type="fig"}). *IS3Bli1*encodes two predicted overlapping CDSs, designated *orfA*and *orfB*in relative translational reading frames of 0 and -1. The presence of a \'slippery heptamer\' motif, AAAAAAG, before the stop codon in *orfA*may indicate that programmed translational frameshifting occurs between these two coding sequences, resulting in a single gene product \[[@B14]\]. The *orfB*gene product harbors the DD(35)E(7)K motif, a highly conserved pattern among insertion sequences. Eight of these *IS3Bli1*elements lie in intergenic regions, and one interrupts the *comP*gene as noted previously \[[@B9]\]. In addition to these insertion sequences, the genome encodes a putative transposase that is most closely related (E = 1.8 × 10^-11^) to one identified in the *Thermoanaerobacter tengcongensis*genome \[[@B15]\]; however, similar transposase genes are also found in the chromosomes of *B. halodurans*\[[@B13]\], *Oceanobacillus iheyensis*\[[@B16]\], *Streptococcus agalactiae*\[[@B17]\] and *Streptococcus pyogenes*\[[@B18]\].
The presence of several bacteriophage lysogens or prophage-like elements was revealed by Smith-Waterman comparisons to other bacterial genomes and by their AT-rich signatures (Figure [3](#F3){ref-type="fig"}, Table [2](#T2){ref-type="table"}). Prophage sequences, designated NZP1 and NZP3 (similar to *B. subtilis*prophages PBSX and φ-105), were discovered by noting the presence of nearby genes that code for the large subunit of terminase, a signature protein that is highly conserved among prophages \[[@B19]\]. Interestingly, a terminase gene was not observed in the third putative prophage, termed NZP2 (similarity to *B. subtilis*phage SPP1); however, its absence may be the result of genome deterioration during evolution. Interestingly, we observed that regions in which the G+C content is less than 39% usually encoded proteins that have no *B. subtilis*ortholog and share identity only with hypothetical and conserved hypothetical genes. Two of these AT-rich segments correspond to the NZP2 and NZP3 prophages.
An isochore plot (Figure [3](#F3){ref-type="fig"}) also revealed the presence of a region with an atypically high (62%) G+C content. This segment contains two hypothetical genes whose sizes (3,831 and 2,865 bp) greatly exceed the size of an average CDS in *B. licheniformis*. The first gene encodes a protein of 1,277 amino acids for which Interpro predicts 16 collagen triple-helix repeats, and the amino acid pattern TGATGPT is repeated 75 times within the polypeptide. The second CDS is smaller, and encodes a protein with 11 collagen triple-helix repeats; the same TGATGPT motif recurs 56 times. The primary translation products from these genes do not contain canonical signal peptides for secretion, and they do not contain motifs for the twin-arginine or sortase-mediated translocation pathways. Therefore, it is not likely that they are exported to the cell surface or the extracellular medium. Interestingly, the chromosomal region (19 kb) adjacent to these genes is clearly non-colinear with the *B. subtilis*genome, and virtually all of the predicted genes encode hypothetical or conserved hypothetical proteins. There are a number of bacterial proteins listed in PIR that also contain collagen triple-helix repeat regions, including two from *Mesorhizobium loti*(accession numbers NF00607049 and NF00607035) and three from *B. cereus*(accession numbers NF01692528, NF01269899 and NF01694666). These putative orthologs share 53-76% amino-acid sequence identity with their counterparts in *B. licheniformis*, and their functions are unknown.
Extracellular enzymes and metabolic activities
----------------------------------------------
In the *Bacillus licheniformis*genome, 689 of the 4,208 gene models have signal peptides forecast by SignalP \[[@B20]\]. Of these, 309 have no transmembrane domain predicted by TMHMM \[[@B21]\] and 134 are hypothetical or conserved hypothetical genes. Based on a manual examination of the remaining 175 genes, at least 82 are likely to encode secreted proteins and enzymes. Moreover, there are 27 predicted extracellular proteins encoded by the *B. licheniformis*ATCC 14580 genome that are not found in *B. subtilis*168. In accordance with its saprophytic lifestyle, the secretome of *B. licheniformis*encodes numerous secreted enzymes that hydrolyze polysaccharides, proteins, lipids and other nutrients.
Cellulose is the most abundant polysaccharide on Earth, and microorganisms that hydrolyze cellulose contribute to the global carbon cycle. Interestingly, two gene clusters involved in cellulose degradation and utilization were discovered in *B. licheniformis*, and there are no counterparts in *B. subtilis*168. The enzymes encoded by the first gene cluster include two putative endoglucanases belonging to glycoside hydrolase families GH9 and GH5, a probable cellulose-1,4-β-cellobiosidase of family GH48, and a putative β-mannanase of family GH5. The β-mannanase (GH5) and endoglucanase (GH9) both harbor carbohydrate-binding motifs. With the exception of the cellulose-1,4-β-cellobiosidase (GH48), all of the gene products encoded in this cluster have secretory signal peptides, and all have homologs in *Bacillus*species other than *B. subtilis*. The overall G+C content of this cluster (48%) does not appear to differ appreciably from that of the genome average (46%). The second gene cluster encodes a putative β-glucosidase (GH1) and three components of a cellobiose-specific PTS transport complex. A second β-glucosidase (GH3) gene is present at an unlinked locus in the genome. Collectively, the genes in these two clusters should enable *B. licheniformis*to utilize cellulose as a carbon and energy source, converting it to cellobiose and ultimately glucose. In this regard, we have confirmed that *B. licheniformis*ATCC 14580 is capable of growth on carboxymethyl cellulose as a sole carbon source (not shown). The chromosome of *B. licheniformis*ATCC 14580 encodes a number of additional carbohydrase activities that may allow the organism to grow on a broad range of polysaccharides. These include xylanase, endo-arabinase and pectate lyase that may be involved in degradation of hemicellulose, α-amylase and α-glucosidase for starch hydrolysis, chitinases for the breakdown of chitooligosaccharides from fungi and insects, and levanase for utilization of β-D-fructans (levans). Several of these activities are marketed as industrial enzymes.
Saprophytic organisms must utilize a variety of nitrogenous compounds as nutrients for growth and metabolism. On the basis of the information encoded in its genome, *B. licheniformis*ATCC 14580 possesses the ability to acquire nitrogen from exogenous proteins, peptides, amino acids, ammonia, nitrate and nitrite. Like *B. subtilis*, the repertoire of extracellular proteases produced by *B. licheniformis*includes serine proteases (*aprE*, *epr*, *vpr*), metalloprotease (*mpr*), and an assortment of endo- and exopeptidases (*yjbG*, *ydiC*, *gcp*, *ykvY*, *ampS*, *bpr*(two copies), *yfxM*, *yuiE*, *yusX*, *ywaD*, *pepT*). However, *B. licheniformis*also has the capacity to produce a number of additional proteases and peptidases that are not encoded in the *B. subtilis*genome. These include a clostripain-like protease, a zinc-metallopeptidase, a probable glutamyl endopeptidase, an aminopeptidase C homolog, two putative dipeptidases and a zinc-carboxypeptidase.
*B. licheniformis*also has the ability to utilize amino and imino nitrogen from arginine, asparagine and glutamine via arginine deiminase, arginase, asparaginase and glutaminase activities. Interestingly, there appear to be two genes each for arginase, asparaginase and glutaminase. Presumably, the arginine deiminase activity is expressed during anaerobic growth on arginine, whereas the arginase activities are predominant during aerobic growth. The occurrence of putative arginase genes is somewhat of an enigma in *B. licheniformis*, because there are no genes encoding urease activity for the hydrolysis of urea that is generated by the arginase reaction. In addition to the absence of urease gene homologs (*ureABC*) in *B. licheniformis*, the glutamine ABC transporters (*glnH*, *glnM*, *glnP*, *glnQ*gene products) are also lacking.
It appears that nitrogen assimilation and transport pathways may be coordinated similarly in *B. licheniformis*and *B. subtilis*owing to the presence of key genes such as *glnA*, *glnR*, *tnrA*and *nrgA*in both species. Likewise, the pathways for nitrate/nitrite transport and metabolism in *B. licheniformis*appear to be analogous to the corresponding pathways in *B. subtilis*as suggested by the presence of *nasABC*(nitrate transport), *narGHIJ*(respiratory nitrate reductase), and *nasDEF*(NADH-dependent nitrite reductase) genes. Unlike *B. subtilis*, *B. licheniformis*evidently possesses the capability for anaerobic respiration using nitric oxide reductase. Moreover, the gene encoding this activity lies in a cluster that includes CDSs for *narK*(nitrite extrusion protein), two putative *fnr*proteins (transcriptional regulators of anaerobic growth), and a *dnrN*-like gene product (nitric oxide-dependent regulator). These observations are consistent with previous findings that certain *B. licheniformis*isolates are capable of denitrification \[[@B22]\]. While denitrification is a process of major ecological importance, the contribution of *B. licheniformis*may be small as the species exists predominantly as endospores in soil \[[@B1]\].
Microbial D-hydantoinase enzymes have been applied to the industrial production of optically pure D-amino acids for synthesis of antibiotics, pesticides, sweeteners and therapeutic amino acids \[[@B23]\]. This enzyme catalyzes the hydrolysis of cyclic ureides such as dihydropyrimidines and 5-monosubstituted hydantoins to *N*-carbamoyl amino acids. Hydantoinase activities have been detected in a variety of bacterial genera, and a cluster of six genes in *B. licheniformis*appears to confer a similar capability. This gene cluster encodes *N*-methylhydantoinase (ATP-hydrolyzing), hydantoin utilization proteins A and B (*hyuAB*homologs), a possible transcriptional regulator (TetR/AcrR family), a putative pyrimidine permease, and a hypothetical protein that contains an Interpro domain (IPR004399) for phosphomethylpyrimidine kinase.
Protein export, sporulation and competence pathways
---------------------------------------------------
Kunst *et al*. \[[@B10]\] listed 18 genes that have a major role in the secretion of extracellular enzymes by the classical (Sec) pathway in *B. subtilis*168. This list includes several chaperonins, signal peptidases, components of the signal recognition particle and protein translocase complexes. All members of this list have *B. licheniformis*counterparts. In addition to the Sec pathway, some *B. subtilis*proteins are directed into the twin-arginine (Tat) export pathway, possibly in a Sec-independent manner. Curiously, the *B. licheniformis*genome encodes three *tat*gene orthologs (*tatAY*, *tatCD*, and *tatCY*), but two others (*tatAC*and *tatAD*) are conspicuously absent. Furthermore, specific proteins may be exported to the cell surface via lipoprotein signal peptides or sortase factors. Lipoprotein signal peptides are cleaved with a specific signal peptidase (Lsp) encoded by the *lspA*gene in *B. subtilis*. An *lspA*homolog can be found in *B. licheniformis*as well, suggesting that this species may possess the ability to export lipoproteins via a similar mechanism. Lastly, surface proteins in Gram-positive bacteria are frequently attached to the cell wall by sortase enzymes, and genome analyses have revealed that more than one sortase is often produced by a given species. In this regard, three possible sortase gene homologs were detected in the genome of *B. licheniformis*ATCC 14580. Together these observations suggest that the central features of the protein export machinery are principally conserved in *B. subtilis*and *B. licheniformis*.
From the list of 139 sporulation genes tabulated by Kunst *et al*. \[[@B10]\], all but six have obvious counterparts in *B. licheniformis*. These six exceptions (*spsABCEFG*) comprise an operon involved in synthesis of a spore coat polysaccharide in *B. subtilis*. In addition, the response regulator gene family (*phrACEFGI*) appears to have a low level of sequence conservation between *B. subtilis*and *B. licheniformis*.
Natural competence (the ability to take up and process exogenous DNA in specific growth conditions) is a feature of few *B. licheniformis*strains \[[@B24]\]. The reasons for variability in competence phenotype have not been explored at the genetic level, but the genome data offer several possible explanations. Although the *B. licheniformis*genome encodes all of the late competence functions ascribed in *B. subtilis*(for example, *comC*, *comEFG*operons, *comK*, *mecA*), it lacks an obvious *comS*gene, and the *comP*gene is punctuated by an insertion sequence element (*IS3Bli1*), suggesting that the early stages of competence development have been pre-empted in *B. licheniformis*ATCC 14580. Whether these early functions can be restored by introducing the corresponding genes from *B. subtilis*is unknown. In addition to an apparent deficiency in DNA uptake, two type I restriction-modification systems were discovered that may also contribute to diminished transformation efficiencies. These are distinct from the *ydiOPS*genes of *B. subtilis*, and could participate in degradation of improperly modified DNA from heterologous hosts used during construction of recombinant expression vectors. Each of these loci in *B. licheniformis*(designated as *BliI*and *BliII*) encode putative HsdS, HsdM and HsdR subunits that share significant amino-acid sequence identity to type I restriction-modification proteins in other bacteria. Curiously, the G+C-content for these loci (37%) is substantially lower than the overall genome average (46%) which may hint that they are the result of gene acquisitions. Lastly, the synthesis of a glutamyl polypeptide capsule has also been implicated as a potential barrier to transformation of *B. licheniformis*strains \[[@B25]\]. While laboratory strains of *B. subtilis*usually do not produce significant capsular material, the genome sequence of *B. subtilis*168 indicates that they may harbor the genes required for synthesis of polyglutamic acid. In contrast, many *B. licheniformis*isolates produce copious amounts of capsular material, giving rise to colonies with a wet or slimy appearance. Six genes were predicted (*ywtABDEF*and *ywsC*orthologs) that may be involved in the synthesis of polyglutamic acid capsular material in *B. licheniformis*.
Antibiotics, secondary metabolites and siderophores
---------------------------------------------------
Bacitracin is a cyclic peptide antibiotic that is synthesized non-ribosomally by some *B. licheniformis*isolates \[[@B26]\]. While there is variation in the prevalence of bacitracin synthase genes among laboratory strains of this species, one study suggested that up to 50% may harbor the *bac*operon \[[@B27]\]. Interestingly, the *bac*operon is not present in the type strain (ATCC 14580) genome. Seemingly, the only non-ribosomal peptide synthase operon encoded by the *B. licheniformis*type strain genome is that responsible for lichenysin biosynthesis. Lichenysin structurally resembles surfactin from *B. subtilis*\[[@B28]\], and their respective biosynthetic operons are highly similar. Surprisingly, we found no *B. licheniformis*counterparts for the *pps*(plipastatin synthase) and polyketide synthase (*pks*) operons of *B. subtilis*. Collectively, these two regions represent sizeable portions (80 kb and 38 kb, respectively) of the chromosome in *B. subtilis*, although they are reportedly dispensable \[[@B29]\].
Unexpectedly, a cluster of 11 genes was found encoding a lantibiotic, with its associated modification and transport functions. We designated this peptide of 75 amino acids as lichenicidin, and its closest homolog is mersacidin from *Bacillus*sp. strain HIL-Y85/54728 \[[@B30]\]. Lantibiotics are ribosomally synthesized peptides that are modified post-translationally so that the final molecules contain rare thioether amino acids such as lanthionine and/or methyl-lanthionine \[[@B31]\]. Like mersacidin, lichenicidin appears to be a type B lantibiotic, comprising a rigid globular peptide with no net charge (7 acidic residues, 7 basic residues) and a leader peptide with a conserved double glycine cleavage site (GG-type leader peptide). These antimicrobial compounds have attracted much attention in recent years as models for the design and genetic engineering of improved antimicrobial agents \[[@B32]\]. However, since several post-translational modifications and product-specific export functions are required, a dedicated expression system is a prerequisite to provide all the factors necessary to synthesize, modify and transport the lantibiotic peptide. With its history of use in industrial microbiology, *B. licheniformis*may be an attractive candidate for the development of such an expression system.
Like *B. subtilis*168, the *B. licheniformis*ATCC 14580 chromosome harbors a siderophore biosynthesis gene cluster (*dhbABCEF*), and the organization of the cluster is similar to the corresponding chromosomal segment in *B. subtilis*. In addition, the *B. licheniformis*genome contains a second gene cluster of four genes (*iucABCD*) that show significant similarity to proteins involved in aerobactin biosynthesis in *E. coli*. Surprisingly, a gene encoding the receptor protein (*iutA*homolog) was not found in *B. licheniformis*. The *B. halodurans*genome also contains genes that are homologous to *iucABCD*, but like *B. licheniformis*, no *iutA*homolog could be found using BLAST or Smith-Waterman searches.
Comparison of the *B. licheniformis*genome with those of other bacilli
----------------------------------------------------------------------
The *B. licheniformis*ATCC 14580 gene models were compared to the list of essential genes in *B. subtilis*\[[@B33]\]. Predictably, all of the essential genes in *B. subtilis*have orthologs in *B. licheniformis*, and most are present in a wide range of bacterial taxa. In pairwise BLAST comparisons, 66% of the predicted *B. licheniformis*genes have orthologs in *B. subtilis*, and 55% of the gene models are represented by orthologous sequences in *B. halodurans*(E-value threshold of 1 × 10^-5^; Figure [4](#F4){ref-type="fig"}). Using a reciprocal BLASTP analysis we found 1,719 orthologs that are common to all three species (E-value threshold of 1 × 10^-5^).
As noted by Lapidus *et al.*\[[@B9]\], there are broad regions of colinearity between the genomes of *B. licheniformis*and *B. subtilis*(Figure [5](#F5){ref-type="fig"}). Less conservation of genome organization exists between *B. licheniformis*and *B. halodurans*, and substantial genomic segments have been inverted in *B. halodurans*with respect to *B. licheniformis*and *B. subtilis*. These observations clearly support previous hypotheses \[[@B8]\] that *B. subtilis*and *B. licheniformis*are phylogenetically and evolutionarily closer to each other than to *B. halodurans*.
Conclusions
===========
In comparisons of shared regions, the genomes of *B. licheniformis*ATCC 14580 and *B. subtilis*168 are approximately 84.6% identical at the nucleotide level and show extensive organizational similarity. Accordingly, their genome sequences represent potentially useful instruments for comparative and evolutionary studies among species within the *subtilis-licheniformis*group, and they may offer new information regarding the evolution and ecology of these closely related species.
Despite the broad colinearity of *B. licheniformis*and *B. subtilis*genomes, there are local regions that are individually unique. These include chromosome segments that comprise prophage and insertion sequence elements, DNA restriction-modification systems, antibiotic synthases, and a number of extracellular enzymes and metabolic activities that are not present in *B. subtilis*. It is tempting to speculate that the presence of these genes forecasts the ability of *B. licheniformis*to grow on an expanded array of substrates and/or in additional ecological niches compared to *B. subtilis*. Together, the similarities and differences may hint at overlapping but non-identical environmental niches for these taxa.
The *subtilis-licheniformis*group of bacilli includes many strains that are used to manufacture industrial enzymes, antibiotics and biochemicals. The availability of a complete genome from *B. licheniformis*should permit a thorough comparison of the biochemical pathways and regulatory networks in *B. subtilis*and *B. licheniformis*, thereby offering new opportunities and strategies for improvement of industrial strains. When considering the safety of *B. licheniformis*as an industrial organism it should be noted that the species is considered neither a human pathogen nor a toxigenic microorganism \[[@B34]\]; however, there are reports in the literature implicating it as a causal agent of food poisoning. In these isolated cases, specific strains were shown to produce a toxin similar to cereulide, the emetic toxin of *B. cereus*\[[@B35]\]. Cereulide is a cyclic depsipeptide synthesized non-ribosomally \[[@B36]\]. Importantly, the only non-ribosomal peptide synthase genes found in the *B. licheniformis*ATCC 14580 genome are those that involved in synthesis of lichenysin. Similarly, we detected no homologs of the *B. cereus*hemolytic and non-hemolytic enterotoxins (Swiss-Prot accession numbers P80567, P80568, P80172, and P81242).
In a comparison of the genotypic and phenotypic characteristics among 182 soil isolates of *B. licheniformis*, Manachini *et al.*\[[@B37]\] observed that while this bacterial species appears to be phenotypically homogeneous, clear genotypic differences are evident between isolates. They postulated the existence of three genomovars for *B. licheniformis*. Similarly, De Clerck and De Vos \[[@B38]\] proposed that this species consists of two lineages that can be distinguished using several molecular genotyping methods. The genome sequence data presented in this work should provide a solid foundation on which to conduct future studies to elucidate the genotypic variation among *B. licheniformis*isolates.
Materials and methods
=====================
Shotgun DNA sequencing and genome assembly
------------------------------------------
The genome of *B. licheniformis*ATCC 14580 was sequenced by a combination of the whole-genome shotgun method \[[@B39]\] and fosmid end sequencing \[[@B40]\]. Plasmid libraries were constructed using randomly sheared and *Mbo*I-digested genomic DNA that was enriched for fragments of 2-3 kb by preparative agarose gel electrophoresis. Approximately 49,000 random clones were sequenced using dye-terminator chemistry (Applied Biosystems) with ABI 377 and ABI 3700 automated sequencers yielding approximately 6× coverage of the genome. A combination of methods was used for gap closure, including sequencing on fosmids \[[@B40]\] and primer-walking on selected clones and PCR-amplified DNA fragments. We also incorporated data from both ends of approximately 1,975 fosmid clones with an average insert size of 40 kb to aid in validating the final assembly. In total, the number of input reads was 62,685, with 78.6% of these incorporated into the assembled genome sequence. Individual nucleotides were called using TraceTuner 2.0 (Paracel), and sequence reads were assembled into contigs using the Paracel Genome Assembler using optimized parameters and the quality score set to \>20. Phrap, Crossmatch and Consed were used for sequence finishing \[[@B41]\].
Prediction and annotation of CDSs
---------------------------------
Protein-coding regions in the assembled genome sequence were identified using a combination of previously described software tools including EasyGene \[[@B42]\], Glimmer \[[@B43]\] and FrameD \[[@B44]\]. EasyGene was used as the primary gene finder in these studies. It searches for protein matches in the raw genome data to derive a good training set, and an HMM with states for coding regions as well as ribosome-binding sites (RBSs) is estimated from the dataset. This HMM is used to score all the predicted CDSs in the genome, and the score is converted to a measure of significance (R-value) which is the expected number of CDSs that would be predicted in 1 Mb of random DNA. Gene models with R-values lower than 10 and a log-odds score of greater than -10 were included/considered significant. The principal advantage of this significance measure is that it properly takes into account the length distribution of random CDSs. EasyGene has been shown to match or exceed other gene finders currently available \[[@B42]\]. Glimmer was used as a secondary gene finder to aid in identification of small genes (\< 100 bp) that were sometimes missed by EasyGene. Glimmer models were post-processed with RBSFINDER \[[@B45]\] to pinpoint the positions of start codons by searching for consensus Shine-Dalgarno sequences. According to the RBS states in the EasyGene HMM model, the bases with the highest probability were AA**AAGGAG**(the bases in bold type had distinctly higher probabilities compared to the initial AA). This motif concurs with the consensus Shine-Dalgarno sequence for *B. subtilis*(AAAGGAGG) \[[@B46]\]. RBSFINDER identified the core AAGGAG motif in around 80% of the cases for Glimmer gene predictions and adjusted the start codon accordingly. Manual inspection and alignments to *B. subtilis*homologs were also used to determine the incidence of specific genes. During the gene-finding process, possible errors and frameshifts were detected by both visual inspection of the CDSs to look for interrupted or truncated genes and by deploying FrameD software \[[@B44]\]. Frameshifts were resolved by re-sequencing of PCR-amplified segments or subclones. After re-sequencing and manual editing a total of 27 frameshifts remain in the genome assembly (excluding those contained in the *IS3Bli1*elements). It is not known at present whether these represent pseudogenes or instances of programmed translational frameshifting. The positions of rRNA operons in the genome assembly were confirmed by long-range PCR amplification using primers that annealed to genes flanking the rRNA genes. These PCR fragments were sequenced to high redundancy and the consensus sequences were manually inserted into the assembly. Among the seven rRNA operons, the nucleotide sequences of 16S and 23S genes are at least 99% identical, differing by only one to three nucleotides in pairwise comparisons. Protein-coding sequences were annotated in an automated fashion with the following software applications. Predicted proteins were compared to the nonredundant database PIR-NREF \[[@B47]\] and the *B. subtilis*genome \[[@B48]\] using BLASTP with a E-value threshold of 1 × 10^-5^. InterProScan was used to predict putative function \[[@B49]\]. The InterPro analysis included comparison to PFAM \[[@B50]\], TIGRFAM \[[@B51]\], Interpro \[[@B52]\] signal peptide prediction using SignalP \[[@B20]\] and transmembrane domain prediction using TMHMM \[[@B21]\]. These CDSs were assigned to functional categories based on the Cluster of Orthologous Groups (COG) database \[[@B53]\] with manual verification as described \[[@B54],[@B55]\]. Phage gene boundaries were predicted using gene finding algorithms and by homology to known bacteriophage genes. Transfer RNA genes were identified using tRNAscan-SE \[[@B56]\]. *B. licheniformis*genes that shared significant homology with *B. subtilis*counterparts were named using the nomenclature in the SubtiList database \[[@B48]\] with updated gene names from the BSORF \[[@B57]\] and UniProt \[[@B58]\] databases.
Comparative analyses
--------------------
VisualGenome software (Rational Genomics) was used for comparisons of ortholog distribution among *B. licheniformis*, *B. subtilis*and *B. halodurans*genomes with precomputed BLAST results stored in a local database.
Accession of genome sequence information
----------------------------------------
The GenBank accession number for the *B. licheniformis*ATCC 14580 genome is CP000002. An interactive web portal for viewing and searching the assembled genome based on the generic genome browser developed by Stein *et al.*\[[@B59]\] is available at \[[@B60]\].
Acknowledgements
================
We thank Alan Sloma, William Widner and Michael D. Thomas for helpful discussions.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Circular representation of the *B. licheniformis*ATCC 14580 chromosome. Circles are numbered from 1 (outermost) to 7 (innermost). Circles 1 and 2 show the locations of predicted CDSs on the + and - strands, respectively; circle 3, %G+C; circle 4, GC skew ((G-C/G+C)); circle 5, homology with *B. subtilis*168; circle 6, homology with *B. halodurans*; circle 7 shows positions of nine copies of insertion sequence element *IS3Bli1*and a putative transposase gene; small green bars inside circle 7 denote the positions of possible prophage elements.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Schematic map of the insertion sequence *IS3Bli1*. Nine identical copies of this 1,285-bp element were found in the genome of *B. licheniformis*ATCC 14580. Features of the *IS3Bli1*element are summarized in the text.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Isochore plot of the *B. licheniformis*ATCC 14580 genome showing G+C content as a function of position. AT-rich peaks (numbered 1-24) are marked on the plot, and a single island that is atypically GC-rich is indicated by number 25. Table 2 lists the specific chromosomal features represented by each numbered peak.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Venn diagram comparing the orthologous gene complements of *B. licheniformis*ATCC 14580, *B. subtilis*168 and *B. halodurans*C-125. Numbers in purple boxes indicate the number of pairwise orthologs between adjacent species (BLAST threshold E = 1 × 10^-5^). Numbers in the outer circles represent the total number of CDSs predicted in each genome, numbers in areas of overlap indicate the number of orthologs predicted by reciprocal BLASTP analysis (threshold E = 1 × 10^-5^), and the number in the center gives the number of orthologous sequences common to all three genomes.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Two- and three-dimensional similarity plots comparing the distribution of orthologs on the chromosomes of *B. licheniformis*, *B. subtilis*and *B. halodurans*. BLAST scores were generated and dots were positioned according to the locations in the genome where orthologs exist in order to view possible regions of possible colinearity. The minimum BLAST expectancy score for each dot in this example was 1 × 10^-50^. **(a)**The plot for *B. licheniformis*versus *B. subtilis*; **(b)***B. halodurans*versus *B. subtilis*; **(c)***B. licheniformis*versus *B. halodurans*; **(d)**a three-dimensional scatter plot comparing the distribution of orthologs among all three species. Dots located on the diagonal are indicative of conserved location of orthologous genes between species, whereas a line of dots that lie perpendicular to the diagonal suggests an inversion of a genomic segment between species.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Features of the *B. licheniformis*genome and comparison with genomes of other *Bacillus*species
:::
Feature *B. licheniformis* *B. subtilis*\* *B. halodurans*^†^ *Oceanobacillus iheyensis*^‡^ *B. anthracis*^§^ *B. cereus*^¶^
---------------------------------- -------------------- ----------------- -------------------- ------------------------------- ------------------- ----------------
Chromosome size (bp) 4,222,336 4,214,630 4,202,352 3,630,528 5,227,293 5,426,909
G+C content (mol%) 46.2 43.5 43.7 35.7 35.4 35.4
Protein coding sequences 4,208 4,106 4,066 3,496 5,508 5,366
Average length (bp) 873 896 879 883 800 835
Percent of coding region 86 87 85 85 84 84
Ribosomal RNA operons 7 10 8 7 11 13
Number of tRNAs 72 86 78 69 95 108
Phage-associated genes 71 268 42 27 62 124
Transposase genes of IS elements 10 0 93 14 18 10
\*Kunst *et al.*\[10\]; ^†^Takami *et al.*\[13\]; ^‡^Takami *et al.*\[16\]; ^§^Read *et al.*\[61\]; ^¶^Ivanova *et al.*\[62\].
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Gene sequences corresponding to isochore peaks shown in Figure 3
:::
Peak Size (kb) \% G+C *B. subtilis*orthologs Annotation
------ ----------- -------- ------------------------ -------------------------------------------------------------------------------------------------------------------------
1 3.2 28 No ABC transporter, conserved hypothetical, and hypothetical genes
2 3.6 38 No Conserved hypothetical and hypothetical genes
3 2.1 37 No Conserved hypothetical and hypothetical genes
4 2.8 37 No Hypothetical genes
5 2.7 37 No Phosphotriesterase, conserved hypothetical genes
6 7.4 37 No Type I restriction-modification system
7 3.5 38 No Hypothetical genes
8 8.4 38 Partial *yybO*, *pucR*, *pucH*, *yurH*, *ycbE*, *yjfA*, *rapG*, carbamate kinase, conserved hypothetical and hypothetical genes
9 10.1 36 No SPP-1 like phage, conserved hypothetical and hypothetical genes
10 4.8 37 Yes Hypothetical genes
11 3.0 33 No Conserved hypothetical and hypothetical genes
12 4.3 34 No Conserved hypothetical and hypothetical genes
13 2.2 34 No Conserved hypothetical and hypothetical genes
14 5.4 36 Partial Conserved hypothetical and hypothetical genes
15 4.4 35 No Conserved hypothetical and hypothetical genes
16 4.6 33 No ABC transporter and hypothetical genes
17 3.5 35 Partial *comP*, *comX*, *comQ*, and *IS3Bli1*
18 6.8 37 No *IS3Bli1*, conserved hypothetical and hypothetical genes
19 3.8 38 No Phage w-105-like genes
20 6.8 35 Yes *tagG*and *tagF*genes
21 3.2 34 No Conserved hypothetical genes
22 1.7 34 No Conserved hypothetical genes
23 1.6 37 No Hypothetical genes
24 16.2 35 No Type I restriction-modification system, conserved hypothetical and hypothetical genes
25 3.3 62 No Hypothetical gene
:::
|
PubMed Central
|
2024-06-05T03:55:51.753050
|
2004-9-13
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545597/",
"journal": "Genome Biol. 2004 Sep 13; 5(10):r77",
"authors": [
{
"first": "Michael W",
"last": "Rey"
},
{
"first": "Preethi",
"last": "Ramaiya"
},
{
"first": "Beth A",
"last": "Nelson"
},
{
"first": "Shari D",
"last": "Brody-Karpin"
},
{
"first": "Elizabeth J",
"last": "Zaretsky"
},
{
"first": "Maria",
"last": "Tang"
},
{
"first": "Alfredo Lopez",
"last": "de Leon"
},
{
"first": "Henry",
"last": "Xiang"
},
{
"first": "Veronica",
"last": "Gusti"
},
{
"first": "Ib Groth",
"last": "Clausen"
},
{
"first": "Peter B",
"last": "Olsen"
},
{
"first": "Michael D",
"last": "Rasmussen"
},
{
"first": "Jens T",
"last": "Andersen"
},
{
"first": "Per L",
"last": "Jørgensen"
},
{
"first": "Thomas S",
"last": "Larsen"
},
{
"first": "Alexei",
"last": "Sorokin"
},
{
"first": "Alexander",
"last": "Bolotin"
},
{
"first": "Alla",
"last": "Lapidus"
},
{
"first": "Nathalie",
"last": "Galleron"
},
{
"first": "S Dusko",
"last": "Ehrlich"
},
{
"first": "Randy M",
"last": "Berka"
}
]
}
|
PMC545598
|
Background
==========
Endogenous retroviruses and long terminal repeat (LTR) retrotransposons (collectively called retroelements) generally comprise a significant portion of higher eukaryotic genomes. Dismissed as parasitic or \'junk\' DNA, these sequences have traditionally received less attention than sequences contributing to the functional capacity of the organism. This perspective has changed with the completion of several eukaryotic genome sequences. The contributions of retroelements to genome content range from 3% in baker\'s yeast to 80% in maize \[[@B1],[@B2]\]. Retroelement abundance has resulted in increased appreciation of the important evolutionary role they play in shaping genomes, fueling processes such as mutation, recombination, sequence duplication and genome expansion \[[@B3]\].
The impact of retroelements on their hosts is not without constraint: the host imposes an environmental landscape (the genome) within which retroelements must develop strategies to persist. Retroelement cDNA insertion directly impacts on the host\'s genetic material, making this step a likely target for regulatory control. Transposable elements (TEs) in some systems utilize mechanisms that direct integration to specific chromosomal sites or safe havens \[[@B4],[@B5]\]. For example, the LTR retrotransposons of yeast are associated with domains of heterochromatin or sites bound by particular transcriptional complexes such as RNA polymerase III \[[@B6]-[@B9]\]. These regions are typically gene poor and may enable yeast retrotransposons to replicate without causing their host undue damage \[[@B10]\]. Non-uniform chromosomal distributions are observed in other organisms as well. For example, many retroelements of *Arabidopsis thaliana*and *Drosophila melanogaster*are clustered in pericentromeric heterochromatin \[[@B11],[@B12]\]. However, beyond the yeast model, it is not known whether retroelements generally seek safe havens for integration.
The genome of *A. thaliana*is ideal for exploring processes that influence the chromosomal distribution of retroelements. *A. thaliana*retroelement diversity has been analyzed previously, preparing the way for this study \[[@B13]-[@B15]\]. In contrast to the genomes of *Saccharomyces cerevisiae*, *Schizosaccharomyces pombe*and *Caenorhabditis elegans*, which have relatively few retroelements, *A. thaliana*has a diverse mobile element population whose physical distribution can be described in detail. Another benefit of *A. thaliana*stems from the fact that in contrast to most other \'completely sequenced\' eukaryotic genomes, the *A. thaliana*genome sequence better represents chromosomal DNA of all types, including sequences within heterochromatin \[[@B11]\]. Here we undertake a comprehensive characterization of the LTR retroelements in the well characterized genome of *A. thaliana*to better understand the factors contributing to their genomic distribution.
Results
=======
Dataset
-------
All reverse transcriptases in the *A. thaliana*genome were identified by iterated BLAST searches (Figure [1](#F1){ref-type="fig"}). The query sequences were representative reverse transcriptases from the Metaviridae, Pseudoviridae and non-LTR retrotransposons (Table [1](#T1){ref-type="table"}). LTRs (if present) were assigned to each reverse transcriptase using the software package RetroMap (Figure [1](#F1){ref-type="fig"}, see also Materials and methods). Although the coding sequences of many elements with flanking LTRs were degenerate, they are referred to as full-length or complete elements (FLE) to indicate that two LTRs or LTR fragments could be identified. 5\' LTRs from FLEs and published *A. thaliana*elements were used to identify solo LTRs in the genome by BLAST searches. The final data set consisted of three insertion subtypes: 376 FLEs, 535 reverse transcriptase (RT)-only hits, and 3,268 solo LTRs (Table [2](#T2){ref-type="table"}). These sequences comprise 3,951,101 bases or 3.36% of the total 117,429,178 bases in The Institute of Genomic Research (TIGR) 7 January 2002 version of the genome. Overall, chromosomal retroelement content ranged from 2.64% (chromosome 1) to 4.31% (chromosome 3). Chromosome 4 contained the fewest FLEs (53) and solo LTRs (449), whereas chromosome 3 had the most (92 FLEs and 1,053 solo LTRs).
Element subtypes (FLE, RT-only and solo LTRs) were sorted into taxonomic groupings using the formal taxonomic nomenclature assigned to retrotransposons \[[@B16],[@B17]\]. Our analysis identified numerous insertions for both the Pseudoviridae (211 FLE/82 RT-only/483 solo LTRs) and Metaviridae (168 FLE/142 RT-only/2,803 solo LTRs). The non-LTR retrotransposons lack flanking direct repeats, and therefore only reverse transcriptase information is provided in this study; 311 non-LTR retrotransposon reverse transcriptases were identified. Unlike the Pseudoviridae, *A. thaliana*Metaviridae elements can easily be divided into sublineages, which are referred to as the *Tat*, *Athila*and *Metavirus*elements \[[@B14],[@B18]\] (Figure [2](#F2){ref-type="fig"}). Our method identified 42 *Tat*FLEs, 38 *Athila*FLEs and numerous divergent *Metavirus*elements (82 FLE). No evidence was found for *BEL*or *DIRS*retroelements.
The Metaviridae make up 2.34% of the *A. thaliana*genome, whereas the Pseudoviridae represent only 1.25% of the total genomic DNA. This difference is accounted for largely by the longer average size of Metaviridae FLEs (8,952 nucleotides) and solo LTRs (447 nucleotides) when contrasted with the Pseudoviridae FLEs (5,336 nucleotides) and solo LTRs (187 nucleotides) (data not shown). Among the subgroups of the Metaviridae, the average length of *Metaviruses*is closer to that of the Pseudoviridae than to the mean lengths of the *Athila*and *Tat*lineages. The Pseudoviridae are also more uniformly sized than the Metaviridae. A second factor contributing to the abundance of Metaviridae is that they have approximately six times more solo LTRs than the Pseudoviridae, even though numbers of complete elements are similar between families (Table [2](#T2){ref-type="table"}). The ratios of solo LTRs to FLEs also clearly differ between the Metaviridae (16.7:1) and Pseudoviridae (2.3:1).
Chromosomal distribution
------------------------
The distribution of retroelements was examined on a genome-wide basis. Upon mapping the retroelement families onto the *A. thaliana*chromosomes, the previously noted pericentromeric clustering of TEs was immediately evident (Figure [3](#F3){ref-type="fig"}) \[[@B11]\]. The Metaviridae appeared to cluster in the pericentromeric regions more tightly than the Pseudoviridae and non-LTR retrotransposons. Distributions of these latter two groups appeared similar, as did the distribution of solo LTRs relative to full-length elements (Figure [4](#F4){ref-type="fig"}).
We assessed statistical support for the apparent clustering of elements by comparing the observed distribution of each lineage to a random uniform distribution model (Table [3](#T3){ref-type="table"}). This model assumes that any location in the genome is expected to have a uniform probability of element insertion. This model was rejected by Kendall-Sherman tests of uniformity for every lineage and chromosome combination. All *p*-values were less than 0.05 and most were less than 0.0001.
We next looked at distribution patterns between element families to determine whether they are similar. On the basis of the retroelement distribution maps (Figure [3](#F3){ref-type="fig"}), we hypothesized that this would not be the case for the Metaviridae because they appeared to be associated with centromeres to a greater degree than the other families. Each family\'s chromosomal distribution, inclusive of all subtypes (for example, FLE, RT-only and solo LTR), was tested for similarity to the distribution of the other families using a permutation test. With the exception of chromosome 3, the distribution of non-LTR retrotransposons was not significantly different from that of the Pseudoviridae. Comparisons of Metaviridae elements with Psedoviridae and/or non-LTR elements differed significantly (*p*\< 0.05) for all combinations.
To assess whether the Metaviridae sublineages contributed equally to the observed distribution bias, we tested a model wherein the three sublineages (*Athila*, *Tat*and *Metavirus*) were expected to have similar distributions. This appears to be true, as significant differences were not detected on any chromosome for these sublineages. We then checked whether the FLEs, RT-only hits or solo LTRs displayed different distributions from one another within their respective families. No consistently significant trends were observed for the Pseudoviridae or the Metaviridae. Oddly, the Metaviridae solo LTR distribution displayed significant differences from the FLEs and RT-only hits for chromosome 3.
A feature of pericentromeric regions in *A. thaliana*is that they are heterochromatic, a state required for targeted integration by the yeast Ty5 retroelement \[[@B19]\]. Because of the observed pericentromeric clustering of retrotransposons in *A. thaliana*, we assessed a simple model that assumes that all elements transpose to heterochromatin (Table [4](#T4){ref-type="table"}). There are several genomic regions that are typically considered heterochromatic in *A. thaliana*- centromeres, knobs (on chromosomes 4 and 5), telomeres and rDNA \[[@B20]-[@B22]\]. We looked for differences between lineages with respect to whether retroelements were within a heterochromatic region, or, if outside, whether differences existed in distances to the nearest heterochromatic domain. All lineage combinations showed highly significant differences in heterochromatic distributions. In the Metaviridae, the *Metavirus*elements are less tightly associated with heterochromatin than are *Tat*and *Athila*, which did not differ significantly from each other. Element subtypes also differed in their distribution with respect to heterochromatin. The major source of differences was the distribution of solo LTRs in the Metaviridae.
Age of insertions
-----------------
LTR retroelements have a built-in clock that can be used to estimate the age of given insertions. At the time an element inserts into the genome, the LTRs are typically 100% identical. As time passes, mutations occur within the LTRs at a rate approximating the host\'s mutation rate. LTR divergence, therefore, can be used to estimate relative ages between elements, assuming that all elements share the same probability of incurring a mutation. Although it is possible to estimate ages for non-LTR retrotransposons by generating a putative ancestral consensus sequence and calculating divergence from the consensus, this method is not directly equivalent to estimating ages by LTR comparisons. Therefore, age comparisons were performed only for the LTR retroelement families. Note that the ages depicted in Figure [5](#F5){ref-type="fig"} are relative, and we do not claim that a particular element is a specific age in this study. Rather, we focus on whether elements are significantly older or younger than each other.
Statistically significant age differences were observed among the Pseudoviridae and three Metaviridae sublineages (*F*= 14.4, df = 3 and 368, *p*\< 0.0001) (Table [5](#T5){ref-type="table"}, Figure [5](#F5){ref-type="fig"}). Overall, the Pseudoviridae are younger than the Metaviridae (*t*= 5.72, df = 368, *p*\< 0.0001). When the Metaviridae sublineages are considered, it is apparent that the *Athila*elements are responsible for much of the increased age of this family. The difference between *Athila*and the other two sublineages is significant, with *p*= 0.0003 being the highest value for sublineage comparisons. Elements within heterochromatic regions were significantly older than those found outside (*F*= 17.19, df = 1 and 368, *p*\< 0.0001). There was suggestive evidence that the mean element ages varied among chromosomes (*F*= 2.73, df = 4 and 368, *p*= 0.0289). However, all pairwise comparisons between chromosomes failed to yield significant results at the 0.05 level using the Tukey-Kramer adjustment (data not shown).
Discussion
==========
Completed genome sequences enable comprehensive analyses of retroelement diversity and the exploration of the impact of retroelements on genome organization. Although most large-scale sequencing projects use the shotgun sequencing method, this method makes it particularly difficult to assemble repetitive sequences and to correctly position sequence repeats on the genome scaffold. Consequently, regions of repetitive DNA such as nucleolar-organizing regions (NORs), telomeres and centromeres tend to be skipped, or are sometimes represented by consensus or sampled sequences. The difficulty of cloning repetitive sequences and the drawbacks noted above result in the under- or misrepresentation of the repetitive content of most genomes. Because retroelements frequently comprise a large proportion of the repetitive DNA, \'completed\' genome sequences are typically not ideal for studies of retroelement diversity and distribution on a genomic scale. In contrast to these cases, the *A. thaliana*genome is reliably sequenced well into heterochromatic regions and work continues to further define these domains \[[@B11],[@B23]\].
Another factor frustrating comprehensive analyses of eukaryotic mobile genetic elements is the inherent difficulty in annotating these sequences. Many mobile element insertions are structurally degenerate, rearranged through recombination or organized in complex arrays. Software tools and databases such as Reputer \[[@B24]\] and Repbase update \[[@B25]\] have been developed to identify and classify repeat sequences, and these tools have proved helpful in several genome-wide surveys of mobile elements. RECON \[[@B26]\] and LTR\_STRUC \[[@B27]\] are software tools that go one step further and consider structural features of mobile elements that can assist in genome annotation. We developed an additional software tool, called RetroMap, to assist in characterizing the LTR retroelement content of genomes. RetroMap delimits LTR retroelement insertions by iterated identification of reverse transcriptases followed by a search for flanking LTRs. The software goes beyond existing platforms and carries out a number of analytic functions, including age assignment, solo LTR identification and visualization of the chromosomal locations of various groups of identified elements on a whole-genome scale.
Data generated by RetroMap are subject to a few caveats. First, because element searches use reverse transcriptase sequences as queries, elements lacking reverse transcriptase motifs (for whatever reason) will not be identified. Second, when RetroMap encounters nested elements, tandem elements, and other complex arrangements, it does not attempt to delimit the element. Rather, the user is notified that a complex arrangement was encountered and the original reverse transcriptase match and any LTR(s) found are logged as separate entities.
For the most part, RetroMap was quite effective in identifying LTR retrotransposon insertions. Our results closely agree with the findings of a parallel study conducted by Pereira \[[@B28]\]. For the Pseudoviridae and two of the three Metaviridae lineages (*Tat*and *Metavirus*), we identified 210 and 128 full-length elements, respectively, whereas Pereira recovered 215 and 130 insertions for these respective element groups. The two studies, however, differed significantly in the number of *Athila*elements identified. We found 38 insertions, whereas Pereira recovered 219. To reconcile these differences, we independently estimated *Athila*copy numbers by conducting iterative BLAST searches with a variety of *Athila*query sequences (data not shown). BLAST hits recovered with each query were then mapped onto the genome sequence. As a result of this analysis, we concluded that RetroMap missed many *Athila*insertions, either because they are highly degenerate or part of complex arrangements. In contrast to Pereira\'s approach, RetroMap requires that a reverse transcriptase reside between LTRs, and in many cases reverse transcriptases were absent or not detectable in *Athila*insertions. This can be resolved in future implementations of RetroMap that enable multiple query sequences to be tested. The *Athila*elements are large, and our underestimate of the number of *Athila*elements resulted in a corresponding underestimate of the total amount of retrotransposon DNA in the *A. thaliana*genome. We calculated 3.36% for this value, whereas Pereira calculated 5.60%. Pereira\'s estimate is likely to be the more accurate of the two.
With the exception of the *Athila*elements, the observed frequency of insertions in complex arrangements was rare. For example, the Pseudoviridae had only eight nested and five unassignable elements. The small observed number of complex element arrangements in *A. thaliana*contrasts sharply with observations in grass genomes, where retroelements are usually found in complex nested arrays \[[@B29],[@B30]\]. This may reflect a difference between species in factors contributing to chromosomal distribution of retroelements, or it may simply be a consequence of the difference in abundance of retroelements between *A. thaliana*(5.60% of the genome) and grasses (up to 80% of some genomes) \[[@B1],[@B28]\].
Genomic distribution of *A. thaliana*retroelements
--------------------------------------------------
Our data on the genomic distribution of retroelements can be considered in the light of theoretical work predicting the distribution of TE populations within genomes. These studies largely focus on the effects of selection and recombination on element insertions \[[@B31],[@B32]\]. Particularly relevant is the recent study by Wright *et al.*\[[@B33]\], which considers the effects of recombination on the genomic distribution of major groups of mobile elements in *A. thaliana*(DNA transposons and retroelements). Our analysis extends this work by considering the genomic distribution of specific retroelement lineages. We investigate a model wherein selection and recombination affect element lineages uniformly, and hypothesize that observed deviations in the genomic distribution of specific element lineages reflect unique aspects of their evolutionary history or survival strategies such as targeted integration.
Ectopic exchange model
----------------------
The ectopic exchange model assumes that inter-element recombination restricts growth of element populations \[[@B31]\]. Elements should be most numerous in regions of reduced recombination such as the centromeres, because of less frequent loss by homologous recombination. A corollary is that element abundance at a genomic location should inversely reflect the recombination rate for that region in the genome. Previous work suggests that this model is not the primary determinant of element abundance in *A. thaliana*. Wright *et al.*\[[@B33]\] examined recombination rate relative to element abundance in detail and found that the abundance of most *A. thaliana*TE families actually had a small but positive correlation with recombination rate, as was also observed in *C. elegans*\[[@B34]\]. Devos *et al.*\[[@B35]\] found ectopic recombination to be very infrequent relative to intra-element recombination, suggesting this process is unlikely to have a significant role in explaining the observed *A. thaliana*retrotransposable element distribution.
The ectopic exchange hypothesis makes two unique predictions for retrotransposons: solo LTRs (a product of recombination) should be observed in higher proportions relative to full-length elements outside of heterochromatin; and heterochromatic elements will show a shift toward greater average age than elements elsewhere in the genome. Our consideration of age assumes that the chance of loss by recombination remains steady or increases with element age. However, old elements will have higher sequence divergence, thereby reducing the likelihood that they will recombine. In considering age, we also assume that all elements evolve at the same rates. This is unlikely to be the case, as local, chromosomal and compartmental locations are increasingly found to have different mutation rates \[[@B36],[@B37]\].
With respect to the distribution of solo LTRs, our data show exactly the opposite bias predicted by the ectopic exchange model: the ratio of Metaviridae solo LTRs to FLEs in heterochromatin was nearly twice that found outside heterochromatin. The frequency of solo LTRs at the centromeres suggests that homologous recombination, at least over short distances (less than 20 kilobases (kb)), occurs frequently in pericentromeric regions.
While we did observe the predicted shift toward older elements within heterochromatin, the data are not consistent with low rates of recombination as the determinant of retrotransposon accumulation at the centromeres. Within the Metaviridae, for example, the *Metaviruses*and *Tat*elements differ significantly in their association with heterochromatin. The ectopic exchange model would predict that the *Tat*elements should be older; however, these two lineages do not differ significantly in age. Although it is possible that recombinational forces could act differentially on different element sublineages, we view this as unlikely. Rather, forces other than ectopic recombination, such as targeted integration (see below), are responsible for the differential genomic distribution of certain element lineages. This is not to say that ectopic exchange has no role; however, it is unlikely to be the sole or prevailing influence.
Deleterious insertion model
---------------------------
The deleterious insertion model hypothesizes that element insertions are generally harmful to the host, and thus elements accumulate in regions of low gene density, where insertions are least likely to have negative effects on the host. According to this model, abundance of all classes of mobile elements should inversely reflect gene density within the genome. This is supported by the observation that elements are over-represented in gene-poor pericentromeric heterochromatin and are rare over much of the chromosome arms. However, we did not observe an increase in element abundance at other gene-poor heterochromatic regions (such as the telomeres and NORs), which would be predicted by the deleterious insertion hypothesis. This model would also predict that element insertions into gene-rich regions that are tolerated by the host should act as founders or safe havens for future element insertions. This could lead to an ever-expanding area of tightly clustered and frequently nested elements in euchromatin, assuming the overall random insertion rate is greater than the rate of sequence loss through recombination. Nested clusters of elements have been reported in cereals such as maize and barley \[[@B29],[@B30]\]. In *A. thaliana*, although numerous potential \'seed\' insertion sites are observed along the chromosome arms, we did not detect dense clusters of nested elements at these locations.
In contrast to the deleterious insertion model, it is important to recognize that some element insertions may provide a selective advantage. Studies in *C. elegans*and rice indicate that many retrotranposons are associated with genes (63% and 20% in these species respectively) \[[@B38],[@B39]\]. In *D. melanogaster*, some retrotransposon-gene associations are preserved in diverse natural populations, consistent with the hypothesis that they confer a positive selective advantage \[[@B40]\]. Furthermore, recent analyses in *S. pombe*suggest that the Tf1 retrotransposons may regulate expression of adjacent genes \[[@B41]\]. We cannot rule out a role for positive selection in the distribution of some *A. thaliana*mobile elements, but identifying such a role would require a more refined analysis of element distribution and gene associations.
Impact of targeted integration
------------------------------
The observation that many LTR retroelements have non-uniform genomic distributions suggested that targeted integration may be a driver of retroelement distribution patterns \[[@B42]\]. Neither the deleterious insertion nor ectopic recombination models address the situation where some or all elements have evolved the ability to bias their distributions through targeted integration. The LTR retroelements of *S. cerevisiae*insert preferentially into heterochromatin or sites occupied by RNA polymerase III, and in the evolutionarily distant *S. pombe*genome, retroelements are located preferentially upstream of genes transcribed by RNA polymerase II \[[@B6]-[@B9]\]. Retroviruses also insert preferentially into transcribed regions, with some retroviruses favoring insertions into promoter regions \[[@B4],[@B43]\].
Targeted integration could contribute significantly to the chromosomal distribution of *A. thaliana*retroelements. As in other systems, targeting may occur because elements recognize a specific chromatin state and actively insert into regions with that type of chromatin. A chromatin-targeting model has the following predictions. First, very few elements will be found outside targeted chromatin domains. For example, all heterochromatic regions such as NORs, knobs and telomeres would be occupied by the same lineage of elements if these regions share a chromatin feature recognized by that lineage. Second, different retroelement lineages may be associated with different regions of the genome if they employ different targeting strategies.
The targeting hypothesis is well supported for the Metaviridae, which on a genome-wide basis differ significantly in their chromosomal distribution from the Pseudoviridae and non-LTR retrotransposons. This is particularly true for the *Athila*and *Tat*lineages, both of which are tightly associated with pericentromeric regions. *Athila*and *Tat*elements are not found in heterochromatin regions around the telomeres, however, suggesting that telomeric and centromeric heterochromatin differ. Targeted integration to pericentromeric heterochromatin may be a general feature of the Metaviridae. Members of the Metaviridae are abundant in pericentromeric heterochromatin in many grass species \[[@B44]\]. Langdon *et al.*\[[@B45]\] suggested that an evolutionary ancient member of the Metaviridae in cereals targets to centromeric domains. Portions of a maize homolog of this element were found to co-precipitate with the centromere-specific histone CENH3, indicating an association of this element with a particular type of chromatin \[[@B46]\].
The Pseudoviridae and non-LTR retrotransposons differ in their genomic organization from the Metaviridae and are more loosely associated with pericentromeric regions. It may be that these element lineages do not target their integration, or they may recognize other chromosomal features, although we did not observe any association with other genome features or gene classes such as tRNA genes (data not shown). *De novo*integration events have been mapped on a chromosomal level for two tobacco Pseudoviridae elements in heterologous hosts - Tto1 in *A. thaliana*and Tnt1 in *Medicago trunculata.*In both cases these elements integrated throughout the genome, displaying some preference for genic regions \[[@B47],[@B48]\]. Whether this observed distribution pattern reflects random integration or recognition of some other subtle chromosomal feature remains to be determined. Because we predict that the Metaviridae recognize pericentromeric heterochromatin, an important dataset for analysis will be maps of the various DNA methylation and histone-modification patterns for the full genome. In-depth characterization of the distribution of retroelements relative to chromatin modifications may reveal additional evidence for targeting and help to understand the impact of targeting on genome organization.
Conclusions
===========
Our analysis of the genomic distribution of the *A. thaliana*LTR retroelements revealed that the distribution of the Pseudoviridae and the Metaviridae is non-uniform and that they tend to cluster at the centromeres. The pericentromeric association of three Metaviridae sublineages (*Metavirus*, *Tat*and *Athila*) was significantly more pronounced than for the Pseudoviridae. Several factors are likely to contribute to the centromeric association of these elements, including target-site bias, selection against euchromatin integration and pericentromeric accumulation of elements due to suppression of recombination. For the *Tat*and *Athila*lineages, however, target-site specificity appears to be the primary factor determining chromosomal distribution. We predict that, like retroelements in yeast, the *Tat*and *Athila*elements target integration to pericentromeric regions by recognizing a specific feature of pericentromeric heterochromatin.
Materials and methods
=====================
RetroMap and the *A. thaliana*retroelement dataset
--------------------------------------------------
Reverse transcriptase amino-acid sequences (as defined by \[[@B49]\], see also Table [1](#T1){ref-type="table"}), were used to query a database of *A. thaliana*chromosomes (TIGR version 7 January 2002) with the tblastn program (E = 1e^-10^, XML output, filtering disabled) \[[@B50]\]. The resulting search report was imported into RetroMap. RetroMap (to be described in detail elsewhere) provides a graphical user interface (GUI) to interactively characterize LTR retrotransposons in targeted genomes or large genomic contigs (Figure [1a](#F1){ref-type="fig"}). RetroMap generates a nonredundant set of database hits from BLAST results generated by a given query sequence set. Hits are merged if they directly overlap or if they align to different portions of the same query sequence. In this study, the nonredundant sequences were used to re-query the chromosome database twice more using tblastx (E = 1e^-10^, XML output, filtering disabled) to identify increasingly divergent or degenerate elements. Unique hits identified in the final round of screening were taken to represent the entire complement of retroelements in *A. thaliana*.
RetroMap assigns putative LTRs where possible for each reverse transcriptase by comparing 10 kb of DNA from each flank. This is accomplished using Blast2Sequences to identify flanking repeats \[[@B51]\] (Figure [1b](#F1){ref-type="fig"}). Direct repeats found closest to the reverse transcriptase, larger than 50 bp and less than 5 kb, are considered to be LTRs. Hits with putative LTRs were considered to be full-length elements (FLE) or complete elements. Twenty-six reverse transcriptase hits were excluded from the FLEs owing to difficulty in automatic LTR assignment (13 each from the Pseudoviridae and Metaviridae). Among these were nested elements and tandem elements sharing a LTR. Reverse transcriptases were assigned to a retroelement lineage (Metaviridae, Pseudoviridae or non-LTR retrotransposon) on the basis of their similarity to the diagnostic reverse transcriptase query sequences. Full-length Metaviridae elements were further subdivided into the classic (*Metavirus*), *Tat*and *Athila*groups on the basis of the highest-scoring match in a BLAST database containing the Metaviridae reverse transcriptase sequences described in \[[@B18]\]. Putative complete elements with a predicted reverse transcriptase failing to significantly match any sequence in this database were removed from further consideration as false positives (two cases).
Solo LTRs and solo LTR fragments were identified with blastn (E \< 1e^-5^) using all predicted 5\' LTRs of known *A. thaliana*elements and the FLEs. RetroMap assigns any putative LTR sequence that fails to match or overlap with a predicted FLE LTR as a solo LTR.
Relative age calculation for full-length elements
-------------------------------------------------
LTRs are identical at the time of retroelement integration, and so relative element ages were estimated from the percentage of identical residues shared between 5\' and 3\' LTRs for FLEs. The age formula used was *T*= *d*/2*k*(time (*T*) = genetic distance (*d*)/ \[2 × substitution rate (*k*)\]), where genetic distance is 1 - (percent identity/100) and the substitution rate is 1.5 × 10^-8^\[[@B52]\].
Assignment of heterochromatin boundaries
----------------------------------------
Chromosome coordinates relative to the left (north) end were used to calculate distances between retroelements and heterochromatic domains. Heterochromatin boundaries were derived from \[[@B20]-[@B22]\] and include the telomeres, heterochromatic knobs, NORs and centromeres. Chromosome end-coordinates were considered as the telomere boundaries. The *A. thaliana*NORs are located at the left (north) ends of chromosomes 2 and 4, and as these regions were only sample sequenced, their boundaries were assigned as the left ends of chromosomes 2 and 4. Heterochromatic knobs and pericentromeric regions were assigned as the outermost physical markers delimiting these regions, as determined by the studies listed above.
Statistical tests
-----------------
A RetroMap-generated datafile was used as the data source for statistical testing. The data file contains chromosomal element coordinates, LTR identity, age and lineage information for all *A. thaliana*retroelement families by element category: reverse transcriptase only (R), full-length (F), and solo LTR (S).
For each element type and each chromosome, a Kendall-Sherman test \[[@B53]-[@B55]\] was conducted to determine if the element positions were randomly distributed across chromosomes according to a uniform distribution. A permutation test \[[@B55]\] was used to assess the statistical significance of observed differences in the chromosomal position distributions for each chromosome across various element categories. The multi-response permutation procedures (MRPP) test is briefly described as follows. The average distance between a pair of elements within a category of interest is determined. A weighted sum of these averages over all categories of interest is computed, with each category weighted in proportion to the number of elements in the category. This weighted sum is the observed value of the test statistic. Next, the test statistic is re-computed for each of 10,000 random permutations of the category labels. For each permutation, the observed chromosomal positions of the elements are held constant while the category labels are randomly shuffled. The proportion of the 10,000 permutation-replicated test statistics that are less than or equal to the original observed test statistic serves as an approximate *p*-value for a test whose null hypothesis is that all element categories of interest have the same chromosomal position distribution. This permutation approach is useful for the chromosomal position data because first, no distributional assumptions are required, second, differences in chromosomal position distributions other than simple location shifts are detectable, and third, the method is not as sensitive to outliers as common parametric approaches.
For FLEs, linear model analyses were used to assess the effects of the factors \'chromosome\', \'lineage/sublineage\', and \'location\' relative to heterochromatin on the response variable \'element age\'. *F*-tests were used to check for interaction between these three factors and to assess the statistical significance of observed differences among the five chromosomes, among the four lineage/sublineage categories (Pseudoviridae and the three Metaviridae sublineages:*Athila*, *Tat*or *Metavirus*), and between elements inside and outside heterochromatin. The square root of age was used as the response variable in the age analysis so that the variance of the response would be roughly constant across categories defined by combinations of chromosome, lineage/sublineage, and location, as required for standard linear model analyses. Outlying observations were present, but the results of the analysis remained essentially the same with or without the outliers. Thus the reported results are based on the full dataset.
Additional data files
=====================
The following additional data are available with the online version of this article: a Microsoft Excel spreadsheet of data generated by RetroMap for each retrotransposon insertion identified; the data in this file was used for all statistical analyses (Additional data file [1](#s1){ref-type="supplementary-material"}). The Java application used to generate the LTR and retrotransposon coordinates and to estimate retrotransposon ages (Additional data file [2](#s2){ref-type="supplementary-material"}). To run RetroMap, version 1.3 or higher of the Java Runtime Environment (JRE <http://java.sun.com>) must be present. To enable searches for LTRs, NCBI\'s BLAST 2 Sequences must be locally installed.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
A Microsoft Excel spreadsheet of data generated by RetroMap for each retrotransposon insertion identified; the data in this file was used for all statistical analyses
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
The Java application used to generate the LTR and retrotransposon coordinates and to estimate retrotransposon ages
:::
::: {.caption}
######
Click here for additional data file
:::
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Assembling the retroelement dataset. **(a)**Flow chart for the generation of the dataset. The shaded region denotes steps coordinated by the RetroMap software. (Eprobe refers to a BLAST query sequence) **(b)**LTR prediction. The innermost direct repeats identified in sequences flanking the original BLAST hit are assigned as LTRs. The repeats delimit the boundaries of the full-length LTR retrotransposons.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
*Arabidopsis thaliana*Metaviridae and Pseudoviridae reverse transcriptase diversity. Phylogenetic trees used in this figure are adapted from \[14,18\]. Each tree is based on ClustalX \[56\] alignments of reverse transcriptase domains for elements in a given family. Neighbor-joining trees (10,000 bootstrap repetitions) were generated using MEGA2 \[57\]. The non-LTR retrotransposon Ta11 served as the root for both trees. The three Metaviridae sublineages are boxed.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Physical distribution of full-length *A. thaliana*retroelements. The five *A. thaliana*chromosomes are designated as Ath1-5. Triangles indicate the location of a particular retroelement on the chromosome. Non-LTR retrotransposons are in black, Pseudoviridae in gray, and Metaviridae in white. Vertical bars on the chromosome show the precise location of the retroelement. Regions of heterochromatin are represented as follows: telomeres and NORs (on Ath2 and Ath4) by rounded chromosome ends; centromeres by hourglass shapes; heterochromatic knobs (on Ath4 and Ath5) by narrowed stretches on chromosome bars. The relatively short chromosome 5 knob is barely visible to the right of the centromere. The inset more clearly depicts heterochromatic regions that are obscured by element insertions. Chromosomes are drawn to scale.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Chromosomal distribution of LTRs for the Metaviridae and Pseudoviridae families in *A. thaliana*. Chromosomes are displayed as in Figure 3. In addition, solo LTRs are drawn as open triangles. The upper chromosome depicts the distribution of Pseudoviridae, the lower the distribution of Metaviridae. In contrast to Figure 3, shading is not used to distinguish between the families.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Relative ages of *A. thaliana*LTR retroelement lineages. **(a)**Box-plot showing the age distribution of Pseudoviridae full-length elements contrasted with those of the Metaviridae. The position of the median is shown as a gray bar in the box that delimits the boundaries of the lower and upper quartiles. Data points more than 1.5 times the inter-quartile range above the upper quartile or below the lower quartile are indicated by individual horizontal lines. Ages were calculated as described in Materials and methods. **(b)**Relative-age box-plots of Metaviridae sublineages. Permutation test *p*-values for the significance of the displayed age distributions are shown below each box-plot.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Retroelement species used as BLAST probes
:::
Element GenBank accession number Host organism Family Genus Length (nucleotides) LTR identity (length in nucleotides)
-------------- -------------------------- ----------------------------- -------- --------------- ---------------------- --------------------------------------
*Athila4*-6 AF296831 *Arabidopsis thaliana* MV *Metavirus* 14,016 98.2 (1747)
Cer1 U15406 *Caenorhabditis elegans* MV *Metavirus* 8,865 100.0 (492)
*Osvaldo* AJ133521 *Drosophila buzzatii* MV *Metavirus* 9,045 99.9 (1196)
*Sushi* AF030881 *Fugu rubripes* MV *Metavirus* 5,645 91.0 (610)
Tf1 M38526 *Schizosaccharomyces pombe* MV *Metavirus* 4,941 100.0 (358)
Ty3 M23367 *Saccharomyces cerevisiae* MV *Metavirus* 5,428 100.0 (340)
Art1 Y08010 *A. thaliana* PV *Pseudovirus* 4,793 99.8 (439)
*Copia* M11240 *Drosophila melanogaster* PV *Hemivirus* 5,416 100.0 (276)
*Endovir1*-1 AY016208 *A. thaliana* PV *Sirevirus* 9,089 99.8 (548)
*SIRE*-1 AF053008 *Glycine max* PV *Sirevirus* 10,444 100.0 (2149)
Tca2 AF050215 *Candida albicans* PV *Hemivirus* 6,428 100.0 (280)
Tca5 AF065434 *C. albicans* PV *Hemivirus* 5,588 100.0 (685)
*Jockey* M22874 *D. melanogaster* NL \- 5154 \-
L1.2 M80343 *Homo sapiens* NL \- 6,050 \-
R1 X51968 *D. melanogaster* NL \- 5356 \-
R2 X51967 *D. melanogaster* NL \- 3,607 \-
Ta11 L47193 *A. thaliana* NL \- 7,808 \-
MV, Metaviridae; PV, Pseudoviridae; NL, non-LTR retrotransposon.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
*A. thaliana*LTR retroelements by chromosome
:::
Chromosome 1 30,080,809 nucleotides Chromosome 2 19,643,621 nucleotides Chromosome 3 23,465,812 nucleotides Chromosome 4 17,549,528 nucleotides Chromosome 5 26,689,408 nucleotides Total 117,429,178 nucleotides
---------------------------------- ------------------------------------- ------------------------------------- ------------------------------------- ------------------------------------- ------------------------------------- -------------------------------
**Pseudoviridae**
RT only 21 19 16 10 16 82
Complete elements 48 42 47 35 38 210
Nucleotides 239,675 211,083 285,207 185,127 199,386 1,120,478
Percentage of total nucleotides 0.88% 1.24% 1.34% 1.21% 0.96% 1.1%
Solo LTRs 84 100 125 89 87 485
Nucleotides 16,516 19,275 23,906 15,500 15,248 90,445
Percentage of total nucleotides 0.13% 0.16% 0.18% 0.15% 0.13% 0.15%
**Metaviridae**
RT only 16 30 41 23 32 142
Complete elements 37 34 45 18 32 166
Nucleotides 309,690 319,802 375,703 161,352 319,535 1,486,082
Percentage of total nucleotides 1.23% 2.82% 2.22% 1.40% 1.59% 1.74%
Solo LTRs 435 500 928 360 560 2,783
Nucleotides 228,115 257,810 326,484 179,500 262,187 1,254,096
Percentage of total nucleotides 1.15% 1.74% 1.71% 1.42% 1.24% 1.42%
***Athila***
Complete elements 7 8 8 4 11 38
Nucleotides 72,094 90,171 93,015 37,339 119,646 412,265
Percentage of total nucleotides 0.38% 0.87% 0.67% 0.41% 0.69% 0.60%
***Tat***
Complete elements 14 10 8 6 8 46
Nucleotides 131,154 102,534 83,327 68,754 103,112 591,944
Percentage of total nucleotides 0.44% 0.54% 0.52% 0.46% 0.56% 0.50%
***Metavirus***
Complete elements 16 16 29 8 13 82
Nucleotides 106,442 127,097 199,361 55,259 96,777 748,231
Percentage of total nucleotides 0.42% 1.03% 1.03% 0.52% 0.33% 0.64%
**Non-LTR retrotransposon** 49 90 69 32 71 311
**Total LTR contribution**
Complete elements 85 76 92 53 70 376
Nucleotides 634,695 798,606 836,968 457,405 679,255 3,331,357
Percentage of total nucleotides 2.11% 4.07% 3.57% 2.61% 2.55% 2.84%
Solo LTRs 519 600 1,053 449 647 3,268
Nucleotides 386,759 373,256 444,804 275,361 364,340 1,844,520
Percentage of total nucleotides 1.29% 1.90% 1.90% 1.57% 1.37% 1.57%
Both
Nucleotides 1,021,454 1,171,862 1,281,772 732,766 1,043,595 5,175,877
Percentage of total nucleotides 3.40% 5.97% 5.46% 4.18% 3.91% 4.41%
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Comparison of genome localization by retroelement lineage
:::
Hypotheses Test Group(s) tested *p*-values by chromosome Accept?
--------------------------------------------------------------------------- ----------------------------------------------------- --------------------------------- -------------------------- ------------ ------------ ------------ ------------ -----
All families are randomly distributed according to a uniform distribution Uniform goodness of fit, 10,000 random permutations MV(F) **0.0000** **0.0000** **0.0000** **0.0000** **0.0000** No
PV(F) **0.0000** **0.0007** **0.0000** **0.0022** **0.0464** No
MV(S) **0.0000** **0.0000** **0.0000** **0.0000** **0.0000** No
PV(S) **0.0000** **0.0000** **0.0000** **0.0000** **0.0000** No
MV(R) **0.0000** **0.0000** **0.0000** **0.0000** **0.0000** No
PV(R) **0.0000** **0.0007** **0.0000** **0.0097** **0.0000** No
NL(R) **0.0000** **0.0000** **0.0000** **0.0002** **0.0000** No
Retroelement family distributions are organized similarly in the genome MRPP, 10,000 random permutations MV(FSR), PV(FSR), NL(R) **0.0000** **0.0000** **0.0000** **0.0000** **0.0000** No
MV(FSR), PV(FSR) **0.0000** **0.0000** **0.0000** **0.0000** **0.0000** No
MV(FSR), NL(R) **0.0000** **0.0000** **0.0000** **0.0000** **0.0000** No
PV(FSR), NL(R) 0.3498 0.8326 **0.0241** 0.1468 0.1417 Yes
All Metaviridae sublineages have similar distributions MRPP, 10,000 random permutations MV *Athila*, *Metavirus*, *Tat* 0.2200 0.1365 0.5676 0.4174 0.2788 Yes
MV *Athila*, *Metavirus* 0.1057 0.3010 0.2657 0.4526 0.4453 Yes
MV *Athila*, *Tat* 0.1687 0.0970 0.7116 0.3773 0.2781 Yes
MV *Metavirus*, *Tat* 0.4903 0.1268 0.7341 0.5753 0.2361 Yes
Metaviridae subtypes have similar distributions MRPP, 10,000 random permutations MV(FSR) 0.7742 0.1247 **0.0000** 0.7425 0.0659 Yes
MV(FS) 0.4544 0.1357 **0.0003** 0.4435 0.7241 Yes
MV(FR) 0.5184 0.9461 0.5750 0.5480 0.2135 Yes
MV(SR) 0.9068 0.1339 **0.0051** 0.8194 **0.0157** Yes
Pseudoviridae subtypes have similar distributions MRPP, 10,000 random permutations PV(FSR) 0.0509 0.2039 0.2199 0.0953 **0.0379** Yes
PV(FS) 0.2732 0.0853 0.2665 0.6567 **0.0453** Yes
PV(FR) **0.0136** 0.5055 0.1185 0.0521 **0.0281** Yes
PV(SR) 0.0743 0.5604 0.2513 **0.0307** 0.3476 Yes
MV, Metaviridae; PV, Pseudoviridae; NL, non-LTR retrotransposon; R, RT-only; S, solo LTR; F, full-length element. *p*-values \< 0.05 are displayed in bold text.
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Association of retroelements with heterochromatin
:::
Hypotheses Test Group(s) tested *p*-values Accept?
--------------------------------------------------------------------------------- ---------------------------------- --------------------------------- ------------ ---------
All families share a similar probability of being in or outside heterochromatin MRPP, 10,000 random permutations MV(FSR), PV(FSR), NL(R) **0.0000** No
MV(FSR), PV(FSR) **0.0000** No
MV(FSR), NL(R) **0.0000** No
PV(FSR), NL(R) **0.0000** No
Metaviridae sublineages have similar heterochromatic distributions MRPP, 10,000 random permutations MV *Athila*, *Metavirus*, *Tat* **0.0011** No
MV *Athila*, *Metavirus* **0.0016** No
MV *Athila*, *Tat* 0.5211 Yes
MV *Metavirus*, *Tat* **0.0105** No
Element subtypes have similar heterochromatic distributions MRPP, 1,000 random permutations MV(SR), PV(SR), NL(R) **0.0000** No
MV(FR), PV(FR), NL(R) 0.3960 Yes
MV(FS), PV(FS) **0.0000** No
Pseudoviridae subtypes have similar heterochromatic distributions Pearson\'s chi-square PV(FSR) **0.0002** No
PV(FS) **0.0001** No
PV(FR) **0.0073** No
PV(SR) 0.9419 Yes
Metaviridae subtypes have similar heterochromatic distributions Pearson\'s chi-square MV(FSR) **0.0001** No
MV(FS) **0.0002** No
MV(FR) 0.5146 Yes
MV(SR) **0.0159** No
MV, Metaviridae; PV, Pseudoviridae; NL, non-LTR retrotransposon; R, RT-only; S, solo LTR; F, full-length element. *p*-values \< 0.05 are displayed in bold text.
:::
::: {#T5 .table-wrap}
Table 5
::: {.caption}
######
Comparison of LTR retroelement age distributions
:::
Hypotheses Test Group(s) tested Chromosome *p*-values Accept?
----------------------------------------------------------------------------------------------------- ---------------------------------- --------------------------------- ----------------------- ---------
Metaviridae and Pseudoviridae have similar age distributions in the genome MRPP, 10,000 random permutations MV(F), PV(F) **0.0000** No
Metaviridae sublineages have similar age distributions MRPP, 10,000 random permutations MV *Athila*, *Metavirus*, *Tat* **0.0000** No
MV *Athila*, *Metavirus* **0.0000** No
MV *Athila*, *Tat* **0.0003** No
MV *Metavirus*, *Tat* 0.4618 Yes
Metaviridae age distributions are similar whether the elements are in or out of heterochromatin MRPP, 10,000 random permutations MV(F) **0.0021** No
Metaviridae sublineage age distributions are similar whether they are in or outside heterochromatin MRPP, 10,000 random permutations MV *Athila* **0.0410** No
MV *Metavirus* 0.5747 Yes
MV *Tat* **0.0457** No
Pseudoviridae age distributions are similar whether the elements are in or outside heterochromatin MRPP, 10,000 random permutations PV(F) **0.0167** No
MV, Metaviridae; PV, Pseudoviridae; NL, non-LTR retrotransposon; R, RT-only; S, solo LTR; F, full-length element. *p*-values \< 0.05 are displayed in bold text.
:::
|
PubMed Central
|
2024-06-05T03:55:51.756928
|
2004-9-29
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545598/",
"journal": "Genome Biol. 2004 Sep 29; 5(10):R78",
"authors": [
{
"first": "Brooke D",
"last": "Peterson-Burch"
},
{
"first": "Dan",
"last": "Nettleton"
},
{
"first": "Daniel F",
"last": "Voytas"
}
]
}
|
PMC545599
|
Background
==========
It has become increasingly clear that the activity of transposable elements (TEs) is a major cause of genome evolution. TEs are ubiquitous components of eukaryotic genomes. For example, 22% of the *Drosophila melanogaster*\[[@B1]\], 45% of the human \[[@B2]\], and up to 80% of the maize \[[@B3]\] genomes consist of TE fossils. TEs have influenced the evolution of cellular gene regulation and function, and have been responsible for chromosomal rearrangements \[[@B4]\]. Variation in genome size and the C-value paradox \[[@B5]\] can be attributed to a large extent to differences in the amount of TEs, particularly of retrotransposons, between the genomes of different species \[[@B6]\]. In plant genomes, large size and structural variation even among closely related species is mainly due to differences in their history of polyploidization \[[@B7]\] and/or amplification of long terminal repeat (LTR)-retrotransposons \[[@B3],[@B8]-[@B10]\]. LTR-retrotransposons (LTR-RTs) are \'copy-and-paste\' (class I) TEs that replicate via an RNA intermediate. Like retroviruses, their (intact) genome consists of two LTRs, which contain the signals for transcription initiation and termination, flanking an internal region (IR) that typically contains genes and other features necessary for autonomous retrotransposition. LTR-RTs are mainly classified into two major families, the Pseudoviridae (also known as *Ty1/Copia*elements) and Metaviridae (*Ty3/Gypsy*).
The evolutionary forces that control copy number and shape the chromosomal distribution of different kinds of TEs in eukaryotic genomes are still poorly understood. Some large plant and animal genomes have expanded owing to an ability to tolerate massive amplification of retrotransposons, whereas in more compact genomes these elements are found in lower copy numbers, non-randomly distributed and mainly confined to heterochromatic regions \[[@B11]-[@B14]\]. TEs have mostly been regarded as parasitic DNA \[[@B15],[@B16]\], and it has been suggested that important epigenetic mechanisms originally evolved to suppress the activity of TEs and other foreign genetic material \[[@B17]\]. Nevertheless, there are examples of individual elements that have been co-opted by, and entire TE families that have become mutualists to, their host genomes \[[@B13]\].
It is often hypothesized that the non-random genomic distribution of TEs in some species reflects the action of purifying selection on the host against the deleterious effects of TE insertions in certain regions. Models differ in the kind of deleterious effects they propose: chromosomal rearrangements due to \'ectopic\' (unequal homologous) recombination \[[@B18]\]; disruption of gene regulation due to insertion near cellular genes \[[@B19]\]; or a burden on cell physiology as a result of the expression of TE-encoded products \[[@B20]\]. In compact genomes, clustering of TE insertions in silent heterochromatin, which has reduced rates of recombination, gene density and levels of transcription, is in principle consistent with a scenario of negative selection and of passive accumulation of TEs where their insertions would be less deleterious. As an alternative to purifying selection, another hypothesis to explain this clustering of TEs involves preferential insertion, or even positive selection for their retention, into heterochromatin \[[@B21]\].
To evaluate these hypotheses, I investigated the evolutionary history of different groups of LTR-RTs in the *Arabidopsis thaliana*genome. The total TE content of the compact genome of *A. thaliana*, with a haploid size of approximately 150 Mbp (million base-pairs), has been previously estimated as around 10%, and is known to cluster around the pericentromeric heterochromatin \[[@B14]\]. Despite the relatively low copy numbers, there is a high diversity of LTR-RTs in *A. thaliana*\[[@B22],[@B23]\]. I have implemented an automated methodology for genome-wide sequence mining of LTR-RTs, and for estimating the age of insertion of different copies. This methodology is capable of identifying nested insertions, which are common in the pericentromeric regions. The technique for dating LTR-RTs has been previously used to reveal a massive amplification of these elements that doubled the size of the maize genome during the last 3 million years, by extrapolation of results found in a 240 kbp stretch of intergenic DNA \[[@B3]\]. Here I report genome-wide age profiles for different groups of LTR-RTs in *A. thaliana*. By comparing the age and chromosomal distributions of young and old insertions it is possible to distinguish between preferential targeting and passive accumulation of elements into heterochromatin. I show that members of the Pseudoviridae have recently been active, that they integrate randomly into the genome (relative to centromere location) and only passively accumulate in proximal regions, as purifying selection eliminates euchromatic insertions. In contrast, the Metaviridae (particularly members of the *Athila*group) preferentially insert into the pericentromeric heterochromatin, and their transpositional activity has declined in the last million years.
Results
=======
Abundance and diversity
-----------------------
Most of the retrieved elements are fragmented and truncated, and nested insertions are common particularly among pericentromeric elements belonging to the *Athila*superfamily, though the core centromere sequences themselves were not available. In fact, the size of the *A. thaliana*genome has been recently estimated as approximately 157 Mbp (around 20% larger than the estimate published with the genome sequence), and the additional size appears to be due to (unsequenced) heterochromatic repetitive DNA in the centromeres, telomeres and nucleolar-organizing regions \[[@B24]\]. Table [1](#T1){ref-type="table"} shows the relative abundance of each superfamily, and the numbers of complete and solo-LTR elements identified in the genome. *Athila*is the most abundant superfamily, followed by the *Copia*-like, *Gypsy*-like, and TRIM (terminal-repeat retrotransposons in miniature). The ratio of solo-LTRs to complete elements is around 2:1. In addition to solo-LTR formation, deletion and fragmentation of retrotransposon DNA in *A. thaliana*also occur via other mechanisms: 36% of the DNA in the *Athila*, 38% in the *Gypsy*-like, 32% in the *Copia*-like, and 21% in the TRIM superfamilies correspond to degraded insertions that are neither \'complete\' elements nor solo-LTRs.
Age distribution
----------------
To obtain the genome-wide age distribution of each superfamily (except TRIM), 564 pairs of intra-element LTRs were (pairwise) aligned and their sequence divergence estimated. Many of the complete TRIM elements have highly divergent LTRs, and I suspect that extensive recombination between inter-element LTRs has occurred. In neighbor-joining trees of LTR sequences (of both complete and solo elements) from the TRIM families *Katydid-At1*and *Katydid-At2*, most intra-element LTR pairs did not cluster. In contrast, when trees were constructed for representatives of the *Athila*(*athila2*), *Gypsy*-like (*atlantys2*), and *Copia*-like (*meta1*, *atcopia49*, *atcopia78*) superfamilies, intra-element LTR pairs always clustered (data not shown), providing evidence for the lack of inter-element recombination in those \'families\'.
The superfamilies differ significantly in their average age of insertions. *Athila*insertions are significantly older than the *Gypsy*-like (Wilcoxon rank-sum test, *p*\< 0.0005), *Gypsy*-like older than *Copia*-like (*p*\< 0.0001). Age distributions are summarized in Figure [1](#F1){ref-type="fig"}.
*Copia*-like insertions are younger than host species
-----------------------------------------------------
Using the rate of 1.5 × 10^-8^substitutions per site per year \[[@B25]\], 97% of 215 complete *Copia*-like elements are younger than 3 million years (Myr), 90% younger than 2 Myr, and only two insertions estimated to be older than 4 Myr. This shows that complete insertions from the known *Copia*-like families in the *A. thaliana*genome are younger than the species itself, whose time of divergence from its closest relatives, such as *A. lyrata*has been estimated (with the same rate of evolution) to be 5.1-5.4 Myr ago \[[@B25]\]. The situation is less clear for *Athila*(and the *Gypsy*-like TEs), as 7% of 219 intra-element LTR pairs were estimated to be older than 5 Myr (3% of the *Gypsy*-like). Furthermore, the *Athila*and *Gypsy*-like superfamilies have an excess of degraded insertions relative to *Copia*-like (Table [1](#T1){ref-type="table"}). Complete elements account for around 50% of the total amount of DNA in *Athila*and *Gypsy*-like, indicating that the majority of insertions remaining in the genome have been degraded or have become solo-LTRs. Some of these are likely to be older than the complete insertions. DNA loss (from LTR-RTs) has been shown to occur in *A. thaliana*\[[@B26]\], and the oldest insertions may have been degraded beyond detection. On the other hand, there is some evidence that synonymous sites in *Arabidopsis*are not evolving in a completely neutral fashion \[[@B27]\]. If this were the case for the chalcone synthase (*Chs*) and alcohol dehydrogenase (*Adh*) loci, their synonymous sites would be evolving more slowly than LTR-RT fossils, and the dating method described above would systematically overestimate the ages of their insertion events.
*Athila*and *Gypsy*-like elements were more active in the past
--------------------------------------------------------------
The age distribution of complete *Copia*-like elements appears to show a recent burst of activity (Figure [1](#F1){ref-type="fig"}), but I provide evidence (below) that the excess of very young elements is the result of the rapid (relative to Metaviridae insertions) elimination of these elements from the genome. In contrast, the age distributions of complete *Athila*and *Gypsy*-like insertions have peaks between 1 and 2 Myr ago (Figure [1](#F1){ref-type="fig"}). Moreover, whereas there are 34 *Copia*-like insertions with their intra-element LTRs identical in sequence, only four such *Athila*and three such *Gypsy*-like insertions are present. These results indicate that levels of transpositional activity of *Athila*and *Gypsy*-like elements have declined since their peak between 1 and 2 Myr ago.
Physical distribution
---------------------
The chromosomal distribution of retrotransposons (and other TEs) in *A. thaliana*has been known to be non-random and dominated by a high concentration of elements in the heterochromatic pericentromeric regions \[[@B14]\]. However, this study has revealed significant differences in the chromosomal locations of the LTR-RT superfamilies. I have analyzed the distribution of complete elements and of solo-LTRs in each superfamily along all the chromosome arms combined, relative to the position of the centromeres (that is, the distribution of the distances between each insertion and the centromere, divided by the length of the respective arm), with results summarized in Figure [2](#F2){ref-type="fig"}.
*Athila*elements are almost exclusively inserted in the pericentromeric regions, and the other superfamilies in significantly and progressively less proximal regions of the chromosome arms (Wilcoxon rank sum tests: *Athila*more proximal than the *Gypsy*-like, *p*\< 0.0001; *Gypsy*-like more proximal than *Copia*-like, *p*\< 0.0001; complete *Copia*-like elements more proximal than complete TRIM elements, *p*\< 0.05; there is no difference between *Copia*-like and TRIM solo-LTRs). Furthermore, except for TRIM, within each superfamily the solo-LTRs are significantly more distal than the complete elements (Wilcoxon rank sum tests, *p*\< 0.001), suggesting that formation of solo-LTRs is more likely to occur in distal regions. The distribution of complete TRIM elements relative to the centromere is not significantly different from random (goodness-of-fit test, χ^2^= 4.22, df = 3, *p*\> 0.2), although sample size is small, while their solo-LTRs are significantly clustered (goodness-of-fit test, χ^2^= 10.70, df = 3, *p*\< 0.02).
Accumulation in proximal regions by distinct evolutionary mechanisms: purifying selection and insertion bias
------------------------------------------------------------------------------------------------------------
The results above indicate that the older a superfamily is, the more its elements are concentrated in the proximal regions. This suggests that insertions into proximal (heterochromatic) regions are more likely to persist for longer periods of time. This interpretation assumes that the neutral mutation rate is the same for both the distal (euchromatic) and proximal (heterochromatic) portions of the genome. Intra-genomic variation in the per-replication mutation rate has been reported between the two sex chromosomes of a flowering plant \[[@B28]\] (although the difference could not be explained their different degree of DNA methylation, a feature often associated with heterochromatin). Given that the dating method used here is based on neutral sequence divergence (between intra-element LTRs), a higher mutation rate in heterochromatin in *A. thaliana*would affect age comparisons among different groups of elements, as they show different degrees of clustering into the pericentromeric heterochromatin. However, older estimates for the age of heterochromatic elements are consistent with the hypothesis that heterochromatin is a \'safe haven\' where TE insertions persist for longer periods of time. Here I show that the mechanisms that led to the accumulation of LTR-RTs in proximal regions are distinct for different groups: elements of the youngest superfamily (*Copia*-like) insert randomly into the genome (relative to the location of the pericentromeric heterochromatin), but there is negative selection (on the host genome) against their insertions in euchromatin; elements of the older superfamilies (*Athila*, *Gypsy*-like) preferentially insert into the pericentromeric regions. These distinct mechanisms become apparent when temporal and spatial data are combined (Figure [3](#F3){ref-type="fig"}), and the chromosomal distribution of young elements compared with the distribution of older elements (within each superfamily).
For complete *Copia*-like elements there is a highly significant negative correlation between relative distance from the centromere and age of the insertions (Spearman rank correlation, ρ = -0.39, *p*\< 0.0001). Furthermore, the distribution along the chromosome arms of 34 *Copia*-like insertions with no divergence between their intra-element LTRs is not significantly different from random (goodness-of-fit test, χ^2^= 3.12, df = 3, *p*\> 0.3). This is evidence that *Copia-*like elements integrate randomly relative to the location of the centromeres, but tend to get eliminated from distal, and passively accumulate in proximal regions.
The average time to fixation (*t*) for a neutral allele is given by *t*= 4*N*~*e*~, where *N*~*e*~is the effective population size. For *A. thaliana t*can be estimated using an average of estimates of nucleotide diversity (*θ*) for 8 different *A. thaliana*genes, *θ* = 9 × 10^-3^\[[@B29]\], and the synonymous rate of substitution per site per generation, *μ* = 1.5 × 10^-8^\[[@B25]\]. *t*= 2*θ*/*μ*, yielding an estimate of *t*≈ 1.2 Myr. This value for *t*is consistent with an independent estimate that placed the time since the divergence between *A. thaliana*and *A. lyrata*between 3.45*t*and 5.6*t*\[[@B30]\]. Given that 75% of all complete *Copia*-like insertions are younger than 1.2 Myr, most of them are likely to be polymorphic. Taken together with the highly significant negative correlation between age and distance from the pericentromeric regions, these results indicate that complete *Copia*-like insertions are less likely to get fixed in the distal, euchromatic portions of the chromosome arms than in the pericentromeric heterochromatin.
In contrast, there is no correlation between age and relative distance from centromeres for complete *Athila*elements (Spearman rank correlation, ρ = 0.01, *p*= 0.9), as both young and old insertions are found only in proximal regions (Figure [3](#F3){ref-type="fig"}), compartmentalized into the pericentromeric heterochromatin. This strongly suggests that elements in the superfamily have evolved to preferentially target the pericentromeric heterochromatin, and their genomic distribution, unlike that of *Copia*-like elements, is not the result of passive accumulation therein. Only if *Athila*insertions were much more deleterious than *Copia*-like ones, so that they would be very rapidly removed by purifying selection, could passive accumulation be the case.
*Gypsy*-like insertions display a similar pattern to *Athila*. Even though there is for complete elements a significant, negative correlation between relative distance from centromeres and age, this is due to an excess of recent insertions near the telomere of the short arm of chromosome II (data not shown). If the arm is excluded from the analysis there is no significant correlation (Spearman rank correlation, ρ = -0.09, *p*\> 0.3). This suggests that for the *Gypsy*-like also there is an insertional bias towards proximal regions. This bias is not as strong as for *Athila*, as complete *Gypsy*-like insertions are not exclusively found around the centromeres, and they cluster (to a much lesser extent) in at least one other heterochromatic region (the telomere of the short arm of chromosome II). Included in the *Gypsy*-like \'superfamily\' is a clade of elements, known as *Tat*, which is a sister group to *Athila*to the exclusion of the remaining *Gypsy*-like elements \[[@B31]\]. The age and physical distribution of *Tat*does not differ from those of the remaining *Gypsy*-like elements (Wilcoxon rank-sum tests, *p*\> 0.4); *Tat*show insertion bias towards the pericentromeric regions, but again to a lesser degree than *Athila*.
Half-life of complete *Copia*-like insertions
---------------------------------------------
Given that *Copia*-like elements have been active until recently but tend to be eliminated by purifying selection, their age distribution (Figure [1](#F1){ref-type="fig"}, bottom) reflects the process of origin and loss of complete elements, when averaged over evolutionary time scales (and over all Pseudoviridae lineages). If this is assumed to be a steady-state process, it can be modeled by the survivorship function: *N*(*K*) = *N*~*o*~e^-*aK*^, where *N(K)*is the number of elements observed with intra-element LTR divergence *K*, and *N*~*o*~and *a*are constants to be fitted. The rate of elimination can then be estimated by linear regression of the log-transformed data (the half-life of insertions is given by ln2/*a*). Figure [4](#F4){ref-type="fig"} shows the fit for all complete *Copia*-like insertions (*R*^2^= 0.94), and for complete insertions outside the proximal regions (i.e. with relative distance from centromeres \>0.2; *R*^2^= 0.95). Complete *Copia-*like elements are eliminated from the genome with a half-life of 648,000 years (*SE*= 48,000 years). Insertions exclusively outside the proximal (heterochromatic) regions are lost more rapidly, with a half-life of 472,000 years (*SE*= 46,000 years).
Discussion
==========
The results above indicate that within a single genome, distinct evolutionary mechanisms can lead to the non-random distribution of retrotransposons, as in *A. thaliana*the accumulation of insertions in the pericentromeric heterochromatin is the result of both insertion bias (for Metaviridae elements) and a lower probability of fixation in euchromatin (Pseudoviridae).
It has recently been shown that most TE lineages in *A. thaliana*were already present in its common ancestor with *Brassica oleracea*(the two species diverged around 15-20 Myr ago), and that copy numbers are generally higher in *B. oleracea*\[[@B32]\]. The authors suggested that differential amplification of TEs between *A. thaliana*and *B. oleracea*was responsible for the larger genome of the latter. Here I have shown that the major LTR-RT families have been active in *A. thaliana*since its divergence from its closest relatives, such as *A. lyrata*. The transpositional activity of *Metaviridae*elements has declined relative to its level between 1 and 2 Myr ago, perhaps suggesting that the host genome has more efficiently suppressed their transposition since. However, Pseudoviridae (*Copia*-like) elements in *A. thaliana*have been subject to constant turnover. They have been recently active and show no insertion bias, and I estimate that the half-life of a complete element inserted in the euchromatic (non-coding) regions of the chromosome arms as around 470,000 years. Most of these *Pseudoviridae*insertions are lost before they reach fixation, and the half-life estimate provides a measure of the pace at which natural selection on the host constrains the genomic distribution and copy number of Pseudoviridae insertions. Turnover of Pseudoviridae insertions, in contrast to the longer persistence of Metaviridae elements that have declined in activity, is consistent with the higher sequence diversity among the Pseudoviridae than the Metaviridae in *A. thaliana*(107 Repbase update (RU) \'families\' represented in 215 complete Pseudoviridaeelements, 25 RU \'families\' in 349 complete Metaviridae elements, where \'families\' were defined on the basis of sequence divergence); frequent reverse transcription during transposition would be likely to lead to faster evolution than that generated by the host genome DNA polymerase error rate on chromosomal insertions.
The lower probability of fixation in euchromatin relative to heterochromatin implies that insertions into euchromatin are more deleterious to the host (and perhaps that purifying selection is less efficient in heterochromatin due to a much reduced rate of recombination). TE density in the *A. thaliana*genome does not correlate with local recombination rate \[[@B33]\], providing some evidence against the ectopic recombination model for the deleterious effects of insertions (if the occurrence of ectopic and meiotic recombination positively correlate). Consistent with my results, the same study supports a model of purifying selection against insertions in intergenic DNA, by inferring that they are less likely to be found near genes \[[@B33]\].
As an alternative to selection, a neutral mutational process that deletes (part of the) insertions could in principle be driving the distribution of *Copia*-like elements, if such a process occurred more often in the euchromatic than in the pericentromeric regions of the genome, and if it were frequent enough. One mechanism that removes LTR-RT DNA from the genome is solo-LTR formation via unequal homologous recombination between intra-element LTRs. However, this mechanism cannot be the driving force shaping the distribution of complete *Copia*-like elements because *Copia*-like solo-LTRs are also non-randomly distributed and clustered in proximal regions (goodness-of-fit test: χ^2^= 13.71, df = 3, *p*\< 0.005). *Copia*-like solo-LTRs are either eliminated faster from distal than proximal regions, like complete elements, or solo-LTR formation on average occurs more slowly than extinction for euchromatic insertions. Despite clustering around the centromeres, *Copia*-like solo-LTRs are significantly more dispersed than complete elements. This suggests that solo-LTRs do form before extinction for distal insertions, but are probably less efficiently eliminated (possibly because they are less deleterious to the host genome) than complete elements. Another known mechanism of (general) DNA loss operates via small deletions due to illegitimate recombination (between short repeats); this has been shown to occur in the *A. thaliana*genome by an analysis of internal deletions in LTR-RTs \[[@B26]\]. In *Drosophila*, rates of spontaneous deletions in euchromatin and heterochromatin do not seem to differ \[[@B34]\]. In *A. thaliana*the relative rates between the two chromatin domains are unknown, but fragmented (that is, neither solo-LTR nor complete) *Copia*-like insertions are as clustered around the centromeres as complete ones (goodness-of-fit test: χ^2^= 80.36, df = 3, *p*\< 0.0001). Therefore small, spontaneous deletions cannot account for the genomic distribution of complete elements. Larger deletions (that remove the entire LTR-RT sequence) occurring primarily in euchromatin would be necessary to explain the observed accumulation pattern; if such a mechanism existed it would be an important force for genome size contraction. As there is no evidence for such mechanism, and given that I estimate that the half-life of (complete) insertions to be less than half the average time to fixation for a neutral allele, a lower probability of fixation in euchromatin relative to the pericentromeric heterochromatin is more likely to be driving the genomic distribution of Pseudoviridae elements.
It is interesting to note that the integrase proteins encoded by LTR-RTs differ between the Pseudoviridae and the Metaviridae in their carboxy-terminal domains, as they have different characteristic motifs \[[@B35],[@B36]\]. This is the least conserved domain of integrase, and has been implicated in the insertion preferences of certain families of LTR-RTs in different organisms \[[@B37]\]. Examples of families of LTR-RTs whose integrase carboxy termini have been shown to interact with chromatin are known for both the Metaviridae \[[@B36]\] and the Pseudoviridae \[[@B38]\], and manipulation of this domain to engineer the targeting specificity of LTR-RTs has also been achieved \[[@B39]\]. *Athila*elements have been known to be present in the *A. thaliana*core centromeric arrays of the 180-bp satellite repeats and are abundant in pericentromeric heterochromatin \[[@B40],[@B41]\]. In this study I have shown that in contrast with the passive accumulation of *Copia*-like elements, the striking compartmentalization of both recent and older *Athila*insertions in the pericentromeric heterochromatin indicates that these elements actively target those regions, and represents an example of a group of retrotransposons that have evolved to colonize a particular \'genomic niche\'. Passive accumulation could not explain the distribution of *Athila*insertions unless they were generally much more deleterious to their host than *Copia*-like ones. Given the absence of complete *Athila*insertions from euchromatin, any one insertion would have to be so deleterious as to be almost immediately eliminated by purifying selection, even from intergenic DNA. Rather, it is likely that *Athila*elements preferentially insert into the pericentromeric heterochromatin and it is possible that this group of elements has been co-opted to play a part in centromere function. There is some evidence that such hypothetical role cannot be that of *cis*-acting sequences \[[@B42]\], but it could be a structural one. Studies on the appearance of neocentromeres \[[@B43]-[@B45]\] point to some degree of epigenetic regulation and function of centromeres via chromatin structuring. Although centromeric sequences are not conserved among plants \[[@B46]\], centromere-specific families of LTR-RTs seem to be common, as they have been found in cereals \[[@B47]-[@B51]\], chickpeas \[[@B52]\] and *A. thaliana*\[[@B40]\].
Both purifying selection (at the host level) against insertions (in euchromatin) and a decline in transpositional activity (of Metaviridae elements) appear to have limited the recent contribution of retrotransposon DNA to genome size expansion in *A. thaliana*. The rapid and recent genome evolution inferred for *A. thaliana*may be a feature common to other higher eukaryotes, in particular those with compact genomes. High turnover of TE insertions in euchromatin also occurs in *Drosophila*and pufferfish \[[@B53]\], for example, and accumulation of TEs into heterochromatin in those genomes may also, as in *A. thaliana*, be due to diverse evolutionary mechanisms.
Materials and methods
=====================
A methodology was developed for the automated mining of sequence data to retrieve the sequence and chromosomal location of genomic \'fossils\' of LTR-RTs, identifying complete elements and solo-LTRs among the retrieved sequence fragments, and estimating the age of the insertion events that gave origin to these elements. This methodology was applied to the genome sequence of *A. thaliana*.
Molecular paleontology of LTR-retrotransposons
----------------------------------------------
Sequences of the organellar and the five nuclear chromosomes (version 200303) were obtained from the Munich Information Center for Protein Sequences (MIPS) \[[@B54]\]. Computational mining for LTR-RT fragments in the *A. thaliana*genome (around 116 Mbp of available sequence) was performed using sequence-similarity search algorithms \[[@B55]\] against a library of representative sequences of LTR-RTs. This reference library was compiled by extracting from Repbase update \[[@B56],[@B57]\] sequences of the LTRs and internal region (IR) of known *A. thaliana*\'families\' of LTR-RTs. The programs RepeatMasker \[[@B58]\] and WU-BLAST \[[@B59]\] were used to search the whole genomic sequence (initially divided into 50 kbp chunks) and obtain the precise coordinates of chromosomal segments homologous to (a part of) the LTR or IR of library elements. The datasets of chromosomal coordinates of the complete LTR-RTs and solo LTRs identified are available as Additional data files 1 and 3.
\'Families\' of LTR-retrotransposons (as classified in Repbase update) are present in low copy numbers; therefore, for the purpose of this analysis they were grouped into three \'superfamilies\': *Athila*, *Gypsy*-like (all \'families\' belonging to the Metaviridae, excluding *Athila*), and *Copia*-like (all \'families\' belonging to the Pseudoviridae). The Metaviridae was split into two groups (*Athila*and *Gypsy*-like), as initial mining of the *A. thaliana*genome revealed that *Athila*elements have been particularly successful in colonizing it. Their copy number is roughly double the number of all other members of the Metaviridae, and higher than the total of all Pseudoviridae elements. *Athila*form a clade and are retroviral-like elements that are likely to have an *envelope*(*env*) gene \[[@B60]\]. Most of the *Copia*- and *Gypsy-*like elements are typical LTR-RTs, although one of the *Copia*-like \'families\' (*metaI*) comprises non-autonomous elements \[[@B22]\] and a few others (*endovir1*\[[@B61]\], *atcopia41-43*\[[@B22]\]) are retroviral-like, featuring a putative *env*gene. A fourth \'superfamily\' was used to include TRIMs. These are short, non-autonomous elements that feature LTRs but no coding genes and cannot currently be classified into either the Pseudoviridae or the Metaviridae; they are described in \[[@B62]\].
The four superfamilies comprise the following \'families\'. ***Athila***(10 families): *athila2 - 5*, *athila4A - D*, *athila6A*, *athila7*, *athila8A*and *B*; ***Gypsy*-like**(15 families): *atgagpol1*, *atgp2*and *3*, *atgp2N*, *atgp5 - 10*, *atgp9B*, *atlantys1 - 3*, *tat1*; ***Copia*-like**(107 families): *atcopia1 - 97*, *atcopia8A*and *B*, *atcopia18A*, *atcopia32B*, *atcopia38A*and *B*, *atcopia65A*, *endovir1*, *TA1-2*, *meta1*; **TRIM**(3 families): *katydid-At1*, *katydid-At2*, *katydid-At3*.
Identification of complete elements and solo-LTRs
-------------------------------------------------
A *Perl*script, LTR\_MINER (available on request), was written to parse all the chromosomal LTR-RT fragments reported by RepeatMasker (WU-BLAST hits of similarity to reference sequences) and identify complete elements and solo-LTRs. LTR\_MINER performs the pattern-recognition function of assembling hits that originated from single LTR-RT insertion events. The algorithm involves: \'defragmentation\' of LTR hits. If a chromosomal LTR fossil contains insertions/deletions (indels) relative to the most similar library sequence, it may be reported as multiple hits (fragments). Defragmentation is the identification of multiple hits that correspond to the same LTR. Parameters were set so that LTR hits were defragmented only when they were separated by no more than 550 bp, belonged to the same family, had the same orientation on the chromosome, and their combined length did not exceed the length of the corresponding family reference sequence by more than 20 bp.
Identification of \'complete\' elements
---------------------------------------
An intact LTR-RT insertion consists of at least three hits: LTR-IR-LTR (an IR from a single element insertion may also yield multiple hits). After LTR defragmentation, LTR\_MINER searches for contiguous patterns of *LTR*, *IR*, *LTR*. In order to check whether the pattern could be straddling a nested insertion of the same family, the search is then recursively extended from each end of the pattern for further contiguous hits to an IR and a LTR (of same family and orientation). The two LTRs of the innermost pattern are classified as a pair of intra-element LTRs.
Identification of \'interrupted\' elements: fossil elements containing insertions between the two LTRs
------------------------------------------------------------------------------------------------------
LTR\_MINER also identifies such elements provided an IR is present between the LTRs. A maximum pairing distance between LTRs was set at 30 kb.
Identification of \'solo-LTRs\'
-------------------------------
LTR\_MINER was set to classify a LTR fragment as a solo-LTR if no other LTR or IR (of same family and orientation) is present within a 5 kbp radius from the fragment\'s ends. The aim was the identification of elements resulting from deletion (of the IR and one LTR) events via homologous recombination between intra-element LTRs, and not to classify as solo-LTRs sequences that are separated from IRs because of insertions.
Dating of insertion events
--------------------------
Nucleotide sequence divergence between pairs of intra-element LTRs was used as a molecular clock, as these pairs are identical at the time of insertion \[[@B63]\]. All mined pairs of intra-element LTR sequences were aligned using ClustalW \[[@B64]\] (with Pwgapopen = 5.0, Pwgapext = 1.0). To ensure correct alignment of any sequences with large indels, pairwise LTR alignments were position-anchored relative to reference sequences: if a chromosomal LTR fossil consisted of multiple hits (of similarity to segments of the reference sequence) then the intervening chromosomal sequence between such hits was replaced by a number of gaps, equal to the length of the region separating the corresponding segments in the reference. The number of nucleotide substitutions per site (*K*) between each intra-element LTR pair was then estimated using Kimura\'s two-parameter model \[[@B65]\]. To reduce sampling bias towards younger elements, elements with truncated LTRs were included in the analysis (provided both LTRs are present), as intact elements are likely to be younger than elements that have accumulated indels.
Alignments with fewer than 80 nucleotides were discarded. As CLUSTAL-W alignments could be poor if LTR sequences were only partially overlapping, for all LTR pairs with *K*greater than 0.2 they were inspected by eye and manually edited if necessary (and *K*then recalculated). Estimates of the ages of insertion were obtained by using the equation *t = K/2r*, where *t*is the age, and *r*is nucleotide substitution rate for the host genome DNA polymerase. The value of 1.5 × 10^-8^substitutions per site per year was used for *r*(1.0 \<*r*\< 2.1 × 10^-8^95% confidence interval), estimated in \[[@B25]\] for the synonymous substitution rate in the *Chs*and *Adh*loci in *Arabidopsis*/*Arabis*species.
Finally, if recombination between LTRs from different insertions had occurred frequently, the dating method above would be invalid for obtaining the age profiles of different families. To detect possible recombination events, multiple alignments of all LTRs (including solos) of certain families were generated using BLASTALIGN \[[@B66]\], a program that can handle datasets that may contain large indels. Neighbor-joining trees of the LTR sequences were then constructed using PAUP\* 4.0b10 \[[@B67]\] with the HKY85 model, to check whether intra-element LTR pairs clustered.
Additional data files
=====================
The following additional data files are available with the online version of this article. Additional data file [1](#s1){ref-type="supplementary-material"} contains the entire dataset of chromosomal coordinates and ages of complete LTR-retrotransposons in *A. thaliana*. Additional data file [2](#s2){ref-type="supplementary-material"} describes the data fields in Additional data file [1](#s1){ref-type="supplementary-material"}. Additional data file [3](#s3){ref-type="supplementary-material"} contains the entire dataset of chromosomal coordinates of solo-LTRs in *A. thaliana*. Additional data file [4](#s4){ref-type="supplementary-material"} describes the data fields in Additional data file [3](#s3){ref-type="supplementary-material"}. Additional data file [5](#s5){ref-type="supplementary-material"} contains the *Perl*script LTR\_MINER, used to de-fragment sequence similarity hits to LTR-retrotransposons, and identify complete and solo-LTR elements. Additional data file [6](#s6){ref-type="supplementary-material"} describes the utility and usage of the *Perl*script in Additional data file [5](#s5){ref-type="supplementary-material"}. Additional data file [7](#s7){ref-type="supplementary-material"} contains the *Perl*script used in conjunction with LTR\_MINER, used to divide long sequences into smaller chunks labeled by their coordinate range. Additional file data [8](#s8){ref-type="supplementary-material"} describes the usage of the *Perl*script in Additional data file [7](#s7){ref-type="supplementary-material"}.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
The entire dataset of chromosomal coordinates and ages of complete LTR-retrotransposons in *A. thaliana*
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
A file describing the data fields in Additional data file 1
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 3
The entire dataset of chromosomal coordinates of solo-LTRs in *A. thaliana*
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 4
A file describing the data fields in Additional data file 3
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 5
The *Perl*script LTR\_MINER, used to de-fragment sequence similarity hits to LTR-retrotransposons, and identify complete and solo-LTR elements
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 6
A file describing the utility and usage of the *Perl*script in Additional data file 5
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 7
The *Perl*script used in conjunction with LTR\_MINER, used to divide long sequences into smaller chunks labeled by their coordinate range
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 8
A file describing the utility and usage of the *Perl*script in Additional data file 7
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
I thank A. Eyre-Walker for original suggestions; D. Bensasson, A. Saez, A. Burt, R. Belshaw, J. Hughes, A. Katzourakis and M. Tristem for critical reading of earlier versions of the manuscript; and an anonymous referee for suggestions. This work was supported by the Natural Environment Research Council, UK.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Age distributions of LTR-retrotransposon superfamilies. *Athila*insertions are on average significantly older, and *Copia*-like ones younger, than those from other superfamilies. There are 34 *Copia*-like, four *Athila*, and three *Gypsy*-like insertions with identical intra-element LTRs. The width of the horizontal boxes above the histograms indicates the middle 50% of age values in each superfamily; the red band indicates 95% confidence limits on the median, and the green stripe the median value.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Differential pericentromeric clustering of complete elements and solo-LTRs along the 10 chromosome arms combined. The vertical axis measures distance from the centromere, divided by the length of the chromosome arm in which a given element is inserted: the value of 0.0 corresponds to the position of the centromeres and 1.0 to telomeres. Box heights indicate the inter-quartile range and widths are proportional to sample size; red bands represent 95% confidence limits on the median; and the green stripe marks the median value of each sample. Coordinates for the approximate centers of the centromeres on the chromosome sequences were set at 14.70 Mbp for chromosome I (total length 30.14 Mbp), at 3.70 Mbp for II (19.85 Mbp), at 13.70 Mbp for III (23.76 Mbp), at 3.10 Mbp for IV (17.79 Mbp), and at 11.80 Mbp for V (26.99 Mbp).
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Relationship between age and physical distributions of complete elements. Insertions into the short arms of chromosomes II and IV were excluded for clarity. These arms contain extensive heterochromatin away from the centromeres, in nucleolar-organizing regions that juxtapose their telomeres, and in a knob \[[@B14]\]. In addition, their short length implies that the pericentromeric heterochromatin, which spans around 1-1.5 Mbp in each arm \[[@B68]\], corresponds to a substantially higher fraction of their total length than in the other eight arms.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Loss of complete *Copia*-like elements. The half-life of complete *Copia*-like elements throughout the whole genome (log-transformed counts marked by blue circles, blue regression line) is estimated as around 650,000 ± 50,000 years. Complete insertions outside the proximal regions (red squares, red regression line) are lost more rapidly, with a half-life estimated as around 470,000 ± 50,000 years.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Relative abundance of LTR-retrotransposons in *Arabidopsis thaliana*
:::
Superfamily Percentage of genome\* Number of complete elements^†^ Percentage DNA in complete elements^†^ Number of solo-LTRs
-------------- ------------------------ -------------------------------- ---------------------------------------- ---------------------
*Athila* 2.73 % 219 50 % 586
*Gypsy*-like 1.32 % 130 53 % 250
*Copia*-like 1.39 % 215 63 % 343
TRIM 0.15 % 28 53 % 58
Total 5.60 % 592 54 % 1,237
\*The \'% of genome\' includes all LTR-RT sequences (in the nuclear genome) for each superfamily, rather than just complete and solo-LTR elements. Fragments of LTR-RTs were also found in the mitochondrial (2.74%) and chloroplast (0.05%) genomes. ^†^Elements containing indels were included as complete elements provided they retain a substantial part of both their LTRs.
:::
|
PubMed Central
|
2024-06-05T03:55:51.762456
|
2004-9-29
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545599/",
"journal": "Genome Biol. 2004 Sep 29; 5(10):R79",
"authors": [
{
"first": "Vini",
"last": "Pereira"
}
]
}
|
PMC545600
|
Background
==========
The Bioconductor project \[[@B1]\] is an initiative for the collaborative creation of extensible software for computational biology and bioinformatics (CBB). Biology, molecular biology in particular, is undergoing two related transformations. First, there is a growing awareness of the computational nature of many biological processes and that computational and statistical models can be used to great benefit. Second, developments in high-throughput data acquisition produce requirements for computational and statistical sophistication at each stage of the biological research pipeline. The main goal of the Bioconductor project is creation of a durable and flexible software development and deployment environment that meets these new conceptual, computational and inferential challenges. We strive to reduce barriers to entry to research in CBB. A key aim is simplification of the processes by which statistical researchers can explore and interact fruitfully with data resources and algorithms of CBB, and by which working biologists obtain access to and use of state-of-the-art statistical methods for accurate inference in CBB.
Among the many challenges that arise for both statisticians and biologists are tasks of data acquisition, data management, data transformation, data modeling, combining different data sources, making use of evolving machine learning methods, and developing new modeling strategies suitable to CBB. We have emphasized transparency, reproducibility, and efficiency of development in our response to these challenges. Fundamental to all these tasks is the need for software; ideas alone cannot solve the substantial problems that arise.
The primary motivations for an open-source computing environment for statistical genomics are transparency, pursuit of reproducibility and efficiency of development.
Transparency
------------
High-throughput methodologies in CBB are extremely complex, and many steps are involved in the conversion of information from low-level information structures (for example, microarray scan images) to statistical databases of expression measures coupled with design and covariate data. It is not possible to say *a priori*how sensitive the ultimate analyses are to variations or errors in the many steps in the pipeline. Credible work in this domain requires exposure of the entire process.
Pursuit of reproducibility
--------------------------
Experimental protocols in molecular biology are fully published lists of ingredients and algorithms for creating specific substances or processes. Accuracy of an experimental claim can be checked by complete obedience to the protocol. This standard should be adopted for algorithmic work in CBB. Portable source code should accompany each published analysis, coupled with the data on which the analysis is based.
Efficiency of development
-------------------------
By development, we refer not only to the development of the specific computing resource but to the development of computing methods in CBB as a whole. Software and data resources in an open-source environment can be read by interested investigators, and can be modified and extended to achieve new functionalities. Novices can use the open sources as learning materials. This is particularly effective when good documentation protocols are established. The open-source approach thus aids in recruitment and training of future generations of scientists and software developers.
The rest of this article is devoted to describing the computing science methodology underlying Bioconductor. The main sections detail design methods and specific coding and deployment approaches, describe specific unmet challenges and review limitations and future aims. We then consider a number of other open-source projects that provide software solutions for CBB and end with an example of how one might use Bioconductor software to analyze microarray data.
Results and discussion
======================
Methodology
-----------
The software development strategy we have adopted has several precedents. In the mid-1980s Richard Stallman started the Free Software Foundation and the GNU project \[[@B2]\] as an attempt to provide a free and open implementation of the Unix operating system. One of the major motivations for the project was the idea that for researchers in computational sciences \"their creations/discoveries (software) should be available for everyone to test, justify, replicate and work on to boost further scientific innovation\" \[[@B3]\]. Together with the Linux kernel, the GNU/Linux combination sparked the huge open-source movement we know today. Open-source software is no longer viewed with prejudice, it has been adopted by major information technology companies and has changed the way we think about computational sciences. A large body of literature exists on how to manage open-source software projects: see Hill \[[@B4]\] for a good introduction and a comprehensive bibliography.
One of the key success factors of the Linux kernel is its modular design, which allows for independent and parallel development of code \[[@B5]\] in a virtual decentralized network \[[@B3]\]. Developers are not managed within the hierarchy of a company, but are directly responsible for parts of the project and interact directly (where necessary) to build a complex system \[[@B6]\]. Our organization and development model has attempted to follow these principles, as well as those that have evolved from the R project \[[@B7],[@B8]\].
In this section, we review seven topics important to establishment of a scientific open source software project and discuss them from a CBB point of view: language selection, infrastructure resources, design strategies and commitments, distributed development and recruitment of developers, reuse of exogenous resources, publication and licensure of code, and documentation.
### Language selection
CBB poses a wide range of challenges, and any software development project will need to consider which specific aspects it will address. For the Bioconductor project we wanted to focus initially on bioinformatics problems. In particular we were interested in data management and analysis problems associated with DNA microarrays. This orientation necessitated a programming environment that had good numerical capabilities, flexible visualization capabilities, access to databases and a wide range of statistical and mathematical algorithms. Our collective experience with R suggested that its range of well-implemented statistical and visualization tools would decrease development and distribution time for robust software for CBB. We also note that R is gaining widespread usage within the CBB community independently of the Bioconductor Project. Many other bioinformatics projects and researchers have found R to be a good language and toolset with which to work. Examples include the Spot system \[[@B9]\], MAANOVA \[[@B10]\] and dChip \[[@B11]\]. We now briefly enumerate features of the R software environment that are important motivations behind its selection.
#### Prototyping capabilities
R is a high-level interpreted language in which one can easily and quickly prototype new computational methods. These methods may not run quickly in the interpreted implementation, and those that are successful and that get widely used will often need to be re-implemented to run faster. This is often a good compromise; we can explore lots of concepts easily and put more effort into those that are successful.
#### Packaging protocol
The R environment includes a well established system for packaging together related software components and documentation. There is a great deal of support in the language for creating, testing, and distributing software in the form of \'packages\'. Using a package system lets us develop different software modules and distribute them with clear notions of protocol compliance, test-based validation, version identification, and package interdependencies. The packaging system has been adopted by hundreds of developers around the world and lies at the heart of the Comprehensive R Archive Network, where several hundred independent but interoperable packages addressing a wide range of statistical analysis and visualization objectives may be downloaded as open source.
#### Object-oriented programming support
The complexity of problems in CBB is often translated into a need for many different software tools to attack a single problem. Thus, many software packages are used for a single analysis. To secure reliable package interoperability, we have adopted a formal object-oriented programming discipline, as encoded in the \'S4\' system of formal classes and methods \[[@B12]\]. The Bioconductor project was an early adopter of the S4 discipline and was the motivation for a number of improvements (established by John Chambers) in object-oriented programming for R.
#### WWW connectivity
Access to data from on-line sources is an essential part of most CBB projects. R has a well developed and tested set of functions and packages that provide access to different databases and to web resources (via http, for example). There is also a package for dealing with XML \[[@B13]\], available from the Omegahat project, and an early version of a package for a SOAP client \[[@B14]\], SSOAP, also available from the Omegahat project. These are much in line with proposals made by Stein \[[@B15]\] and have aided our work towards creating an environment in which the user perceives tight integration of diverse data, annotation and analysis resources.
#### Statistical simulation and modeling support
Among the statistical and numerical algorithms provided by R are its random number generators and machine learning algorithms. These have been well tested and are known to be reliable. The Bioconductor Project has been able to adapt these to the requirements in CBB with minimal effort. It is also worth noting that a number of innovations and extensions based on work of researchers involved in the Bioconductor project have been flowing back to the authors of these packages.
#### Visualization support
Among the strengths of R are its data and model visualization capabilities. Like many other areas of R these capabilities are still evolving. We have been able to quickly develop plots to render genes at their chromosomal locations, a heatmap function, along with many other graphical tools. There are clear needs to make many of these plots interactive so that users can query them and navigate through them and our future plans involve such developments.
#### Support for concurrent computation
R has also been the basis for pathbreaking research in parallel statistical computing. Packages such as *snow*and *rpvm*simplify the development of portable interpreted code for computing on a Beowulf or similar computational cluster of workstations. These tools provide simple interfaces that allow for high-level experimentation in parallel computation by computing on functions and environments in concurrent R sessions on possibly heterogeneous machines. The *snow*package provides a higher level of abstraction that is independent of the communication technology such as the message-passing interface (MPI) \[[@B16]\] or the parallel virtual machine (PVM) \[[@B17]\]. Parallel random number generation \[[@B18]\], essential when distributing parts of stochastic simulations across a cluster, is managed by *rsprng*. Practical benefits and problems involved with programming parallel processes in R are described more fully in Rossini *et al.*\[[@B19]\] and Li and Rossini \[[@B20]\].
#### Community
Perhaps the most important aspect of using R is its active user and developer communities. This is not a static language. R is undergoing major changes that focus on the changing technological landscape of scientific computing. Exposing biologists to these innovations and simultaneously exposing those involved in statistical computing to the needs of the CBB community has been very fruitful and we hope beneficial to both communities.
### Infrastructure base
We began with the perspective that significant investment in software infrastructure would be necessary at the early stages. The first two years of the Bioconductor project have included significant effort in developing infrastructure in the form of reusable data structures and software/documentation modules (R packages). The focus on reusable software components is in sharp contrast to the one-off approach that is often adopted. In a one-off solution to a bioinformatics problem, code is written to obtain the answer to a given question. The code is not designed to work for variations on that question or to be adaptable for application to distinct questions, and may indeed only work on the specific dataset to which it was originally applied. A researcher who wishes to perform a kindred analysis must typically construct the tools from scratch. In this situation, the scientific standard of reproducibility of research is not met except via laborious reinvention. It is our hope that reuse, refinement and extension will become the primary software-related activities in bioinformatics. When reusable components are distributed on a sound platform, it becomes feasible to demand that a published novel analysis be accompanied by portable and open software tools that perform all the relevant calculations. This will facilitate direct reproducibility, and will increase the efficiency of research by making transparent the means to vary or extend the new computational method.
Two examples of the software infrastructure concepts described here are the `exprSet` class of the *Biobase*package, and the various Bioconductor metadata packages, for example *hgu95av2*. An `exprSet` is a data structure that binds together array-based expression measurements with covariate and administrative data for a collection of microarrays. Based on `R data.frame` and `list` structures, `exprSets` offer much convenience to programmers and analysts for gene filtering, constructing annotation-based subsets, and for other manipulations of microarray results. The exprSet design facilitates a three-tier architecture for providing analysis tools for new microarray platforms: low-level data are bridged to high-level analysis manipulations via the exprSet structure. The designer of low-level processing software can focus on the creation of an exprSet instance, and need not cater for any particular analysis data structure representation. The designer of analysis procedures can ignore low-level structures and processes, and operate directly on the `exprSet` representation. This design is responsible for the ease of interoperation of three key Bioconductor packages: *affy*, *marray*, and *limma*.
The *hgu95av2*package is one of a large collection of related packages that relate manufactured chip components to biological metadata concerning sequence, gene functionality, gene membership in pathways, and physical and administrative information about genes. The package includes a number of conventionally named hashed environments providing high-performance retrieval of metadata based on probe nomenclature, or retrieval of groups of probe names based on metadata specifications. Both types of information (metadata and probe name sets) can be used very fruitfully with `exprSets`: for example, a vector of probe names immediately serves to extract the expression values for the named probes, because the `exprSet` structure inherits the named extraction capacity of `R data.frames`.
### Design strategies and commitments
Well-designed scientific software should reduce data complexity, ease access to modeling tools and support integrated access to diverse data resources at a variety of levels. Software infrastructure can form a basis for both good scientific practice (others should be able to easily replicate experimental results) and for innovation.
The adoption of designing by contract, object-oriented programming, modularization, multiscale executable documentation, and automated resource distribution are some of the basic software engineering strategies employed by the Bioconductor Project.
#### Designing by contract
While we do not employ formal contracting methodologies (for example, Eiffel \[[@B21]\]) in our coding disciplines, the contracting metaphor is still useful in characterizing the approach to the creation of interoperable components in Bioconductor. As an example, consider the problem of facilitating analysis of expression data stored in a relational database, with the constraints that one wants to be able to work with the data as one would with any exprSet and one does not want to copy unneeded records into R at any time. Technically, data access could occur in various ways, using database connections, DCOM \[[@B22]\], communications or CORBA \[[@B23]\], to name but a few. In a designing by contract discipline, the provider of `exprSet` functionality must deliver a specified set of functionalities. Whatever object the provider\'s code returns, it must satisfy the `exprSets` contract. Among other things, this means that the object must respond to the application of functions exprs and `pData` with objects that satisfy the R matrix and data.frame contracts respectively. It follows that `exprs`(*x*) `[i,j]`, for example, will return the number encoding the expression level for the *i*th gene for the *j*th sample in the object *x*, no matter what the underlying representation of *x*. Here *i*and *j*need not denote numerical indices but can hold any vectors suitable for interrogating matrices via the square-bracket operator. Satisfaction of the contract obligations simplifies specification of analysis procedures, which can be written without any concern for the underlying representations for exprSet information.
A basic theme in R development is simplifying the means by which developers can state, follow, and verify satisfaction of design contracts of this sort. Environment features that support convenient inheritance of behaviors between related classes with minimal recoding are at a premium in this discipline.
#### Object-oriented programming
There are various approaches to the object-oriented programming methodology. We have encouraged, but do not require, use of the so-called S4 system of formal classes and methods in Bioconductor software. The S4 object paradigm (defined primarily by Chambers \[[@B12]\] with modifications embodied in R) is similar to that of Common Lisp \[[@B24]\] and Dylan \[[@B25]\]. In this system, classes are defined to have specified structures (in terms of a set of typed \'slots\') and inheritance relationships, and methods are defined both generically (to specify the basic contract and behavior) and specifically (to cater for objects of particular classes). Constraints can be given for objects intended to instantiate a given class, and objects can be checked for validity of contract satisfaction. The S4 system is a basic tool in carrying out the designing by contract discipline, and has proven quite effective.
#### Modularization
The notion that software should be designed as a system of interacting modules is fairly well established. Modularization can occur at various levels of system structure. We strive for modularization at the data structure, R function and R package levels. This means that data structures are designed to possess minimally sufficient content to have a meaningful role in efficient programming. The `exprSet` structure, for example, contains information on expression levels (`exprs` slot), variability (`se.exprs`), covariate data (`phenoData` slot), and several types of metadata (slots `description`, `annotation` and `notes`). The tight binding of covariate data with expression data spares developers the need to track these two types of information separately. The `exprSet` structure explicitly excludes information on gene-related annotation (such as gene symbol or chromosome location) because these are potentially volatile and are not needed in many activities involving `exprSets`. Modularization at the R function level entails that functions are written to do one meaningful task and no more, and that documents (help pages) are available at the function level with worked examples. This simplifies debugging and testing. Modularization at the package level entails that all packages include sufficient functionality and documentation to be used and understood in isolation from most other packages. Exceptions are formally encoded in files distributed with the package.
#### Multiscale and executable documentation
Accurate and thorough documentation is fundamental to effective software development and use, and must be created and maintained in a uniform fashion to have the greatest impact. We inherit from R a powerful system for small-scale documentation and unit testing in the form of the executable example sections in function-oriented manual pages. We have also introduced a new concept of large-scale documentation with the *vignette*concept. Vignettes go beyond typical man page documentation, which generally focuses on documenting the behavior of a function or small group of functions. The purpose of a vignette is to describe in detail the processing steps required to perform a specific task, which generally involves multiple functions and may involve multiple packages. Users of a package have interactive access to all vignettes associated with that package.
The *Sweave*system \[[@B26]\] was adopted for creating and processing vignettes. Once these have been written users can interact with them on different levels. The transformed documents are provided in Adobe\'s portable document format (PDF) and access to the code chunks from within R is available through various functions in the *tools*package. However, new users will need a simpler interface. Our first offering in this area is the vignette explorer `vExplorer` which provides a widget that can be used to navigate the various code chunks. Each chunk is associated with a button and the code is displayed in a window, within the widget. When the user clicks on the button the code is evaluated and the output presented in a second window. Other buttons provide other functionality, such as access to the PDF version of the document. We plan to extend this tool greatly in the coming years and to integrate it closely with research into reproducible research (see \[[@B27]\] for an illustration).
#### Automated software distribution
The modularity commitment imposes a cost on users who are accustomed to integrated \'end-to-end\' environments. Users of Bioconductor need to be familiar with the existence and functionality of a large number of packages. To diminish this cost, we have extended the packaging infrastructure of R/CRAN to better support the deployment and management of packages at the user level. Automatic updating of packages when new versions are available and tools that obtain all package dependencies automatically are among the features provided as part of the reposTools package in Bioconductor. Note that new methods in R package design and distribution include the provision of MD5 checksums with all packages, to help with verification that package contents have not been altered in transit.
In conclusion, these engineering commitments and developments have led to a reasonably harmonious set of tools for CBB. It is worth considering how the S language notion that \'everything is an object\' impacts our approach. We have made use of this notion in our commitment to contracting and object-oriented programming, and in the automated distribution of resources, in which package catalogs and biological metadata are all straightforward R objects. Packages and documents are not yet treatable as R objects, and this leads to complications. We are actively studying methods for simplifying authoring and use of documentation in a multipackage environment with namespaces that allow symbol reuse, and for strengthening the connection between session image and package inventory in use, so that saved R images can be restored exactly to their functional state at session close.
### Distributed development and recruitment of developers
Distributed development is the process by which individuals who are significantly geographically separated produce and extend a software project. This approach has been used by the R project for approximately 10 years. This was necessitated in this case by the fact no institution currently has sufficient numbers of researchers in this area to support a project of this magnitude. Distributed development facilitates the inclusion of a variety of viewpoints and experiences. Contributions from individuals outside the project led to the expansion of the core developer group. Membership in the core depends upon the willingness of the developer to adopt shared objectives and methods and to submerge personal objectives in preference to creation of software for the greater scientific community.
Distributed development requires the use of tools and strategies that allow different programmers to work approximately simultaneously on the same components of the project. Among the more important requirements is for a shared code base (or archive) that all members of the project can access and modify together with some form of version management system. We adopted the Concurrent Versions System \[[@B28],[@B29]\] and created a central archive, within this system, that all members of the team have access to.
Additional discipline is needed to ensure that changes by one programmer should not result in a failure of other code in the system. Within the R language, software components are naturally broken into packages, with a formal protocol for package structure and content specified in the R Extensions manual \[[@B30]\]. Each package should represent a single coherent theme. By using well defined applications programming interfaces (APIs) developers of a package are free to modify their internal structures as long as they continue to provide the documented outputs.
We rely on the testing mechanisms supported by the R package testing system \[[@B30]\] to ensure coherent, non-regressive development. Each developer is responsible for documenting all functions and for providing examples and possibly other scripts or sets of commands that test the code. Each developer is responsible for ensuring that all tests run successfully before committing changes back to the central archive. Thus, the person who knows the code best writes the test programs, but all are responsible for running them and ensuring that changes they have made do not affect the code of others. In some cases changes by one author will necessitate change in the code and tests of others. Under the system we are using these situations are detected and dealt with when they occur in development, reducing the frequency with which error reports come from the field.
Members of the development team communicate via a private mailing list. In many cases they also use private email, telephone and meetings at conferences in order to engage in joint projects and to keep informed about the ideas of other members.
### Reuse of exogenous resources
We now present three arguments in favor of using and adapting software from other projects rather than re-implementing or reinventing functionality. The first argument that we consider is that writing good software is a challenging problem and any re-implementation of existing algorithms should be avoided if possible. Standard tools and paradigms that have been proven and are well understood should be preferred over new untested approaches. All software contains bugs but well used and maintained software tends to contain fewer.
The second argument is that CBB is an enormous field and that progress will require the coordinated efforts of many projects and software developers. Thus, we will require structured paradigms for accessing data and algorithms written in other languages and systems. The more structured and integrated this functionality, the easier it will be to use and hence the more it will be used. As specific examples we consider our recent development of tools for working with graph or network structures. There are three main packages in Bioconductor of interacting with graphs. They are *graph*, *RBGL*and *Rgraphviz*. The first of these provides the class descriptions and basic infrastructure for dealing with graphs in R, the second provides access to algorithms on graphs, and the third to a rich collection of graph layout algorithms. The *graph*package was written from scratch for this project, but the other two are interfaces to rich libraries of software routines that have been created by other software projects, BOOST \[[@B31],[@B32]\] and *Graphviz*\[[@B23]\] respectively, both of which are very substantial projects with large code bases. We have no interest in replicating that work and will, wherever possible, simply access the functions and libraries produced by other projects.
There are many benefits from this approach for us and for the other projects. For bioinformatics and computational biology we gain rapid access to a variety of graph algorithms including graph layout and traversal. The developers in those communities gain a new user base and a new set of problems that they can consider. Gaining a new user base is often very useful, as new users with previously unanticipated needs tend to expose weaknesses in design and implementation that more sophisticated or experienced users are often able to avoid.
In a similar vein, we plan to develop and encourage collaboration with other projects, including those organized through the Open Bioinformatics Foundation and the International Interoperability Consortium. We have not specifically concentrated on collaboration to this point in part because we have chosen areas for development that do not overlap significantly with the tools provided by those projects. In this case our philosophy remains one of developing interfaces to the software provided by those projects and not re-implementing their work. In some cases, other projects have recognized the potential gains for collaboration and have started developing interfaces for us to their systems, with the intent of making future contributions \[[@B33]\].
Another argument in favor of standardization and reuse of existing tools is best made with reference to a specific example. Consider the topic of markup and markup languages. For any specific problem one could quickly devise a markup that is sufficient for that problem. So why then should we adopt a standard such as XML? Among the reasons for this choice is the availability of programmers conversant with the paradigm, and hence lower training costs. A second reason is that the XML community is growing and developing and we will get substantial technological improvements without having to initiate them. This is not unusual. Other areas of computational research are as vibrant as CBB and by coordinating and sharing ideas and innovations we simplify our own tasks while providing stimulus to these other areas.
### Publication and licensing of code
Modern standards of scientific publication involve peer review and subsequent publication in a journal. Software publication is a slightly different process with limited involvement to date of formal peer review or official journal publication. We release software under an open-source license as our main method of publication. We do this in the hope that it will encourage reproducibility, extension and general adherence to the scientific method. This decision also ensures that the code is open to public scrutiny and comment. There are many other reasons for deciding to release software under an open-source license, some of which are listed in Table [1](#T1){ref-type="table"}.
Another consideration that arose when determining the form of publication was the need to allow an evolutionary aspect to our own software. There are many reasons for adopting a strategy that would permit us to extend and improve our software offerings over time. The field of CBB is relatively volatile and as new technologies are developed new software and inferential methods are needed. Further, software technology itself is evolving. Thus, we wanted to have a publication strategy that could accommodate changes in software at a variety of levels. We hope that that strategy will also encourage our users to think of software technology as a dynamic field rather than a static one and to therefore be on the lookout for innovations in this arena as well as in more traditional biological ones.
Our decision to release software in the form of R packages is an important part of this consideration. Packages are easy to distribute, they have version numbers and define an API. A coordinated release of all Bioconductor packages occurs twice every year. At any given time there is a release version of every package and a development version. The only changes allowed to be made on the release version are bug fixes and documentation improvements. This ensures that users will not encounter radical new behaviors in code obtained in the release version. All other changes such as enhancements or design changes are carried out on the development branch \[[@B34]\].
Approximately six weeks before a release, a major effort is taken to ensure that all packages on the development branch are coordinated and work well together. During that period extensive testing is carried out through peer review amongst the Bioconductor core. At release time all packages on the development branch that are included in the release change modes and are now released packages. Previous versions of these packages are deprecated in favor of the newly released versions. Simultaneously, a new development branch is made and the developers start to work on packages in the new branch. Note that these version-related administrative operations occur with little impact on developers. The release manager is responsible for package snapshot and file version modifications. The developers\' source code base is fairly simple, and need not involve retention of multiple copies of any source code files, even though two versions are active at all times.
We would also like to point out that there are compelling arguments that can be made in favor of choosing different paradigms for software development and deployment. We are not attempting at this juncture to convince others to distribute software in this way, but rather elucidating our views and the reasons that we made our choice. Under a different set of conditions, or with different goals, it is entirely likely that we would have chosen a different model.
### Special concerns
We now consider four specific challenges that are raised by research in computational biology and bioinformatics: reproducibility, data evolution and complexity, training users, and responding to user needs.
#### Reproducible research
We would like to address the reproducibility of published work in CBB. Reproducibility is important in its own right, and is the standard for scientific discovery. Reproducibility is an important step in the process of incremental improvement or refinement. In most areas of science researchers continually improve and extend the results of others but for scientific computation this is generally the exception rather than the rule.
Buckheit and Donoho \[[@B35]\], referring to the work and philosophy of Claerbout, state the following principle: \"An article about computational science in a scientific publication is not the scholarship itself, it is merely advertising of the scholarship. The actual scholarship is the complete software development environment and that complete set of instructions that generated the figures.\"
There are substantial benefits that will come from enabling authors to publish not just an advertisement of their work but rather the work itself. A paradigm that fundamentally shifts publication of computational science from an advertisement of scholarship to the scholarship itself will be a welcome addition. Some of the concepts and tools that can be used in this regard are contained in \[[@B36],[@B37]\].
When attempting to re-implement computational methodology from a published description many difficulties are encountered. Schwab *et al.*\[[@B38]\] make the following points:
\"Indeed the problem occurs wherever traditional methods of scientific publication are used to describe computational research. In a traditional article the author merely outlines the relevant computations: the limitations of a paper medium prohibit complete documentation including experimental data, parameter values and the author\'s programs. Consequently, the reader has painfully to re-implement the author\'s work before verifying and utilizing it\.... The reader must spend valuable time merely rediscovering minutiae, which the author was unable to communicate conveniently.\"
The development of a system capable of supporting the convenient creation and distribution of reproducible research in CBB is a massive undertaking. Nevertheless, the Bioconductor project has adopted practices and standards that assist in partial achievement of reproducible CBB.
Publication of the data from which articles are derived is becoming the norm in CBB. This practice provides one of the components needed for reproducible research - access to the data. The other major component that is needed is access to the software and the explicit set of instructions or commands that were used to transform the data to provide the outputs on which the conclusions of the paper rest. In this regard publishing in CBB has been less successful. It is easy to identify major publications in the most prestigious journals that provide sketchy or indecipherable characterizations of computational and inferential processes underlying basic conclusions. This problem could be eliminated if the data housed in public archives were accompanied by portable code and scripts that regenerate the article\'s figures and tables.
The combination of R\'s well-established platform independence with Bioconductor\'s packaging and documentation standards leads to a system in which distribution of data with working code and scripts can achieve most of the requirements of reproducible and replayable research in CBB. The steps leading to the creation of a table or figure can be clearly exposed in an Sweave document. An R user can export the code for modification or replay with variations on parameter settings, to check robustness of the reported calculations or to explore alternative analysis concepts.
Thus we believe that R and Bioconductor can provide a start along the path towards generally reproducible research in CBB. The infrastructure in R that is used to support replayability and remote robustness analysis could be implemented in other languages such as Perl \[[@B39]\] and Python \[[@B40]\]. All that is needed is some platform-independent format for binding together the data, software and scripts defining the analysis, and a document that can be rendered automatically to a conveniently readable account of the analysis steps and their outcomes. If the format is an R package, this package then constitutes a single distributable software element that embodies the computational science being published. This is precisely the compendium concept espoused in \[[@B36]\].
#### Dynamics of biological annotation
Metadata are data about data and their definition depends on the perspective of the investigator. Metadata for one investigator may well be experimental data for another. There are two major challenges that we will consider. First is the evolutionary nature of the metadata. As new experiments are done and as our understanding of the biological processes involved increases the metadata changes and evolves. The second major problem that concerns metadata data is its complexity. We are trying to develop software tools that make it easier for data analysts and researchers to use the existing metadata appropriately.
The constant changing and updating of the metadata suggests that we must have a system or a collection process that ensures that any metadata can be updated and the updates can be distributed. Users of our system will want access to the most recent versions. Our solution has been to place metadata into R packages. These packages are built using a semi-automatic process \[[@B41]\] and are distributed (and updated) using the package distribution tools developed in the *reposTools*package. There is a natural way to apply version numbers so users can determine if their data are up to date or if necessary they can obtain older versions to verify particular analyses. Further, users can synchronize a variety of metadata packages according to a common version of the data sources that they were constructed from.
There are a number of advantages that come from automating the process of building data packages. First, the modules are uniform to an extent that would not be possible if the packages were human written. This means that users of this technology need only become acquainted with one package to be acquainted with all such packages. Second, we can create many packages very quickly. Hence the labor savings are substantial. For microarray analyses all data packages should have the same information (chromosomal location, gene ontology categories, and so on). The only difference between the packages is that each references only the specific set of genes (probes) that were assayed. This means that data analysts can easily switch from one type of chip to another. It also means that we can develop a single set of tools for manipulating the metadata and improvements in those tools are available to all users immediately. Users are free to extend data packages with data from other, potentially proprietary, sources.
Treating the data in the same manner that we treat software has also had many advantages. On the server side we can use the same software distribution tools, indicating updates and improvements with version numbering. On the client side, the user does not need to learn about the storage or internal details of the data packages. They simply install them like other packages and then use them.
One issue that often arises is whether one should simply rely on online sources for metadata. That is, given an identifier, the user can potentially obtain more up-to-date information by querying the appropriate databases. The data packages we are proposing cannot be as current. There are, however, some disadvantages to the approach of accessing all resources online. First, users are not always online, they are not always aware of all applicable information sources and the investment in person-time to obtain such information can be high. There are also issues of reproducibility that are intractable as the owners of the web resources are free to update and modify their offerings at will. Some, but not all, of these difficulties can be alleviated if the data are available in a web services format.
Another argument that can be made in favor of our approach, in this context, is that it allows the person constructing the data packages to amalgamate disparate information from a number of sources. In building metadata packages for Bioconductor, we find that some data are available from different sources, and under those circumstances we look for consensus, if possible. The process is quite sophisticated and is detailed in the *AnnBuilder*package and paper \[[@B41]\].
#### Training
Most of the projects in CBB require a combination of skills from biology, computer science, and statistics. Because the field is new and there has been little specialized training in this area it seems that there is some substantial benefit to be had from paying attention to training. From the perspective of the Bioconductor project, many of our potential users are unfamiliar with the R language and generally are scientifically more aligned with one discipline than all three. It is therefore important that we produce documentation for the software modules that is accessible to all. We have taken a two-pronged approach to this, we have developed substantial amounts of course material aimed at all the constituent disciplines and we have developed a system for interactive use of software and documentation in the form of vignettes and more generally in the form of navigable documents with dynamic content.
Course materials have been developed and refined over the past two to three years. Several members of the Bioconductor development team have taught courses and subsequently refined the material, based on success and feedback. The materials developed are modular and are freely distributed, although restrictions on publication are made. The focus of the materials is the introduction and use of software developed as part of the Bioconductor project, but that is not a requirement and merely reflects our own specific purposes and goals.
In this area we feel that we would benefit greatly from contributions from those with more experience in technical document authoring. There are likely to be strategies, concepts and methodologies that are standard practice in that domain that we are largely unaware of. However, in the short term, we rely on the students, our colleagues and the users of the Bioconductor system to guide us and we hope that many will contribute. Others can easily make substantial contributions, even those with little or no programming skills. What is required is domain knowledge in one field of interest and the recognition of a problem that requires additional domain knowledge from another of the fields of interest.
Our experience has been that many of these new users often transform themselves into developers. Thus, our development of training materials and documentation needs to pay some attention to the needs of this group as well. There are many more software components than we can collectively produce. Attracting others to collaboratively write software is essential to success.
#### Responding to user needs
The success of any software project rests on its ability to both provide solutions to the problems it is addressing and to attract a user community. Perhaps the most effective way of addressing user needs is through an e-mail help list and one was set up as soon as the project became active. In addition it is important to keep a searchable archive available so that the system itself has a memory and new users can be referred there for answers to common questions. It is also important that members of the project deal with bug reports and feature requests through this public forum as it both broadcasts their intentions and provides a public record of the discussion. Our mailing list (mailto:<bioconductor@stat.math.ethz.ch>) has been successful: there are approximately 800 subscribers and about 3,000 email messages per year.
Attracting a user community itself requires a method of distributing the software and providing sufficient training materials to allow potential users to explore the system and determine whether it is sufficient for their purposes. An alternate approach would be to develop a graphical user interface (GUI) that made interactions with the system sufficiently self-explanatory that documentation was not needed. We note that this solution is generally more applicable to cases where the underlying software tasks are well defined and well known. In the present case, the software requirements (as well as the statistical and biological requirements) are constantly evolving. R is primarily command-line oriented and we have chosen to follow that paradigm at least for the first few years of development. We would of course welcome and collaborate with those whose goal was in GUI development but our own forays into this area are limited to the production of a handful of widgets that promote user interaction at specific points.
Users have experienced difficulties downloading and installing both R and the Bioconductor modules. Some of these difficulties have been caused by the users\' local environments (firewalls and a lack of direct access to the internet), and some by problems with our software (bugs) which arise in part because it is in general very difficult to adequately test software that interacts over the internet. We have, however, managed to help every user, who was willing to persist, get both R and Bioconductor properly installed. Another substantial difficulty that we had to overcome was to develop a system that allowed users to download not just the software package that they knew they wanted, but additionally, and at the same time, all other software packages that it relies on. With Bioconductor software there is a much larger inter-reliance on software packages (including those that provide machine learning, biological metadata and experimental data) than for most other uses of R and the R package system. The package, reposTools contains much of the necessary infrastructure for handling these tasks. It is a set of functions for dealing with R package repositories which are basically internet locations for collections of R packages.
Once the basic software is installed, users will need access to documentation such as the training materials described above and other materials such as the vignettes, described in a previous section. Such materials are most valuable if the user can easily obtain and run the examples on their own computer. We note the obvious similarity with this problem and that described in the section on reproducible research. Again, we are in the enjoyable situation of having a paradigm and tools that can serve two purposes.
Other open-source bioinformatics software projects
--------------------------------------------------
The Open Bioinformatics Foundation supports projects similar to Bioconductor that are nominally rooted in specific programming languages. BioPerl \[[@B42]\], BioPython \[[@B43]\] and BioJava \[[@B44]\] are prominent examples of open-source language-based bioinformatics projects. The intentions and design methodologies of the BioPerl project have been lucidly described by Stajich and colleagues \[[@B45]\].
### BioPerl
In this section we consider commonalities and differences between BioPerl and Bioconductor. Both projects have commitments to open source distribution and to community-based development, with an identified core of developers performing primary design and maintenance tasks for the project. Both projects use object-oriented programming methodology, with the intention of abstracting key structural and functional features of computational workflows in bioinformatics and defining stable application programming interfaces (API) that hide implementation details from those who do not need to know them. The toolkits are based on highly portable programming languages. These languages have extensive software resources developed for non-bioinformatic purposes. The repositories for R (Comprehensive R Archive Network, CRAN) and Perl (Comprehensive Perl Archive Network, CPAN) provide mirrored WWW access to structured collections of software modules and documents for a wide variety of workflow elements. Development methodologies targeted at software reuse can realize large gains in productivity by establishing interfaces to existing CPAN or CRAN procedures instead of reimplementing such procedures. For reuse to succeed, the maintainer of the external resource must commit to stability of the resource API. Such stability tends to be the norm for widely-used modules. Finally, both languages have considerable interoperability infrastructure. One implication is that each project can use software written in unrelated languages. R has well-established interfaces to Perl, Python, Java and C. R\'s API allows software in R to be called from other languages, and the *RSPerl*package \[[@B46]\] facilitates direct calls to R from Perl. Thus there are many opportunities for symbiotic use of code by Bioconductor and BioPerl developers and users. The following script illustrates the use of BioPerl in R.
\> library(RSPerl)
\> .PerlPackage(\"Bio::Perl\")
\> x \<- .Perl(\"get\_sequence\", \"swiss\",
\"ROA1\_HUMAN\")
\> x\$division()
\[1\] \"HUMAN\"
\> x\$accession()
\[1\] \"P09651\"
\> unlist(x\$get\_keywords())
\[1\] \"Nuclear protein\" \"RNA-binding\"
\[3\] \"Repeat\" \"Ribonucleoprotein\"
\[5\] \"Methylation\" \"Transport\"
\...
The .PerlPackage command brings the BioPerl modules into scope. .Perl invokes the BioPerl get\_sequence subroutine with arguments \"swiss\" and \"ROA1\_HUMAN\". The resulting R object is a reference to a perl hash. RSPerl infrastructure permits interrogation of the hash via the \$ operator. Note that *RSPerl*is not a Bioconductor-supported utility, and that installation of the BioPerl and *RSPerl*resources to allow interoperation can be complicated.
Key differences between the Bioconductor and BioPerl projects concern scope, approaches to distribution, documentation and testing, and important details of object-oriented design.
#### Scope
BioPerl is clearly slanted towards processing of sequence data and interfacing to sequence databases, with support for sequence visualization and queries for external annotation. Bioconductor is slanted towards statistical analysis of microarray experiments, with major concerns for array preprocessing, quality control, within- and between-array normalization, binding of covariate and design data to expression data, and downstream inference on biological and clinical questions. Bioconductor has packages devoted to diverse microarray manufacturing and analysis paradigms and to other high-throughput assays of interest in computational biology, including serial analysis of gene expression (SAGE), array comparative genomic hybridization (arrayCGH), and proteomic time-of-flight (SELDI-TOF) data. We say the projects are \'slanted\' towards these concerns because it is clear that both projects ultimately aim to support general research activities in computational biology.
#### Distribution, documentation and testing
BioPerl inherits the distribution paradigm supported by CPAN. Software modules can be acquired and installed interactively using, for example `perl -MCPAN -e shell`. This process supports automated retrieval of requested packages and dependencies, but is not triggered by runtime events. Bioconductor has extended the CRAN distribution functionalities so that packages can be obtained and installed \'just in time\', as required by a computational request. For both Perl and R, software modules and packages are structured collections of files, some of which are source code, some of which are documents about the code. The relationship between documentation and testing is somewhat tighter in Bioconductor than in BioPerl. Manual pages and vignettes in Bioconductor include executable code. Failure of the code in a man page or vignette is a quality-control event; experimentation with executable code in manual pages (through the example function of R) is useful for learning about software behavior. In Perl, tests occupy separate programs and are not typically integrated with documentation.
#### Details of object-oriented procedure
Both R and Perl are extensible computer languages. Thus it is possible to introduce software infrastructure supporting different approaches to object-oriented programming (OOP) in various ways in both languages.
R\'s core developers have provided two distinct approaches to OOP in R. These approaches are named S3 and S4. In S3, any object can be assigned to a class (or sequence of classes) simply by setting the class name as the value of the object\'s class attribute. Class hierarchies are defined implicitly at the object level. Generic methods are defined as ordinary functions and class-specific methods are dispatched according to the class of the object being passed as an argument. In S4, formal definition of class structure is supported, and class hierarchy is explicitly defined in class definitions \[[@B12]\]. Class instances are explicitly constructed and subject to validation at time of construction. Generic methods are non-standard R functions and metadata on generic methods is established at the package level. Specific methods are dispatched according to the class signature of the argument list (multiple dispatch). Overall, the OOP approach embodied in S4 is closer to Dylan or Scheme than to C++ or Java. Bioconductor does not require specific OOP methodology but encourages the use of S4, and core members have contributed special tools for the documentation and testing of S4 OOP methods in R.
OOP methodology in Perl has a substantial history and is extensively employed in BioPerl. The basic approach to OOP in Perl seems to resemble S3 more than S4, in that Perl\'s bless operation can associate any perl data instance with any class. The CPAN `Class::Multimethod` module can be used to allow multiple dispatch behavior of generic subroutines. The specific classes of objects identified in BioPerl are targeted at sequence data (Seq, LocatableSeq, RelSegment are examples), location data (Simple, Split, Fuzzy), and an important class of objects called interface objects, which are classes whose names end in \'I\'. These objects define what methods can be called on objects of specified classes, but do not implement any methods.
### BioJava, BioPython, GMOD and MOBY
Other open bioinformatics projects have intentions and methods that are closely linked with those of Bioconductor.
BioJava \[[@B44]\] provides Dazzle, a servlet framework supporting the Distributed Annotation System specification for sharing sequence data and metadata. Version 1.4 of the BioJava release includes java classes for general alphabets and symbol-list processing, tools for parsing outputs of blast-related analyses, and software for constructing and fitting hidden Markov models. In principle, any of these resources could be used for analysis in Bioconductor/R through the *SJava*interface \[[@B46]\].
BioPython \[[@B43]\] provides software for constructing python objects by parsing output of various alignment or clustering algorithms, and for a variety of downstream tasks including classification. BioPython also provides infrastructure for decomposition of parallelizable tasks into separable processes for computation on a cluster of workstations.
The Generic Model Organism Database (GMOD) project targets construction of reusable components that can be used to reproduce successful creation of open and widely accessible databases of model organisms (for example, worm, fruitfly and yeast). The main tasks addressed are genome visualization and annotation, literature curation, biological ontology activities, gene expression analysis and pathway visualization and annotation.
BioMOBY \[[@B47]\] provides a framework for developing and cataloging web services relevant to molecular biology and genomics. A basic aim is to provide a central registry of data, annotation or analysis services that can be used programmatically to publish and make use of data and annotation resources pertinent to a wide variety of biological contexts.
As these diverse projects mature, particularly with regard to interoperability, we expect to add infrastructure to Bioconductor to simplify the use of these resources in the context of statistical data analysis. It is our hope that the R and Bioconductor commitments to interoperability make it feasible for developers in other languages to reuse statistical and visualization software already present and tested in R.
Using Bioconductor (example)
----------------------------
Results of the Bioconductor project include an extensive repository of software tools, documentation, short course materials, and biological annotation data at \[[@B1]\]. We describe the use of the software and annotation data by description of a concrete analysis of a microarray archive derived from a leukemia study.
Acute lymphocytic leukemia (ALL) is a common and difficult-to-treat malignancy with substantial variability in therapeutic outcomes. Some ALL patients have clearly characterized chromosomal aberrations and the functional consequences of these aberrations are not fully understood. Bioconductor tools were used to develop a new characterization of the contrast in gene expression between ALL patients with two specific forms of chromosomal translocation. The most important tasks accomplished with Bioconductor employed simple-to-use tools for state-of-the-art normalization of hundreds of microarrays, clear schematization of normalized expression data bound to detailed covariate data, flexible approaches to gene and sample filtering to support drilling down to manageable and interpretable subsets, flexible visualization technologies for exploration and communication of genomic findings, and programmatic connection between expression platform metadata and biological annotation data supporting convenient functional interpretation. We will illustrate these through a transcript of the actual command/output sequence. More detailed versions of some of the processing and analysis activities sketched here can be found in the vignettes from the *GOstats*package.
The dataset is from the Ritz laboratory at the Dana Farber Cancer Institute \[[@B48]\]. It contains data from 128 patients with ALL. Two subgroups are to be compared. The first group consists of patients with a translocation between chromosomes 4 and 11 (labeled ALL1/AF4). The second group consists of patients with a translocation between chromosomes 9 and 22 (labeled BCR/ABL). These conditions are mutually exclusive in this dataset.
The Affymetrix HGu95Av2 platform was used, and expression measures were normalized using *gcrma*from the *affy*package. The output of this is an object of class *exprSet*which can be used as input for other functions. The package *hgu95av2*provides biological metadata including mappings from the Affymetrix identifiers to GO, chromosomal location, and so on. These data can, of course be obtained from many other sources, but there are some advantages to having them as an R package.
After loading the appropriate packages we first subset the ALL `exprSet` to extract those samples with the covariates of interest. The design of the `exprSet` class includes methods for subsetting both cases and probes. By using the square-bracket notation on ALL, we derive a new `exprSet` with data on only the desired patients.
\> data(\"ALL\")
\> eset \<- ALL\[, ALL\$mol %in%
c(\"BCR/ABL\", \"ALL1/AF4\")\]
Next we find genes which are differentially expressed between the ALL1/AF4 and BCR/ABL groups. We use the function `lmFit` from the *limma*package, which can assess differential expression between many different groups and conditions simultaneously. The function `lmFit` accepts a model matrix which describes the experimental design and produces an output object of class `MArrayLM` which stores the fitted model information for each gene. The fitted model object is further processed by the `eBayes` function to produce empirical Bayes test statistics for each gene, including moderated *t*-statistics, *p*-values and log-odds of differential expression. The log~2~-fold changes, average intensites and Holm-adjusted *p*-values are displayed for the top 10 genes (Figure [1](#F1){ref-type="fig"}).
We select those genes that have adjusted *p*-values below 0.05. The default method of adjusting for multiple comparisons uses Holm\'s method to control the family-wise error rate. We could use a less conservative method such as the false discovery rate, and the multtest package offers other possibilities, but for this example we will use the very stringent Holm method to select a small number of genes.
\> selected \<- p.adjust(fit\$p.value\[, 2\])
\< 0.05
\> esetSel \<- eset \[selected, \]
There are 165 genes selected for further analysis. A heat map produced by the heatmap function from R allows us to visualize the differential action of these genes between the two groups of patients. Note how the different software modules can be integrated to provide a very rich data-analysis environment. Figure [2](#F2){ref-type="fig"} shows clearly that these two groups can be distinguished in terms of gene expression.
We can carry out many other tests, for example, whether genes encoded on a particular chromosome (or perhaps on a specific strand of a chromosome) are over-represented amongst those selected by moderated *t*-test. Many of these questions are normally addressed in terms of a hypergeometric distribution, but they can also be thought of as two-way or multi-way tables, and alternate statistical tests (all readily available in R) can be applied to the resulting data.
We turn our attention briefly to the use of the Gene Ontology (GO) annotation in conjunction with these data. We first identify the set of unique LocusLink identifiers among our selected Affymetrix probes. The function `GOHyperG` is found in the *GOstats*package. It carries out a hypergeometric test for an overabundance of genes in our selected list of genes for each term in the GO graph that is induced by these genes (Figure [3](#F3){ref-type="fig"}).
The smallest *p*-value found was 1.1e-8 and it corresponds to the term, \"MHC class II receptor activity\". We see that six of the 12 genes with this GO annotation have been selected. Had we used a slightly less conservative gene selection method then the number of selected genes in this GO annotation would have been even higher.
Reproducing the above results for any other species or chip for which an annotation package was available would require almost no changes to the code. The analyst need only substitute the references to the data package, *hgu95av2*, with those for their array and the basic principles and code are unchanged.
Similarly, substitution of other algorithms or statistical tests is possible as the data analyst has access to the full and complete source code. All tools are modifiable at the source level to suit local requirements.
Conclusions
===========
We have detailed the approach to software development taken by the Bioconductor project. Bioconductor has been operational for about three years now and in that time it has become a prominent software project for CBB. We argue that the success of the project is due to many factors. These include the choice of R as the main development language, the adoption of standard practices of software design and a belief that the creation of software infrastructure is an important and essential component of a successful project of this size.
The group dynamic has also been an important factor in the success of Bioconductor. A willingness to work together, to see that cooperation and coordination in software development yields substantial benefits for the developers and the users and encouraging others to join and contribute to the project are also major factors in our success.
To date the project provides the following resources: an online repository for obtaining software, data and metadata, papers, and training materials; a development team that coordinates the discussion of software strategies and development; a user community that provides software testing, suggested improvements and self-help; more than 80 software packages, hundreds of metadata packages and a number of experimental data packages.
At this point it is worth considering the future. While many of the packages we have developed have been aimed at particular problems, there have been others that were designed to support future developments. And that future seems very interesting. Many of the new problems we are encountering in CBB are not easily addressed by technology transfer, but rather require new statistical methods and software tools. We hope that we can encourage more statisticians to become involved in this area of research and to orient themselves and their research to the mixture of methodology and software development that is necessary in this field.
In conclusion we would like to note that the Bioconductor Project has many developers, not all of whom are authors of this paper, and all have their own objectives and goals. The views presented here are not intended to be comprehensive nor prescriptive but rather to present our collective experiences and the authors\' shared goals. In a very simplified version these can be summarized in the view that coordinated cooperative software development is the appropriate mechanism for fostering good research in CBB.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Limma analysis of the ALL data. The leftmost numbers are row indices, ID is the Affymetrix HGU95av2 accession number, M is the log ratio of expression, A is the log average expression, and B is the log odds of differential expression.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Heat map (produced by the Bioconductor function heatmap()) of the ALL leukemia data.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Hypergeometric analysis of molecular function enrichment of genes selected in the analysis described in Figure 1.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Reasons for deciding to release software under an open-source license
:::
--------------------------------------------------------------------------------------------------------------------------------
To encourage reproducibility, extension and general adherence to the scientific method
To ensure that the code is open to public scrutiny and comment
To provide full access to algorithms and their implementation
To provide to users the ability to fix bugs without waiting for the developer, and to extend and improve the supplied software
To encourage good scientific computing and statistical practice by exhibiting fully appropriate tools and instruction
To provide a workbench of tools that allow researchers to explore and expand the methods used to analyze biological data
To ensure that the international scientific community is the owner of the software tools needed to carry out research
To lead and encourage commercial support and development of those tools that are successful
To promote reproducible research by providing open and accessible tools with which to carry out that research
--------------------------------------------------------------------------------------------------------------------------------
:::
|
PubMed Central
|
2024-06-05T03:55:51.765575
|
2004-9-15
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545600/",
"journal": "Genome Biol. 2004 Sep 15; 5(10):R80",
"authors": [
{
"first": "Robert C",
"last": "Gentleman"
},
{
"first": "Vincent J",
"last": "Carey"
},
{
"first": "Douglas M",
"last": "Bates"
},
{
"first": "Ben",
"last": "Bolstad"
},
{
"first": "Marcel",
"last": "Dettling"
},
{
"first": "Sandrine",
"last": "Dudoit"
},
{
"first": "Byron",
"last": "Ellis"
},
{
"first": "Laurent",
"last": "Gautier"
},
{
"first": "Yongchao",
"last": "Ge"
},
{
"first": "Jeff",
"last": "Gentry"
},
{
"first": "Kurt",
"last": "Hornik"
},
{
"first": "Torsten",
"last": "Hothorn"
},
{
"first": "Wolfgang",
"last": "Huber"
},
{
"first": "Stefano",
"last": "Iacus"
},
{
"first": "Rafael",
"last": "Irizarry"
},
{
"first": "Friedrich",
"last": "Leisch"
},
{
"first": "Cheng",
"last": "Li"
},
{
"first": "Martin",
"last": "Maechler"
},
{
"first": "Anthony J",
"last": "Rossini"
},
{
"first": "Gunther",
"last": "Sawitzki"
},
{
"first": "Colin",
"last": "Smith"
},
{
"first": "Gordon",
"last": "Smyth"
},
{
"first": "Luke",
"last": "Tierney"
},
{
"first": "Jean YH",
"last": "Yang"
},
{
"first": "Jianhua",
"last": "Zhang"
}
]
}
|
PMC545601
|
Background
==========
Cellular proteins differ widely in their lability, ranging from those that are completely stable to those with half-lives measured in minutes. Proteins with a short half-life are among the most critical to the cell. Regulated degradation of specific proteins contributes to the control of signal transduction pathways, cell-cycle control, transcription, apoptosis, antigen processing, biological clock control, differentiation and surface receptor desensitization \[[@B1],[@B2]\]. Rapid turnover makes it possible for the cellular level of a protein to change promptly when synthesis is increased or reduced \[[@B3]\]. Furthermore, degradation rate is itself subject to regulation. For instance, inflammatory stimuli cause the rapid degradation of IκBα, the inhibitor of NFκB, resulting in the activation of that transcription factor \[[@B4]-[@B6]\].
Analysis of labile proteins has been time-consuming and labor-intensive. The most definitive form of analysis requires pulse-chase labeling cells and immunoprecipitation extracts. *In vitro*assay of degradation is simpler than *in vivo*analysis, but an *in vitro*assay system may not fully mimic the degradation of proteins in the cells. Genome-wide functional screening and systemic characterization of cellular short-lived proteins has received little attention \[[@B7]\]. GFP, the green fluorescent protein from the jellyfish *Aequorea victoria*, has been widely used to monitor gene expression and protein localization \[[@B8]\]. Recently, we demonstrated that fusion of GFP to the degradation domain of ornithine decarboxylase \[[@B9]\], a labile protein, can destabilize GFP \[[@B10]\] and that the degradation of an IκB-GFP fusion protein can be monitored by GFP fluorescence \[[@B11]\]. These studies demonstrate that introducing GFP as a fusion within the context of a rapidly degraded protein does not alter the degradation properties of the parent molecule, and that the GFP moiety of the fusion protein is degraded along with the rest of the protein. GFP fluorescence, which provides a sensitive, rapid, precise and non-destructive assay of protein abundance, can therefore be used to monitor protein degradation \[[@B12]\]. Furthermore, fluorescence associated with single cells can be analyzed using fluorescence-activated cell sorting (FACS), a technology easily adapted to high-throughput screening \[[@B13]\].
We developed a GFP-based, genome-wide screening method for short-lived proteins. We made a GFP fusion expression library of human cDNAs and introduced the library into mammalian cells. Transfected cells were FACS-fractionated into subpopulations of uniform fluorescence. Individual subpopulations were treated with cycloheximide (CHX) to inhibit protein synthesis and re-sorted after 2 hours of treatment. Sorting was gated to recover cells with a fluorescent signal that was diminished compared to the population mode. Repeated application of this process resulted in a high yield of clones that encode labile fusion proteins.
Results
=======
The selection scheme is shown in Figure [1](#F1){ref-type="fig"}. GFP-cDNA expression libraries were transfected into mammalian cells and cells fractionated into subpopulations, each with a narrow range of fluorescence intensities. Subpopulations were then twice enriched for cells with the desired characteristics. Plasmid DNAs were recovered from the selected cells, subjected to sequence analysis and functionally verified. We made the expression libraries with modified pEGFP C1/C2/C3 vectors by cloning the cDNAs downstream of EGFP. The titer of the library was found to be high: around 10^6^cell transformants per microgram of DNA. In addition, we confirmed by PCR amplification that 95% of clones contained a cDNA insert larger than 800 base-pairs (bp) (data not shown). The libraries were thus deemed to be useful for screening short-lived proteins in mammalian cells. We used 293T cells as the recipient. These cells offer two advantages. First, they express the SV40 large T antigen. This allows the library plasmids, which contain an SV40 origin of replication, to be highly replicated. Plasmids can therefore be recovered easily. Second, 293T cells have high transfection efficiency.
After we introduced the GFP-fusion libraries into the mammalian cells, the transfected cells were easily separated by FACS from non-transfected cells or cells transformed by non-productive constructs. We imposed selection for cells that became less bright within 2 hours of exposure to cycloheximide (CHX), a protein synthesis inhibitor. We chose a short treatment time to avoid selecting cells that became dimmer as a result of secondary responses other than rapid turnover of the GFP tagged proteins. To enrich for cells that are susceptible to CHX treatment, we started with a cell population that has an approximately log-normal fluorescence histogram distribution, with a working range of 1.5 to 4.5 logs. We used FACS fractionation to divide this population into five subpopulations (R2, R3, R4, R5, R6) of ascending brightness, gating each on successive one-half log~10~intervals of fluorescence (Figure [2](#F2){ref-type="fig"}). Each subpopulation (R2-R6) was divided into two; one portion was treated with 100 μg/ml CHX for 2 hours and the other left untreated. Subpopulations were then reanalyzed to determine whether they had retained a distribution consistent with the gating criteria used to obtain this narrow subpopulation and were susceptible to CHX treatment. We found that subpopulations R3 and R4 were susceptible to CHX treatment (Figure [3](#F3){ref-type="fig"}), whereas R5 and R6 did not change their fluorescence properties in response to CHX (data not shown). The fluorescence intensity of R2 was too low to detect after CHX treatment. The lack of susceptibility of the brighter R5-R6 subpopulations was most likely the result of their expressing predominantly stable proteins, which would be expected to provide more intense fluorescence.
We selected R4 for further screening in this study. We collected 10^6^cells from the shifted population, the left shoulder of the population observed in the CHX-treated but not in the untreated R4 cells (Figure [3](#F3){ref-type="fig"}). Plasmid DNAs were recovered from the sorted cells and were propagated in *Escherichia coli*, resulting in a total of 400 clones. The individual clones were stored in 15% glycerol LB medium in a 96-well format.
To perform second-round selection, we grouped the 400 clones into 12 pools, each composed of approximately 33 clones. The individual pools of clones were cultured and used for plasmid preparation. We transfected these 12 groups of plasmid DNA into 293T cells and again subjected them to FACS analysis and gating as before. The EGFP-C1 vector was used as a control. Because enhanced green fluorescent protein (EGFP) is a stable protein, its fluorescence intensity would not be changed by treatment with CHX. We found that eight of the 12 groups showed a decrease of the fluorescence intensity peak by 30-50% (compared to untreated cells) after 2 hours of CHX treatment. In four out of 12 groups, no change in fluorescence intensity was detected.
To isolate individual clones with the desired property, we randomly chose one of the eight CHX-responsive groups and characterized individual clones. We analyzed 30 clones from this group by individually transfecting them into 293T cells and determining the half-life by FACS-based analysis of CHX chase kinetics. We found out that 22 clones showed a decrease in fluorescence intensity ranging from 30 to 90% after treatment with CHX for 2 hours. Assuming first order kinetics of turnover, this single-time-point experiment implies that the proteins corresponding to these 22 clones have a range of half-lives ranging from about half an hour to 3-4 hours (Table [1](#T1){ref-type="table"}). The 22 clones were partially sequenced and BLAST used to search for similar protein sequences in the National Center for Biotechnology Information (NCBI) public database. Of these, 19 corresponded to annotated genes in GenBank and the remaining three to unknown genes. Sequencing analysis also indicated that the inserts of these clones corresponded to full-length or near full-length translation reading frames.
As no data are available on the intracellular turnover kinetics of the 19 identifiable proteins, we picked three clones - splicing factor SRp30c, a guanine nucleotide-binding regulatory protein (G protein), and cervical cancer 1 proto-oncogene protein - and examined their turnover by CHX chase and western blot analysis. These three clones (Table [1](#T1){ref-type="table"}, numbers 5, 19 and 26) were estimated in the fluorescence-based screen to have diverse turnover kinetics; two of them have a half-life of less than 1 hour while the third turns over somewhat more slowly. To confirm these estimates of turnover by a means independent of GFP fluorescence, 293T cells were transfected with these clones, treated with CHX and periodically sampled over the next 3 hours. Western blot analysis of cell extracts with antibody to GFP showed that the abundance of all three fusion proteins diminished in the presence of CHX (Figure [4a](#F4){ref-type="fig"}). The half-life of the proteins determined by western blot analysis was similar to that determined by FACS analysis. Two of the proteins showed a half-life of about 1 hour, while the proto-oncogene protein appears to initiate abrupt degradation within about 2 hours of treatment with CHX. The results for all three proteins are thus consistent with those observed using the fluorescence-based screening method. As positive and negative controls, we similarly analyzed cells expressing a destabilized version of EGFP, d1EGFP, whose short half-life has been previously characterized \[[@B10]\], and a stable EGFP protein (Figure [4b](#F4){ref-type="fig"}).
Sequencing analysis indicated that these three GFP fusion cDNAs do not contain a full-length coding sequence. SRp30c cDNA is missing 17 amino acids at its amino terminus, G protein 20 amino acids, and proto-oncogene p40 three amino acids. To exclude the possibility that the missing amino acids or the fused GFP domain contribute artifactually to protein liability, we amplified the full-length coding sequences of these three genes and expressed them as Myc fusion proteins. Their turnover was examined by CHX chase and western blot analysis with antibody to the Myc tag (Figure [5](#F5){ref-type="fig"}). Turnover rates assessed in this way were similar to those of the GFP fusion proteins obtained from library screening, ruling out the presence of these artifacts.
This technology is subject to two kinds of false-positive results. First, fusion to a detection tag such as GFP or Myc may affect the folding of tagged proteins, which could accelerate their turnover. Second, expression of the fusion proteins under the control of viral promoter elements could result in overexpression, with concomitant misfolding or failure to associate with endogenous interaction partners. To rule out these artifacts, we measured the degradation of native non-fusion endogenous counterparts of two of the proteins we identified, those for which antibodies were available. Turnover of the proteins associated with clone 19 and clone 25 was measured by CHX chase and western blot analysis. The results (Figure [6](#F6){ref-type="fig"}) demonstrated that the half-life of clone 19, a guanine nucleotide-binding regulatory protein (G protein), was less than 1 hour and the half-life of clone 25, heat-shock 70 kD protein (hsp70), was about 1 hour. The turnover of the native proteins is thus at least as fast as that of the corresponding clones analyzed in the screen, suggesting that the technology can accurately identify short-lived proteins.
Discussion
==========
The abundance of a given cellular protein is determined by the balance between its rate of synthesis and degradation. The two are of equal importance in their effect on the steady-state level. Furthermore, degradation determines the rate at which a new steady state is reached when protein synthesis changes \[[@B3]\]. Despite its importance, degradation, the \'missing dimension\' in proteomics \[[@B7]\], has received far less comprehensive attention than synthesis. This deficiency has arisen because developing the tools for a proteome-wide study of protein turnover is technically challenging. Proteins that are labile tend to be present at low abundance, and methods for characterizing turnover time are laborious.
We have developed an efficient and rather specific screen by combining GFP fluorescence, as a high-throughput measure of protein abundance, with pharmacologic shutoff of protein synthesis. Of 30 clones that were recovered from the screen (Figure [1](#F1){ref-type="fig"}) and individually examined by CHX treatment and FACS analysis, 22 (73%) are associated with proteins with a half-life of less than 4 hours. Given the relative rarity of rapidly degraded proteins in the proteome \[[@B14]\], this result demonstrates the specificity of the screening method. We have so far analyzed a restricted subset of the clones that were recovered in our screening procedure - 30 clones present in one of eight positive pools (among 12) from the R4 population. A second population, R3, appears to be equally rich in clones responsive to CHX. Extrapolation from this small sample implies that perhaps 300-400 (that is, 22 × 8 × 2) clones within the GFP-cDNA library may be found to be associated with proteins that are labile according to our secondary screening criterion. In contrast to the results with the less bright R3 and R4 cell populations, the failure to detect a CHX-sensitive subpopulation among the brighter R5-R6 cells is consistent with the expectation that labile proteins tend to be of lower abundance than more stable proteins.
For some of the proteins uncovered in this survey, rapid turnover can be rationalized as intrinsic to their cellular function. SRp30c factor (accession number U87279) is responsible for pre-mRNA splicing. Alterative splicing is a commonly used mechanism to create protein isoforms. It has been proposed that organisms regulate alternative splice site selection by changing the concentration and activity of splicing regulatory proteins such as SRp30c in response to external stimuli \[[@B15]\]. The finding that SRp30c is a short-lived protein is consistent with its postulated regulatory function.
The G proteins are a ubiquitous family of proteins that transduce information across the plasma membrane, coupling receptors to various effectors \[[@B16],[@B17]\]. About 80% of all known hormones, neurotransmitters and neuromodulators are estimated to exert their cellular regulation through G proteins. The G protein (accession number M69013) shown here to short-lived is a G protein α subunit that transduces signals via a pertussis toxin-insensitive mechanism \[[@B18]\]. Like other pertussis toxin-insensitive G proteins such as the Ga12 class, it causes the activation of several cytoplasmic protein tyrosine kinases: Src, Pyk2 (proline-rich tyrosine kinase 2) and Fak (focal adhesion kinase) \[[@B19]\]. However, it is not known how this G protein is regulated. Its rapid turnover suggests a testable mechanism of its regulatory activation. Cervical cancer 1 proto-oncogene protein p40 (accession number AF195651), is a third protein shown here to turn over rapidly, but its function is unknown. Further studies of its turnover may provide important information on its function and regulation.
In mammalian cells, proteasomes have the predominant role in the degradation of short lived proteins, whereas lysosomal degradation appears to be quantitatively less important \[[@B20]\]. Determining the mechanism that cells use to degrade the proteins uncovered by the method described here will require the use of specific inhibitors \[[@B21]\]. Before degradation, most short-lived proteins are covalently coupled to multiple copies of the 76-amino-acid protein ubiquitin \[[@B22]\], a reaction catalyzed by a series of enzymes \[[@B23]\]. These ubiquitinated proteins are recognized by the 26S proteasome and degraded within its hollow interior \[[@B24]\]. This system of regulated degradation is central to such processes as cell-cycle progression, gene transcription and antigen processing. A few proteins have been found to be exceptions \[[@B25],[@B26]\]; like ODC, they do not require ubiquitin modification for degradation by the proteasome. In most cases it is not clear how short-lived proteins are selected to be modified and degraded. Some rapidly degraded proteins have been shown to contain an identifiable \'degradation domain\'. Removal of this degradation domain makes such proteins stable, and appending this domain to a stable protein reduces its stability. Such a degradation domain has been identified in a number of short-lived proteins, including the carboxy terminus of mouse ODC \[[@B6],[@B27]\] and the destruction box of cyclins \[[@B28]\]. In some cases, the signal is a primary sequence - like the PEST sequence \[[@B29],[@B30]\]. However, the identifiable structural features of such degradation domains are not sufficiently uniform to provide a reliable guide to identifying labile proteins. The method we have described does not use ubiquitin conjugation as a search criterion. This approach thus has the potential to discover labile proteins regardless of whether ubiquitin modification plays a role in their turnover. Once a large and representative sample of short-lived proteins is identified, a search for structural motifs among these proteins may facilitate the discovery of those motifs which correlate to protein degradation.
Conclusions
===========
In this study we have developed an innovative technology to identify labile proteins using GFP-fusion expression libraries. Using this technology we have discovered short-lived proteins in a high-throughput format. This technology will greatly facilitate the discovery and study of short-lived proteins and their cellular regulation.
Materials and methods
=====================
Construction of GFP-cDNA expression libraries
---------------------------------------------
Messenger RNAs from brain, liver, and the HeLa cell line (Clontech) were used as templates for cDNA synthesis, using a cDNA synthesis kit from Stratagene according to the manufacturer\'s recommendation, with some modifications. First-strand cDNA was synthesized using an oligo(dT) primer-linker containing an *Xho*I restriction site and with StrataScript reverse transcriptase. Synthesis was performed in the presence of 5-methyl dCTP, resulting in hemimethylated cDNA, which prevents endogenous cutting within the cDNA during cloning. Second-strand cDNA was synthesized using *E. coli*DNA polymerase and RNase H. Adaptors containing *Eco*RI cohesive ends were introduced into the double-stranded cDNA, which were then digested with *Xho*I. The cDNAs contained two different sticky ends: 5\' *Eco*RI and 3\' *Xho*I. The cDNAs were separated on a 1% SeaPlaque GTG agarose gel in order to collect those larger than 800 bp. After extracting cDNAs from the agarose gel with AgarACE-agarose-digesting enzyme followed by ethanol precipitation, the cDNAs were directionally cloned into EGFP-C1/2/3 expression vectors with three open reading frames (ORFs) (Clontech). The vectors were modified within the multiple cloning sites in order to be compatible with the cDNA orientation. By this means, cDNA ORFs were aligned to the carboxy terminus of EGFP. The host cell used for plasmid transfection and expression, 293T, expresses the SV40 large T antigen. Therefore, the cDNA EGFP-C1/2/3 vector containing the SV40 origin of replication can replicate independently from chromosome DNA in the host cells, which facilitates the recovery of plasmid DNAs from the host cells.
Transfection of the libraries into 293T cells
---------------------------------------------
293T cells were cultured at 37°C in DMEM (Invitrogen) supplemented with 10% FBS, 1% nonessential amino acids and 100 U/ml penicillin, 0.1 mg/ml streptomycin. One day before transfection, cells were seeded in 10-cm plate in 10 ml growth medium without antibiotics. Transfection was performed using Lipofectamine 2000 reagent according to the manufacturer\'s instructions. Samples (25 μg) of a cDNA library were diluted in 1.5 ml Opti-MEM (Invitrogen). Lipofectamine 2000 was diluted in 1.5 ml Opti-MEM and mixed with diluted DNA. After 20 min incubation, the DNA-Lipofectamine 2000 complex was added to the cells. The cells were incubated for 16 h before analysis.
FACS analysis of GFP-expressing cells
-------------------------------------
Cells were harvested by trypsinization, washed, and resuspended in DMEM. Cytometric analysis and sorting were performed using a hybrid cell sorter combining a Becton Dickinson FACStarPLUS optical bench with Cytomation Moflo electronics (Stanford Beckman Center shared facility). Green fluorescence was measured using a 525/50 band pass filter. Gates were set to exclude cellular debris and the fluorescence intensity of events within the gated regions was quantified. Fluorescence-activated cell sorting was performed with a lower forward scatter threshold to detect transfected cells while ensuring that debris and electronic noise were not captured as legitimate events. Transfection efficiency was so high that normal voltages for detecting GFP were reduced. For fractionation, the cell population was gated on the basis of the fluorescence intensity. Cells were sorted at a rate of 8,000 events/sec. 10^6^cells were collected in 12 × 75 mm glass tubes containing 200 μl serum to enhance the cell survival rate. For short-lived protein screening, sorted cells were recultured in a 12-well plate and treated with or without 100 μg/ml CHX for 2 h. The cells then were collected and subjected to FACS analysis and sorting. The cells showing a decrease in fluorescence intensity with CHX treatment were collected for further analysis.
Plasmid recovery
----------------
Plasmid DNA was extracted from sorted cells using a Qiagen mini-plasmid preparation kit. Plasmid DNAs were eluted in water and transformed into electro-competent DH10B *E. coli* (Invitrogen). Bacterial colonies were transferred to 96-well plates containing LB with 50 μg/ml kanamycin and 30% glycerol. After overnight growth at 37°C, the colonies are stored at -80°C. Plasmid DNAs were prepared from individual clones, sequenced and BLAST searches performed against the NCBI database.
Construction of Myc-tagged full-length coding sequences of genes
----------------------------------------------------------------
To obtain full-length coding sequence of the genes, we amplified them with a human full-length cDNA kit (Panomics) according to the manufacturer\'s instructions. The full-length coding sequences of cDNAs were then cloned into the pCMV-Myc vector (Clontech) for expression in 293T cells.
Western blot analysis of protein degradation
--------------------------------------------
The plasmid DNAs of individual clones were prepared and transfected into 293T cells. The transfected cells, with or without CHX treatment, were collected in PBS and cell lysates were prepared by sonication. Proteins were resolved by SDS-polyacrylamide gel electrophoresis and transferred to a membrane. Fusion proteins were detected using a polyclonal antibody against GFP (Clontech), a monoclonal antibody against the Myc epitope (Sigma), a polyclonal antibody against G protein (Santa Cruz) or an antibody against Hsp70 (Santa Cruz). Bands were visualized with SuperSignal West Pico kit (Pierce).
Additional data files
=====================
Additional data file [1](#s1){ref-type="supplementary-material"} contains the original data used to perform this analysis and is available with the online version of this paper.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
The original data used to perform this analysis
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank the staff of the FACS service center at the Beckman Center, Stanford University, for their technical support, and Robert Lam and Shanmei Li at Panomics for their help. This work was supported by NIH SBIR grant R43 GM64036 to X.L. and NIH grant RO1 45335 to P.C.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Schematic diagram of the four steps of the screening procedure. **(a)**Fractionate by FACS cells transfected with an EGFP-cDNA expression library according to their fluorescence intensities; **(b)**refractionate those cells made dimmer by cycloheximide (CHX) treatment; **(c)**recover plasmids, clone in bacteria, pool clones and select CHX-responsive pools by FACS analysis; **(d)**recover and characterize individual cDNA clones.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Fractionation of 293T cells transfected with a GFP-cDNA expression library. Cells were subjected to FACS analysis and fractionated into five subpopulations: R2, R3, R4, R5 and R6.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
FACS analysis of fractionated cells treated with CHX or untreated. The fractionated subpopulations R3 and R4 treated with or without CHX were subjected to FACS analysis. The log-normal fluorescence histogram distributions from **(a)**R3 and **(b)**R4 populations are shown. The gray curve represents cell populations not treated with CHX and the black curve represents the treated cells. The shaded area represents cells from the populations left-shifted by CHX that were used for plasmid recovery.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
CHX chase analysis by western blot of three labile EGFP-fused library clones. **(a)**Cells were individually transfected with GFP-cDNA clones representing splicing factor SRp30c, guanine nucleotide-binding regulatory protein or cervical cancer proto-oncogene p40. Transfected cells were treated with CHX and were collected immediately thereafter or after 1, 2 or 3 hours for western blot analysis using anti-EGFP polyclonal antibody. The mobility of protein markers is indicated. **(b)**Cells were transfected with constructs expressing EGFP or d1EGFP, a destabilized form of GFP, and analyzed as in (a).
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Cycloheximide chase analysis by western blot of three full-length myc-tagged cDNAs. Cells were transiently transfected to express splicing factor SRp30c, guanine nucleotide-binding regulatory protein or cervical cancer proto-oncogene p40, each with an amino-terminal myc epitope tag. Transfected cells were treated with CHX and samples subjected to western blot analysis using anti-myc antibody. The mobility of protein markers is indicated.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
Cycloheximide chase analysis by western blot of two endogenous proteins. 293T cells were treated with CHX and samples subjected to western blot analysis using antibodies against G protein or Hsp70. The mobility of protein markers is indicated.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
The estimated half-lives of 22 labile proteins
:::
Clone Accession number Gene description Stability Estimated half-life (h)
------- ------------------ ----------------------------------------------- ------------- -------------------------
2 BC005843 Similar to SH3-containing protein Short-lived 2
3 AF176555 A-kinase anchoring protein Short-lived 3
4 AF209502 Calpain Short-lived 1
5 U87277 Splicing factor SRp30c Short-lived 0.5
9 BC000804 Adaptor-related protein complex Short-lived 2
10 BC011384 ATP synthase, H^+^transporting Short-lived 2
12 BC005380 Apolipoprotein A-I Short-lived 2
14 BC007513 H19, imprinted maternally expressed Short-lived 2
18 AP000014 Function unknown Short-lived 2
19 M69013 Guanine-nucleotide-binding regulatory protein Short-lived 0.5
21 BC003128 CGI-89 protein Short-lived 2
22 AL109795 Function unknown Short-lived 2
23 X89399 Ins(1,3,4,5)P~4~-binding protein Short-lived 2
25 BC057397 Heat shock 70 kDa protein 1A Short-lived 2
26 NM\_015416 Cervical cancer 1 proto-oncogene protein p40 Short-lived 2
27 AF248272 Gag-Pro-Pol precursor protein gene. Short-lived 3
28 X07868 Insulin-like growth factors II Short-lived 2
29 NM\_000062 Serine (or cysteine) proteinase inhibitor Short-lived 2
30 AF068706 Gamma2-adaptin (G2AD) Short-lived 2
32 AK054590 Function unknown Short-lived 3
33 BC000404 Thyroid hormone receptor interactors 13 Short-lived 2
The clones were recovered as described in the text and their half-lives were estimated by FACS-based analysis of CHX chase kinetics. All 22 clones were partially sequenced and BLAST analysis performed to identify similar protein sequences in the NCBI public database.
:::
|
PubMed Central
|
2024-06-05T03:55:51.772568
|
2004-9-28
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545601/",
"journal": "Genome Biol. 2004 Sep 28; 5(10):R81",
"authors": [
{
"first": "Xin",
"last": "Jiang"
},
{
"first": "Philip",
"last": "Coffino"
},
{
"first": "Xianqiang",
"last": "Li"
}
]
}
|
PMC545602
|
Background
==========
*MuDR/Mu*transposable elements are widely used for mutagenesis and as tags for gene cloning in maize \[[@B1],[@B2]\]. The high efficiency of *Mu*insertional mutagenesis regulated by *MuDR*in highly active Mutator lines reflects four features of this transposon family. First, a plant typically has 10-50 copies of the mobile *Mu*elements \[[@B3]\], although some plants have over 100 copies. Second, they insert late in the maize life cycle, generating diverse mutant alleles transmitted in the gametes of an individual Mutator plant \[[@B1]\]. Third, they exhibit a high preference for insertion into genes \[[@B1]\]. And fourth, most maize genes are targets as judged by the facile recovery of *Mu*insertion alleles in targeted screens \[[@B1],[@B4]-[@B6]\]. In directed tagging experiments, the frequency of *Mu*-induced mutations for a chosen target gene is 10^-3^-10^-5^\[[@B7]\]. Interestingly, a *bronze1*exon \[[@B8]\] and the 5\' untranslated region of *glossy8*\[[@B9]\] contain hotspots for *Mu*insertion in specific regions, which may explain the higher frequency of mutable allele recovery for these genes.
Somatic mutability, visualized as revertant sectors on a mutant background, is indicative of transposon mobility. By monitoring maintenance of a mutable phenotype, it was established that the Mutator transposon system is subject to abrupt epigenetic silencing, which affects some individuals in most families \[[@B10],[@B11]\]. A molecular hallmark of silencing is that both the non-autonomous *Mu*elements and the regulatory *MuDR*element become hypermethylated \[[@B12],[@B13]\]. Without selection for somatic instability of a visible reporter allele and/or hypo-methylation, Mutator lines inevitably lose *Mu*element mobility.
The high efficiency of *Mu*mutagenesis has been exploited in several reverse genetics strategies. The first protocol described used PCR to screen plant DNA samples to find *Mu*insertions into specific genes using one primer reading out from the conserved *Mu*terminal inverted repeats (TIRs) and a gene-specific primer \[[@B14]-[@B17]\]. Alternatively, survey sequencing of maize genomic DNA flanking *Mu*insertions yields a list of tagged genes in each plant \[[@B18],[@B19]\]. A third method uses *RescueMu*, *a Mu1*element containing a pBluescript plasmid, to conduct plasmid rescue by transformation of *Escherichia coli*with total maize DNA samples. To identify insertions in genes of interest, *RescueMu*plasmids can be screened or the contiguous host genomic DNA can be sequenced using primers permitting selective sequencing from the right or left TIRs of *Mu1*\[[@B20]\].
Here we describe the initial results of a large scale *RescueMu*tagging effort conducted by the Maize Gene Discovery Project. The tagging strategy employed grids of up to 2,304 plants organized into 48 rows and 48 columns. Plasmid rescue was undertaken from individual pools of up to 48 plants per row or column. Genomic sequences next to *RescueMu*insertion sites were obtained for all the rows and for a subset of columns of six grids. Maize genomic sequences were subsequently assembled into 14,887 unique genomic loci using computational approaches. These loci were analyzed for gene content, the presence of repetitive DNA and correspondence to mapped maize genes and ESTs. Gene models were built by co-assembling the genomic sequence with ESTs and cDNAs by spliced alignment and by *ab initio*gene prediction. Identified gene models were tentatively classified using gene ontology terms of potential homologs \[[@B21]\].
Many features of *Mu*element behavior have been examined previously using hundreds of tagged alleles or by analyzing the population of *Mu*elements in particular plants and a few descendants. With single founder individuals for the analyzed tagging grids, we could examine the distribution of new insertion sites of *RescueMu*in large progeny sets. The contiguous genomic sequences were analyzed to determine if there were insertion hotspots, preferential insertion site motifs, routine generation of the expected 9-base-pair (bp) direct target sequence duplication (TSD) and evidence of pre-meiotic insertion events.
Like other *Mu*elements, *RescueMu*exhibits a strong bias for insertion into or near genes, as few insertions were recovered in retrotransposons or other repetitive DNA. In addition, for the set of *RescueMu*insertions into confirmed genes, a bias for insertions into exons (rather than introns) was observed, consistent with the well-established use of Mutator as a mutagen. The gene-enrichment exhibited by *RescueMu*was compared against two physical methods of gene enrichment, methyl filtration \[[@B22]\] and high *C~0~t*genome fractionation \[[@B23]\].
Results
=======
*RescueMu*transposition in active Mutator lines
-----------------------------------------------
In standard Mutator lines, *Mu1*elements maintain copy number through successive outcrosses, indicating that some type of duplicative transposition occurs \[[@B24]\] in the absence of genetic reversion \[[@B25]\]. Most new mutations are independent and occur late in the life cycle \[[@B26],[@B27]\]. Consequently, a single pollen donor can be used to generate thousands of progeny with diverse *Mu*insertion events (Figure [1](#F1){ref-type="fig"}). Initially *RescueMu*germinal insertions were sought by direct mobilization of elements from transgene arrays containing multiple copies of the original *35S:RescueMu:Lc*plasmid and the plasmid conferring resistance to the herbicide Basta used for selection of transformed callus \[[@B20]\]. Using eight different transgene arrays crossed with diverse active Mutator lines, the average germinal transposition frequency through pollen was only 0.07 (Table [1](#T1){ref-type="table"}, grid A); lines with a single *MuDR*element had no transposed *RescueMu*(*trRescueMu*).
Materials were selected from the progeny of grid A plants for grids B through E using two criteria: there were visible seedling mutations in around 10% of progeny characteristic of a very active Mutator line \[[@B26]\] and the presence of *trRescueMu*. By DNA blot hybridization of individuals within grids B through E, the *RescueMu*transposition frequencies ranged from 0.1 to 0.26 (Table [1](#T1){ref-type="table"}). By sequence analysis after plasmid rescue, *trRescueMu*were identified that had inserted into likely maize genes and generated the diagnostic 9-bp TSD characteristic of *Mu*transposition (data not shown). There were also events initially scored as transposition by blot hybridization that represented *RescueMu*rearrangements within the transgene array, and deleted forms of *RescueMu*were detected by blot hybridization and gel electrophoretic sizing of rescued plasmids (data not shown). Although *RescueMu*insertion frequency was low, overall *Mu*movement was very high in these grids; visible, independent seedling mutations were identified in 10.1-28.3% of the selfed progeny (Table [1](#T1){ref-type="table"}), as high as the most active Mutator lines described to date \[[@B28]\].
In an effort to increase transposition frequency, lines with *trRescueMu*but no transgene array were selected. Plants with a verified *trRescueMu*were crossed to *r-g*and colorless kernels selected - these lack red spotting from *RescueMu*somatic excision from the *35S:RescueMu:Lc*transgene. During subsequent plant growth Basta-sensitivity was scored as a second indicator that the transgene array was absent \[[@B20]\] and DNA blot hybridization then confirmed that a *trRescueMu*but not the Basta-resistance transgene was present in the plant. To guard against Mutator silencing, plants were also screened by DNA blot hybridization to verify that they contained unmethylated *Mu1*and *MuDR*elements after digestion of genomic DNA with the methylation-sensitive enzymes *Hin*fI and *Sst*I, respectively (data not shown). Four plants each with a single *trRescueMu*were identified by these criteria and crossed to *r-g*. A DNA blot hybridization screen was conducted on 393 progeny of these four individuals. Seven progeny were identified with two new *trRescueMu*, seven plants were identified with three events, and 33 plants had a single *trRescueMu*; the original, parental *trRescueMu*elements were shown to segregate as Mendelian factors in the populations screened (data not shown). The 14 plants with two or three new *trRescueMu*were each crossed by an anthocyanin tester and also crossed multiple times as pollen parents to tester lines to generate sufficient progeny to construct one grid from each founder plant. Inexplicably, in sampling seedling progeny from each outcross ear, some lineages had very few new *trRescueMu*. The lines with the highest transposition frequencies had two *trRescueMu*and were used in grids G through J; DNA blot hybridization analysis of 30-200 grid plants was used to estimate transposition frequencies within each grid, which ranged from 0.38 to 0.66 (Table [1](#T1){ref-type="table"}), with an average of 0.58 per plant and 0.29 per parental *RescueMu*element. The two parental *trRescueMu*elements were shown to be segregating 1:1 and independently (Figure [2](#F2){ref-type="fig"} for grid G, and data not shown for other families).
Subsequently, surveys within each grid were used to identify plants with two or three newly *trRescueMu*and no evidence of Mutator silencing for construction of the next tagging populations. In this manner, the frequency of *trRescueMu*was increased in some grids to 1.0-1.4 per plant (Table [1](#T1){ref-type="table"}) reflecting a frequency of 0.5-0.7 per parental element.
Library plate preparation and gene representation
-------------------------------------------------
As shown schematically in Figure [1](#F1){ref-type="fig"}, the *trRescueMu*insertion sites have been immortalized by preparing libraries from each of the row and column leaf pools from 16 grids, with three additional grid libraries under construction (Table [1](#T1){ref-type="table"}). Briefly, total maize DNA was digested with *Bam*HI and *Bgl*II, both of which recognize sites outside of *RescueMu*, and the fragment mixture was used to transform *E. coli*(see Materials and methods). The resulting library plates contain 56-96 individual row and column libraries representing the diversity of germinal *trRescueMu*and a sampling of somatic events present in the harvested leaf tissue (each well in a library plate is a pool of 20-48 plants from a row or column). The parental *RescueMu*insertion sites inherited from the grid founder(s) are present in every library.
Library plates contain a high diversity of genomic sequences. In a row of 48 plants, assuming random insertion, two segregating founder elements and a transposition frequency of 1.0, there will be 50 different plasmid types in the heritable class. Including heritable and somatic insertions, we estimate that each row or column library contains about 100-200 distinct plasmid types. Given these parameters, a library plate from a 48 row × 48 column grid with an average of 150 somatic plasmids per row or column library would contain 14,400 somatic insertion sites plus 2,304 germinal events and the two parental insertion sites. Because *RescueMu*shows a strong bias for insertion into genes \[[@B20]\], each library plate contains a substantial fraction of the predicted 50,000 genes of maize \[[@B29]\], provided the insertion sites are random. Ultimately, library plates for 19 grids derived from 33,000 plants and containing an estimated 30,108 heritable *trRescueMu*insertion sites (grid size × transposition frequency from Table [1](#T1){ref-type="table"}) will be available online from the Maize Gene Discovery project through MaizeGDB \[[@B30]\].
Plasmid recovery analysis and identification of probable germinal insertions (PGIs)
-----------------------------------------------------------------------------------
Based on gel electrophoretic analysis of nearly 1,000 rescued plasmids, the genomic DNA flanking *RescueMu*averaged 3.5 kilobases (kb), with a range of 0.4-15 kb (data not shown). To accommodate the large size of some plasmids, a PCR template preparation protocol was devised to amplify genomic inserts of up to 16 kb for high-throughput sequencing \[[@B31]\]; primers were designed to amplify from within the right and left TIRs reading outward into the maize genomic DNA such that high quality sequence would be available to identify the TSDs flanking *RescueMu*insertion sites. Plasmids from all rows plus several columns of a grid were sequenced, with a routine yield of 80-92% success. A subset of plasmids could not be bidirectionally sequenced, because they lacked the TIRs at one or both ends. Deleted forms of *trRescueMu*were detected in several percent of the individuals surveyed by DNA blot hybridization (see Figure [2](#F2){ref-type="fig"} for an example). If such derivatives retained the origin of replication and ampicillin-resistance marker, they could be cloned by plasmid rescue; if the TIRs were absent, they could not be sequenced.
Previous analysis of *trRescueMu*demonstrated that somatic insertion events, typically found in a tiny leaf sector, were sequenced just once from a leaf DNA sample while multiple instances of the germinal events could be recovered \[[@B20]\]. Out of 28,988 non-parental plasmids sequenced, 41% (11,749) were recovered once (new *trRescueMu*somatic plus germinal insertion events) for each grid, and 59% (17,239) were recovered multiple times (probable new *trRescueMu*germinal insertion events). In addition, a total of 24,875 parental plasmids were transmitted from the founder plants. The percentage of parental plasmids within each grid varied from 17% for grid G to 61% for grid P. Some grids had more parentals than other grids and some parental plasmids were preferentially sequenced for unknown reasons. The parental insertion sites include the two or three known parental sites that each segregated into 50% of the progeny. Somatic sectors in the tassel or ear of the parental plant that generated plasmids found in multiple individuals within the grid are analyzed in a later section.
Grid sequence data were used to cross-check the transposition frequency estimated from DNA blot hybridization (Table [1](#T1){ref-type="table"}) using both a row and column matching method and a more general multiple recovery method. Analysis of 80 individuals from six contributing outcross ears in grid G identified 54 that were newly *trRescueMu*, equivalent to a frequency of 0.68 new insertions per plant. Using a Poisson model based on this transposition frequency for an individual grid (Table [1](#T1){ref-type="table"}), the sequencing goal was established to reach a depth sufficient to insure that with 95% confidence, each probable germinal insertion would be recovered at least once. In the Poisson model, the 5% probability for the zero class (in other words, the 95% probability of finding all PGIs at least once) occurs when the observed mean is -ln(0.05) or approximately 3. After sequencing several rows and at least one column for a grid, multiple occurrences of PGIs were counted and used to project the sequences required to obtain the desired average of 3 occurrences of each PGI. As a cross-check of this coverage using the row and column matching method, the sequenced row plasmids were compared to the sequences available from four columns of grid G and 149 matches were found. This is equivalent to a transposition frequency of 0.81 based on 149/(4 × 46 plants per row), somewhat higher than the estimate of 0.68 based on blot hybridization analysis of individual plants. Recovery in both a row and a column is highly indicative of a probable germinal insertion because the row and column plasmids were obtained from different leaves and only germinal insertions would be found throughout a plant. The results for each analyzed grid are shown in Table [2](#T2){ref-type="table"}. The low column sampling in grid K (only 192 plasmids were attempted for each of three columns) and grid M (96 plasmids for two columns and 192 plasmids for a third column) resulted in a lower than expected number of germinal insertions. Grid P had a low germinal insertion count with this method because a portion of the column sequences was from rows generated from different parental plants and subsequently excluded from the analysis.
Analysis of the row and column sequence data within grids demonstrates that the row sequencing was too shallow to recover some probable germinal insertions more than once and that a fraction of germinal insertions were not sequenced. For example, within grid G, 385 plasmids were identified twice in the available column data but were missing from the row sequences; this is over twice the number of plasmids identified by row and column matching. From the number of plasmids successfully sequenced per row within grid G, we estimated a 70-95% probability of sequencing the likely germinal insertion events at least once in the rows. For other grids, the sampling efficiency ranged from 30 to 95% per row. Grids in which some rows had sampling efficiency less than 60% are listed as partial in Table [1](#T1){ref-type="table"}; sequencing was terminated in portions of these grids because of technical difficulties such as an excess representation of a parental insertion site, a large number of rearranged *RescueMu*elements that could not be sequenced with the standard protocol, or poor yield of *RescueMu*plasmids for unknown reasons.
The second method of identifying probable germinal insertions includes plasmids that were recovered multiple times, regardless of whether a column sequence was present. Almost all somatic insertions should only be recovered once due to their occurrence in just a few cells. The results using this method for each grid are shown in Table [3](#T3){ref-type="table"}.
What these data mean in practice is that the 3,138 probable germinal insertions identified after sequencing the same *RescueMu*plasmid at least twice is not a comprehensive list of the heritable insertion events. On the basis of the number of grid plants and estimated transposition frequencies (Table [1](#T1){ref-type="table"}), 8,311 probable germinal insertions were expected from the six grids (see Table [3](#T3){ref-type="table"}). From this we estimate that the majority of the heritable insertion events are represented by only a single sequenced *RescueMu*plasmid. It is likely that nearly half of the plasmids recovered just once represent a germinal insertion (0.44 = (8,311-3,138/11,749)). By PCR screening of library plates containing the immortalized row and column plasmids, plants containing a specific insertion event can be verified (Figures [1](#F1){ref-type="fig"} and [3](#F3){ref-type="fig"}). Selection against specific plasmids in *E. coli*probably contributed to non-recovery of certain insertion sites as sequencing templates, and these plasmids may also be under-represented in library plates.
Verification of germinal transmission
-------------------------------------
Individual grid plants with probable germinal insertions were identified on the basis of recovery of the same plasmid in both a row and a column. In addition, library plates containing all of the row and column libraries can be screened using PCR, with one primer designed to the *Mu1*TIRs present in *RescueMu*and a second primer in the gene of interest, as illustrated in Figure [3](#F3){ref-type="fig"}. A probable germinal insertion plasmid should yield the same size product in at least one row and one column library of that grid plate; the row and column identifiers specify the address of the plant(s) containing this insertion. To test this method, 11 instances of duplicate plasmid recovery in grid G (N. Arnoult and G-L.N., unpublished data) and 14 such cases in grid H (K. Goellner and V.W., unpublished data) were verified to be represented in both a row and a column library by PCR screening of the corresponding library plate. Seedling progeny from the identified row and column plants were evaluated for the presence of the expected *RescueMu*insertion site. A germinal insertion was verified for 16/16 cases examined by DNA blot hybridization and/or PCR of individual progeny plants in the family (see Additional data file 2 for methods and for plants used to verify germinal transmission \[[@B31]\]).
Mutational spectrum of *RescueMu*
---------------------------------
As shown in Figure [4](#F4){ref-type="fig"}, *RescueMu*insertions occur in diverse gene types. Illustrating the utility of *Mu*tagging, insertions are found in housekeeping genes, such as actin, as well as in regulatory genes such as those for transcription factors and protein kinases. Using the database of mapped maize genes and expressed sequence tags (ESTs) \[[@B30]\], *RescueMu*insertions are identified in genes on all 10 maize chromosomes \[[@B32]\]. These data confirm earlier studies tracking *Mu*insertions using DNA blot hybridization that established that these elements insert throughout the genome and do not show a measurable bias for insertion locally \[[@B1]\]. In addition, about 85% of *RescueMu*insertion sites that match maize ESTs correspond to genes of unknown function, suggesting the discovery of novel genes.
Of the 14,887 *RescueMu*insertion sites identified in six grids (multiple insertions into a gene from the same grid being counted only once because the majority are the same insertion event), 88% represent single instances of transposon insertion locations. There were 596 instances of a specific genomic sequence having two or more *RescueMu*insertion events. If the maize genome contains 50,000 distinct genes that are targets of *Mu*insertional mutagenesis, then far fewer cases of duplicate recovery would be expected by chance alone, given the number of events analyzed (*p*\< 0.001); therefore, *RescueMu*exhibits some preference for particular genes.
To determine if there were \'hotspots\' for *RescueMu*insertion within particular genes, data were compared between grids with independent founder individuals. As summarized in Table [4](#T4){ref-type="table"}, 90% of the *RescueMu*insertion sites were found in just one grid. This was true for both probable germinal insertion events (plasmids found two or more times within a grid) as well as for singlet sites (a mixture of germinal and somatic events). The 10% of insertion sites found in two or more grids represent independent recovery of a *RescueMu*insertion into the same locus.
In addition to the computational comparison in which an overlap of 50 bases (95% identity) was scored as insertion into the same gene, over 730 insertion sites were examined manually for 250 cases of genes with insertions from more than one grid. Of these insertion sites, 80% were at different locations within the same locus; we found 85 cases of insertions within a 1-10 bp region and 67 cases of insertions at the same base. Previously, Dietrich *et al*. \[[@B9]\] reported that 62 of 75 *Mu*insertions at *glossy8*were in the 5\' untranslated region, with 15 insertions at the same base; similarly, the beginning of exon 2 within *bronze1*is the most frequent site of *Mu*insertion in that gene \[[@B8]\].
One *RescueMu*contig from the Genomic Survey Sequencing (GSS) section of GenBank, ZM\_RM\_GSStuc03-10-31.4765 \[[@B33]\], is a hotspot for *RescueMu*insertion, with six plasmids sequenced from row 42 of grid G and one each from grids H, I, and M. Insertion sites were identical across the grids. Sequences generated to both the left and right of the *RescueMu*element were aligned as demonstrated in Figure [3a](#F3){ref-type="fig"}. Many maize ESTs matching a maize acetohydroxyacid synthase were found near this insertion site; the closest (GenBank GI: 4966438) is less than 50 bp away. Because this *RescueMu*insertion site was recovered multiple times in grid G, a heritable insertion may exist. After PCR screening of grid G plasmid libraries, summarized in Figure [3a](#F3){ref-type="fig"}, the plant at row 42, column 22 was identified. To assess heritability of this *RescueMu*insertion site, total leaf DNA was extracted from selfed seed of this plant, namely G 42-22, obtained from the Maize Genetics Cooperation Stock Center. PCR screening of the DNA (Figure [3c](#F3){ref-type="fig"}) indicated that plant 5 is homozygous for the insertion and plant 7 is homozygous wild type. DNA blot hybridization with a 0.6-kb purified PCR probe amplified with primer pair 1 + 5 confirmed plant 5 to contain the homozygous insertion allele, plant 7 to be wild-type, and the rest to be heterozygous for the insertion (Figure [3e](#F3){ref-type="fig"}). Various mutant phenotypes were observed in plant 5 (Figure [3f](#F3){ref-type="fig"}), including retarded seedling growth, reduced plant height, discolored streaks on adult leaves and sterile tassel and ear. Because there are multiple *Mu*elements in this line, further characterization of selfed progeny of its heterozygous siblings will be performed to determine the true phenotype caused by this insertion.
Analysis of 9-bp TSD and insertion site preferences
---------------------------------------------------
Because a 9-bp TSD is characteristic of *Mu*insertion events, the 9 bp next to the left and right TIRs of an individual *RescueMu*plasmid were used to join the right and left flanking sequence provided they were complementary (Figures [1](#F1){ref-type="fig"}, [3](#F3){ref-type="fig"}); note that the sequences are complementary because they were generated from different strands. For non-parental plasmids, left and right sequence data were available for 13,966 plasmids, and the 9 bp was readily identified computationally for 47.2% (6,596) of these. The remaining non-parental plasmids did not have both right and left sequence data and/or the 9-bp motif could not be verified; 5.7% (1,816) contain only post-ligation sequences. Possible explanations for incomplete sequencing results include deletions next to *Mu1*elements that remove a portion of the TIR as well as flanking host sequence \[[@B34],[@B35]\]; these events occur with about a 10^-2^frequency at existing insertion sites and if they occurred during or subsequent to *RescueMu*insertion they would preclude identification of the 9-bp repeat. Alternatively, the lack of a 9-bp TSD could reflect sequencing error. Manual inspection of 300 of the unmatched cases indicated that for nearly 90% there was an 8/9-base repeat match with the mismatch being an undetermined base (an \'N\') or a single missing or additional base. Given that all sequences were single pass but of high average quality (phred 35, equivalent to one base-calling error in 3,160 bases), we consider that 9-bp TSDs exist in virtually all *trRescueMu*insertion sites. A few cases showed anomalies in the TSDs, which probably reflect rearrangements near *RescueMu*.
Several groups have reported weak consensus insertion site preferences for *Mu*based on smaller data sets \[[@B9],[@B18],[@B20]\]. We have derived a site-specific frequency profile of the bases from 3,999 *RescueMu*insertion regions \[[@B32]\]. The profile is in agreement with what has been reported earlier by Dietrich *et al*. \[[@B9]\], showing a strong bias for high G/C content in the 9-bp TSD within a flanking dyad-symmetrical consensus: CCT-(TSD)-AGG. The non-random insertion pattern strongly suggests that *RescueMu*targeting is at least partially dependent on sequence features. In addition, we have compared the profiles derived independently from insertion sites within confirmed exons, introns and uncharacterized regions, respectively, and found the same base preferences in all three sets (data not shown).
Of 14,887 genomic loci, 62% matched maize or other plant EST/cDNAs. As more genomic sequence becomes available that can be assembled with ESTs to annotate the non-coding portions of maize genes, it will be interesting to determine if the *RescueMu*insertion sites that do not match an EST or gene in another species represent introns or other non-coding genic regions. On the basis of the gene structure annotated by maize EST matching, we have located 968 TSD sites within genes. Of these, 849 are inside exons. To check if *RescueMu*has a preference for insertion into exons (that is, the above observed high frequencies of exon insertions is not the result of potential high exon proportion in the maize genes), a standard binomial test with normal approximation was performed. On the basis of the matching to ESTs, the lengths of all exons and introns observed from all *RescueMu*contigs were counted as 2,182,954 bp and 439,403 bp, respectively. Assuming that *RescueMu*does not have a preference to insert into exons (null hypothesis), the probability of observing an exon insertion event is proportional to the length of exons (single binomial trial probability 0.832). The probability of observing at least 849 exon insertion events was calculated (less than 0.001; reject the null hypothesis). This result suggests that *RescueMu*has some preference to target exon regions within genes.
As outlined in Materials and methods, the *RescueMu*GSS sequences were scanned and masked for repetitive elements as collected in The Institute for Genomic Research (TIGR) Cereal Repeat Database \[[@B36]\]. The repeat content was compared with results for GSS sequences derived by methylation filtration (MF) and high *C~0~t*selection (HC) using the same repeat-masking criteria \[[@B36]\]. The percentage of masked nucleotides was 16.5, 24.5 and 16.2% for *RescueMu*, MF and HC, respectively.
Therefore, on the nucleotide level, *RescueMu*shows similar repeat content as the physical enrichment methods. However, after we assembled the *RescueMu*GSS sequences to remove redundancy, only about 3% of the *RescueMu*loci are composed of repetitive DNA (equal or greater than 75% masked, Table [5](#T5){ref-type="table"}). If the maize genome is two-thirds retroelements \[[@B37]\], then there is an approximately eightfold insertion bias by *RescueMu*against this component of the genome. We also downloaded the latest MF and HC contigs (version 3.0) from TIGR \[[@B38]\] and applied the same repeat masking on those contigs. Our results show that 28% of the MF and 6% of HC contigs are repetitive DNA. Thus, *RescueMu*and HC have similar bias against repetitive DNA, superior to the MF bias. It should be noted, however, that the MF and HC GSS sequencing has generated, on average, much longer contigs than *RescueMu*(see Additional data file 2).
In addition, only 0.4% of the *RescueMu*insertions were found in either the approximately 10,000 copies of the 9.1 kb 28S + 18S rRNA genes \[[@B39]\] comprising 3.6% of the 2.5 gigabase (Gb) maize genome, or in the large number of tRNA and 5S rRNA genes in the maize genome (Table [5](#T5){ref-type="table"}). These results demonstrate a strong bias against insertion into genes transcribed by RNA polymerases I and III.
Also shown in Table [5](#T5){ref-type="table"}, about 62% of the *RescueMu*loci match strongly to maize or other plant ESTs or appear to encode proteins with high similarity to known proteins. In addition, about another 5% of the loci were predicted to be genic regions with high stringency by *ab initio*gene prediction programs. As a control, we matched ESTs to contigs assembled from unfiltered (random) maize GSS sequences \[[@B38]\]. From about 33,000 of those unfiltered contigs, less than 20% of them show significant matching to ESTs. This shows that *RescueMu*contigs contain more than threefold enrichment of genic regions than random sequencing. This is consistent with our expectation that *RescueMu*preferentially inserts into genes. It is worth pointing out that plant EST collections contain ESTs from repetitive elements. Although we masked contigs using the annotated TIGR repeat database \[[@B38]\], it is possible that some contigs still contain unidentified repetitive elements, which might overestimate the number of genic regions by matching the same ESTs to different copies of repetitive elements. In particular, 18% of the EST matched regions show high similarity to transposon coding regions based on BLAST searches against the GenBank nucleotide and protein databases, suggesting that at most 14% of unfiltered contigs include protein-coding genes. The numbers of genic sequences from MF and HC was reported to be 27% and 22%, respectively \[[@B36]\]. However, these numbers are not directly comparable to our *RescueMu*results, because these authors used much higher stringency for the EST spliced alignments with the BLAT program \[[@B40]\], requiring 95 and 80% identity, respectively, when matching to the TIGR maize gene index or other plant indices. We used the GeneSeqer program for spliced alignment of the *RescueMu*data, which tolerates less sequence matching without compromising gene structure prediction accuracy \[[@B41]\]. The results using GeneSeqer for *RescueMu*, MF, and HC are very similar (data not shown).
Palmer *et al*. \[[@B42]\] evaluated the gene discovery rates of MF, EST sequencing and *RescueMu*by comparing the respective sequence sets to rice gene models. They concluded that unique gene discovery is most efficient with MF at a sequencing depth when EST sampling saturates. However, their reported low gene discovery rate for *RescueMu*does not reflect the *RescueMu*insertion bias, because their dataset included all sequences deposited in GenBank. That is, they did not remove the redundancy resulting from multiple sequencing of parental insertions.
Multiply recovered *RescueMu*insertion sites in the progeny of a single founder plant
-------------------------------------------------------------------------------------
Probable germinal insertions involve plasmids recovered several times within a sequenced row and/or column, but in addition, some *RescueMu*insertion sites were found in two or more row libraries (Table [6](#T6){ref-type="table"}). Although these could represent hotspots for *Mu*insertion at exactly the same base, we consider it more likely that they reflect the known ability of *Mu*elements to insert pre-meiotically, resulting in several progeny with the same newly generated mutation present as a sector on an ear indicative of a single insertion event \[[@B43],[@B44]\]. Robertson estimated that 20% of *Mu*transpositions occur pre-meiotically, 60% occur during meiosis or immediately afterwards, and 20% occur after the mitosis that separates the two sperm in haploid pollen \[[@B1]\]. We infer that multiple row recovery of the same insertion site within a grid was indicative of a likely pre-meiotic insertion; in contrast, authentic hotspots have the same insertion site among grids. A second line of evidence is that DNA blot hybridization surveys to calculate transposition frequency within a grid identified many instances of a particular fragment size shared in two or more progeny (data not shown). Finally, phenotypic screening of grid progeny families identified numerous instances of identical phenotypes segregating in related families \[[@B45]\]; each such phenotypic class was counted just once in calculating the percentage of families with a new visible phenotypic mutation (Table [1](#T1){ref-type="table"}).
To calculate the extent and timing of pre-meiotic sectors, the sequenced plasmids from grids G, H, I, K, M and P were classified as occurring in a single row or in multiple rows. The development of the tassel and ear must be considered when evaluating these data. An insertion event that occurs during meiosis can be represented in two haploid cells. During microgametophyte (haploid plant) ontogeny, both of these cells survive, resulting in two pollen grains with the same event. In contrast, only one megagametophyte develops after megaspore meiosis; therefore, female meiotic and subsequent events in the haploid megagametophyte are always represented in just one progeny plant. Most grid plants resulted from male transmission of *RescueMu*and a minority (about 10%) from female transmission. Given that the founder plants produced copious pollen, there is a low probability that two grains carrying the same meiotic insertion will both result in seed; therefore, the same *RescueMu*insertion site found in two rows should usually be from a pre-meiotic transposition event. For all events found in three or more rows, the insertion event must be pre-meiotic.
The 103 insertions sites found in three or more rows of grid G must be pre-meiotic events (see Table [6](#T6){ref-type="table"}). They represent 9% of the probable germinal insertion events (103/1,091) identified by the criterion of recovery of the same plasmid twice or more (see Table [3](#T3){ref-type="table"}). The percentage was similar for all six grids: there were 321 events identified in three or more rows out of 3,138 probable germinal insertions. Surprisingly, 138 contigs were found in four or more rows in these six grids, including 34 events in 10 or more rows (Table [6](#T6){ref-type="table"}). Therefore, occasionally there is a *RescueMu*insertion event very early in the somatic development of the inflorescence or in the apical meristems. The majority of *trRescueMu*insertion sites are found in only one row (92% of germinal plus somatic insertion sites, Table [6](#T6){ref-type="table"}).
As a cross-check on the analysis of pre-meiotic events presented in Table [6](#T6){ref-type="table"}, we evaluated the actual number of individual plants containing the same insertion site for a subset of each grid, using the sequence data from columns. Using this method we confirmed that among 184 plants in grid G with both row and column sequence data, there were 65 cases of insertion sites found in two or more rows or in two or more columns (Table [7](#T7){ref-type="table"}). Similar results were obtained for the other five grids. From these calculations and the data in Table [6](#T6){ref-type="table"} it appears that *RescueMu*insertions must occur routinely before meiosis and that, although rare, there are a significant number of early somatic insertion events that are transmitted to multiple progeny.
Discussion
==========
*RescueMu*was introduced into maize by particle bombardment resulting in complex transgene loci containing multiple copies of the transposon and the Basta-resistance plasmid used for selection of transgenic lines \[[@B20]\]. After crossing with an active Mutator line, *RescueMu*exhibited somatic excision from a *35S:Lc*reporter allele resulting in a red-spotted aleurone but the heritable insertion frequency was very low. Progeny screening identified individuals containing two or three *trRescueMu*elements lacking the original transgene array by genetic segregation and unmethylated *Mu1*and *MuDR*elements. Some of these individuals and subsequent derivatives with the same characteristics were used as founder plants to construct grids of plants organized into rows and columns for efficient generation and analysis of germinal mutations. Tagging maize sequences with *RescueMu*followed by plasmid rescue and sequencing of the flanking host DNA has identified 3,138 insertion locales from 17,239 plasmids (see Table [3](#T3){ref-type="table"}). These plasmids represent 59.5% (17,239/28,988) of the total non-parental plasmids of the genomic loci found in each grid. Because sequencing depth was too shallow to identify all likely germinal insertions, the 40.5% of non-parental plasmids recovered just once (11,749 from Table [3](#T3){ref-type="table"}) represent a mixture of somatic and germinal events. On the basis of the estimation of germinal insertion frequency from DNA blot hybridization, the six grids should contain more than 8,000 heritable *trRescueMu*insertion sites, but the sequencing depth was too shallow to identify all of these by multiple recovery of the same plasmid two or more times.
*RescueMu*is suited for both reverse and forward genetic strategies. Given the genomic sequence contiguous to any *trRescueMu*, a PCR screen can be designed to identify which plant contains the insertion of interest using 96-well plates containing the immortalized collection of row and column rescued plasmids. The row and column plant address can be used to order seed for further genetic and phenotypic analysis as illustrated by the *RescueMu*insertion into the acetolactate synthase gene (Figure [3](#F3){ref-type="fig"}). Alternatively, the phenotype database, which is organized by individual plant, can be searched to identify individuals segregating for mutations of interest. Active Mutator lines with multiple mobile *Mu*elements were used so most mutations will be caused by these *Mu*elements because they increase mutation frequency 50-100-fold above spontaneous levels \[[@B1]\]. The high forward mutation frequency reflects the copy number of the elements and their preference for insertion into or near transcription units \[[@B1]\]. From the DNA hybridization blots (data not shown) used to verify that grid founder plants had unmethylated *Mu*elements, the copy number of unmethylated *Mu*elements was estimated at 20-40 per founder; therefore, two mobile *RescueMu*elements would be expected to account for 5-10% of the newly generated mutations. Seed was ordered through the Maize Genetics Cooperation Stock Center \[[@B46]\] for further characterization.
*RescueMu*insertions were found in genes and ESTs mapped to all 10 maize chromosomes \[[@B31]\], and were found in all of the gene classifications for maize (Figure [4](#F4){ref-type="fig"}). These data confirm the empirical observations of maize geneticists that *MuDR/Mu*transposons are general and efficient mutagens for maize genes \[[@B1]\]. Analysis of 14,887 loci defined by *RescueMu*insertions demonstrates that transposition is highly preferential for RNA polymerase II transcription units: about 62% of the sites match maize or plant ESTs. Because the EST collections are incomplete and lack intron and promoter sequences, it is likely that an even higher proportion of *RescueMu*insertion sites are in or near genes but cannot be currently assigned to a specific gene. Given the current efficiency, large tagging populations in excess of 200,000 plants would be required in order to recover *RescueMu*mutations in all maize genes (estimation is based on the calculation method in \[[@B47]\]). The numerous grids evaluated for phenotypic characteristics should approach saturation of visible mutations, although most of the mutations are caused by standard *Mu*elements.
Given that the maize genome comprises approximately 70% retrotransposons and other highly repetitive sequences, including around 10,000 copies of the rRNA genes \[[@B37]\], these components of the maize genome are significantly under-represented in *RescueMu*insertion sites. Only about 8% of the *RescueMu*insertion sites match repetitive elements and few insertions (0.4%) were recovered in genes transcribed by RNA polymerase I or III. These results suggest that a chromatin component associated with polymerase II transcription units or the absence of a structure in other classes of genes is important in targeting *RescueMu*and other *Mu*elements to maize genes. Similarly, recombination during meiosis and transcription *per se*is targeted to genes. It is likely that the parasitic *Mu*elements exploit an element of host gene packaging that evolved for other reasons to facilitate transposition into genes.
The biological specificity for maize genes exhibited by *RescueMu*is close to methyl filtration and high *C~0~t*fractionation. The probable germinal insertion class defines a collection of mutations of enormous potential for the phenotypic characterization of maize with specifically disrupted functions. However, the low cost of template production is a distinct advantage of both physical enrichment methods compared to the high cost of designing, sampling and self-pollinating tagging grids. Current levels of sample sequencing from the physical enrichment templates highlight the desired redundancy of the *RescueMu*method, which is important for distinguishing somatic from germinal insertions at individual loci. The physical enrichment methods are considerably below one times coverage of the transcriptome of around 250 Mb; hence the current efficiency of generating novel sequence (the likelihood that the next clone sequenced is new) is much higher with these methods than with *RescueMu*.
Using the *RescueMu*insertion site data, several parameters of *Mu*transposition behavior were investigated. We confirm that a 9-bp TSD is characteristic of virtually all *Mu*insertion sites. We confirm that a small percentage of *trRescueMu*suffer deletions, including loss of a TIR, as noted in previous studies of *Mu1*\[[@B35]\]. Through evaluation of several hundred *Mu*insertion sites \[[@B9],[@B18]\], consensus motifs have been proposed for insertion sites. The sequence profile derived from the much larger population of *RescueMu*insertion sites is consistent with the previously proposed motifs. A bias exists for G+C-rich sequence, reflecting the composition of maize exons. We confirm that there are hotspots for *Mu*insertion, identified by finding identical *trRescueMu*insertion sites in independent grids. A few loci were recovered in four or more of the six grids analyzed, and many more in two (1,295 genes) or three (233 genes) grids. There is no strong DNA consensus motif at these hotspots, and we consider it more likely that a specific DNA structure or a protein associated with genes establishes conditions for efficient *Mu*insertion at particular sites. It is important to note that active transcription is not a requirement for *Mu*element insertion; otherwise *Mu*would preferentially insert into genes active late in floral development and in gametophytes.
The *trRescueMu*insertion sites represent a mixture of non-heritable somatic insertions present in leaves, germinal insertions in single grid individuals, insertion events in pre-germinal sectors within flowers, and parental elements. Parental elements identified in a grid founder plant segregated 1:1 in the progeny as expected. In addition, some insertion events were found in three or more grid rows, and hence in three or more individuals, and must be pre-meiotic transposition events in the founder. This class represented 10.2% (321/3,138) of all the likely germinal insertions identified (calculated from Table [6](#T6){ref-type="table"}). Given the clonal analysis model of the pattern of cell divisions establishing the ear and tassel of maize \[[@B48]-[@B50]\], the earliest events within the apical meristem could affect up to half of the ear or tassel, with subsequent events affecting progressively narrower portions of the inflorescence. The majority of the pre-meiotic events are consistent with *RescueMu*transposition in the floral cells a few cell divisions before the onset of meiosis, that is, in precursor cells that are still proliferating and could generate at least two and up to approximately 50 meiocytes. A smaller fraction of new insertions events occurred early enough to be represented in many progeny of a particular plant. These rare, early transposition events generate very large sectors within the developing inflorescence.
*Mu*transposon mutagenesis is highly efficient, primarily because the transposon targets genes and it is usually found in 10-50 copies per genome. How does the plant tolerate the large number of mutations generated by this agent? Within the diploid somatic tissues, most new mutations lack a phenotype; however, the haploid gametophytes are subject to stringent selection. Unlike animals, in which the phenotypes of the sperm and egg are set by previous gene activity in the parent, many characteristics of the haploid phase of the plant life cycle reflect haploid genetic activity, which requires overlapping but distinctive suites of genes in the mega- and microgametophytes \[[@B51]\]. Consequently, the late timing of new *Mu*insertions generates gamete diversity, but the unfit genotypes are culled from the population before fertilization. Coe *et al*. \[[@B52]\] describe the general problem that lethals occur much more frequently in pollen than in the megagametophyte. Any method that relies on pollen transmission will therefore fail to recover certain types of mutations that would be recovered through female transmission. For this reason, a subset of maize genes required in both types of gametophyte is refractory to knockout mutagenesis.
Conclusions
===========
A public resource of transposon-tagged maize alleles was constructed and evaluated. *RescueMu*is an efficient tag for mutagenizing and cloning maize genes, because 66% of insertion sites appear to be in genes. Sequencing from immortalized plasmid libraries organized into row and column plates reflecting the organization of fields of plants permit identification of probable germinal insertions; the library plates can be searched by PCR to verify germinal insertions and subsequently acquire seed of the corresponding plant. Alternatively, a searchable database of segregating plant phenotypes in seed, seedling, or adult tissues can be used to find plants carrying mutations of interest. Although *RescueMu*can target most, if not all, RNA polymerase II transcription units in the nuclear genome, the transposon does exhibit hotspots in particular genes. Neither the hotspots nor other insertion sites contain a motif(s) defining predictable insertion locations. *RescueMu*properties confirm attributes established with smaller populations of standard *Mu*elements.
Materials and methods
=====================
Biological materials
--------------------
*RescueMu*contains all of *Mu1*plus a 400-bp segment of *Sinorhizobium meliloti*and pBluescript (Stratagene), as described previously by Raizada *et al*. \[[@B20]\]. The complete sequence of *RescueMu*was obtained in this study using PCR primers to amplify overlapping sections of the element \[[@B31]\] for bidirectional sequencing (GenBank accession AY301066). In the construct used to make transgenic plants, the *RescueMu*transposon was placed in the 5\' untranslated region of a *35S:Lc*expression plasmid where it blocked expression \[[@B20]\]. *Lc*is a member of the R family of transcriptional regulators of the anthocyanin pathway \[[@B53]\]. Transgenic maize lines in the A188 × B73 (*r-r/r-g*, *A1*, *Bz1*, *Bz2*) hybrid background were crossed to *r-g*testers and subsequently with *r-g*Mutator lines containing multiple copies of *MuDR*to visualize *RescueMu*somatic excision as red anthocyanin sectors in an otherwise white aleurone. The tagging populations used here were developed by screening for transposition of *RescueMu*from the original, complex transgene arrays to diverse genomic locations. Using DNA blot hybridization, these once-transposed *RescueMu*(*trRescueMu*) were closely monitored for subsequent transposition, and lines were monitored for *Mu1*and/or *MuDR*methylation in the TIRs, a sign of incipient Mutator silencing. Details of line development and evaluation, including DNA blot hybridization methods, will be presented elsewhere. The anthocyanin tester lines (recessive for *r-g*, *a1*, *bz1*or *bz2*) were in inbreds W23, K55, A188, or hybrid combinations of these lines. Some *RescueMu*lines used in tagging grids were crossed to inbreds A619 or B73, which are both *r-g*, *A1*, *Bz1*, *Bz2*. Grid backgrounds are presented in detail at \[[@B31]\].
Plasmid rescue and DNA sequencing
---------------------------------
Detailed protocols are presented at \[[@B54]\], and a schematic is provided in Figure [1](#F1){ref-type="fig"}. Briefly, leaf tissue was collected from all plants in each row and from a different leaf in each column of a grid. A separate plasmid rescue library was constructed after *Bam*HI plus *Bgl*II digestion of the genomic DNA preparations. These libraries were immortalized in library plates available from the project \[[@B31]\]. Plated colonies were picked, grown overnight in liquid media, and sequencing templates prepared by a direct PCR method suitable for amplifying genomic inserts of up to 16 kb. Cycle sequencing was performed using Big Dye Terminator chemistry to read out from a position around 110 within the left or right terminal inverted repeat (TIR) of *RescueMu*; although the primers were selective for one TIR, there was some cross-priming. All grid rows plus several columns were sequenced. Three 96-well plates were normally sequenced for each row or column to obtain sequence information for a desired minimum of 200 plasmids; additional sequencing reactions were conducted if necessary. Matches of row and column sequences are designated as probable germinal insertions, because they represent an insertion site present in two leaves of that plant (designated by its row and column address); when only row sequences were available from a particular plasmid, probable germinal insertions were designated after recovery of the same sequence two or more times. Plasmid types recovered just once are a mixture of heritable and strictly somatic insertions. Parental *RescueMu*insertion sites present in a grid founder plant segregated in the grid progeny, and these insertion sites were expected to be found in all rows and columns. In some cases, particular parental plasmids were over-represented in the sequenced plasmid population. To reduce their contribution and increase recovery of new insertion sites, a rare-cutting restriction enzyme site was identified in the parental plasmid and the corresponding enzyme was included in the genomic DNA preparation to bias against recovery of that parental plasmid.
PCR screening of a library plate to quantify a *RescueMu*insertion hotspot Six gene primers plus one *RescueMu*left readout primer were used in this study:
1\. 5\'-TTGGGAGGTTGAAGGTAAAGACAT-3\'
2\. 5\'-GTGCTG GATTGGTTACTCCG-3\'
3\. 5\'-CGATGATTCTAGTTGAGCGTCTG-3\'
4\. 5\'-ACTCGCACCAACATGAATACC-3\'
5\. 5\'-GTTTCCGAGGACGCAGAGG-3\'
6\. 5\'-AGCGCCAGGGCCAGGGGATTC-3\'
L. 5\'-CAT TTC GTC GAA TCC CCT TCC-3\' (*RescueMu*)
Locations and directions with respect to the insertion site of *RescueMu*are shown in Figure [3a](#F3){ref-type="fig"}. PCR conditions were as follows: 5-20 ng of each plasmid library, 2.0-2.5 mM Mg^2+^, 0.4 mM dNTPs, 0.8-1.0 μM gene primer and 4-5 μM *RescueMu*L primer in a 50 μl reaction was first denatured for 2 min at 95°C followed by 35 cycles of 30 sec at 95°C, 30 sec at 55°C and 2 min at 72°C, and a final 2 min extension at 72°C. The same PCR conditions were used for screening using 5-100 ng samples of maize total genomic DNA.
DNA blot hybridization
----------------------
Total genomic DNA was extracted from leaf tissues using a modified urea method \[[@B55]\]. After overnight digestion, the restricted DNA was separated on a 0.8% agarose gel and transferred onto Hybond-N+ membrane (Amersham Biosciences) in 0.4 M NaOH. Blots were hybridized with non-radioactive probes labeled with AlkPhos DIRECT system (Amersham Biosciences) for chemiluminescence detection on X-ray film.
Initial clustering and assembly of genomic sequences
----------------------------------------------------
The sequences were screened to remove the TIR sequences using the program crossmatch \[[@B56]\] and then trimmed to achieve a minimum phred score \>15 in sliding windows over 40 bases. Overall the quality scores averaged phred \>35, or less than one error in 3,160 bases. The average length of the trimmed, high quality genomic sequence entering the assembly was 378 bases. The right-TIR primer yielded 22% more successful sequence than the left-TIR primer resulting in an excess of right side sequences. Trimmed sequences were then assembled into contigs using phrap \[[@B56]\] with the following parameters: -minmatch 35 -minscore 30 -node\_seq 14 -node\_space 9. The member sequences for each contig were extracted from the phrap output files and assigned to a row or column of a grid. Within each contig, only a single sequence from a plasmid was used to determine the row and column representation. For example, if both the left- and right-flanking sequence from a plasmid assembled into one contig, this was considered one recovery of the plasmid. If the left-flanking sequence from one plasmid and the right-flanking sequence from a separate plasmid assembled into the same contig, this was considered two independent recoveries of the same genomic locus. In the latter case, if the right- flanking sequence was from a different row, then the sequence was recovered in multiple rows as well. All sequences were deposited into the Genomic Survey Sequencing (GSS) section of GenBank \[[@B57]\].
Assembly of *RescueMu*-derived genomic sequence data
----------------------------------------------------
As shown in Figure [1](#F1){ref-type="fig"}, using the 9-bp TSD characteristically generated during *Mu*element insertion \[[@B1]\], the sequences to the right and left of a particular *RescueMu*element can be assembled into a continuous sequence. To do this, trimmed *RescueMu*GSS sequences were downloaded from GenBank \[[@B58]\], for comparison to raw GSS sequences containing the *Mu1*TIR sequences. The TIRs were masked by the cross\_match program \[[@B56]\] to determine the flanking 9-bp TSD sequences. The TSDs are the end-overlaps between GSS sequences generated from the left and right side of *RescueMu*insertion. Merging through TSDs using the reverse-complementary strand of the left and right sequences recovers the original genomic sequences flanking the *RescueMu*insertion. A special consideration in the assembly of the genomic sequences flanking the right- and left-TIRs of *RescueMu*is the presence of a GGATCC (*Bam*HI), AGATCT (*Bgl*II), or a GGATCT (*Bgl*II/*Bam*HI) or AGATCC (*Bam*HI/*Bgl*II) motif. The two restriction digestion sites represent a true ligation site of sequence that was non-contiguous in the maize genome, but the post-restriction site sequences can unambiguously be assigned to the right or the left of *RescueMu*. On the other hand, the GGATCT or AGATCC motif could be contiguous genomic sequence or could have been generated during the ligation step of the plasmid rescue. Consequently, assignment of the position of the sequence beyond the GGATCT or AGATCC motif is ambiguous. If the *RescueMu*insertion site matched EST sequence across and beyond the GGATCT or AGATCC motif, the post-ligation sequence could be properly assigned (Figure [1](#F1){ref-type="fig"}). In the *RescueMu*plasmid sequences considered here, the average number of sequences reported to GenBank was 2.3 (131,364/57,022) per plasmid.
The 131,364 *RescueMu*GSS sequences deposited at GenBank were screened for vector sequences against the UniVec database at the National Center for Biotechnology Information (NCBI) \[[@B59]\] using the crossmatch program: -mismatch 12 -penalty -2 -minscore 20. The resulting 130,861 vector-trimmed sequences were then screened against the maize repeat database annotated by TIGR \[[@B60]\] using the Vmatch program \[[@B61]\] with the parameters -l 50 -h 3 -identity 95. The 127,708 repeat-free sequences were then used to identify parental insertions. Any given *RescueMu*-transformed plant contains the parental *RescueMu*elements that were recovered at a high frequency during sequencing (from every sequenced row or column). Because our goal is to analyze the gene discovery by newly inserted *RescueMu*(that is, we are interested in where those non-parentals inserted into the maize genome), we decided to filter out the parental sequences as much as possible. We used Vmatch to cluster near-identical left and right sequences for each grid. A parental cluster contains sequences from nearly all the row or column sequences. A total of 59,069 parental sequences were identified and were excluded from the subsequent assembly. All the non-parental sequences were first preassembled for each plasmid using the left and right 9-bp TSD overlap. The merged GSSs were first clustered by PaCE \[[@B62]\] (minimum exact match 36 bp, minimum score threshold 30%) and then consensus sequences (contigs) for each cluster were generated by CAP3 \[[@B63]\] (overlap 40 bp; 90% identity cutoff). Because PaCE and CAP3 only pair sequence with the minimal overlap required to establish statistically significant identity, the number of contigs is probably an overestimate of the number of independent *RescueMu*insertion sites. For the particular case where TSDs were not recovered during sequencing, the left and right sequences could not be assembled together, even though they were from the same plasmid. Therefore, a Perl script was developed to conduct single-linkage clustering based on clone-pair constraints to assemble the GSS to the same \'genomic loci\' if they were derived from the same plasmid clone.
Classification of insertion site context
----------------------------------------
To be successful as a gene-discovery tool, the transposon insertions must be predominantly into the genic regions of the maize genome. To quantify the potential enrichment of the *RescueMu*flanking sequences for genic regions, we matched all assembled contig sequences against various classes of known repetitive sequences, including retrotransposons, DNA transposons, centromeric and telomeric repeats, rRNA genes and plastid DNA. For this analysis, the non-parental sequences were used in their original form, with only vector sequences but not repeat sequences trimmed. The sequences previously discarded for analysis because they consist almost entirely of repetitive elements were assembled using the same procedure as described above for the repeat-trimmed sequences. Note, however, that this number of loci is unreliable and probably an underestimate of the true number of loci recovered because of the intrinsic difficulty with assembling repetitive DNA. To identify the repetitive elements in the contigs, Vmatch (-seedlength 14 -hxdrop3 -l 30 -identity 70) was used in combination with the TIGR cereal repeat database (version 2 consisting of maize, rice, barley, sorghum and wheat repeats). The contigs were also scanned from tRNA genes by tRANscan-SE program \[[@B64]\] with its default parameters.
Gene discovery in GSS contigs
-----------------------------
Both similarity-based and *ab initio*approaches have been used to detect gene structures of the GSS contigs. For the similarity-based approach, GeneSeqer \[[@B65]\] programs were used to match plant EST contigs and cDNAs to GSS contigs. The plant EST contigs were regularly assembled by PlantGDB \[[@B66]\]. For the *ab initio*prediction, GENSCAN \[[@B67]\] (with default parameter settings for maize) was used and only high exon score predications (≥0.90) were selected. The GSS contigs were compared against SPTR \[[@B68]\], a nonredundant protein data set collected by the European Bioinformatics Institute (EBI), using BLASTX \[[@B69]\] with an E-value ≤ e-20. The BLASTX top protein hits were used to assign putative functions to the unique regions and for classification into functional categories based on annotation in the Gene Ontology \[[@B21]\] database.
The genetically mapped maize ESTs were retrieved from MaizeGDB \[[@B70]\]. These ESTs were spliced-aligned to GSS contigs using GeneSeqer as described above. The matched GSS contigs were then plotted on the maize IBM Neighbor genetic map \[[@B30]\].
Analysis of 9-bp TSD and insertion site preferences
---------------------------------------------------
For the analysis of *RescueMu*target sites, we retrieved the 9-bp TSD sequences from the confirmed insertion sites where both the left and right sequences match on the 9-bp TSD. We also retrieved the 20 bp up- and downstream sequences around the TSD. Then a 15-base long profile (9-base TSD and its three up- and downstream neighbors) was derived from the sequences and their reverse-complement orientation determined using the Expectation Maximization Algorithm \[[@B71]\].
Analysis of tentative unique contigs containing GSS sequences from multiple grids
---------------------------------------------------------------------------------
The GSS seqeunces present in each tentative unique contig (TUCs) were extracted from \[[@B31]\] and assigned to a row or column within a grid. A sample of TUCs with GSS sequences from multiple grids was then selected for detailed analysis. For each GSS in the TUC (excluding post-ligation sequences), the exact location of the TSD was determined by visual examination of the sequence alignment file for the TUC and the untrimmed GSS sequence data. The number of GSS sequences for each grid at each transposition site was recorded.
Phenotypic analysis
-------------------
Grid plants were self-pollinated unless male or female-sterile. The resulting F1 families were evaluated by inspection of ears and kernels, at weekly intervals for five weeks after germination in a sand bench in a greenhouse, and at weekly intervals throughout the life cycle in the field. Phenotypes observed were recorded and are assembled into a searchable database at \[[@B31]\]. Unique phenotypes were documented with a digital image, and there are links to corresponding *RescueMu*flanking sequences where established. Instructions on how to obtain seed of grid plants is also provided.
Additional data files
=====================
The following additional data are available with the online version of this article: a table listing the internal primers used in sequencing *RescueMu*(Additional data file [1](#s1){ref-type="supplementary-material"}), supplementary material for this paper, including details of methods (Additional data file [2](#s2){ref-type="supplementary-material"}).
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
A table listing the internal primers used in sequencing *RescueMu*
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
Supplementary material for this paper, including details of methods
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank the Maize Gene Discovery team for the development of stocks and extensive DNA blot hybridization data that preceded construction of *RescueMu*tagging grids and for the phenotypic evaluation of seed and seedling mutations. Diane Chermak generated all of the *RescueMu*library plates and most of the sequencing templates; we thank China Lunde for grid H templates and Laura Roy for grid S templates. We thank Xiaowu Gai and Trent Seigfried, who contributed to the development of ZmDB. This research was supported by a plant genome research program contract from the National Science Foundation (98-72657), which initiated the Maize Gene Discovery project. An REU supplement supported undergraduate students Warren Chen and Justin Schaffer at Stanford and Fred Oakley and Laura Schmitt at ISU in 2001.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Schematic diagram of *RescueMu*grid tagging and sequencing (*RescueMu*not to scale). Step 1: *RescueMu*is introduced into embryogenic callus followed by crossing of regenerated plants to active Mutator lines. Lines are screened for transposed *RescueMu*elements in plants lacking the original transgene array. Pollen from one *RescueMu*donor plant is crossed to multiple ears of a non-*RescueMu*line to generate tagging grids of up to 48 rows × 48 columns of *trRescueMu*plants in the field. Step 2: plant DNA prepared from pools of row or column leaves is used to generates transformed bacterial libraries of *RescueMu*plasmids. These are used as sequencing templates and for construction of a library plate representing the diverse insertion sites in grid plants. Step 3: genomic DNA is digested using two restriction enzymes (*Bam*HI, *Bgl*II), religated into plasmids and transformed into *E. coli*. Step 4: after transformation, *RescueMu*plasmid-containing *E. coli*colonies are selected by plating onto carbenicillin agar plates and picked into 384-well plates with growth/freezing media. Overnight incubation is followed by a PCR reaction designed to amplify longer inserts with lengths up to 16 kb. Using the PCR product, eight 96-well sequencing plates (four for sequence from the left TSD and four from the right TSD) are created. Step 5: priming strategy and relative locations of PCR and sequencing primers within the *RescueMu*element. The sequencing reactions are read out from the TSDs to recover the germinal insert sequence. Although a *Bam*HI and *Bgl*II double-restriction digest produces a shorter, easier-to-sequence insert length, it also increases the ambiguity in interpreting the sequence during analysis. Given successful sequencing in both directions, two GSS sequences may be submitted for every plasmid (sequence flanking the left and right TIRs). Two additional GSS sequences may be submitted for a plasmid when a *Bam*HI, *Bgl*II or *Bam*HI-*Bgl*II ligation site is encountered. Each of these occurrences yields sequence that was not necessarily contiguous *in vivo*. Dubious GSS sequences are designated with the suffix .1EL (re-created enzyme ligation site) or .2EL (re-created enzyme ligation of two restriction sites not encountered *in vivo*). Sequence flanking TIRs *in vivo*is submitted as GSS sequences with no suffix except the .x or .y (right or left) direction designation.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
DNA blot hybridization analysis of *trRescueMu*elements in grid G. Total DNA was prepared from individual grid G plants in rows 1 and 5, as listed at the top of the lanes; these rows represent two ears crossed by the same founder *RescueMu*pollen source. DNA samples were digested with *Hin*dIII, a unique site 0.5 kb from the internal end of the left TIR of the *RescueMu*element, and the resulting gel blot was hybridized with an ampicillin-resistance gene fragment to visualize *RescueMu*. The two parental *trRescueMu*had been identified in the founder plant, and these size classes are marked along the right side of the autoradiogram. Hybridizing bands corresponding to new *trRescueMu*are indicated with a black square; the hybridizing band too small to be a full-length *trRescueMu*is marked with a white arrow. GP, grid G parental insertion sites 1 and 2 shown to be segregating in the progeny.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
*RescueMu*plasmid library plate screening for a gene with multiple insertion sites. **(a)**Schematic diagrams of a *RescueMu*insertion hotspot: demonstration of the assembly of flanking genomic sequences; locations and directions of all primers used in this study; EST alignment to genomic sequence assembly showing introns. **(b)**An ethidium bromide-stained agarose gel of the PCR products from columns 1 and 22 and row 42 plasmid libraries, using primer pair 2 + L. **(c)**An ethidium bromide-stained agarose gel of the PCR products with leaf DNA extracted from G42-22(x) progeny 1 to 8, using primer pair 3 + 6. **(d)**An ethidium bromide-stained agarose gel of the PCR products with the same DNA used in **(c)**, except using primer pair 3 + L (column B is blank). **(e)***Nco*I-digested DNA blot from plants 1 and 3 to 8 probed with a fragment spanning a 0.6-kb PCR product amplified with primer pair 1 + 5. **(f)**Phenotypes at several developmental stages (from left to right): 10-day-old seedlings (1 to 10 from left to right) of the G42-22(x) progeny; a side-by-side comparison of plants 5 and 6 at 10 days, including their root mass; adult plants at 1 month showing plant 5 in the foreground of the picture with two siblings on either side; a close-up of the plant 5 adult leaf phenotype.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Functional spectrum of genes targeted by *trRescueMu*. Functional spectrum of probable proteins, identified by BLASTX of GSS contigs against the SPTR database, for trRescueMu targeted genes. Functional categories were derived from the Gene Ontology (GO) database.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Grid organization and analysis of mutant phenotypes segregating among selfed progeny of grid plants
:::
Grid\* Year^†^ Grid size^‡^(row × col) Plasmid rescued Libraries sequenced^§^ Transposition frequency^¶^ Independent mutations (% of families)^¥^
-------- --------- ------------------------- ----------------- ------------------------ ---------------------------- ------------------------------------------ ------
A 1999H 34 × 48 No No 0.07 7.2 4.5
B 1999SD 52 × 48 No No 0.10 8.6 10.1
C 1999B 40 × 40 No No 0.13 8.3 28.3
D 1999S 48 × 48 No No 0.26 8.7 15.1
E 2000H 40 × 48 No No 0.25 8.6 27.0
F 2000AZ 41 × 41 No No 0.57 6.6 19.5
G 2000S 46 × 48 Yes Yes 0.68 5.0 11.9
H 2000B 38 × 36 Yes Yes 0.62 7.5 6.9
I 2000B 38 × 34 Yes Yes 0.62 9.5 9.8
J 2000SD 38 × 45 Yes Survey 0.38 9.8 11.1
K 2001H 30 × 30 Yes Yes 0.66 8.0 20.3
L 2001H 36 × 20 Yes Yes 0.66 12.8 17.4
M 2001AZ 40 × 40 Yes Partial 1.30 8.2 ND
N 2001B 32 × 44 In progress No 0.20 6.3 ND
O 2001S 47 × 48 Yes Survey 0.50 5.2 ND
P 2002H 48 × 48 Yes Yes 1.40 5.9 ND
Q 2002H 48 × 24 Yes Yes 1.00 2.7 ND
R 2002AZ 36 × 36 Yes Survey 0.72 3.7 ND
S 2002SD 48 × 48 Yes Survey 1.00 12.7 ND
T 2002H 48 × 46 Yes Survey 1.00 ND ND
U 2002H 48 × 48 Yes Partial \>1.30 ND ND
AA 2002S 48 × 48 Yes Yes 0.60 ND ND
BB 2001B 34 × 48 Yes Survey 0.60 6.2 ND
V 2003AZ 45 × 45 In progress Survey 1.00 ND ND
X 2003SD 44 × 44 In progress Survey 1.00 ND ND
\*Grids with a single letter contain mainly plants with a *RescueMu*pollen parent plus the seed from the ear of the founder male crossed by a non-Mutator line. In grids with a double letter, both parents contained *RescueMu*. ^†^Summer nurseries are designated by year and location: A, Tucson, AZ; B, Berkeley, CA; SD, San Diego, CA; S, Stanford, CA. H indicates the winter Hawaii nursery. ^‡^Vandalism, animals, and environmental damage in the field resulted in some losses compared to expectation of the ear harvest. Ears with fewer than 100 kernels and those from outcross pollinations of male or female sterile grid plants were not assessed for mutation frequency; these lines are being propagated at the Maize Coop by sib pollination to establish a permanent line for later evaluation and distribution. ^§^Yes, indicates that all rows plus four columns were sequenced with the goal of coverage to a depth such that there was a 80-95% probability that plasmids representing germinal insertions would be identified at least once. Grids listed as partial have limited (40%-80%) depth from some rows. Survey sequencing was performed on several rows and columns on the indicated grids to verify that plasmids organized into library plates contained authentic *trRescueMu*. Library plates will be available from all grids, including V and X, during 2004 as listed at \[31\]. ^¶^Frequency of newly transposed *RescueMu*per plant based on DNA blot hybridization, sampling 30-200 plants per grid. For grid A only, the data are from plants sibling to those in the grid. ^¥^Progeny families generated by self-pollination of grid plants were examined for kernel defects before shelling, and seedling traits were scored on up to 30 surviving individuals grown from each family. A minimum of 200 families were scored for the seedling forward-mutation frequency, and all selfed ears were scored for the seed defects. Mutations were scored as independent if they were not segregating in multiple families from the same founder. Phenotypic descriptions are available at \[31\], and it is expected that the grids not yet analyzed (ND) and the summer 2003 grids V, W, and X will be scored during 2004 and 2005 for reporting through the project database. Most mutations are caused by standard *Mu*elements.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Probable germinal insertions (PGI) based on row and column matches
:::
Grid Rows (r) Columns (c) Transposition frequency (τ)\* Expected PGI (e)^†^ Row + column matches (m) Percentage of expected^‡^ Transposition frequency (using row + column)^§^
------- ---------- ------------- ------------------------------- --------------------- -------------------------- --------------------------- -------------------------------------------------
G 46 4 0.68 125.1 149 119% 0.81
H^¶^ 36 4 0.62 89.3 115 129% 0.80
I 38 5 0.62 117.8 128 109% 0.67
K 30 3 0.66 59.4 32 54% 0.36
M^¶^ 40 3 1.30 156.0 33 21% 0.28
P 37 4 1.40 207.2 71 34% 0.48
Total 754.8 528 70%
\*Expected frequency of PGI was determined from DNA gel blot analysis of frequency of newly transposed *RescueMu*per plant as stated in Table 1; ^†^expected = r × c × τ; ^‡^percentage of expected = 100 × m/e; ^§^transposition frequency = m/(r × c); ^¶^for grids H and M, rows were considered columns and vice versa to simplify calculations.
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Probable germinal insertions (PGI) based on multiply recovered plasmids
:::
Grid Multiple recovery (m) Single recovery (s) Percentage PGI\* Expected frequency (τ)^†^ Expected PGI (e)^‡^ Percentage of expected^§^ Plasmids in multiple recoveries
------- ----------------------- --------------------- ------------------ --------------------------- --------------------- --------------------------- ---------------------------------
G 1,091 3,801 22% 0.68 1,501 73% 5,535
H 535 2,142 20% 0.62 848 63% 2,945
I 544 2,000 21% 0.62 801 68% 3,162
K 228 1,000 19% 0.66 594 38% 1,202
M 330 1,075 23% 1.3 2,080 16% 2,053
P^¶^ 410 1,731 19% 1.4 2,486^¶^ 16% 2,342
Total 3,138 11,749 21% 8,311 38% 17,239
Single recoveries are also shown. \*Percentage of PGI = m/(m + s); ^†^expected frequency of PGI was taken from Table 1; ^‡^expected PGI = τ × rows × columns (see Table 2); ^§^percentage of expected = 100 × m/e; ^¶^based on 37 rows only.
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Detailed analysis of insertion sites recovered multiple times
:::
Number of same-base insertions that occurred in the indicated number of grids Number of contigs with the indicated number of different insertion sites
------------------------------------------------------------------------------- -------------------------------------------------------------------------- ------ ------- -----
1 572 90% 1 48
2 60 9% 2 71
3 6 1% 3 89
4 1 0% 4 32
5 7
6 2
7 1
Total 639 100% Total 250
:::
::: {#T5 .table-wrap}
Table 5
::: {.caption}
######
Matching of *RescueMu*genomic loci to other available databases to determine percentage of genic and repeat loci
:::
Category Number of genomic loci
----------------------------------------------- ------------------------
Total maize genomic loci discovered 14,265\*
Number of genic loci identified by:
Maize EST/cDNA 7,555
Plant EST/cDNA 1,253
Protein database 84 (62%)
GENSCAN prediction 708
Number of genic loci^†^(percentage of total) 9,600 (67.3%)
Number of loci matching repeats:
Retrotransposon 1,074
DNA transposon 212
MITEs 193
Centromere-related repeats 57
Telomere-related repeats 3
Unknown repeats 221
Other repeats 8
45S ribosomal DNA (18S + 28S) 23
5S ribosomal DNA 10
Transfer DNA 25
Number of repeat loci^‡^(percentage of total) 1,113 (8%)
\*The 14,887 unique loci were collapsed into 14,265 unique loci by linking forward/reverse sequence pairs. This provides a more conservative estimate, but in some cases may have incorrectly combined separate loci. ^†^Numbers are cumulative: that is, GSSs were first matched to maize EST/cDNAs, then the unmatched GSSs were screened against other plant EST/cDNAs, and so on. ^‡^Numbers are not cumulative: that is, some loci could match to both retrotransposon and DNA transposon sequences.
:::
::: {#T6 .table-wrap}
Table 6
::: {.caption}
######
Single and multiple recovery of specific *RescueMu*insertion sites within the sequenced rows of grids G, H, I, K, M, and P
:::
Grid Single recovery Multiple recoveries
------ ----------------- --------------------- ------- ----- ----- ---- ---- ---- ---- --------
G 3,801 22 640 326 62 20 17 2 2 4,892
H 2,142 13 331 136 31 9 6 3 6 2,677
I 2,000 10 348 124 39 3 9 6 6 2,544
K 1,000 6 155 50 11 2 1 2 1 1,228
M 1,075 3 225 74 13 5 5 3 2 1,405
P 1,731 8 246 100 27 5 22 1 1 2,141
All 11,749 62 1,945 810 183 44 60 17 17 14,887
Counts are of contigs containing sequences from the indicated number of rows. \*The zero class represents plasmids identified in column sequencing that were not identified in any row. ^†^The 1 column data include singlet plasmids as well as plasmids recovered two or more times but within a single row.
:::
::: {#T7 .table-wrap}
Table 7
::: {.caption}
######
Insertions found in at least two rows or columns among plants with both row and column sequence data
:::
Grid Plants from sequenced rows + columns Insertions found in 2+ rows or 2+ columns
------- -------------------------------------- -------------------------------------------
G 184 65
H 144 35
I 190 35
K 90 6
M 120 6
P 148 23
Total 876 177
:::
|
PubMed Central
|
2024-06-05T03:55:51.775917
|
2004-9-23
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545602/",
"journal": "Genome Biol. 2004 Sep 23; 5(10):R82",
"authors": [
{
"first": "John",
"last": "Fernandes"
},
{
"first": "Qunfeng",
"last": "Dong"
},
{
"first": "Bret",
"last": "Schneider"
},
{
"first": "Darren J",
"last": "Morrow"
},
{
"first": "Guo-Ling",
"last": "Nan"
},
{
"first": "Volker",
"last": "Brendel"
},
{
"first": "Virginia",
"last": "Walbot"
}
]
}
|
PMC545603
|
Background
==========
The fruitfly *Drosophila melanogaster*has been the prime genetic model organism for almost a century. This success story is mainly founded on countless so-called forward genetic screens designed to elucidate gene functions on the basis of their mutant phenotypes. Many of those screens reached a scale that has been termed \'saturating\' as they identify all nonredundant genes involved in a certain phenotypic trait. However, forward genetic screens are limited in that they are only capable of uncovering functions that are easily measurable or visible. Furthermore, genes having a redundant or nonessential role are less likely to be found by forward genetics.
The reverse genetic approach to unravel gene function starts with the DNA sequence. Mutations within the gene are induced and identified by various techniques and only subsequently is the mutant phenotype analyzed \[[@B1]\]. Reverse genetics may be undirected or directed, the undirected approach involving random mutagenesis, commonly by transposable elements or by chemicals, the establishment of mutant collections, and the identification of mutations in the gene of interest \[[@B2]-[@B5]\]. In contrast, directed reverse genetics is based on techniques that allow for specific inactivation of a gene. These include specific knockdown of gene activities through RNA-mediated interference (RNAi) \[[@B6],[@B7]\] and targeted gene disruption \[[@B8],[@B9]\].
Both undirected and directed reverse genetic techniques have certain advantages and drawbacks. Transposon-based mutagenesis tends to be nonrandom because of the occurrence of hotspots for transposon integration. The use of transposable elements of different origin, such as P-elements and piggyBac, which exhibit a different insertion bias, can partly circumvent this problem. However despite large-scale efforts, the ultimate goal of covering the whole *Drosophila*genome by insertion mutagenesis is far from being achieved \[[@B10],[@B11]\]. Moreover, while null mutants of P-element-tagged genes (P-elements have the tendency to integrate 5\' to a gene) can easily be generated by imprecise excision, piggyBac transposons only excise precisely \[[@B10]\].
RNAi and small interfering RNA (siRNA) screens provide a powerful tool to dissect the function of genes at a genome-wide scale \[[@B12]-[@B14]\], but the technique is most easily applied to cell cultures and is thus limited to cell-biological problems. Large-scale RNAi screens in multicellular organisms have been done only in *C. elegans*\[[@B15]\] and for technical reasons a similar approach in *Drosophila*is not feasible.
Targeted gene knockout in *Drosophila*allows for generation of both null as well as hypomorphic mutations \[[@B16]\]. However, the technique is time-consuming and technically challenging and hence not applicable on a large scale.
Random mutagenesis in reverse genetics generally relies on well-established techniques and commonly used mutagens, such as ethylmethansulfonate (EMS) \[[@B5],[@B17]\] and *N*-ethyl-*N*-nitrosourea (ENU) \[[@B18]\]. Those chemicals primarily induce single-nucleotide polymorphisms, which can most efficiently be detected by sequencing \[[@B19]\], by denaturing high-pressure liquid chromatography (DHPLC) \[[@B5],[@B17]\], or by enzymatic cleavage of heteroduplex DNA with single-strand-specific endonucleases such as Cel-I \[[@B18],[@B20]-[@B22]\]. Mismatch-cleavage analysis and DHPLC require special machinery and DHPLC is not very well suited for high-throughput analysis.
Fast neutrons have also been used to introduce small DNA lesions, which can simply be resolved by agarose electrophoresis after PCR amplification \[[@B23]\]. This kind of mutagenesis may be limited to seeds or to labs in the vicinity of a reactor.
We reasoned that it would be worthwhile to establish a generally applicable reverse genetic technique based on an unbiased and practicable random mutagenesis and an efficient mutation-detection performed on standard laboratory equipment. Here we introduce a novel mutagenesis protocol utilizing the cross-linking drug hexamethylphosphoramide (HMPA) \[[@B24]\], streamlined fly genetics and high-throughput fragment analysis on sequencers to demonstrate the feasibility of our reverse genetics approach.
Results and discussion
======================
Fly genetics
------------
There are two ways to handle mutagenized progeny. Either large collections are established and maintained, which then are systematically and continuously screened for mutations of interest, or mutagenized progeny are screened directly and only animals exhibiting a desired trait are kept. The first method is in practice an F3 screen, which requires balancing of mutagenized chromosomes and maintenance of many stocks. This approach is far more labor-intensive than a simple F1 screen of progeny and thus is more suited to stock centers. Moreover, balancer chromosomes have many DNA sequence polymorphisms to wild-type chromosomes (our unpublished data), which will interfere with detection of mutagen-induced sequence polymorphisms.
To circumvent the inherent problems with balancers, we devised an alternative genetic strategy, which had to fulfill the following criteria. First, mutagenized chromosomes have to be passed on in an unrecombined form such that mutations cannot be lost. Second, the mutagenized chromosomes should be brought into an isogenic background for mutation detection. Third, for economic reasons stock-keeping should be kept at an absolute minimum.
We generated a fly strain (KNF306) isogenic to our *yw*wild-type laboratory strain but containing the same dominant marker on the two major autosomes. Both chromosome 2 and chromosome 3 are carrying *white*^+^marked P-element insertions, which were chosen because *white*^+^expression is restricted to different subregions of the eye (Figure [1a](#F1){ref-type="fig"}). Chromosome 2 is marked by an insertion in the *CG31666*locus, which results in *white*^+^expression only in the posterior part of the eye. Chromosome 3 harbors an insertion in the promoter of *CG32111*, and this transgene causes dorsal *white*^+^expression. The combined expression patterns of both show a \'pie-slice\' eye-color appearance (Figure [1a](#F1){ref-type="fig"}). Thus, the same marker permits us to distinguish between linkage on chromosomes 2 or 3. Neither of the transgenes affects viability, and the line can be kept as a homozygous stock.
Mutagenized chromosomes of strain KNF306 were passaged only via males, which were mated to the parental *yw*strain background. Thus, the marked autosomes remained unrecombined and could be unambiguously assigned because of the dominant character of the *white*^+^transgenes (Figure [1b](#F1){ref-type="fig"}). Mutagen-fed F0 males were mass-mated and F1 males were mated in single crosses (see Materials and methods). After 4 to 5 days, nonsterile F1 males were recovered, pooled in groups of five, and their DNA extracted and analyzed. If a pool gave a positive signal, the crosses were traced back and F2 progeny carrying the mutant chromosome (as judged by the eye-color pattern) of each of the five crosses were individually re-tested. If this re-test was positive, a single F2 male of the respective cross was taken to establish a balanced stock. Non-positive crosses were discarded.
Like any other genomic locus, the *white*^+^coding regions of both transgenes constitute targets for mutagenesis, and mutagenic events can be easily scored in the F1 progeny as a loss of the characteristic expression pattern. As discussed later, effectiveness of mutagenesis can be assessed from the occurrence of *white*^-^progeny, and as an internal control mutation rates at the two loci should be comparable.
The crossing scheme and analysis procedure illustrated was optimized for autosomal genetics. We have generated another strain, KNF307, which in addition carries X chromosomes marked by a characteristic enhancer trap insertion at the *omb*locus (data not shown). However, analysis of X-chromosomal loci would require additional handling of F1 females or mutagenesis of F0 females and hence we did not carry out X-chromosomal screens.
Mutagenesis
-----------
EMS has been used as a deletion-inducing chemical in large-scale screens \[[@B25]\], but unbiased evaluation of its properties suggests that EMS-induced deletions are exceptional \[[@B26]\]. On the other hand, the deletions found by Liu *et al*. \[[@B25]\] ranged in size between 545 base-pairs (bp) and 1,902 bp and would not have been detected by Greene *et al*. \[[@B26]\]. The cross-linking carcinogen hexamethylphosphoramide (HMPA) has been shown to predominately induce deletions that were either in the range 2-315 bp or reached cytologically visible dimensions \[[@B24]\]. As our analysis method restricted the size of PCR fragments to about 800 bp, we chose HMPA as a mutagen, because EMS-induced deletions are likely to affect at least one of the primer-binding sites and would hence be undetectable.
We modified the original HMPA mutagenesis protocol to administer a shorter, but more intense pulse of HMPA (\[[@B24]\], see also Materials and methods). A dose was applied that causes a similar rate of X-linked recessive lethals as standard EMS treatment, but only moderate male sterility (Table [1](#T1){ref-type="table"} and data not shown). We also did not add *N*,*N*-dimethylbenzylamine, which in our hands potentiated the sterilizing activity of HMPA.
It has been reported that F1 progeny may exhibit mosaicism for mutagenized tissue \[[@B27]\]. Mosaic flies could generate a primary positive signal, but might not transmit the mutated gene. We have seen mosaicism at the *white*^+^loci and we have found positive F1 pools that did not yield mutant F2 progeny (Table [1](#T1){ref-type="table"} and data not shown). However, we were unable to determine whether some of the primary positives were due to mosaicism or to PCR artifacts.
Mutation detection
------------------
DNA from pools was prepared by a novel high-throughput extraction protocol allowing for up to 2,000 PCRs per pool (see Materials and methods). As HMPA is reported to induce deletions as small as 2 bp and as a mutated allele is diluted 10-fold as a result of our pooling of five flies, we decided to analyze PCR fragments on a sequencer offering maximal resolution and high sensitivity. We have also evaluated the \'poison-primer technique\' which is reported to preferentially amplify alleles with a deletion at the poison-primer binding site from large pools \[[@B28]\]. However, the small deletion alleles we have tested did not outperform the amplification of the wild-type allele to the extent previously reported, implicating that the technique is more suited to large deletions and not generally applicable (data not shown).
PCR products were analyzed on either a gel-based or a capillary sequencer (see Materials and methods). To increase efficiency of mutation detection on gels, we pooled up to three PCR products. These were labeled with different fluorescent tags, partly because they were of similar size (Figure [2a](#F2){ref-type="fig"}).
Screening
---------
The efficiency of HMPA mutagenesis could be assessed from the rate of *white*^-^mutations at the transgenes on chromosomes 2 and 3. Overall, we found 24 mutations in about 62,700 male and female flies. Two flies were mosaic for the mutations. Given that mosaicism can only be scored in eyes and there only in nonoverlapping expression domains, the mutation rates discussed below may be slightly underestimated (Table [1](#T1){ref-type="table"}). Male sterility was 25.4 %.
Aguirrezabalaga *et al*. \[[@B24]\] reported a mutation rate of 2.8 × 10^-4^at the *vermilion*locus scoring early and late progeny. The rate reached 3.7 × 10^-4^when only late progeny was regarded. After a few rounds of screening we have stopped screening early progeny (brood 1 flies, see Materials and methods), because we did not recover any *white*^-^mutation. As sperm development takes up to 10 days \[[@B27]\], we also consider it unlikely that brood 1 from our crossing scheme will yield appreciable efficiency. Disregarding brood 1, we obtained an average rate of 2.25 × 10^-4^mutations at the *white*^+^loci, which are about twice as large as the *vermilion*locus. Our mutagenesis procedure involving an overnight incubation with HMPA rather than a 3-day incubation with HMPA and *N*,*N*-dimethylbenzylamine is therefore not much less efficient than the original protocol.
There was no difference in the frequency of induced *white*^-^mutations between brood 2 and brood 3 (Table [1](#T1){ref-type="table"}). The small difference between mutation frequency on the identical *mini-white*genes located on chromosomes 2 and 3 may be attributed to statistical variance, to position effects, to different size of the enhancers driving *white*^+^expression or to systemic errors due to the smaller expression domain of the insertion on chromosome 3.
The following additional parameters can be utilized to estimate mutant recovery. The *white*gene for which the mutation rate has been assessed encodes a protein of 688 amino acids from an open reading frame (ORF) of 2,064 bp. We assume that any deletion within the ORF would generate a null phenotype. Only 14 out of 31 HMPA-induced deletions selected at the *vermilion*locus would have been scoreable by our PCR approach, because the remaining 17 mutations were caused by large deletions affecting both primer-binding sites \[[@B24]\].
We designed PCR primers for each gene to be scored such that they encompass the first coding exon and the PCR products are between 450 and 807 bp in size. The average weighted length of our PCR fragments was 710 bp (including two primers of 20 nucleotides each). We thus expect one mutation in 30,317 flies (1/(2.25 × 10^-4^× 14/31 × (710 - 2 × 20)/2,064)) or one mutation in 6,063 pools, respectively. Taking into account the fact that two mosaic flies may not have transmitted (reducing the mutation rate to 2.0 × 10^-4^), the estimate would be one mutation in 33,883 flies or one in 6,777 pools.
We have scored 16,902 F1 males at two to 11 loci and recovered two transmitting mutations from about 20,900 analyzed PCR reactions (see Additional data file 2). According to the estimate we would have expected three.
The first mutation detected was a 41-bp deletion in the first exon of *CG15000*, which during the course of this study turned out to be the second exon of the *dNAB*locus (Figure [2c](#F2){ref-type="fig"}, and see \[[@B29]\]). The deletion causes a frameshift and very probably constitutes a null mutation. As shown in Figure [2a,b](#F2){ref-type="fig"}, the mutation was identified on a gel-based sequencer in a pool of PCR products labeled with the fluorophore NED (Applied Biosystems) and propagated in one of the five F2 crosses. The mutant chromosome is currently purified by separating the *CG15000/dNAB*allele - easily traceable by a restriction-fragment length polymorphism - from the *white*^+^marker (P. Geuking and K.B., unpublished work).
Second, we detected a mutation in *CG17367*on the capillary sequencer (Figure [3a,b](#F3){ref-type="fig"}). The net 11-bp deletion (19-bp deletion, 8-bp insertion) is situated in the first intron and 5\' to the start codon. The allele is viable over a deficiency uncovering the *CG17367*locus.
This study focused on implementing HMPA mutagenesis for reverse genetics. As discussed above, HMPA efficacy has been assessed from mutations at the *white*^+^loci, which have been selected on the basis of phenotype rather than sequence. Thus, our modified HMPA protocol may also prove to be valuable for forward genetic approaches. At the molecular level we could also identify deletions in the *white*^+^genes (data not shown), but we have not systematically investigated all of the *white*^-^mutations.
Conclusions
===========
While the analysis of PCR fragment-length polymorphisms on our sequencers was very efficient, HMPA mutagenesis turned out to be the limiting parameter. It is about 28-fold less efficient than EMS mutagenesis when it is assumed that all HMPA hits are deleterious (3.2 × 10^-3^nucleotide substitutions at the 1 kb *awd*locus \[[@B5]\] for EMS compared to 2.25 × 10^-4^deletions per 2 kb *white*^+^locus for HMPA), but mutagen dose cannot be increased further because of the concomitant increase in male sterility.
The new techniques that we have introduced increase the diversity of the toolkit available to laboratories interested in conducting reverse genetic screens. The pros and cons of the critical parameters are next considered individually.
### EMS or ENU versus HMPA as mutagen
HMPA-induced deletions are very likely to induce null mutations when hitting an exon. EMS, on the other hand, primarily induces GC-to-AT transitions, but is not well suited for introducing small deletions. A considerable fraction of the transitions will not affect protein function. In *Arabidopsis*, about 44% of the mutations after EMS mutagenesis were silent, 51% were missense mutations and 5% were nonsense mutations \[[@B26]\]. Similarly, in a zebrafish ENU screen, only 15 out of 270 mutants (5.5%) were truncation mutants \[[@B18],[@B19]\]. Recently, Guo *et al*. \[[@B30]\] determined the tolerance of a protein to random amino-acid changes and determined that about two thirds of amino-acid substitutions were neutral and only 34% were disruptive. Assuming that all truncation mutations are deleterious, it can be concluded that about 22% (34% of 51% plus 5%) of EMS-induced mutations negatively influence protein function. Of those amino-acid substitutions an unknown fraction will retain partial function. Thus, allelic series can be generated through EMS \[[@B22]\] and the generation of partial loss-of-function alleles may be a potential asset of EMS mutagenesis. Overall, HMPA is maximally sixfold less effective at inducing loss-of-function mutations (22% of 28%) than a high dose of EMS, but this disadvantage is compensated for by a more straightforward mutant analysis.
### Mutant analysis
Mutant analysis depends critically on the mutagen and vice versa. Currently, the most effective way to screen for EMS-induced polymorphisms is the TILLING approach, which, however, requires a second round of PCR, specialized chemistry of the secondary primers, and an enzymatic reaction on the secondary product. TILLING cannot easily be performed on standard sequencers: we have tried to analyze Cel-I cleaved fluorescent heteroduplex DNA on an ABI 3730 sequencer, but did not obtain satisfactory sensitivity (data not shown). HMPA induced mutations can be detected by fragment-length analysis of primary PCR products on standard sequencers. Hence, screening for small deletions reduces PCR costs by a factor of 2 and spares the effort of secondary assays.
### Mutant handling
Mutant handling is independent of the mutagenesis protocol and may be combined with either EMS or HMPA mutagenesis. For example, TILLING can be performed both on large mutant collections and on a continuous supply of freshly generated mutants.
Finally, given the genotoxic properties of HMPA in both prokaryotes and higher eukaryotes \[[@B31],[@B32]\], both the mutagenesis and the mutation-detection procedures described in this study may be directly transferred to other model organisms.
Materials and methods
=====================
HMPA mutagenesis
----------------
About 150 1-3-day-old F0 KNF306 (*y*, *w*; *CG31666-white*^+^; *CG32111*-*white*^+^) males were starved for 4 to 6 hours in a plastic bottle containing three layers of water-soaked LS14 filter papers (Schleicher & Schüll). A 1.1 ml sample of HMPA solution (5% sucrose, 0.1 M NaPO~4~, 25 mM HMPA, optional 0.05% bromophenol blue) was carefully applied to the filters using a syringe with a long needle (21G2) inserted through the foam stopper. The starved males were exposed to the HMPA solution overnight. Bromophenol blue does not affect mutagenicity detectably, but stains the guts of the flies blue and thus enables mutagen uptake to be monitored and controlled. Freshly eclosed flies do not ingest enough mutagen. HMPA-contaminated plasticware must be disposed of by thermal waste treatment.
Fly work and crossing procedure
-------------------------------
In six bottles containing standard corn medium, each 25 mutagenized KNF306 F0 males (Figure [1a](#F1){ref-type="fig"}) were allowed to mate to 15 to 20 virgin *yw*females (brood 1). After 2 days males were taken out and crossed to *yw*virgins in new bottles (brood 2A) and this cross was transferred after 3 days (brood 2B). After another 2 days F0 males were recovered and mated to fresh *yw*virgins (brood 3A). F1 males of broods 2A, 2B and 3A were collected and mated individually to three *yw*virgins in about 650 separate crosses per week. Five hundred non-sterile males were removed after 4 to 5 days and five males were pooled for DNA extraction. Fertilized females were returned, and unsuccessful crosses were discarded. If analysis of PCR fragments indicated a primary positive pool, crosses were traced back and kept for further analysis; the other crosses were discarded. From each of the five crosses of primary positive pools a single F2 male or female containing the chromosome of interest as manifested by the typical eye-color pattern was collected for DNA extraction. If PCR analysis yielded a secondary positive result in one of the five F2 flies, a single F2 male containing the chromosome of interest was taken out from the respective cross for balancing (Figure [1b](#F1){ref-type="fig"}).
The whole crossing scheme requires 6 weeks and was organized such that a mutagenesis was performed every second week (see Additional data file 1).
Large-scale DNA extraction, PCR and fragment analysis
-----------------------------------------------------
DNA was extracted in bulk by squishing pools of each five flies through mechanic force in a vibration mill (Retsch MM30) programmed to shake for 20 sec at 20 strokes per second. Flies were placed into wells of a 96-well deep-well plate. Each well was then filled with 500 μl squishing buffer (10 mM Tris-Cl pH 8.2, 1 mM EDTA, 0.2% Triton X-100, 25 mM NaCl, 200 μg/ml freshly added proteinase K) and one tungsten carbide bead (Qiagen). The deep-well plate was then sealed with a rubber mat (Eppendorf) and clamped into the vibration mill. (Tungsten carbide beads can be recycled: after an overnight incubation in 0.1 M HCl and thorough washing in double-distilled water (ddH~2~O) the beads were virtually free of contaminating DNA.) Debris was allowed to settle for about 5 min and each 50 to 100 μl of supernatant were transferred into a 96-well PCR plate. The reactions were incubated in a thermocycler for 30 min at 37°C, and finally for 5 min at 95°C to heat-inactivate proteinase K.
A Tecan pipeting robot was used for PCR setup. To 5 μl of template DNA, master-mix was added and PCR was performed on an MJR thermocycler that was integrated into the robot. The master-mix per reaction was composed of 20.48 μl ddH~2~O, 0.6 μl MgCl~2~(25 mM), 0.6 μl dNTPs (10 mM), 0.1 μl fluorescently labeled primer 1 (100 μM), 0.1 μl primer 2 (100 μM), 0.12 μl hot-start Taq polymerase (HotStar, Qiagen, 5 U/μl), 3 μl 10× buffer containing MgCl~2~(Qiagen). Cycling conditions were 95°C 15 min, 35 × (95°C 20 sec, 60°C 30 sec, 72°C 1 min), 72°C 2 min, 4°C.
Three differently labeled PCR reactions (oligos were 5\' labeled with Applied Biosystems\' fluorophors FAM, NED and VIC, respectively) were then pooled. To facilitate sizing of fragments we also added ROX1000 size marker (Applied Biosystems) to five DNA pools. Samples of 1.5 μl pooled DNA were mixed with 1.5 μl loading buffer (consisting of one part 25 mM EDTA pH 8.0 with 50 mg/ml blue dextran and five parts HiDi formamide (Applied Biosystems)). The reactions were incubated for 3 min at 95°C, cooled down, and 1.5 μl each were loaded onto a 96-lane ABI 377 sequencer. Run conditions were as follows: 1 h pre-run at 1,000 V, 35 mA, 51°C and 10 h run at 2,400 V, 50 mA, 51°C. Gel images recorded at four different color channels by the GeneScan software were analyzed visually.
Slight modifications to this protocol were introduced for analysis performed on an ABI 3730 capillary sequencer. First, DNA was diluted 20-fold before PCR. Second, after PCR, reactions were diluted 100-fold and 2 μl of diluted PCR products were added to each 15 μl HiDi formamide (Applied Biosystems). PCR product was diluted on a Tecan pipeting robot. Diluted DNA was denatured for 2 min at 95°C before analysis. Sample injection (10 sec) and analysis (12,000 scans) was done according to standard protocols. Identification of deletion fragments was then performed by visual inspection of gel-images generated by the Data Collection Software (Array Viewer option, Applied Biosystems). No internal size standard was used, as deletion fragments were identified relative to wild-type PCR product.
Additional data files
=====================
The following additional files are available with the online version of this paper. Additional data file [1](#s1){ref-type="supplementary-material"} contains the time schedule of mutagenesis, fly work, and screening. The whole procedure takes six weeks and is organized such that one mutagenesis has to be performed every second week to generate a continuous supply of mutagenized progeny. Additional data file [2](#s2){ref-type="supplementary-material"} contains information on the 10 other genes scored. Gene names, fluorescent labels, fragment lengths and the number of analyzed F1 flies are given. Labeled primers were ordered from Applied Biosystems. Primer sequences are available upon request.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
The time schedule of mutagenesis, fly work, and screening
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
Information on the 10 other genes scored
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank Anni Strässle for performing fragment analysis, Yüksel Kocaman for fly work, and Melanie Greter, Philip Gast and Angela Baer for excellent technical assistance. We also thank Oleg Georgiev and Dieter Egli for the *CG31666*insertion line.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Fly genetics. **(a)** The fly strain used in this study is isogenic to a *yw*standard lab strain, but carries neutral *white*^+^transgene insertions on the two major autosomes. The P-element insertion on chromosome 2 localizes to the promoter of *CG31666*and the other transposon is situated 5\' to *CG32111*on chromosome 3. The *white*^+^expression domains are restricted to the anterior and dorsal parts of the eye in the respective strains, and the patterns overlap such that the genotypes can be unambiguously assessed from eye appearance. **(b)** The genetic scheme ensures that mutagenized F0 chromosomes are passed unrecombined, because they are transmitted via males only, and that flies carrying a mutation can easily be singled out on the basis of the eye phenotype. The mutagenized flies are always crossed back to the parental wild type, and only chromosomes from F2 progeny carrying a confirmed mutation are balanced.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Isolation of a *CG15000/dNAB*null allele. **(a)** Three differently labeled PCR reactions were pooled and concomitantly analyzed on an ABI 377 sequencer. Green, HEX-labeled *white*fragment (control reaction); yellow, NED-labeled *CG15000*products; blue, FAM-labeled *CG33273/DILP5*fragment. Color channels are not completely tight and strong products exhibit translucence. Arrows at the screenshot and the magnified inset mark a shorter product specific for one reaction. **(b)** One F2 fly from each cross (1 to 5) of the primary positive pool (P) was analyzed in the context of further F1 screening. Number 5 harbors the deletion allele identified in the pool. The larger, but also specific, fragment in the pool might represent a PCR artifact. The specific band marked by an asterisk turned out to be a PCR artifact as well. **(c)** Genomic organization of wild-type (WT) and mutant *dNAB/CG15000*. Exons are depicted and coding region is symbolized by filled rectangles. Small half arrows indicate primer-binding sites of the PCR-amplified region. The deleted region is filled yellow and the sequence of the wild-type and the mutant alleles starting at an ATG originally annotated as the first codon in the *CG15000*ORF are aligned below.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Identification of an HMPA-induced deletion at the *CG17367*locus. **(a)** Electropherograms of the primary positive pool (top), of one of the four wild-type F2 flies (center), and of a mutant F2 fly (bottom) are shown. The *x*-axis shows fragment size in bp; the *y*-axis, arbitrary intensity units. For unknown reasons the deletion allele runs faster when amplified from pool DNA than from a heterozygous fly. **(b)** Genomic organization of *CG17367/LNK*. Coding (filled rectangles) and noncoding (open) exons are shown and primer-binding sites are indicated by the half arrows. The deletion/insertion is situated 16 bp 5\' to a splice site of the first intron, as can be ascertained from the sequence alignment.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Mutation rates induced by HMPA treatment as assessed from *white*^-^mutations at the transgene insertions on chromosomes 2 and 3
:::
*white*^+^loci Brood Mutation rate (× 10^-4^) Mosaicism at *white*^+^loci
----------------- ---------------- -------------------------- -----------------------------
Both loci 2A + 2B and 3A 2.25
On chromosome 2 2A + 2B and 3A 2.68 1/14
On chromosome 3 2A + 2B and 3A 2.0 1/10
Both loci 2A + 2B 2.35
Both loci 3A 2.33
Numbers are based on 19 mutations recovered after screening of brood 1 has ceased. Two out of a total of 24 *white*^-^mutants (five mutants were recovered while brood 1 was screened) were mosaic for the mutation. Mutations on chromosome 2 were slightly more common than mutations on chromosome 3; brood-2- and brood-3-derived mutations were equally frequent.
:::
|
PubMed Central
|
2024-06-05T03:55:51.784836
|
2004-9-28
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545603/",
"journal": "Genome Biol. 2004 Sep 28; 5(10):R83",
"authors": [
{
"first": "Knud",
"last": "Nairz"
},
{
"first": "Peder",
"last": "Zipperlen"
},
{
"first": "Charles",
"last": "Dearolf"
},
{
"first": "Konrad",
"last": "Basler"
},
{
"first": "Ernst",
"last": "Hafen"
}
]
}
|
PMC545604
|
Background
==========
Many methods for high-throughput, experimental elucidation of gene function (functional genomics) depend on the availability of full-length cDNA clone collections \[[@B1]\]. These clones provide access to the protein-coding open reading frames (ORFs) and facilitate expression of large numbers of proteins in the native form or as fusion proteins. The value of ORF-containing full-length cDNA clone collections (ORF clones) has now been amply demonstrated by studies in model organisms, in particular in the area of protein interaction mapping using methods based on yeast two-hybrid or mass spectrometry \[[@B2]-[@B8]\].
Extension of functional genomic approaches to mammalian genomes requires development of adequate ORF clone collections. Several projects based on complete sequencing of clones isolated from cDNA libraries are in place to generate these collections for mouse \[[@B9]\] and human \[[@B10]-[@B13]\]. Additional efforts have also focused on subsequent manipulation and exploitation of the full-length clones using versatile recombinational cloning systems so that the ORFs are formatted for expression \[[@B14]-[@B16]\]. However, obtaining a complete set of human clones has been hampered by the inadequacies of cDNA libraries and uncertainty over the true number and identity of all protein-coding genes. Approaches based on cDNA libraries have two major limitations in mammals. The first is the difficulty in obtaining full-length cDNA clones and the complexities of alternate and partial splice forms. Hence, many clones have to be sampled to obtain a canonical full-length version of each cDNA. The second is that these projects inevitably reach a point of diminishing return when it is no longer financially viable to continue to sequence more clones from the same library or from different tissues in order to add small numbers of new full-length cDNAs to the collection. Therefore, it is pertinent to ask how complete the cDNA collections currently are, and whether they can be supplemented or replaced by other approaches in order to develop complete ORF clone sets.
There is still uncertainty over the exact number of human and mouse genes and hence over the completeness of the existing cDNA collections. Therefore, we have investigated a defined subset of genes, namely the full-length protein-coding genes defined in our current annotation of human chromosome 22 \[[@B17]\]. In this study, we have found that in the currently available major cDNA collections, a total of 60% of chromosome 22 protein-coding genes are represented by complete ORF clones, although no single collection contains more than 48% (Table [1](#T1){ref-type="table"}). This leaves a sizeable fraction of the genes unavailable. Thus there are still considerable challenges to be faced in identifying and isolating full-length cDNAs and ORFs for functional analyses.
To extend the coverage of full-length ORF clones, we have developed an alternative method which exploits knowledge of gene structure based on genomic sequence. It involves the specific amplification of a targeted ORF plus short regions of the 5\' and 3\' untranslated regions from a mixed pool of cDNAs. Amplified fragments are cloned into a standard plasmid sequencing vector and their identity and integrity confirmed by DNA sequencing. The aim of the method is to provide cDNA clones containing confirmed full-length ORFs, which can later be manipulated into suitable vector systems such as Gateway (Invitrogen) or Creator (BD Biosciences) for functional genomics. We have applied this method to the same set of chromosome 22 protein-coding genes and have shown that we can obtain clones representing 70% of the targeted genes with a limited range of experimental conditions. We have also demonstrated a reasonable expectation that we can isolate clones for 83%.
Results and discussion
======================
Analysis of full-length cDNA collections
----------------------------------------
We have previously described a gene annotation of chromosome 22 \[[@B17]\] and its characterization \[[@B18]\]. In this annotation, 546 genes were defined as protein-coding genes, 387 being full length and the remainder (159) being partial, mostly as a result of unconfirmed 5\' ends, incomplete genomic sequence or partial gene duplication events. We subsequently identified and removed two full-length genes which we now consider to be antisense transcripts and have extended 13 genes to full length to give a total of 398 full-length protein-coding genes (see \[[@B19]\] for details of the chromosome 22 ORFs). In the other cases of partial annotations we have not been able to extend the annotation sufficiently to allow identification of a complete ORF suitable for cloning. Therefore, for the purposes of this paper, where the aim is to identify clones containing complete ORFs, we only consider genes annotated as full-length protein coding as targets because of the difficulty of defining success for the partial genes.
We first considered the completeness of available full-length human cDNA collections, by comparing the DNA sequences of available cDNA library clones with our targeted set of 398 ORFs. For this analysis we used cDNA sequences downloaded from the major collections in January 2004. The publicly available cDNA collections analyzed were those from the Mammalian Gene Collection (MGC) \[[@B11]\], the full-length long Japan collection (FLJ) \[[@B12]\], the German cDNA Consortium (DKFZ) \[[@B10]\] and the Kazusa cDNA project (KIAA) \[[@B13]\]. In addition, we analyzed a commercially available set of cDNAs from Invitrogen. We aligned each of our target chromosome 22 ORFs to the available cDNA sequences to assess whether clones representing the entirety or any part of each of the chromosome 22 ORFs existed in each collection (Table [1](#T1){ref-type="table"}, and see Materials and methods). This analysis showed that 240 out of 398 ORFs (60%) were represented by a cDNA clone with more than 95% identity over the full length of the ORF in at least one of the collections. In addition, a further 25 ORFs were covered by cDNA clones with gapped matches. However, only 227 (57% of the total ORFs) of these clones maintain the correct reading frame at the amino acid level. Examining the matches from individual cDNA clone collection showed that 80% of the full-length matches were provided by the MGC. This probably reflects the selection process in this program whereby initial sequencing of the ends of cDNA clones was used to select the optimal clone for complete sequencing. The KIAA collection provided full-length matches at approximately the same rate as the MGC, given the number of sequences available (1.25% chromosome 22 full-length matches out of the total MGC collection compared with 1.38% for KIAA) and notably provided the five largest clones matched that maintained the complete ORF (sizes between 4,719 base-pairs (bp) and 3,516 bp), reflecting the emphasis on long clones in the KIAA program. The FLJ and DKFZ collections gave rates of 0.28% and 0.27% respectively, presumably because a smaller proportion of full-length clones were sequenced. Analysis of the chromosome 22 genes from these collections shows that length, but not GC content, of the ORF is a significant factor in cloning success for these collections (Mann Whitney test, *p*\< 0.0007), that is, there is bias against longer ORF clones.
In summary, there is currently a 60% chance of obtaining a full-length cDNA clone from one of these collections, based on a sample of 1% of the human genome. The best single collection (MGC) provides 48% of the clones. This analysis of coverage, based on the subset of full-length protein-coding genes on chromosome 22, mimics the situation occurring in a positional cloning type strategy where one might want to obtain clones for a region identified by genetic mapping. However, it does not assess whether the collections are enriched or depleted for specific classes of gene by function, tissue distribution or level of expression. As chromosome 22 is particularly GC-rich, and compared to other human chromosomes the set of genes we have used for this assessment may be biased towards housekeeping genes with widespread or ubiquitous expression which are known to be enriched in GC-rich regions of the genome. Hence, results for specific classes of genes will differ. In any case, one can expect to obtain roughly half of the clones required from one of these collections. This is testimony to the considerable effort that has gone into constructing the resources, but is also frustrating, because other sources are required to make up the substantial remainder. To investigate whether other approaches could be used to address the completeness of cDNA clone resources, we developed an alternative method which is complimentary to cDNA library sequencing, and tested this approach on the same set of chromosome 22 ORFs.
Strategy for assembling a chromosome 22 ORF clone collection
------------------------------------------------------------
Previous efforts in human to obtain cDNA clones suitable for future functional genomics studies have started by isolating the longest possible cDNA clones \[[@B10]-[@B13]\]. In *Caenorhabditis elegans*, an alternative strategy has been developed that is directly tailored to clone ORFs defined by gene annotations from cDNA libraries into Gateway vectors ready for functional genomics \[[@B20]\]. The strategy we have developed (Figure [1](#F1){ref-type="fig"}) uses genome annotation to define the full-length ORFs of interest. We then aim to amplify the ORF bracketed by short sequences at either end from uncloned primary cDNA (rather than cDNA libraries) using reverse transcription (RT) PCR with modifications to allow efficient and high-throughput application. The overall aim is to obtain cDNA clones containing the defined set of ORFs more efficiently than by cDNA library screening and to access ORFs not present in existing cDNA library collections. This strategy enables a single protocol to be used for all genes, and therefore does not require the import of any previously existing cDNA clones which might be from multiple laboratories and in several vector systems. In addition, it avoids potential biases associated with cloned cDNA libraries by utilizing uncloned cDNA. We chose not to format the ORF directly for a specific recombinational cloning system because this might compromise our ability to isolate some ORFs by RT PCR. Furthermore ORFs cloned into a generic vector will be useful for those who do not want to use a specific vector format. ORFs in clones derived and verified by this method can be readily transferred into recombinational cloning systems by PCR with appropriately designed oligonucleotides.
For the 398 targets, a nested set of two pairs of PCR oligonucleotide primers surrounding each ORF and including a short region of the 5\' and 3\' untranslated regions was identified. As these primers were to be used to extract a fragment containing the ORF from an extremely complex cDNA template, design was not restricted to the sequences at the start and stop of the ORF. A highly processive, proof-reading thermostable DNA polymerase was use to amplify the ORF from a pool of cDNA derived from various tissues using two rounds of PCR. In 76% of cases amplification with KOD Hot Start polymerase was successful in generating a PCR product of expected size under one set of amplification conditions (see Additional data file 2). However, where the expected-sized PCR fragment was not obtained, we were often able to obtain a fragment by subsequent repeat of the procedure with slight modifications including increasing the annealing temperature, using *Pfu*-turbo DNA polymerase as an alternative enzyme for one or both rounds of PCR, or using a cDNA template from a single tissue rather than the pooled cDNA. Fragments of the correct size were cloned into a T-tailed plasmid and the inserts were verified by complete sequencing using vector primers and anticipated gene specific primers. Assembled sequence for each clone was then compared with the expected gene sequence. Clones were accepted as correct versions of the ORF if identical to the expected sequence or if they contained only base changes that were known to be single-nucleotide polymorphisms (SNPs) or resulted in silent codon changes. Clones were also accepted with an alternative splicing event that maintained the ORF. Clones were rejected (for this study) if they contained a nonsynonymous base change that could not be confirmed as a known SNP (\'unconfirmed bases\') or if they resulted from an alternative splice or partially processed mRNA that did not maintain the ORF. When a clone generated from a fragment of the correct size failed validation because of the presence of unconfirmed bases, or retention of a small intron, an alternative clone was picked and sequenced until a correct version was obtained. If alternative splicing or partial processing events gave unacceptable clones, a further round of reamplification was undertaken in order to obtain a correct fragment. Finally, if clone inserts were repeatedly unacceptable as a result of mispriming events, annotation error or amplification of a related gene, a new set of nested oligonucleotide primers were designed.
Process error rate and SNPs
---------------------------
One possible concern with a strategy that involves reverse transcription and multiple rounds of PCR amplification followed by cloning of a single molecule is that the process will introduce base errors that alter the sequence of the final cloned ORF. Analysis of error rate here is complicated by the frequency of SNPs in humans and the fact that the starting cDNA template is a mix of cDNA from multiple human donors. We estimated the error rate from reverse transcription, PCR and the cloning process by sequencing 48 clones (covering 70,656 bases) containing the ORF of the NAGA gene. These were derived by our cloning protocol using cDNA from 10 lymphoblastoid cell lines as a template, as polymorphism would be easier to identify where each cDNA mix could only be one of two haplotypes. We categorized observed base changes as known SNPs if they were found to exist in dbSNP, in ESTs or in independently sequenced cDNA clones. Base changes were categorized as putative errors if no equivalent sequence could be identified. From this analysis we identified six putative base errors, giving an overall estimate of 0.085 errors per kilobase (kb), or one error per 7.8 clones assuming a mean ORF size of 1.5 kb.
Chromosome 22 ORF clone collection
----------------------------------
Applying the strategy outlined above to the 398 chromosome 22 ORFs, we were able to clone and confirm 278 (70%) of the targeted chromosome 22 ORFs (see Additional data file 1). Sequences of the valid ORF clones are available \[[@B19]\], and have been submitted to the EMBL database (accession numbers CR456339 to CR456616). Of these, 253 (91%) were derived from fragments generated with KOD polymerase. The remainder were generated using either an alternative polymerase (16; 6%) or a combination of polymerases (9; 3%) (see Additional data file 2). The universal cDNA pool was used for 249 (90%) of the clones, with 29 (10%) of clones derived from lower-complexity cDNA templates from single tissues. Of the accepted clones, 239 (86%) were the predicted splice form, with the remainder being an alternative splice which maintained the ORF; 183 (66%) clones matched the genomic DNA exactly. Of the 162 deviations from the genomic sequence (from 95 clones), 144 (89%) are previously identified SNPs either in dbSNP or dbEST, and 11 (7%) were not identified as known SNPs but did not alter the amino acid (see Additional data file 3). Seven changes were insertion/deletion events (see below). Of the 144 confirmed SNPs in a total of 372,916 bases (1 SNP every 2,590 bases), 81 were synonymous and 63 were nonsynonymous codon changes. Individual clones contained between one and eight SNPs (see Additional data file 3).
Insertions or deletions that retained the ORF were observed in five clones. None of these significantly altered the ORF, as four cases involved three bases while one involved 12 bases. We also observed a polymorphism in MSE55 which involved the insertion or deletion of six amino acid repeat units and exists in three different alleles. We amplified and sequenced genomic DNA fragments across this region from 152 chromosomes of European ancestry and found that all three alleles are common and in Hardy-Weinberg equilibrium. In this case the clone chosen for the ORF collection was the same allele as seen in the publicly available genomic sequence.
In three cases we obtained clones with insertion/deletion polymorphisms that altered the ORF but were supported by available chromosome 22 sequence. To determine whether to accept these clones as ORF cDNAs, we examined all three in more detail. The clone obtained for gene *APOL4*contains a 2-bp insertion compared to the canonical genomic sequence annotation. This results in a frameshift that substantially extends the ORF from 127 amino acids to 348 amino acids. We designed a PCR reaction to directly interrogate the insertion/deletion and sequenced 144 chromosomes of European ancestry. Both alleles are common in this population, and are in Hardy-Weinberg equilibrium, with the 348-amino acid form being the minor allele at 46.5%. For bK216E10.6 we obtained an ORF clone with a 2-bp insertion compared to the genomic annotation, which results in an ORF that contains an extra 318 amino acids. Using the same strategy we sequenced 150 chromosomes and showed that the sequence producing the shorter peptide is the minor allele with a frequency of 20%, and the alleles are again in Hardy-Weinberg equilibrium. In this case we do not have an accepted clone, as the insertion increased the ORF length beyond the primer sequence. The third gene is *TXN2*which shows a 2-bp insertion compared to the genomic sequence which is also found in an EST (AA586375), but has not been studied further. An insertion/deletion polymorphism that alters the ORF has previously been observed in *MICA*on chromosome 6 \[[@B21]\]. From these examples we concluded that insertion/deletion polymorphisms in ORFs that alter amino acid sequence may be relatively common, and can result in altered proteins. Complete ORF collections for outbred organisms like humans should ultimately address this issue and obtain examples of all common forms of the ORF.
In addition, we were able to amplify a PCR fragment which could be identified as originating from the correct gene for an additional 53 ORFs, but have not yet been able to obtain an acceptable clone because of the presence of unconfirmed bases, or problems with splice forms including partially processed transcripts. In most cases, only one or two amino acids are changed, which could make these clones usable under some circumstances, perhaps after site-directed mutagenesis. It is also possible that these are rarer SNPs that are not currently present in dbSNP. This suggests that by sequencing more examples we will be able to obtain clones for these ORFs in the near future. Thus the clone collection would cover 83% (331) of the targeted ORFs.
Process failure
---------------
In total, we initiated the amplification and cloning process 538 times, excluding initial pilot trials. These 538 events break down as follows. For 180 (45%) targeted ORFs an acceptable clone was generated at the first attempt. Further rounds of clone-picking, reamplification or primer redesign generated a further 99 acceptable clones, 83 clones containing an unconfirmed base alteration, 54 clones containing an alternative splice which lost the ORF, 23 clones containing a rearrangement or erroneous amplification event, 19 clones with retained intron sequences, four clones containing unresolved sequencing problems and 36 clones which were not the expected gene. For 41 genes we were unable to amplify a suitable product or failed to clone the fragment. Hence the efficiency of the process in terms of the return of acceptable clones is approximately 52% (278/540).
A significant area of concern is where we were unable to generate a PCR product at all corresponding to the targeted gene. To find explanations for this type of failure, we examined both the sequence characteristics of the targeted ORF and elements of the experimental design. First we examined the crude differences between the classes of ORFs that we could and could not amplify. Figure [2a](#F2){ref-type="fig"} shows a plot of the distributions of these two classes by GC content and length of ORF. Both GC content and length are significant predictors of success/failure to amplify (Mann Whitney test *p*\< 0.0001), although logistic regression indicates there is no significant interaction between them. This suggests that alternative amplification protocols using different polymerases or PCR additives might result in additional ORFs being obtained. However, we have tested three additional enzymes or mixes (*Pfu*Ultra (Stratagene), Phusion (Finnzymes) and Expand 20 kb+ PCR (Roche)) and additives including DMSO, glycerol and betaine so far without identifying a design that solves the problem.
Next, we explored whether it was possible to amplify any part of the failed target cDNAs from the universal mix. For 51 of the genes where we failed to amplify the expected fragment, we designed additional nested oligonucleotide primer pairs to amplify a short (100-274 bp) sequence across a splice junction. In 39 cases (74%) we amplified a fragment of the correct size and sequence under our standard nested PCR conditions, suggesting that template is present in the cDNA mix for these ORFs (data not shown). Therefore, in most cases it is possible to amplify part of the targeted ORF from the cDNA mix using this protocol, indicating that the level of target in the mix is not limiting in these cases. Given that we know we can amplify parts of many of the problematic genes, one variation that could improve access to larger ORFs in the future would be to amplify larger transcripts in pieces that can then be reassembled into a single clone using appropriate restriction enzyme digestion and ligation or PCR cloning methods.
We also examined whether successful amplification was biased towards genes expressed in many tissues. Su *et al*. \[[@B22]\] have generated microarray data indicating the distribution of expression for many human genes over 47 tissues. We downloaded these data \[[@B23]\] and were able to obtain tissue-distribution data for 206 of our 398 targeted genes. Codifying the diversity of tissues in which the genes were expressed as the proportion of positive tissues, and analyzing for the success or failure of amplification by logistic regression, indicated that the probability of amplifying a gene is not significantly affected by the diversity of its expression (data not shown).
We also examined diversity of expression by analyzing serial analysis of gene expression (SAGE) data derived from 242 *Nla*III SAGE libraries downloaded from the SAGEmap resource \[[@B24]\]. SAGE tags could be uniquely mapped to 315 of the 398 ORFs targeted. Using the number of SAGE libraries in which a SAGE tag for an ORF was found to represent the diversity of tissues in which the gene was expressed, no significant relationship was found with the probability of amplifying a gene (Mann Whitney test, *p*= 0.84). Furthermore, because the SAGE tag data also gives an indication of expression level, we examined whether the mean expression level found by SAGE (mean normalized tags per million SAGE reads) affected probability of expression and again found no significant relationship (Mann Whitney test, *p*= 0.79). Taken together these analyses indicate that the success of our amplification strategy is not significantly influenced by either the range of tissues in which a gene is expressed or the level of expression. Clearly there will be some genes expressed at low levels, at specific times or in specific tissues that will need special treatment, but these data suggest that these cases may be few.
Comparison of the chromosome 22 ORF collection with other cDNA sources
----------------------------------------------------------------------
Returning to the cDNA clone collections, of the 331 targeted genes for which we can obtain either an acceptable clone (278) or a clone of the correct ORF but currently with a problem in its sequence (53), 208 genes also have clones in the cDNA clone collections we analyzed; 123 genes only have clones in the new chromosome 22 ORF set described here. In addition, for 19 genes which are represented in the cDNA clone collections we were unable to isolate a clone (Figures [2b](#F2){ref-type="fig"}, [3](#F3){ref-type="fig"}). This means that 88% (350) of the full-length protein-coding genes on chromosome 22 have cDNA clones. This also suggests that achieving 88% coverage of the readily accessible human ORFeome should be possible with an approach that combines the existing cDNA collections with directed RT-PCR as implemented in this analysis. Of course, because the actual number of human genes is still unknown and a significant number of genes have only partial annotation, there is still an indeterminate number of genes for which there is insufficient annotation to attempt the current strategy.
We analyzed the four classes of genes (isolated by us and in the cDNA collections (BOTH), isolated only here (SANGER), isolated only by the cDNA collections (OTHER) and not isolated (NOT)) by GC content, length and diversity of expression as defined above for microarray data and SAGE using nonparametric analysis of variance (Figure [3](#F3){ref-type="fig"}, and Additional data file 5). ORF length was significantly higher (*p*\< 0.001) for genes not isolated (NOT) as compared to those isolated by us (SANGER) or those isolated both by us and the cDNA collections (BOTH). This suggests, as expected, that longer ORFs are harder to amplify or clone. A significant influence (*p*\< 0.05) was also found for higher GC content in the genes that were either not isolated (NOT) or found only in the cDNA collections (OTHER) compared with the SANGER or BOTH classes, reflecting the influence of GC content on the ability to amplify a cDNA target as discussed above. The only significant difference (*p*\< 0.05) for diversity of expression was between genes cloned only by us (SANGER) and those present in both our set and the cDNA collections (BOTH), with less diversely expressed genes slightly enriched in the SANGER class. This result was seen only in the microarray data, although the effect was also present in the SAGE data at just below significance. This suggests that the method described here may be able to access less widely expressed genes than have been sampled by existing cDNA library sequencing, although the effect is small. Finally, analysis of the mean level of expression of the genes in the four classes based on the normalized SAGE tag count showed no significant difference, indicating that level of expression is not a significant factor for this set of genes.
Conclusions
===========
Even given a high-quality human genome sequence, we still face considerable challenges in identifying and isolating full-length cDNAs and ORFs in order to construct genome-wide clone sets for functional analyses. The method we have described here offers an alternative approach to obtaining full-length ORF clones compared with sequencing or amplifying from cDNA libraries. We have demonstrated that we can readily obtain clones for 70% of the full-length protein-coding genes on chromosome 22, increasing to 83% if we include the largely correct clones which have not passed the confirmation criteria. In addition, a small number of clones (19) that we could not obtain are present in the cDNA collections analyzed, and when these are included, the overall coverage of the known full-length protein-coding genes reaches 88%. While this represents a substantial gain over cDNA sequencing alone, it is clear that complete coverage may require further modification of the approach or additional strategies as well.
The quality control that is introduced by starting with annotated genes on the genomic sequence allows identification of SNPs and artifacts within the clones, and allows confirmation or rejection of each clone as it is generated. The checking process also provides verification of gene structures annotated from assembled ESTs, and in a few cases revealed errors. Our approach also has some advantages for scale-up to whole genomes. The starting point is a single PCR reaction using a universal template, which could be adapted to standard automation platforms. Subsequent steps, including ligation, transformation, clone picking, sequencing and sequence analysis, are all amenable to existing robotic approaches or automation. At present, the gel-purification step of the amplified PCR fragment might be difficult to automate. It is also likely that the final sign-off on the sequence alignment of clones will require human intervention in much the same way as finishing genomic sequences does. However, application to whole genomes demands a high-quality gene annotation to be available for the whole genome.
We have generated a set of quality-controlled ORFs surrounded by a short stretches of 5\' and 3\' untranslated sequence in a uniform vector. The ORF portions of these intermediary clones are currently being amplified and subcloned in frame into a mammalian expression vector which fuses the amino-terminal T7 phage major capsid protein to the amino or carboxy terminus of the protein. We have successfully performed subcellular localization studies using immunofluorescence microscopy with these clones. We are also transferring the ORFs into Gateway pDONR clones (Invitrogen) and subsequently using GFP fusion destination vectors for subcellular localization. The availability of the ORF in a generic vector provides flexibility in the future downstream formats in that the endogenous Kozak sequence and the translation start and stop are maintained, and without additional amino acids from recombination sites. Finally, it is worth noting that this approach could also be applied to amplifying and cloning the many alternatively spliced forms of genes, or ORFs from different individuals or haplotypes. The ability to access the many additional variants beyond the canonical ORFeome could prove a valuable tool for future studies.
Materials and methods
=====================
cDNA sequence sources and websites
----------------------------------
cDNA sequences were downloaded from the websites of the following publicly available cDNA collections in January 2004. For the Mammalian Gene Collection (MGC \[[@B11],[@B25]\], 15,454 sequences were downloaded on 16 January 2004 \[[@B26]\]. For the full-length long Japan collection (FLJ \[[@B12]\]), 25,696 sequence accession numbers were obtained on 16 January 2004 \[[@B27]\] and the sequences were downloaded from the EMBL sequence database. For the German cDNA Consortium (DKFZ \[[@B10]\]) we identified 9,271 sequence accessions on 16 January 2004 \[[@B28]\] and sequences were downloaded from the EMBL database. For the Kazusa cDNA project (KIAA \[[@B13],[@B29]\]), 2,037 sequence accession numbers were obtained on 26 January 2004 \[[@B29]\], and sequences were downloaded from the EMBL database, although two cDNA sequences were missing (KIAA0013 and KIAA0302). In addition, we downloaded 4,361 of the commercially available Invitrogen cDNAs on 8 December 2003 \[[@B30]\] (file datestamp 20 October 2003).
Amplification and cloning of ORFs
---------------------------------
Chromosome 22 gene annotations containing full-length ORFs, as defined in Collins *et al.*\[[@B17]\], but not including the genes described as possible antisense, and 13 genes subsequently completed, provided 398 complete chromosome 22 gene sequences. Nested sets of two pairs of PCR primers surrounding each ORF were designed using Primer3 (Steve Rozen, Helen J. Skaletsky (1996, 1997), Primer3, Code available at \[[@B31]\]) and Perl (version 5.004) scripts to automate the process (see Additional data file 4 for primer pairs designed). Fragments were amplified with the outer primer pair from either 0.1 ng of a pool of cDNAs from 37 tissues (Human Universal QUICK-Clone cDNA, Clontech), or cDNA from a single tissue (cervix, liver, brain, testis, fetal liver or fetal brain obtained as RNA from Stratagene or QUICK-Clone cDNA from Clontech), or cDNA from lymphoblastoid cell lines (European Collection of Cell Cultures, Porton Down, UK HRC collection, cell lines lines C0043, C0092, CO118, C0127, C0139, C0143, C0155, C0167, C0179, C0259, C0573). For the lymphoblastoid cells lines, total RNA was extracted from tissue culture cells with TRIzol reagent (GibcoBRL/Invitrogen). Total RNA was reverse transcribed into cDNA with Superscript II (Invitrogen) according to the manufacturer\'s instructions. The first-round amplification protocol used KOD Hot Start DNA polymerase (Novagen), *Pfu*-turbo Hotstart DNA polymerase (Stratagene) or *Pfu*DNA polymerase (Stratagene) using the manufacturers\' recommended cycling profiles for 30 or 35 cycles in a 25 μl reaction. Fragments were then diluted 1 in 50 with sterile water and 5 μl used as template for a second 25 μl amplification using the inner primer pair (see Additional data file 2 for variant amplification conditions). Additional enzymes including *Pfu*Ultra (Stratagene), Phusion (Finnzymes) and Expand 20 kb+ PCR (Roche) were trialed according to the manufacturers\' recommendations. Fragments of the expected size were gel-purified, extracted with QIAquick spin columns (Qiagen), 3\'-tailed with an adenosine residue using Amplitaq polymerase (Perkin Elmer) and subcloned using the pGEM-T Easy Vector System (Promega). Sequencing template was prepared either by plasmid miniprep or, in the majority of cases, by amplifying clone inserts with vector primers and cleaning the amplified fragment with either QIAquick Gel Extraction Kit or Shrimp alkaline phosphatase (1 unit, Amersham) and ExonucleaseI (1 unit, Amersham) (see below). Sequencing was performed with BigDye terminator v3 Cycle Sequencing Kits (Applied Biosystems) using vector primers, the inner nested primer pair and pairs of primers designed at 600-base intervals along the predicted gene sequence. Sequence was assembled using the contig assembly program CAP3 \[[@B32]\], aligned against the predicted transcript sequence and checked manually.
Sequence comparison and analysis
--------------------------------
The 398 annotated ORF sequences were matched by blastn (version 2.0 MP-WashU, 10 April 2004 \[[@B33]\]) to cDNA collection databases MGC, FLJ, DKFZ, KIAA and Invitrogen. MSPcrunch \[[@B34]\] was used to parse blastn output and exclude matches with lower than 95% identity. ORFs were extracted from each of the matching cDNA sequences using the EMBOSS program getorf \[[@B35]\] and compared to the annotated ORFs using cross\_match (P. Green, unpublished work). The GC content of the ORFs was calculated using the EMBOSS program geecee \[[@B35]\] and Perl scripts (version 5.004) were written to analyze and summarize data.
Microarray data indicating the distribution of expression for many human genes over 47 tissues using the Affymetrix human U95A array \[[@B22]\] was downloaded \[[@B23]\]. Tissue distribution data for 206 genes was obtained and a gene was called as expressed in a tissue sample if the average difference was \> 200 \[[@B22]\]. The tissue expression diversity of a gene was defined as the proportion of positive tissues. Where replicate experiments existed, the highest tissue-expression diversity value was used.
For SAGE data, the 398 target ORF sequences were matched by blastn \[[@B33]\] against Unigene (Homo sapiens, 12 May 2004 Build 170 \[[@B36]\]) and the best (highest identity greater than 99%) full-length matching UniGene cluster was assigned to each ORF. SAGEmap data \[[@B24]\] was downloaded \[[@B37]\] together with a file of tag frequencies \[[@B38]\].
A Perl program was then used to search these files for SAGE tags mapping to each UniGene cluster and the tag counts for each *Homo sapiens Nla*III library (GPL4) were determined. Tag counts were normalized to tags per million for each library, and then averaged to give a mean expression level. Diversity of expression was defined as the number of libraries in which a tag occurred.
Amplification and sequencing of genomic DNA for insertion/deletion analysis
---------------------------------------------------------------------------
Fifty nanograms of genomic DNA from 78 unrelated individuals (ECACC Human Random Control Panel) was amplified in 15 μl reactions containing: 6.7 mM MgCl~2~, 67 mM Tris-HCl, 16.7 mM (NH~4~)~2~SO~4~pH 8.8, 170 μg/ml BSA, 10 mM 2-mercaptoethanol, 500 μM each dATP, dCTP, dGTP, dTTP, 0.04 units/μl Amplitaq, 0.75 μM each primer, 10.13% sucrose, 0.0029% Cresol Red (sodium salt). Reactions were cycled in an MJ thermocycler at 94°C for 5 min, followed by 35 cycles of 30 seconds at 94°C; 30 sec at 65-66°C; 30 sec at 72°C, followed by a final 72°C for 5 min. PCR reactions were treated with 1 unit of shrimp alkaline phosphatase and 1 unit of exonuclease I in reaction buffer supplied by the manufacturer (USB, 10 × buffer - 200 mM Tris-HCl pH 8, 100 mM MgCl~2~) for each 10 μl PCR reaction. Reactions were heated at 37°C for 30 min followed by 80°C for 15 min to inactivate the enzymes. PCR products were then sequenced from both ends using the primers used from the amplification step and BigDye terminator v3 Cycle Sequencing Kits (Applied Biosystems). Sequences were analyzed using GAP4 \[[@B39]\].
Additional data files
=====================
The following additional data files are available with the online version of this article. Additional data file [1](#s1){ref-type="supplementary-material"} lists the 278 successfully cloned ORFs (see also \[[@B19]\]); Additional data file [2](#s2){ref-type="supplementary-material"} lists enzymes and templates used to amplify ORFs; Additional data file [3](#s3){ref-type="supplementary-material"} lists the sequence variation between ORF clone and genomic sequence; Additional data file [4](#s4){ref-type="supplementary-material"} lists the nested oligonucleotide primers designed for the 398 targeted genes; Additional data file [5](#s5){ref-type="supplementary-material"} contains the results of nonparametric ANOVA (Kruskal-Wallis Test) for chromosome 22 genes isolated as cDNA by the method described here only (SANGER), found in the cDNA collections only (OTHER), isolated by both ourselves and the cDNA collections (BOTH) or not isolated (NOT). Mean rank differences and *p*-values are given after Dunn\'s multiple comparisons test. Additional data file [6](#s6){ref-type="supplementary-material"} contains a list of the 278 cloned ORFs and Additional data file [7](#s7){ref-type="supplementary-material"} contains a list of the 398 target ORFs; both files are also available at \[[@B19]\].
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
The 278 successfully cloned ORFs
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
Enzymes and templates used to amplify ORFs
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 3
The sequence variation between ORF clone and genomic sequence
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 4
The nested oligonucleotide primers designed for the 398 targeted genes
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 5
The results of nonparametric ANOVA (Kruskal-Wallis Test) for chromosome 22 genes isolated as cDNA by the method described here only (SANGER), found in the cDNA collections only (OTHER), isolated by both ourselves and the cDNA collections (BOTH) or not isolated (NOT)
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 6
A list of the 278 cloned ORFs
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 7
A list of the 398 target ORFs
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank Graeme Bethel for comments on the manuscript and Jorn Scharlemann and Chris Greenman for advice on statistics. B.A. was funded by the UK Medical Research Council and M.M. was supported by a Medical Research Council Studentship. This work was supported by the Wellcome Trust.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Summary of the ORF cloning method.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Sequence characteristics of cloned ORFs. **(a)**Plot of the distribution of the 398 chromosome 22 ORFs by GC content (%) and length (bases). Closed circles are the 331 ORFs that were isolated as acceptable clones (278) or as clones with the correct ORF but currently with a problem in the sequence (53). Dotted circles are the rest of the ORFs which were not amplifiable or clonable (67). **(b)**Overlap of chromosome 22 ORF clones isolated here with cDNA collections. Analysis of GC content and length for 398 chromosome 22 ORFs, split according to whether the gene has been isolated only by the strategy described here (SANGER, red circles), only in the cDNA collections (OTHER, green triangles), in both (BOTH, black circles), or not at all (NOT, yellow triangles).
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Schematic Venn diagram showing the relationships of the set of ORF clones isolated here compared with the full-length cDNA clones in current high-throughput clone collections (227 maintain the correct reading frame at the amino acid level from Table 1) for the 398 annotated full-length chromosome 22 ORFs. The four different classes of genes are labeled as in the text and Figure 2b.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Analysis of genome-wide collections
:::
cDNA collection Total cDNAs available Matches to 398 chromosome 22 ORFs at more than 95% identity\*
----------------- ----------------------- --------------------------------------------------------------- ---- ---- ---- ----
MGC 15,454 193 14 21 23 17
FLJ 25,696 72 24 25 75 25
DKFZ 9,271 25 10 3 49 16
KIAA 2,035 28 1 1 18 13
Invitrogen 4,361 16 0 61 1 17
Combined 56,817 240^†^ 25 27 39 14
\*For definitions of match types see Materials and methods. Values are not significantly altered by raising the identity required to \>99%. ^†^Only 227 (57% of the total ORFs) of these clones maintain the correct reading frame at the amino acid level.
:::
|
PubMed Central
|
2024-06-05T03:55:51.787382
|
2004-9-30
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545604/",
"journal": "Genome Biol. 2004 Sep 30; 5(10):R84",
"authors": [
{
"first": "John E",
"last": "Collins"
},
{
"first": "Charmain L",
"last": "Wright"
},
{
"first": "Carol A",
"last": "Edwards"
},
{
"first": "Matthew P",
"last": "Davis"
},
{
"first": "James A",
"last": "Grinham"
},
{
"first": "Charlotte G",
"last": "Cole"
},
{
"first": "Melanie E",
"last": "Goward"
},
{
"first": "Begoña",
"last": "Aguado"
},
{
"first": "Meera",
"last": "Mallya"
},
{
"first": "Younes",
"last": "Mokrab"
},
{
"first": "Elizabeth J",
"last": "Huckle"
},
{
"first": "David M",
"last": "Beare"
},
{
"first": "Ian",
"last": "Dunham"
}
]
}
|
PMC545646
|
Background
==========
Genome-wide association studies for complex diseases such as asthma, schizophrenia, diabetes, and hypertension will soon produce genotypes on hundreds of thousands of single nucleotide polymorphisms (SNPs). Due to the large number of SNPs tested and the potential for both genetic and environmental interactions, determining which SNPs modify the risk of disease is a methodological challenge. While the number of genotypes produced by candidate gene approaches will be somewhat less daunting, on the order of hundreds to thousands of SNPs, it will still be a considerable challenge to weed out the noise and identify the SNPs contributing to complex traits.
A logical first approach to dealing with massive numbers of SNPs is to first conduct univariate association tests on each individual SNP, in order to screen-out those with no evidence for disease association. The primary goal of such a procedure is not to prove that a particular variant or set of variants influences disease risk, but to prioritize SNPs for further study. Using a univariate test at this stage will result in low power for SNPs with very small marginal effects in the population, even if the SNPs have large interaction effects. Of course, in addition to taking all individual SNPs, all SNP pairs could also be tested for association. However, when dealing with multiple thousands of SNPs at the outset, such an approach is cumbersome, and raises the question of where to stop: why not all sets of three, four, or even five SNPs as well?
Many model-building methods exist for dealing with large numbers of predictors. For example, stochastic search variable selection (SSVS) \[[@B1]\], a form of Bayesian model selection, has been explored as a tool to discover joint effects of multiple loci in the context of genetic linkage studies \[[@B2]-[@B4]\]. However, these methods are limited in the number of predictors that can be included at one time, causing some researchers to resort to a two-stage approach, in which only main effects are considered in a first stage, and interactions between loci with strong main effects are considered in a second stage. This approach can lead to the loss of important interactions with only weak main effects.
Multivariate adaptive regression splines (MARS) models have also been explored in the context of genetic linkage and association studies \[[@B5],[@B6]\] with some degree of success. However, these and other model selection methods appear to be limited in the number of predictors that can reasonably be accommodated in one analysis, and the types of possible interactions that are allowed must be specified in advance. They are not suited to the initial task of identifying from a massive set of SNPs a subset for further analyses.
Combinatorial partitioning and multifactor dimensionality reduction \[[@B7]-[@B10]\] are closely related methods developed specifically to detect higher-order interactions among polymorphisms that predict trait variation. However, these methods are meant to identify interactions among small sets of SNPs, and have minimal power in the presence of genetic heterogeneity \[[@B10]\]. They are therefore inappropriate for use as a screening tool for searching through thousands of SNPs to identify those contributing to phenotypes in the context of whole-genome association studies. The problem remains: how do we reasonably weed down from thousands or hundreds of thousands of SNPs to a number that can be used by available modeling methods, without losing the interactions that we hope to model in the first place?
An additional concern to be considered is genetic heterogeneity. We define genetic heterogeneity to mean that there are multiple possible ways to acquire a disease or trait that can involve different subsets of genes. Traditional regression models are limited in their ability to deal with underlying genetic heterogeneity (see, *e.g*., \[[@B11]\]). If genetic heterogeneity also leads to phenotypic heterogeneity, then methods that classify individuals into phenotypic subgroups for further analysis can be successful. Likewise, if heterogeneity in genetic etiology is primarily due to ethnic background, sub-dividing samples by self-reported ethnicity or genetically defined subgroups can be a powerful antecedent to data analyses for the identification of complex disease genes. However, even in the realm of Mendelian genetic diseases, heterogeneity is rarely so simple. For example, multiple polymorphisms in each of two different genes are responsible for familial breast cancer in the relatively homogeneous sub-population of Ashkenazi Jewish women \[[@B12]\]. When the root of the heterogeneity is not known *a priori*, traditional regression models, which lump all individuals into a single group and estimate average effects over the entire sample, are unlikely to successfully identify the genetic causes of diseases.
Classification trees and random forests
---------------------------------------
Tree-based methods consist of non-parametric statistical approaches for conducting regression and classification analyses by recursive partitioning (see, e.g., Hastie et al. \[[@B13]\]). These methods can be very efficient at selecting from large numbers of predictor variables such as genetic polymorphisms those that best explain a phenotype. Tree methods are useful when predictors may be associated in some non-linear fashion, as no implicit assumptions about the form of underlying relationships between the predictor variables and the response are made. They are well-adapted to dealing with some types of genetic heterogeneity, as separate models are automatically fit to subsets of data defined by early splits in the tree.
The ease of interpretation of classification trees, along with their flexibility in accommodating large numbers of predictors and ability to handle heterogeneity, has resulted in increasing interest in their application to genetic association and linkage studies. Classification trees have been adapted for use with sibling pairs to subdivide pairs into more homogenous subgroups defined by non-genetic covariates \[[@B14]\], thus increasing the power to detect linkage under heterogeneity \[[@B15]\]. They have also shown promise for the dissection of complex traits for both linkage and association \[[@B16],[@B17]\], and for exploring interactions \[[@B6]\]. A related adaptive regression method has also shown promise in selecting a small number of predictive SNPs from a set of hundreds of potential predictors \[[@B18]\]. Tree methods have also been used to identify homogeneous groups of cases for further analyses \[[@B19]\], and as an adjunct to more traditional association methods \[[@B20]\].
Classification trees are grown by recursively partitioning the observations into subgroups with a more homogeneous categorical response \[[@B21]\]. At each node, the explanatory variable (e.g., SNP) giving the most homogeneous sub-groups is selected. Choosing alternative predictors that produce slightly sub-optimal splits can result in very different trees that have similar prediction accuracy. The Random Forests methodology \[[@B22]\] builds on several other methods using multiple trees to increase prediction accuracy \[[@B23]-[@B25]\]. A random forest is a collection of classification or regression trees with two features that distinguish it from trees built in a deterministic manner. First, the trees are grown on bootstrap samples of the observations. Second, a random selection of the potential predictors is used to determine the best split at each node. For each tree, a bootstrap sample is obtained by drawing a sample with replacement from the original sample of observations. The bootstrap sample has the same number of individuals as the original sample, but some individuals are represented multiple times, while others are left out. The left-out individuals, sometimes called \"out-of-bag\", are used to estimate prediction error. Because a different bootstrap sample is used to grow each tree, there is a different set of out-of-bag individuals for each tree. With a forest of classification trees, each tree predicts the class of an individual. For each individual, the predictions, or \"votes\", are counted across all trees for which the individual was out-of-bag, and the class with the most votes is the individual\'s predicted class. Random forests produce an importance score for each variable that measures its importance. This score can be used to prioritize the variables, much as p-values from test statistics are used.
Using ensembles of trees built in this manner increases the probability that some trees will capture interactions among variables with no strong main effect. Unlike variable selection methods, interactions among predictors do not need to be explicitly specified in order to be utilized by a forest of trees. Instead, any interactions between variables serve to increase the importance of the individual interacting variables, making them more likely to be given high importance relative to other variables. Thus, random forests appear to be particularly well-suited to address a primary problem posed by large scale association studies. In preliminary studies, we have shown the potential of random forests in the context of linkage analysis \[[@B26]\]. Other investigators are beginning to recognize the potential of the Random Forest methodology for studying SNP association \[[@B27]\] and classification \[[@B28]\].
To fully understand the basis of complex disease, it is important to identify the critical genetic factors involved, and to understand the complex relationships between genotypes, environment, and phenotypes. The few successes to date in identifying genes for complex disease suggest that despite carefully collected large samples, novel approaches are needed in the pursuit to dissect the multiple and varying factors that lead to complex human traits. Ultimately, the challenge in identifying polymorphisms that modulate the risk of complex disease is to find methods that can seamlessly handle large numbers of predictors, capitalize on and identify interactions, and tease apart the multiple heterogeneous etiologies. Here, we explore the use of the Random Forest methodology \[[@B22],[@B29]\] as a screening tool for identifying SNPs associated with disease in the presence of interaction, heterogeneity, and large amounts of noise due to unassociated polymorphisms.
Results
=======
Genetic models
--------------
We simulated complex diseases with sibling recurrence risk ratio for the disease (*λ*~s~) fixed at 2.0 and population disease prevalence *K*~*p*~equal to 0.10. These values are consistent with or lower than estimates from known complex genetic traits, such as Alzheimer disease, where estimates of cumulative prevalence in siblings of affected range from 30--40%, compared to a population prevalence of 10% at age 80 \[[@B30]\]. Such traits are understood to be caused by multiple interacting genetic and environmental factors. Our genetic models incorporate both genetic heterogeneity and multiplicative interaction as defined by Risch \[[@B31]\]: we simulate sets of 4, 8, 16, and 32 risk SNPs (\"rSNPs\") in linkage equilibrium, interacting in independent pairs or quartets to increase disease risk, and contributing equally to the overall sibling recurrence risk ratio of 2 and population disease prevalence of 0.10. For simplicity, we simulated the models such that each rSNP pair or quartet accounts for the same proportion of the genetic risk, and each SNP within a pair/quartet is responsible for an equal proportion of the genetic risk. Thus, all of the rSNPs simulated for a model have the same allele frequency and the same observed marginal effect in the population. We denote the models using the shorthand HhMm, where H = h (=2, 4, 8, 16) is the number of heterogeneous systems, and M = m (=2 or 4) is the number of multiplicatively interacting SNPs within each system. For example, 16 loci are responsible for the total *λ*~s~= 2 and *K*~*p*~= .10 for models H4M4 and H8M2, and 32 loci are responsible for models H8M4 and H16M2. Table [1](#T1){ref-type="table"} presents relevant features of our models. The Methods section describes the genetic models in more detail.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Genetic models used for simulating case-control data.
:::
Risk SNPs Case Genotype Correlation
------- ----------- ------- ------ ------ --------- ------ --------------------------- ------ -------
H2M2 4 0.207 2.85 4.71 2.4E-02 0.51 1 0.30 -0.32
H4M2 8 0.160 1.99 2.96 3.9E-04 0.50 1 0.35 -0.12
H8M2 16 0.104 1.66 2.18 8.0E-06 0.56 1 0.32 -0.05
H16M2 32 0.069 1.46 1.78 1.4E-05 0.59 1 0.28 -0.02
H4M4 16 0.282 1.63 1.79 1.2E-08 0.79 1 0.17 -0.06
H8M4 32 0.214 1.34 1.40 2.8E-03 0.86 1 0.14 -0.02
:::
Simulation and analysis
-----------------------
All analyses were performed on 100 replicate data sets of 500 cases and 500 controls. In addition to the rSNPs contributing to the trait, we simulated noise SNPs (\"nSNPs\"), independent of disease status, with allele frequencies distributed equally across the range .01--.99. To simulate the results of an association study, in which we do not expect to be lucky enough to genotype all polymorphisms related to a trait, we included only a subset of the total number of rSNPs in each analysis. We denote the analysis design using the shorthand KkSsNn, where K = k is the total number of rSNPs genotyped in the analysis, S = s is the number of SNPs within each interaction system genotyped, and N = n is the total number of SNPs genotyped in the design. Thus, N-K is the total number of nSNPs in the analysis. For example, suppose the genetic model is H8M4, and the design is K4S2N100. Then out of the total of 8 × 4 = 32 rSNPs that contribute to the trait, four are genotyped: two interacting SNPs from within each of two heterogeneity systems. Six heterogeneity systems are not represented at all in the analysis. In addition to the four genotyped rSNPs, 100-4 = 96 total nSNPs are also genotyped in the design.
### Comparison of raw and standardized importance scores
Random forests version 5 software \[[@B29]\] produces both raw (*I*~*T*~) and standardized (*Z*~*T*~) variable importance scores (see Methods section for definitions of the scores). Little is known about the properties of importance indices under different distributions of the predictor variables. We use simulation to investigate their properties in the context of discrete predictors such as genetic polymorphisms conferring susceptibility to a complex trait.
We first compared the raw and standardized scores, in order to determine whether one might outperform the other in screening. We considered a K4S2N100 analysis design for each genetic model described in Table [1](#T1){ref-type="table"}. *I*~*T*~and *Z*~T~are highly correlated; the average correlation coefficient over 100 replicate data sets ranged from .93 (H8M4) to \>.99 (H2M2 and H4M2) (Table [2](#T2){ref-type="table"}). The average correlation between the ranks based on *I*~*T*~and *Z*~*T*~for the 100 SNPs over the 100 replicate data sets was 0.98 for each of the six models (Table [2](#T2){ref-type="table"}). Comparing the ranks of the four rSNPs among all SNPs, neither importance measure outperforms the other for all models (Table [3](#T3){ref-type="table"}). The mean ranks of the rSNPs for the two measures are significantly different only for the H16M2 and H8M4 models. For H16M2, the average rank of the rSNPs is higher for *Z*~*T*~than for *I*~*T*~. The opposite is true for H8M4 (see Table [3](#T3){ref-type="table"}).
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Summary of the correlation between *I*~*T*~and *Z*~*T*~(\"raw\") and rank(*I*~*T*~) and rank(*Z*~*T*~) (\"rank\") for four rSNPs and 96 nSNPs over 100 replicate data sets: K4S2N100 analysis design.
:::
H2M2 H4M2 H8M2 H4M4 H16M2 H8M4
------ ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- -------
Mean 0.996 0.983 0.990 0.982 0.975 0.982 0.957 0.982 0.964 0.982 0.933 0.982
SD 0.001 0.006 0.003 0.006 0.009 0.006 0.013 0.007 0.013 0.006 0.018 0.007
Min 0.990 0.963 0.980 0.957 0.941 0.960 0.921 0.955 0.926 0.962 0.891 0.953
Max 0.998 0.993 0.996 0.992 0.989 0.993 0.984 0.992 0.983 0.994 0.970 0.992
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Comparison of ranks based on *Z*~*T*~and *I*~*T*~for the four rSNPs over 100 replicate data sets: K4S2N100 analysis design.
:::
H2M2 H4M2 H8M2 H4M4 H16M2 H8M4
----------- ------ ------ ------ ------ ---------- ---------- ------ ------ ------ ------ ------- ------
Mean 2.5 2.5 2.5 2.5 2.51 2.52 2.64 2.61 5.16 5.94 9.35 8.69
SD 1.12 1.12 1.12 1.12 1.13 1.16 1.74 1.62 8.06 8.33 13.67 13.9
Max 4 4 4 4 6 8 23 21.5 77 62.5 83 88.5
p-value\* 0.94 1.00 0.77 0.12 1.32E-20 2.91E-12
\*p-value for the Wilcoxon signed-rank test comparing the rSNP ranks based on *I*~*T*~and *Z*~*T*~
:::
### Ranking SNPs based on Z~T~and Fisher p-value
We next compared the ranking of rSNPs by importance score (Z~T~) to ranking by Fisher Exact test p-value using K4S2N100 and K4S2N1000 analysis designs, where two SNPs from each of the first two interaction systems are in the analysis. Figure [1](#F1){ref-type="fig"} shows the proportion of replicates for which the top ranked 1, top 2, top 3, and top 4 SNPs are the four genotyped rSNPs in the data set for each of the four most complex genetic models. For N100, the random forest Z~T~criterion ranks the four rSNPs as the most significant SNPs more often than the univariate Fisher Exact test association p-value under all genetic models. The difference between the random forest and association p-value ranking is less extreme for N1000. For the H8M4 genetic model, the results do not suggest that one ranking system is better than the other overall. Figure [2](#F2){ref-type="fig"} shows the proportion of replicates for which all rSNPs are among the top N SNPs. In other words, it is the proportion of data sets for which none of the genotyped rSNPs are screened out, if the top ranking N SNPs are chosen for further study. For N100, a consistently higher proportion of replicates ranked using Z~T~contain all of the rSNPs. Thus, for a given probability of retaining all of the rSNPs, more SNPs can be eliminated using the Z~T~criterion than the Fisher exact test p-value. For example, for model H16M2, only 15 SNPs must be retained to have 80% probability that the 4 rSNPs are in the retained set, while 44 SNPs must be retained if the p-value criterion is used. The difference is less dramatic for H8M4: 37 SNPs give 80% probability that the four genotyped rSNPs are in the retained set for the Z~T~criterion, compared to 43 for the p-value criterion. For N1000, the advantage of the Z~T~criterion is clear for the H8M2 and H16M2 models. For H4M4, the advantage of Z~T~is minor, while for H8M4, ranking by Z~T~appears to give poorer results than the p-value criterion for the higher cutoff values of N. A second interpretation of Figure [2](#F2){ref-type="fig"} is that, for any number of retained (not screened-out) SNPs, the probability that all of the genotyped rSNPs are retained is higher for the Z~T~criterion than for the univariate p-value criterion for all but the H8M4 model with 1000 total SNPs.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Proportion of replicates for which the most significant 1, 2, 3, and 4 SNPs are all rSNPs for K4S2N100 and K4S2N1000 analysis designs. Genetic models are listed on the plots. \"RF\" and \"Fisher\" refer to the random forest importance index Z~T~and the Fisher Exact test p-value. See text for notation description.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Proportion of replicates for which all rSNPs are among the top-ranking N SNPs for K4S2N100 and K4S2N1000 analysis designs. Other notation as in Figure 1.
:::
:::
Noticing that the analyses with all SNPs from an interacting system (e.g., the H16M2K4S2 simulations) had a more substantial improvement in ranking using Z~T~over p-value than the analyses with subsets of SNPs from an interacting system, we hypothesized that the interactions among the pairs of analyzed rSNPs influenced the improved ranking performance of the random forests over the univariate tests. To confirm this, we used the H8M4 genetic model and analyzed the data in the following manner. For a constant number of analyzed rSNPs included in the model (K = 4, 8, or 16) and a constant 96 nSNPs, we looked at the effect of increasing S, the number of rSNPs from each interaction system that were genotyped. Thus, for K8S1, along with 96 nSNPs, one SNP from each of the first 8 systems was included in the analysis, while for K8S4, all four SNPs in the first two systems were included in the analysis. For K8S3, three SNPs from the first two systems, and one from the third were included. Assuming that the random forest analysis was taking advantage of the interactions among the rSNPs, and that this was responsible for the improved performance of the random forests over the univariate tests, we expected the Fisher p-values and random forest importance Z~T~to perform similarly when only a single rSNP was genotyped from each system, and the random forests to perform increasingly better than the univariate Fisher tests as S increased from 1 to 4. Figures [3](#F3){ref-type="fig"} and [4](#F4){ref-type="fig"} show the results, which are consistent with this hypothesis. For the Fisher p-values, the proportion of replicates for which the N most significant SNPs were rSNPs is similar for each S. For the random forest importance Z~T~, the S = 1 analyses for each K were similar to the Fisher results, while for each increase in S, the proportion of replicates for which the N most significant SNPs were rSNPs increases (Figure [3](#F3){ref-type="fig"}) and the proportion of replicates for which all rSNPs are present at any cutoff point increases (Figure [4](#F4){ref-type="fig"}). The differences can be substantial: for the H8M4 model, with K = 4 rSNPs in the analysis, the number of most significant SNPs required to have 80% probability that all four rSNPs are included is 50, 34, 22, and 5, respectively for S1, S2, S2, and S4. We conclude that for a given number of rSNPs within a set of potential predictors, the more interacting SNPs there are, and the larger the groups of SNPs that interact, the greater the performance increase of the random forest analysis as compared to a univariate analysis.
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Proportion of replicates for which the most significant N SNPs are all rSNPs. H8M4 genetic model. Analysis designs include 96 noise SNPs; K and S are listed on the plots. Other notation as in Figure 1.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Proportion of replicates for which all rSNPs are among the top-ranking N SNPs for H8M4 genetic model. Analysis designs include 96 noise SNPs; K and S are listed on the plots.
:::
:::
### Magnitude of difference
Beyond simply ranking SNPs, we may wish to use the magnitude of the difference in importance or p-value to determine which subset of top-ranked SNPs should be prioritized for further study. Thus, particularly for the cases where the rSNPs are among the top-ranked SNPs, we want to determine not just that Z~T~ranks interacting rSNPs higher than the univariate test, but also that the differences in rank correspond to differences in magnitude of Z~T~that are meaningful. In other words, we want to know how much \"better\" in terms of Z~T~(or p-value) the rSNPs are than the nSNPs. Toward this goal, we computed the difference between the importance Z~T~of the top ranked rSNP and the top ranked nSNP:
D~max~(Z~T~) = max~*rSNP*~(*Z*~*T*~) - max~*nSNP*~(*Z*~*T*~),
as well as the lowest ranked rSNP and the top ranked nSNP:
D~min~(Z~T~) = min~*rSNP*~(*Z*~*T*~) - max~*nSNP*~(*Z*~*T*~).
Thus, D~min~(Z~T~) is positive when the lowest ranked rSNP is larger than the highest ranked nSNP, and negative when the lowest ranked rSNP is smaller than the highest ranked nSNP.
We computed the analogous quantities, D~max~(-log p) and D~min~(-log p), for the -log~10~transformed Fisher Exact test p-values. In Figure [5](#F5){ref-type="fig"}, we have plotted box plots of these differences for several models using analysis designs K4S2N100 and K4S2N1000. P-values for a paired T-test of whether the mean difference is equal to 0 are also placed on the plot. For H8M2, the lowest ranking rSNP has Z~T~that is significantly higher than the highest ranking nSNP for both N100 and N1000, while the difference in -log10(p) is not significantly different from 0 for N100, and is significantly less than 0 for N1000. These plots illustrate that the positive differences are typically more extreme for Z~T~than for -log p, and that the negative differences are less extreme.
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Distribution of difference in importance ZT between the top ranked rSNP and the top ranked nSNP (Dmax(ZT), and lowest ranked rSNP (Dmin(ZT)) and top ranked nSNP. Dmax(-log p) and Dmin(-log p): differences using -log10 p-value from the Fisher exact test. Beside each boxplot is the p-value for the test of whether the mean difference over the 100 replicates is significantly different from 0. Genetic models listed in plot. Analysis design: K4S2, with N100 and N1000 shown on plot.
:::
:::
Discussion
==========
A key advantage of the random forest approach is that the investigator does not have to propose a model of any kind. This is important in an initial genome-wide or candidate region association study, where little is known about the genetic architecture of the trait. If interactions among SNPs exist, they will be exploited within the trees, and the variable importance scores will reflect the interactions. Therefore, we expect that when unknown interactions between true risk SNPs exist, the random forest approach to screening large numbers of SNPs will outperform a univariate ranking method in finding the risk SNPs among a large number of irrelevant SNPs. Our genetic models for simulation feature both multiplicative interaction and genetic heterogeneity. The multiplicative interaction results in a marginal effect in the population, the size of which is dependent on many factors, including the amount of heterogeneity. Thus, we have highlighted models in which univariate tests still have power, and shown that the random forest analysis can outperform these tests for selecting subsets of SNPs for further study. For models with genetic heterogeneity and interactions resulting in no main effect, similar to the models described by Ritchie et al. \[[@B10]\], the performance of random forests compares considerably more favorably to univariate tests (data not shown), since the univariate tests have no power when main effects are absent. Further investigation of how to determine a cutoff for SNPs to keep for further analysis is needed. Unfortunately, this task is likely to be strongly dependent on information that is impossible for an investigator to know *a priori*, such as the underlying genetic model and the ratio of associated risk SNPs to noise SNPs in an analysis.
Our results from analyses with four risk SNPs among 1000 SNPs suggest that even when a high proportion of the analyzed SNPs are unassociated, a random forest can rank interacting SNPs considerably higher than a univariate test, and that the proportional difference in importance between the risk SNPs and the best of the noise SNPs can be larger on average for a random forest. In our scale-up from 100 to 1000 total SNPs, we kept the number of risk SNPs constant. In practice, as we increase the number of SNPs genotyped, we expect that we will also increase the number of risk SNPs (or SNPs in linkage disequilibrium with risk SNPs) that are captured in an analysis. Thus, as a larger and large proportion of the genome or candidate region is captured by a scan, the more likely we will be to have all or most of sets of SNPs that interact, and thus the more likely we are to be in situations where random forest screening will outperform univariate screening of SNP data.
It is important to consider the tuning parameters for such analyses. Consistent with the recommendations made by Breiman and Cutler \[[@B22],[@B29]\], the number of variables randomly selected at each split seems to have minimal effect over a wide range of values surrounding the square root of the number of covariates (SNPs). Breiman and Cutler do not recommend a method to determine the number of trees necessary for an analysis. The documentation examples typically use on the order of 100--1000 trees, but these examples are primarily in the context of prediction, without computing estimates of variable importance. In our experience with the simulated data sets presented here, in which the truly associated covariates are outnumbered considerably by those that are noise, multiple thousands of trees must be used in order to get stable estimates of the variable importance. In practice, we recommend building several forests for a data set with a given number of trees. If the ranking of variables by importance does not change significantly from forest to forest, then the number of trees is adequate.
We have examined the use of random forests in the context of association studies for complex disease with uncorrelated SNP predictors. Random forests can also be used when predictors are correlated, as is the case with multiple SNPs in linkage disequilibrium within a small genetic region. For any analysis procedure, the more highly correlated variables are, the more they can serve as surrogates for each other, weakening the evidence for association for any one of the correlated variables to the outcome. In a random forest analysis, limited simulations suggest that correlated variables lead to diminished variable importance for each correlated risk SNP (data not shown). One way to limit the problems presented by SNPs in linkage disequilibrium is to use haplotypes instead of SNPs as predictor variables in a random forest. Future challenges include quantifying more completely the effect of linkage disequilibrium among SNPs submitted to a random forest analysis, and developing random forests in the context of haplotypes.
Conclusions
===========
With the increasing size of association studies, two-stage analyses, in which in the first stage a subset of the loci are retained for further analyses, are becoming more common. The most frequently voiced concern for these analyses is that variables that interact to increase disease risk but have minimal main effects in the population will be missed. Random forest analyses address this concern by presenting a summary importance of each SNP that takes into account its interactions with other SNPs. Current implementations of random forests can accommodate up to one thousand of SNPs in one analysis with the computation of importance. Further, there is no reason to restrict the input variables to SNPs. Potential environmental covariates can also be easily accommodated, allowing for SNPs with no strong main effect, but environmental interactions, to be distinguished from unassociated SNPs. We have shown that when unknown interactions among SNPs exist in a data set consisting of hundreds to thousands of SNPs, random forest analysis can be significantly more efficient than standard univariate screening methods in ranking the true disease-associated SNPs highly. After identifying the top-ranked SNPs and other variables, and weeding out those unlikely to be associated with the phenotype, more thorough statistical analyses, including model building procedures, can be performed.
Methods
=======
Variable importance
-------------------
Rather than selecting variables for modeling, a random forest uses all available covariates to predict response. Here, we use measures of variable importance to determine which covariates (SNPs, in our case) or sets of covariates are important in the prediction. Breiman \[[@B22]\] proposed to quantify the importance of a predictor variable by disrupting the dependence between the variable and the response and measuring the change in the tree votes compared to the original observations. In practice, this is achieved by permuting the variable values among all observations in the out-of-bag sample of each tree. If the variable is predictive of the response, it will be present in a large proportion of trees and be near the root of those trees. Observations with a changed variable value may be directed to the wrong side of the tree, leading to vote changes from the right to the wrong class. Conversely, if the variable is not related to the response, it will be present in few trees and, when present, it will be near a terminal node, so that few tree votes will be changed. In Random Forests (version 5) Breiman and Cutler \[[@B29]\] define the importance index as follows. For individual i, let **X**~i~represent the vector of predictor variable values, y~i~its true class, V~j~(**X**~i~) the vote of tree j and t~ij~an indicator taking value 1 when individual i is out-of-bag for tree j and 0 otherwise. Let **X**^(A,j)^= (**X**~1~^(A,j)^,\..., **X**~N~^(A,j)^) represent the vector of predictor variables with the value of variable **A**randomly permuted among the out-of-bag observations for tree j, and **X**^(A)^the collection of **X**^(A,j)^for all trees where N is the total number of individuals in the sample. Letting 1(C) denote the indicator function taking value 1 when the condition C is true and 0 otherwise, the importance index averages over the trees of the forest, and is defined as:
where *N*~*j*~represents the number of out-of-bag individuals for tree j and T is the total number of trees.
The importance index can be standardized by dividing by a standard error derived from the between-tree variance of the raw index *I*~*T*~, . The standardized index is defined as:
The variance  represents the tree to tree variance of *I*~*T*~, rather than the variance of *I*~*T*~due to the sampling of the individuals from a population: the magnitude of *Z*~*T*~increases with the number of trees in the forest, and the number of trees is limited only by computing time. Thus, this standardized index cannot be treated as a Z-score in the traditional sense.
Simulation models and methods
-----------------------------
For simplicity, assume each locus has the same effect, and let (q~0~, q~1~, q~2~) represent the penetrance factors for 0, 1, or 2 risk alleles for an individual locus in a given model. Let G = {g~11~, g~12~, . ., g~HM~} be the multilocus genotype for an individual, where g~hm~(=0, 1, 2 risk alleles) indicates the individual\'s genotype at locus m (=1, . ., M) of heterogeneity system h (=1, . ., H). Then the penetrance for genotype G is defined as:
For example, for model H2M2, an individual with genotype G = {0101} would have penetrance *P*~*G*~= 1 - (1 - *q*~0~*q*~1~)^2^= 2*q*~0~*q*~1~- (*q*~0~*q*~1~)^2^.
The penetrance factors (q~0~, q~1~, q~2~) and risk allele frequencies, as well as other features of our genetic models, are listed in Table [1](#T1){ref-type="table"}. For a given model type, such as H4M4, and a given *λ*~s~and *K*~*p*~, there is a unique allele frequency when we make the assumption that each SNP subunit has equal effect (the given penetrance factor vector) in the population. We chose penetrance factors such that the risk alleles at each locus for the H2M2, H4M2, H8M2, and H16M2 models are approximately additive in effect on the penetrance factor scale. For H4M4 and H8M4, we chose penetrance factor vectors such that the risk alleles show a moderate degree of dominance.
The marginal genotype relative risks (GRRs) listed in Table [1](#T1){ref-type="table"} are the relative penetrances for heterozygote and homozygote carriers of each risk allele, as compared to non-carriers in the population, which would be observed if only a single rSNP were considered at a time. Thus, this is a measure of the observed effect size of each of the rSNPs in the population. The marginal GRRs are modest, in line with what might be expected when there are a large number of small effects contributing to a complex phenotype. For cases, the genotypes for pairs/quartets of SNPs within an interacting system are positively correlated, while SNPs from distinct systems are negatively correlated. The magnitude of the correlations decreases with increasing number of heterogeneity systems and increasing number of equal-effect SNPs interacting within each system.
Analysis
--------
All analyses were performed on 100 replicate data sets of 500 cases and 500 controls. We treated the SNPs as ordinal predictors. Random forests have one primary tuning parameter: \"mtry\" the number of randomly picked covariates to choose among for each split. The Random Forest v5 manual \[[@B29]\] recommends trying the square root of the number of predictors, along with values smaller and larger than the square root, and choosing the value that minimizes the out of bag prediction error rate. We considered both the prediction error and the stability of the variable importance estimates when determining the values of mtry to use and the number of trees to grow. We found that the prediction error rate was very stable over a wide range of mtry for the number of trees we required for consistent measures of importance. We analyzed each replicate data set with 4--16 rSNPs and 96 nSNPs using a random forest of 5000 trees, choosing the best split from among a different randomly-selected set of 35 SNPs at each node (mtry = 35). On average, each replicate data set with 100 total SNPs took 40 minutes to complete on a 2.6 Ghz Intel Xeon processor. For data sets with 4 rSNPs and 996 nSNPs (1000 SNPs total), we used 15000 trees, and chose from among 125 SNPs at each node (mtry = 125). Analysis of each replicate of these data sets took 123 minutes on average. User time could potentially be substantially decreased by parallel-processing: trees could be grown on separate nodes, and combined for analysis of importance. However, parallel tree-building is not yet available in the Random Forest progream. To compare the performance of random forests with that of a univariate, one-SNP-at-a-time approach, we tested for association between genotypes for each individual SNP and disease status using a Fisher Exact test \[[@B32]\].
Ranking of rSNPs
----------------
The random forest analysis produces the raw and standardized importance indices (*I*~*T*~, *Z*~*T*~,), which can be used to rank order the importance of SNPs much as p-values from association tests are used. Using either method, the ranking of the SNPs in an analysis can be used to prioritize which sets of SNPs (or genome regions) should be followed up with further genotyping and/or additional analyses. We use the convention that a rank of 1 is the highest ranking SNP.
We compare the ranking of the raw and standardized importance measures, and further compare these with the rank based on p-values from a test of association, the Fisher Exact test, to determine whether the random forest can better discriminate susceptibility SNPs from SNPs unrelated to disease status when there is interaction and heterogeneity among and between SNPs.
Authors\' contributions
=======================
KLL conceived of and led the design of the study, coordinated all phases of simulation and analysis, and drafted the manuscript. LBH and JS automated the simulation and random forest analysis procedures, and participated in the design and analysis of study results. PVE added critical insight to the development of the genetic models and participated in the interpretation of study results. All authors provided comments on a draft manuscript and read and approved the final manuscript.
Acknowledgments
===============
We thank Kathleen Falls for her help with earlier versions of this work.
|
PubMed Central
|
2024-06-05T03:55:51.791208
|
2004-12-10
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545646/",
"journal": "BMC Genet. 2004 Dec 10; 5:32",
"authors": [
{
"first": "Kathryn L",
"last": "Lunetta"
},
{
"first": "L Brooke",
"last": "Hayward"
},
{
"first": "Jonathan",
"last": "Segal"
},
{
"first": "Paul",
"last": "Van Eerdewegh"
}
]
}
|
PMC545647
|
Background
==========
In 1979 the Canadian task Force on the Periodic Health Examination published one of the first efforts to explicitly characterise the level of evidence underlying healthcare recommendations and the strength of recommendations \[[@B1]\]. Since then a number of alternative approaches has been proposed and used to classify clinical practice guidelines \[[@B2]-[@B28]\].
The original approach used by the Canadian Task Force was based on study design alone, with randomised controlled trials (RCTs) being classified as good (level I) evidence, cohort and case control studies being classified as fair (level II) evidence and expert opinion being classified as poor (level III) evidence. The strength of recommendation was based on the level of evidence with direct correspondence between the two; e.g. a strong recommendation (A) corresponded to there being good evidence. A strength of the original Canadian Task Force approach was that it was simple; the main weakness was that it was too simple. Because of its simplicity, it was easy to understand, apply and present. However, because it was so simple there were many implicit judgements, including judgements about the quality of RCTs, conflicting results of RCTs, and convincing results from non-experimental studies.
For example:
• Should a small, poorly designed RCT be considered level I evidence?
• Should RCTs with conflicting results still be considered level I evidence?
• Should observational studies always be considered level II evidence, regardless of how convincing they are?
The original approach by the Canadian Task Force also did not include explicit judgements about the strength of recommendations, such as how trade-offs between the expected benefits, harms and costs were weighed and taken account of in going from an assessment of how good the evidence is to determining the implications of the results for practice.
The GRADE Working Group is an informal collaboration of people with an interest in addressing shortcomings such as these in systems for grading evidence and recommendations. We describe here a critical appraisal of six prominent systems and the results of the critical appraisal.
Methods
=======
We selected systems for grading the level of evidence and the strength of recommendations that we considered prominent and that included features not captured by other prominent systems. These were selected based on the experience and knowledge of the authors through informal discussion. A description of the most recent version (as of summer 2000) of each of these systems (Appendix 1 to 6), was prepared by one of the authors familiar with the system, and used in this exercise. The following six systems were appraised: the American College of Chest Physicians (ACCP, \[see [Additional file 1](#S1){ref-type="supplementary-material"}\]) \[[@B21]\], Australian National Health and Medical Research Council (ANHMRC, \[see [Additional file 2](#S2){ref-type="supplementary-material"}\]) \[[@B17]\], Oxford Centre for Evidence-Based Medicine (OCEBM, \[see [Additional file 3](#S3){ref-type="supplementary-material"}\]) \[[@B16]\], Scottish Intercollegiate Guidelines Network (SIGN, \[see [Additional file 4](#S4){ref-type="supplementary-material"}\]) \[[@B18]\], US Preventive Services Task Force (USPSTF, \[see [Additional file 5](#S5){ref-type="supplementary-material"}\]) \[[@B22]\], US Task Force on Community Preventive Services (USTFCPS, \[see [Additional file 6](#S6){ref-type="supplementary-material"}\]) \[[@B25]\].
These descriptions of the systems were given to the twelve people who independently appraised the six systems, all of the authors minus GEV appraised the six systems, three of the authors (DH, SH and DO\'C) appraised as a group and reported as one (see contributions). The 12 assessors all had experience with at least one system and most had helped to develop one of the six included systems. Twelve criteria described by Feinstein \[[@B29]\] provided the basis for assessing the sensibility of the six systems.
Criteria used to assess the sensibility of systems for grading evidence and recommendations
-------------------------------------------------------------------------------------------
1\. To what extent is the approach applicable to different types of questions? -effectiveness, harm, diagnosis and prognosis (No, Not sure, Yes)
2\. To what extent can the system be used with different audiences? -patients, professionals and policy makers (Little extent, Some extent, Large extent)
3\. How clear and simple is the system? (Not very clear, Somewhat clear, Very clear)
4\. How often will information not usually available be necessary? (Often, Sometimes, Seldom)
5\. To what extent are subjective decisions needed? (Often, Sometimes, Seldom)
6\. Are dimensions included that are not within the construct (level of evidence or strength of recommendation)? (Yes, Partially, No)
7\. Are there important dimensions that should have been included and are not? (No, Partially, Yes)
8\. Is the way in which the included dimensions are aggregated clear and simple? (No, Partially, Yes)
9\. Is the way in which the included dimensions are aggregated appropriate? (No. Partially, Yes)
10\. Are the categories sufficient to discriminate between different levels of evidence and strengths of recommendations? (No, Partially, Yes)
11\. How likely is the system to be successful in discriminating between high and low levels of evidence or strong and weak recommendations? (Not very likely, Somewhat likely, Highly likely)
12\. Are assessments reproducible? (Probably not, Not sure, Probably)
No training was provided and we did not discuss the 12 criteria prior to applying them to the six systems.
Our independent appraisal of the six systems were summarised and discussed. The discussion focused on differences in the interpretation of the criteria, disagreement about the judgements that we made and sources of these disagreements, the strengths and weaknesses of the six systems, and inferences based on the appraisals and subsequent discussion.
In order to identify important systems that we might have overlooked following our appraisal of these six systems we also searched the US Agency for Health Care Research and Quality (AHRQ) National Guidelines Clearing House for organisations that have graded two or more guidelines in the Clearing House using an explicit system \[[@B30]\]. These systems were compared with the six systems that we critically appraised.
Results
=======
There was poor agreement among the 12 assessors who independently assessed the six systems. A summary of the assessments of the sensibility of the six approaches to rating levels of evidence and strength of recommendation is shown in Table [1](#T1){ref-type="table"}.
Discussion
==========
The poor agreement among the assessors likely reflects several factors. Some of us had practical experience using one of the systems or used additional background information related to one or more grading systems, and we may have been biased in favour of the system with which we were most familiar. Each criterion was applied to grading both evidence and recommendations. Some systems were better for one of these constructs than the other and we may have handled these discrepancies differently. In addition each criterion may have been assessed relative to different judgements about the evidence, such as an assessment of the overall quality of evidence for an important outcome (across studies) versus the quality of an individual study. Some of the criteria were not clear and were interpreted or applied inconsistently. For example, a system might be clear and not simple or visa versa. We likely differed in how stringently we applied the criteria. Finally, there was true disagreement.
There was agreement that the OCEBM system works well for all four types of questions. There was disagreement about the extent to which the other systems work well for questions other than effectiveness. It was noted that some systems are not intended to address other types of questions and it is not clear that it is important that a system should address all four types of questions that we considered (effectiveness, harm, diagnosis, prognosis), although criteria for assessing individual studies must take this into account \[[@B31],[@B32]\].
Most of us did not find that any of the systems are likely to be suitable for use by patients. Almost all agreed that the ACCP system was suitable for professionals and most considered that the USPSTF system was suitable for professionals. There was not much agreement about the suitability of any of the other systems for professionals or about the suitability of any of the systems for policy makers, although most assessed the USTFCPS system to be suitable for policy makers.
There was no agreement that any of the systems are clear and simple, although USPSTF, ACCP and SIGN systems were generally assessed more favourably in this regard. It was generally agreed that the clearer a system was the less simple it was; e.g. the OCEBM system is clear but not simple for categorising the level of evidence. There was some confusion regarding whether we were assessing how clear and simple the system was to guideline developers (as some interpreted this criterion) or how clear and simple the outcome of applying the system was to guideline users (as others interpreted this criterion). Either way, the simpler a system is the less clear it is likely to be.
Most of us judged that for most of the systems necessary information would not be available at least sometimes. The OCEBM system came out somewhat better than the other systems and lack of availability of necessary information was considered to be less of a problem for the USTFCPS system. However, the OCEBM and USTFCPS systems were considered by most to be missing dimensions which may, in part, explain why missing information was considered to be less of a problem. This would be the case to the extent the missing dimensions were the ones for which information would often or sometimes not be available. The dimension for which we considered that information would most often be missing was trade-offs; i.e. knowledge of the preferences or utility values of those affected. Additional problems were identified in relationship to complex interventions and counselling, particularly with the USTFCPS and USPSTF systems. It was pointed out that the USTFCPS system addressed this problem by including availability of information about the intervention as part of its assessment of the quality of evidence.
Most of the systems were assessed to require subjective decisions at least to some extent. The OCEBM system again stood out as being assessed more favourably, although it may be related to omission of dimensions that require more subjective decisions. Judgement is clearly needed with any system. The aim should be to make judgements transparent and to try to protect against bias in the judgements that are made by being systematic and explicit.
Inclusion of dimensions that are not within the constructs being graded was not considered a problem for most of the systems by most of us. Several people considered that it might be a problem for the USTFCPS and USPSTF systems. On the other hand, all of the systems were evaluated to be missing at least one important dimension by at least one person. The challenge of missing dimensions were considered less of a problem for the ACCP and ANHMRC systems. There was not agreement about any of the systems having a clear and simple approach to aggregating the dimensions, although this was considered to be less of a problem for the ACCP, SIGN and USTFCPS systems.
There was also not agreement on the appropriateness of how the dimensions were aggregated. This was considered to be more of a problem for the ANHMRC and USTFCPS systems than the other four systems, all of which were considered to have taken an approach to aggregating the dimensions that was at least partially inappropriate by more than half of us.
Most of us considered that most of the systems had sufficient categories, with the exception of the ANHMRC system. There was almost agreement that the USPSTF system has sufficient categories. We agreed that it is possible to have too many categories as well as too few, the OCEBM system being an example of having too many categories.
There was not agreement that any of the systems are likely to discriminate successfully, although everyone thought that the ACCP, SIGN and USPSTF systems are somewhat to highly likely to discriminate. Lastly, we largely agreed that we were not sure how reproducible assessments are using any of the systems, although half of us considered that assessments using the ANHMRC system are unlikely to be reproducible and about 1/3 considered that assessments using the OCEBM and ACCP systems are likely to be reproducible.
We identified 22 additional organisations that have produced 10 or more practice guidelines using an explicit approach to grade the level of evidence or strength of recommendations. Another 29 have produced between two and nine guidelines using an explicit approach. These systems include a number of minor variations of the six systems that we appraised in detail.
There was generally poor agreement between the individual assessors about the scoring of the six approaches using the 12 criteria. However, there was general agreement that none of these six prominent approaches to grading the levels of evidence and strength of recommendations adequately addressed all of the important concepts and dimensions that we thought should be considered. Although we limited our appraisal to six systems all of the additional approaches to grading levels of evidence and strength of recommendations that we identified were, in essence, variations of the six approaches that we had critically appraised. Therefore we are confident that we did not miss any important grading systems available at the time when these assessments were undertaken.
Based on discussions following the critical appraisal of these six approaches, we agreed on some conclusions:
• Separate assessments should be presented for judgements about the quality of the evidence and judgements about the balance of benefits and harms.
• Evidence for harms should be assessed in the same way as evidence for benefits, although different evidence may be considered relevant for harms than for benefits; e.g. local evidence of complication rates may be considered more relevant than evidence of complication rates from trials for endarterectomy.
• Judgements about the quality of evidence should be based on a systematic review of the relevant research.
• Systematic reviews should not be included in a hierarchy of evidence (i.e. as a level or category of evidence). The availability of a well-done systematic review does not correspond to high quality evidence, since a well-done review might include anything from no studies to poor quality studies with inconsistent results to high quality studies with consistent results.
• Baseline risk should be taken into consideration in defining the population to whom a recommendation applies. Baseline risk should also be used transparently in making judgements about the balance of benefits and harms. When a recommendation varies in relationship to baseline risk, the evidence for determining baseline risk should be assessed appropriately and explicitly.
• Recommendations should not vary in relationship to baseline risk if there is not adequate evidence to guide reliable determinations of baseline risk.
Conclusions
===========
Based on discussions of the strengths and limitations of current approaches to grading levels of evidence and the strength of recommendations, we agreed to develop an approach that addresses the major limitations that we identified. The approach that the GRADE Working Group has developed is based on the discussions following the critical appraisal reported here and a pilot study of the GRADE approach \[[@B33]\]. Based on the pilot testing and the discussions following the pilot, the GRADE Working Group has further developed the GRADE system to its present format \[[@B34]\].
The GRADE Working Group has continued to grow as an informal collaboration that meets one or two times per year. The group maintains web pages <http://www.gradeworkinggroup.org> and a discussion list.
Competing interests
===================
DA has competing interests with the US Preventive Services Task Force (USPSTF), PAB has a competing interest with the US Task Force on Community Preventive Services (USTFCPS), GHG and HS have competing interests with the American College of Chest Physicians (ACCP), DH, SH and DO\'C have competing interests with the Australian National Health and Medical Research Council (ANHMRC), BP has competing interests with the Oxford Centre for Evidence-Based Medicine (OCEBM). Most of the other members of the GRADE Working Group have experience with the use of one or more systems of grading evidence and recommendations.
Contributions
=============
DA, PAB, ME, SF, GHG, DH, SH, AL, DO\'C, ADO, BP, HS, TTTE, GEV & JWW Jr as members of the GRADE Working Group have contributed to the preparation of this manuscript and the development of the ideas contained herein, participated in the critical assessment, and read and commented on drafts of this article. GHG and ADO have led the process. GEV has had primary responsibility for coordinating the process.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1472-6963/4/38/prepub>
Supplementary Material
======================
::: {.caption}
###### Additional File 1
American College of Chest Physicians (ACCP), a brief description of the ACCP approach.
:::
::: {.caption}
######
Click here for file
:::
::: {.caption}
###### Additional File 2
Australian National Health and Medical Research Council (ANHMRC), a brief description of the ANHMRC approach.
:::
::: {.caption}
######
Click here for file
:::
::: {.caption}
###### Additional File 3
Oxford Centre for Evidence-based Medicine (OCEBM), a brief description of the OCEBM approach.
:::
::: {.caption}
######
Click here for file
:::
::: {.caption}
###### Additional File 4
Scottish Intercollegiate Guidelines (SIGN), a brief description of the SIGN approach.
:::
::: {.caption}
######
Click here for file
:::
::: {.caption}
###### Additional File 5
U.S. Preventive Services Task Force (USPSTF), a brief description of the USPSTF approach.
:::
::: {.caption}
######
Click here for file
:::
::: {.caption}
###### Additional File 6
U.S. Task Force on Community Preventive Services (USTFCPS), a brief description of the USTFCPS approach.
:::
::: {.caption}
######
Click here for file
:::
Acknowledgements
================
We wish to thank Peter A Briss for participating in the critical assessment and for providing constructive comments on the process. The institutions with which members of the Working Group are affiliated have provided intramural support. Opinions expressed in this paper do not necessarily represent those of the institutions with which the authors are affiliated.
Figures and Tables
==================
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Summary of assessments of the sensibility of six approaches to rating levels of evidence and strength of recommendation
:::
Criteria^1^ ACCP ANHMRC^2^ USTFCPS OCEBM SIGN USPSTF^3^
---------------------------------------- ------ ----------- --------- ------- ------ ----------- --- --- ---- --- --- ---- --- --- ---- --- ---- ----
1\. Applicable to different questions:
Effectiveness 12 2 8 1 11 12 1 11 2 9
Harm 1 11 5 5 1 7 4 1 11 1 3 8 2 2 7
Diagnosis 7 3 2 4 4 2 9 3 12 5 2 5 2 2 7
Prognosis 6 3 3 2 5 3 9 2 1 11 4 3 5 3 3 5
2\. Can be used by:
Professionals 1 11 1 5 3 7 4 1 6 5 5 7 3 8
Policy makers 1 5 6 1 5 3 1 2 9 3 7 2 2 6 4 1 4 6
Patients 4 5 3 5 5 6 3 3 9 3 7 5 4 6 1
3\. Clear and simple 1 5 6 2 6 1 2 8 2 2 4 5 1 5 6 1 4 7
4\. Information not available 8 4 1 5 3 1 6 5 4 8 1 7 4 1 9 2
5\. Subjective decisions 2 10 5 2 2 5 5 2 7 5 5 7 2 9
6\. Inappropriate dimensions 1 3 8 1 6 2 4 6 1 10 1 2 8 1 4 6
7\. Missing dimensions 1 6 5 2 2 4 5 4 3 9 3 1 5 4 3 2 5 4
Aggregation of dimensions:
8. Clear and simple 1 5 6 4 1 2 2 2 7 3 4 4 6 6 2 7 2
9. Appropriate 6 5 3 1 1 3 4 4 2 5 4 1 4 6 1 5 5
10\. Sufficient categories 1 4 6 4 2 1 5 7 2 2 7 1 2 9 1 10
11\. Likely to discriminate 7 5 2 5 1 1 9 2 2 4 6 5 7 4 7
12\. Assessments reproducible 1 8 3 4 4 2 7 2 7 4 1 8 2 10
^1^See Criteria described in Methods.
^2^Two people did not assess the ANHMRC because it was more descriptive and others responded not applicable for some questions.
^3^One person did not assess the USPST and one person had two responses on questions 3 and 4.
:::
|
PubMed Central
|
2024-06-05T03:55:51.795631
|
2004-12-22
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545647/",
"journal": "BMC Health Serv Res. 2004 Dec 22; 4:38",
"authors": [
{
"first": "David",
"last": "Atkins"
},
{
"first": "Martin",
"last": "Eccles"
},
{
"first": "Signe",
"last": "Flottorp"
},
{
"first": "Gordon H",
"last": "Guyatt"
},
{
"first": "David",
"last": "Henry"
},
{
"first": "Suzanne",
"last": "Hill"
},
{
"first": "Alessandro",
"last": "Liberati"
},
{
"first": "Dianne",
"last": "O'Connell"
},
{
"first": "Andrew D",
"last": "Oxman"
},
{
"first": "Bob",
"last": "Phillips"
},
{
"first": "Holger",
"last": "Schünemann"
},
{
"first": "Tessa Tan-Torres",
"last": "Edejer"
},
{
"first": "Gunn E",
"last": "Vist"
},
{
"first": "John W",
"last": "Williams"
}
]
}
|
PMC545648
|
Background
==========
Cluster sample study designs are a cost-effective way of sampling difficult to reach populations. Examples include sampling schools to obtain cluster samples of students or medical practitioners to sample patients\[[@B1]\]. Cluster samples violate the simple random sample assumption of independence of observations, since observations are sampled from within the selected cluster -- defined as the primary sampling unit. Observations within a cluster may be more alike than observations across clusters. This intra-cluster correlation leads to increased variation between clusters compared to the variation within clusters. Failure to account for intra-cluster correlation when designing a study where participants are recruited within clusters will lead to an under-powered study. To allow for any loss in power and precision, a cluster sample requires a larger sample size to answer the same research question as a study using simple random sampling \[[@B2]-[@B4]\].
Both the size of the intra-cluster correlation and the number of observations sampled within each cluster influence the power of the study. Even for a small intra-cluster correlation, as is often found in general practice and community samples, the loss of power can be appreciable, particularly if the size of the cluster is large\[[@B1],[@B4]\].
Estimates of the size of intra-cluster correlations come from post hoc examination of studies that have used either allocation or sampling by cluster and a number of intervention studies have published observed intra-cluster correlation coefficients \[[@B4]-[@B6]\]. Many intervention studies however, still fail to report intra-cluster correlation coefficients\[[@B7]\] and there is even less information reported on survey studies that employ a cluster sample\[[@B1],[@B8]\]. The lack of published estimated intra-cluster correlations continues to hamper the design of studies that employ a cluster sample\[[@B9]\].
Intra-cluster correlation varies within a study and depends on the outcome under analysis\[[@B1],[@B4],[@B6]\]. The intra-cluster correlation of the same outcome may also vary across studies depending on the primary sampling unit, and whether outcomes are reported as prevalence rates or modeled in association with other variables\[[@B1],[@B4]\]. Researchers need reliable estimates of the intra-cluster correlations, specific to the primary sampling unit and selected outcomes of interest when making sample size calculations. These estimates will assist in deciding the trade off between cluster number and cluster sub-sample size in a study design\[[@B10]\]. There is however, little published on the estimated intra-cluster correlation coefficients in the Australian context, especially for primary health surveys where the health practitioner is the primary sampling unit.
Research questions
------------------
In one Australian study Carlin and Hocking\[[@B1]\] examined the intra-cluster correlation in two cross-sectional cluster surveys of school children that used the school as the primary sampling unit. The researchers observed that design effects for sociodemographic variables were larger than for morbidity related variables. Furthermore intra-cluster correlation was greater for descriptive outcomes such as prevalence estimates, means and proportions than for measures of association between variables such as regression coefficients and odds ratios. We wanted to examine whether these patterns could be generalised to other large cluster survey studies in the primary care setting.
This paper reports some of the intra-cluster correlations observed in the Bettering the Evaluation and Care of Health (BEACH) program, a large cross-sectional survey of general practice patient encounters in Australia, where a random sample of general practitioners was used as the primary sampling unit.
The BEACH study draws a new random sample of Australian general practitioners (GPs) each year, and this provided an opportunity to assess the stability of intra-cluster correlation coefficients across successive samples. If a population is re-sampled using the same cluster survey design, will the intra-cluster correlation coefficient for a particular outcome be the same across samples? This analysis takes an applied approach, examining the observed intra-cluster correlations for a range of demographic, morbidity and treatment outcomes.
Methods
=======
The BEACH program is a continuous study of general practice activity commenced in 1998. The BEACH method is described in detail elsewhere and a brief summary is reported below\[[@B11]\].
Cluster sample design
---------------------
A random sample of approximately 1,000 general practitioners (GPs) is drawn each year from the Health Insurance Commission\'s sampling frame of the population of GPs in Australia. The GP population is randomly ordered into a list and GPs are recruited sequentially from the list, with re-randomisation of the sampling frame every three years\[[@B11],[@B12]\]. GPs are sampled without replacement and have one chance of selection over three years. Sampling is continuous across the year, with around 20 GPs participating in the study in any one week. Each GP completes details of 100 consecutive patient encounters. The GP is the primary sampling unit (PSU), while the primary unit of inference is the patient encounter.
Data elements
-------------
A single page encounter form contains elements including:
• Patient age and sex.
• Whether English was the main language spoken at home.
• Whether the patient holds an Australian health care concession card.
• The problems managed by the GP at the encounter (up to four problems per encounter).
• Treatments received at the encounter, including medications, other procedures, referrals and orders for pathology and imaging tests.
Although sample weights are calculated each year for population estimates\[[@B13]\], the outcomes reported in this paper are unweighted to allow us to calculate estimates of the intra-cluster correlation based on the observed variance in the sample data.
Descriptive outcomes
--------------------
Descriptive outcomes were defined as rates, means and percentages of single variables, e.g. mean age, per cent of encounters with female patients, per cent of encounters where at least one respiratory problem was managed. BEACH samples the GP-patient encounter, not independent patients. If a patient returns to the GP in the sampling period then that patient contributes two (or more) encounters to the sample. Therefore BEACH estimates are not true \"prevalence\" rates because the denominator, the population of GP-patient encounters, is many times larger than the population of all general practice patients. To avoid misunderstanding in this paper we have used the term \"descriptive\" rather than \"prevalence\" to report single variable estimates and their accompanying intra-cluster correlation coefficients. Descriptive rates are interpreted for example as \"Proportion of patients at encounter who are female\".
### Demographic variables
Demographic variables include patient sex, patient age, whether the patient held a health care concession card and whether the main language spoken at home was not English.
### Morbidity variables
Problems were classified using the International Classification of Primary Care (ICPC-2)\[[@B14]\]. The upper level of ICPC-2 classifies problems according to the body system involved, for example skin problems, respiratory problems, cardiovascular problems and problems of the digestive system etc. There are an additional three chapters for psychological problems, social problems and problems of a general or unspecified nature. Morbidity estimates are expressed as the percent of patient encounters where at least one problem from the chapter was managed. The total number of problems managed by the GP at the encounter was also included as an outcome.
### Treatment outcomes
Treatment outcomes included the proportion of encounters that resulted in at least one medication, the proportion that received at least one referral, the proportion that received at least one order for an imaging test and the proportion receiving at least one order for a pathology test.
Association outcomes
--------------------
Intra-cluster correlation coefficients were calculated for associations between variables using logistic regression e.g.: the effect of patient age (predictor) on the rate of cardiovascular problems (outcome).
Design effect
-------------
Obtaining the sample size for cluster designs involves calculating the sample size under the assumption of simple random sampling and then inflating the number of observations to allow for the design effect of the cluster sample. The design effect (Deff) of an outcome has been defined as the ratio of the variance taking into account the cluster sample design and the variance of a simple random sample (srs) design with the same number of observations\[[@B1]\].
Deff = Variance~(clustersample)~/Variance~(srs)~
Intra-cluster correlations and their standard errors for the outcome variables were calculated using the method described by Carlin & Hocking\[[@B1]\]. Specifically STATA 7 was used to calculate the design effects using the \"survey estimator\" procedures, which were purposefully designed to analyse complex survey data. STATA 7 calculates the design effect directly from the ratio of the estimated variances\[[@B15]\].
The intra-cluster correlation coefficient (ICC) was then calculated from the design effect using the formula:
*ICC*= (*Deff*- 1)/(*k*- 1)
and the approximate standard error (SE) of the intra-cluster correlation was calculated using the formula\[[@B1],[@B6]\]:
where *m*= number of clusters, *k*= mean number of observations per cluster.
The intra-cluster correlations and respective 95% confidence intervals for the second BEACH year sample from the period April 1999 to March 2000 were compared against those in the year 5 sample (April 2002 to March 2003) to assess whether the intra-cluster correlations were consistent across samples over time. All calculations specified the GP as the primary sampling unit.
Results
=======
From April 1999 to March 2000, 1,047 GPs were recruited, recording a sample of 104,700 patient encounters. From April 2002 to March 2003, 1,008 GPs were recruited and 100,800 encounters recorded. Table [1](#T1){ref-type="table"} shows the age and sex distribution of the two samples of GPs compared with the sampling frame of the population of Australian GPs in the year April 2002 to March 2003\[[@B13]\]. The two GP samples were comparable to the GP population in terms of distribution by age, sex and state.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Comparison of GP participants and all active recognised Australian GPs.
:::
**BEACH April 1999--March 2000 % (95%CI) (N = 1,047)** **BEACH April 02--March 03 % (95%CI (N = 1,008)** **Australian GPs April 02 to March 03 % (N = 17,884)\[13\]**
------------------- -------------------------------------------------------- --------------------------------------------------- --------------------------------------------------------------
**Males** 69.6 (66.8,72.4) 64.8 (61.8,67.7) 66.8
**Age group**
\<35 8.4 (6.7,10.1) 7.3 (5.7,9.0) 9.7
35--44 32.4 (29.6,35.3) 26.6 (23.9,29.3) 25.1
45--54 32.4 (29.6,35.3) 35.2 (32.3,38.2) 33.1
55+ 26.7 (24.1,29.4) 30.9 (28.0,33.7) 32.0
**State**
NSW 37.4 (34.5,40.4) 39.6 (36.7,42.7) 33.6
Victoria 20.1 (17.7,22.5) 18.8 (16.4,21.3) 24.5
Queensland 20.2 (17.8,22.6) 21.2 (18.7,23.8) 18.5
South Australia 9.1 (7.3,10.8) 6.2 (4.7,7.6) 8.7
Western Australia 8.8 (7.1,10.5) 8.9 (7.2,10.7) 9.5
Tasmania 2.4 (1.5,3.3) 2.8 (1.8,3.8) 2.9
ACT 1.1 (0.5,1.8) 1.4 (0.6,2.0) 1.5
NT 0.9 (0.3,1.4) 1.1(0.4,1.7) 0.8
:::
The two samples of patient encounters were similar in terms of demographics (Table [2](#T2){ref-type="table"}) In the year 1999--00, 59.0% of encounters were with female patients compared with 59.3% in 2002--03. The samples were comparable in terms of the mean age of patients, the proportion of health care card holders, and encounters with patients from a non-English speaking background.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Descriptive parameters of demographic, morbidity and treatment variables with design effects (Deff), intra-cluster correlation coefficients (ICC) and standard errors of ICC (SE) for sample year April 1999 to March 2000 (N = 1,047 general practitioners): compared with ICC and SE for sample April 2002 to March 2003 (N = 1,008 GPs).
:::
**1999--2000 (N = 1,047 GPs)** **2002--2003 (1,008 GPs)**
----------------------------------------- -------------------------------- ---------------------------- ------------- ------- -------------
**Demographics**
Sex (% female) 59.0 (.39) 6.4 .055 (.003) 59.3 .066 (.003)
Age (years) -- mean 44.5 (.31) 16.6 .159 (.006) 45.4 .153 (.006)
Holds health care card (%) 40.1 (.70) 21.4 .206 (.007) 42.7 .209 (.008)
Patient language ^(*c*)^(%) 7.0 (.53) 45.6 .451 (.011) 8.8 .423 (.011)
**Morbidity**
Number of problems (per 100 encounters) 149.5 (.86) 13.6 .127 (.005) 148.7 .141 (.006)
*Problem by ICPC-2 chapter^(b)^*
Cardiovascular (%) 15.2 (.29) 6.7 .057 (.003) 15.3 .056 (.003)
Respiratory (%) 21.0 (.26) 4.2 .032 (.002) 19.0 .040 (.002)
Psychological (%) 10.6 (.25) 6.8 .059 (.003) 10.6 .061 (.003)
Endocrine/Metabolic (%) 8.8 (.18) 4.2 .032 (.002) 10.1 .031 (.002)
Blood (%) 1.7 (.08) 3.7 .027 (.002) 1.4 .007 (.001)
Digestive (%) 9.6 (.12) 1.8 .008 (.001) 9.7 .010 (.001)
Eye (%) 2.8 (.06) 1.5 .005 (.001) 2.6 .003 (.001)
Musculoskeletal (%) 16.3 (.23) 4.1 .032 (.002) 16.5 .045 (.002)
Skin (%) 16.1 (.19) 2.7 .017 (.001) 15.9 .042 (.002)
General unspecified (%) 14.0 (.22) 4.3 .034 (.002) 15.8 .043 (.002)
**Treatment (% of encounters)**
Any medications 67.0 (.37) 6.6 .056 (.003) 64.4 .068 (.003)
Any referrals 11.2 (.20) 4.1 .031 (.002) 12.0 .033 (.002)
Any pathology tests ordered 14.7 (.26) 5.8 .048 (.002) 16.0 .046 (.002)
Any imaging tests ordered 6.9 (.15) 3.8 .028 (.002) 7.8 .029 (.002)
^(*a*)^Average number of observations per cluster *k*= 100, except age (*k*= 99.2) and sex (*k*= 98.8).
^(*b*)^Per cent of encounters where at least one problem from the chapter was managed.
^(*c*)^Patient speaks a language other than English at home.
:::
Descriptive ICCs (March 1999--April 2000)
-----------------------------------------
### Demographics
For descriptive estimates of demographic variables the intra-cluster correlation ranged from 0.055 for sex of patient at encounter to 0.451 for language spoken by the patient at home. (Table [2](#T2){ref-type="table"}). With a standard cluster size of 100 encounters this produced design effects ranging from 6.4 for patient sex to 45.6 for non-English speaking background.
### Morbidity (ICPC body chapter)
For descriptive estimates of the management rates of morbidity problems, the intra-cluster correlations ranged from 0.005 for estimates of eye problems to 0.059 for estimates of psychological problems, with design effects of between 1.5 and 6.8 respectively.
### Treatments
The intra-cluster correlation coefficients for treatments received ranged from 0.028 for any imaging tests ordered to 0.056 for any medications.
Association ICCs
----------------
For bivariate relationships between an outcome and predictor, the association ICCs were considerably smaller than the descriptive ICCs (Table [3](#T3){ref-type="table"}). This pattern was observed for both demographic and morbidity outcomes. When analysing the association between holding a health care card and other demographic variables, the ICCs ranged from 0.012 for patient sex to 0.128 for language background (Table [3](#T3){ref-type="table"}), which were smaller than for the descriptive estimate of the percentage holding a health care card (Table [2](#T2){ref-type="table"}).
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Associations between demographic and morbidity variables, measured as odds ratios, with design effect (Deff) and intra-cluster correlation coefficients (ICC) with standard errors (SE) for sample year April 1999 to March 2000 (N = 1,047 general practitioners): and ICC and SE for sample April 2002 to March 2003 (N = 1,008 GPs).
:::
**1999--2000 (N = 1,047 GPs)** **2002--2003 (1,008 GPs)**
-------------------------------- ------------------------- -------------------------------- ---------------------------- ------------- -------------
**a) Demographic**
Patient holds health care card Female patient 1.07 2.2 .012 (.001) .018 (.001)
Age (years) 1.03 6.5 .056 (.003) .073 (.003)
Patient language^(*b*)^ 1.26 13.7 .128 (.005) .114 (.005)
Patient language^(*b*)^ Female patient 0.94 3.7 .028 (.002) .024(.001)
Age (years) 1.00 11.2 .104 (.004) .098 (.004)
**b) Morbidity Chapters**
Cardiovascular Female patient 0.90 1.3 .003 (.001) .004 (.001)
Age (years) 1.05 2.1 .011 (.001) .017 (.001)
Holds health care card 2.55 2.4 .014 (.001) .018 (.001)
Patient language^(*b*)^ 1.17 5.2 .042 (.002) .034 (.002)
Respiratory Female patient .86 1.3 .003 (.001) .003 (.001)
Age (years) .99 2.5 .015 (.001) .022 (.001)
Holds health care card .89 2.1 .011 (.001) .017 (.001)
Patient language^(*b*)^ 1.17 2.9 .020 (.001) .025 (.002)
Psychological Female patient 1.13 2.2 .013 (.001) .008 (.001)
Age 1.01 3.3 .024 (.001) .022 (.001)
Holds health care card 1.89 2.4 .014 (.001) .020 (.001)
Patient language^(*b*)^ .74 5.0 .040 (.002) .026 (.002)
Endocrine/metabolic Female patient .95 1.6 .006 (.001) .004 (.001)
Age 1.03 2.1 .011 (.001) .012 (.001)
Holds health care card 1.63 2.4 .014 (.001) .009 (.001)
Patient language^(*b*)^ 1.58 3.3 .023 (.001) .019 (.001)
\* Each predictor is fitted alone, each line represents a separate model.
^(*a*)^Average number of observations per cluster *k*= 100, except age (*k*= 99.2) and sex (*k*= 98.8).
^(*b*)^Patient speaks a language other than English at home.
:::
When analysing the association between cardiovascular problems as the outcome and selected demographic variables, the ICCs ranged from 0.042 (patient language as the predictor) to 0.003 (patient sex as the predictor)(Table [3](#T3){ref-type="table"}) compared with the larger ICC of 0.057 when describing the rate of cardiovascular problems (Table [2](#T2){ref-type="table"}).
Comparison of year 2 (April 1999 to March 2000) and year 5 (April 2000 to March 2003)
-------------------------------------------------------------------------------------
For descriptive outcomes the intra-cluster correlations for year 2 and year 5 samples there was consistency in the patterns of ICCs across samples. (Table [2](#T2){ref-type="table"} and Figure [1](#F1){ref-type="fig"}). One exception was for the management of problems related to the blood system, where the descriptive ICC in 1999--00 was 0.027 (95% CI: 0.024--0.030), three times that observed in 2002--03 (0.007, 95% CI: 0.005--0.008). This was influenced by one GP in the 1999--00 sample who managed blood-related problems at more than 50% of encounters. When this GP was removed, the descriptive ICC for blood related problems in 1999--00 was 0.011 (95%CI: 0.009--0.013), much closer to the ICC observed in 2002--03.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Intra-cluster correlation(ICC) and 95% confidence intervals for descriptive and morbidity outcomes in two BEACH samples, April 1999--March 2000 (N = 1047 GPs) and April 2002--March 2003(N = 1008 GPs) \* Total problems = the number of problems managed at the current encounter.
:::
:::
The intra-cluster correlation for associations between morbidity outcomes and demographic predictors are shown in Table [3](#T3){ref-type="table"} and Figure [2](#F2){ref-type="fig"}. Although the intra-cluster correlations for associations between variables across each year were statistically significantly different for some outcomes, in these instances the ICCs were very small and the difference between samples was less than 0.01.
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Intra-cluster correlation (ICC) and 95% confidence interval for association between morbidity outcomes with health care card status as predictor in two BEACH samples, April 1999--March 2000 (N = 1,047 GPs) and April 2002--March 2003 (N = 1,008 GPs)
:::
:::
Discussion
==========
The pattern of intra-cluster correlation and design effects observed in the BEACH study agree with Carlin and Hocking\'s observations in other cluster sample surveys\[[@B1]\]. Generally we found that sociodemographic variables had larger intra-cluster correlation coefficients than morbidity or treatment variables and outcomes fitted with explanatory variables had smaller intra-cluster correlation coefficients than outcomes reported as descriptive rates. Therefore when designing cluster sample surveys, the effect of the intra-cluster correlation on power calculations, depends on whether the main outcomes of interest are demographic or morbidity variables, and whether the main aims of the study are descriptive or predictive\[[@B1]\].
We further demonstrated that for a large range of variables the size and patterns of intra-cluster correlation coefficients for particular outcomes were mostly consistent over different sample periods. This indicates that intra-cluster correlation is quite stable when re-sampling a population using the same primary sampling unit, where the number of clusters is sufficiently large. This repeatability demonstrates the validity of using published intra-cluster correlation coefficients to predict intra-cluster correlation in future studies of similar design.
Precision can be an issue for estimating intra-cluster correlation, especially for studies with a small number of clusters\[[@B10]\]. The large number of clusters in this study gave good precision in the estimated intra-cluster correlation coefficients\[[@B10]\]. There are no other published studies in general practice in Australia with such a large sample of clusters and a large balanced sample of observations per cluster, thus estimating intra-cluster correlation with a high degree of precision.
The BEACH study also has the advantage of being a nationwide survey of general practice where the generalisability to Australian general practice has been well-described\[[@B11]\]. Most research in primary care in Australia is done through general practice, so estimating the intra-cluster correlation for a range of outcomes is important for future researchers who intend to use the GP as the primary sampling unit. The good representation of general practice in the BEACH study, the large sample of clusters and the large cluster size, allow the intra-cluster correlation coefficients reported here to be generalisable to other general practice surveys. These reported intra-cluster correlation coefficients are also likely to be useful for intervention studies that use the GP as the unit of randomisation\[[@B1]\].
Treatments received at the encounter are outcomes that arise as a result of the GP-patient interaction. Treatments are directly related to GPs\' behaviour and so might be expected to be highly correlated within clusters. However we found that the intra-cluster correlation coefficients for medications, referrals, imaging and pathology orders were of a similar order to those for health problems managed.
The difference across samples in the intra-cluster correlation coefficients for the management of blood system problems indicates that, even in large samples, intra-cluster correlation may be influenced by GPs in the sample who specialise in particular areas of health.
Demographic variables are collected in the BEACH study for the purpose of understanding health status and health service use and these variables are likely to be correlated to a patient\'s choice of GP. Furthermore a patient can be sampled more than once if they return to the GP during the survey period. Therefore the intra-cluster correlation estimated for demographic variables may be larger than those that have been reported in community based surveys\[[@B1],[@B8]\].
Conclusions
===========
As with cluster randomised trials, researchers in primary health care need access to a range of estimates of intra-cluster correlation for the successful planning of cluster survey study designs. We have reported relatively stable intra-cluster correlation coefficients for a range of outcomes across two independent random samples in a large-scale representative survey of general practice in Australia. The demonstrated precision and reliability of the estimated intra-cluster correlations indicate that these coefficients will be useful for calculating sample sizes in future general practice surveys that use the GP as the primary sampling unit.
Abbreviations
=============
GP: General Practitioner
ICC: Intra-cluster correlation coefficient
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors contributions
=====================
SK conceived the research questions, undertook the analysis and wrote the main draft of the manuscript. PC participated in formulating the research questions and the design of the analysis, undertook a literature search and assisted in the writing of the main draft and subsequent revisions of the manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2288/4/30/prepub>
Acknowledgements
================
The authors wish to thank the GPs who participated in the study and all BEACH staff past and present who made this work possible. The organisations who contributed financially to the conduct of the BEACH study from April 1999 to March 2003 were: The Australian Department of Health and Ageing; The Department of Veteran Affairs; The National Occupational Health and Safety Division; AstraZeneca Pty Ltd (Australia); Aventis Pharma Pty Ltd; Roche Products Pty Ltd, Janssen-Cilag Pty Ltd; Merck Sharp and Dohme (Australia) Pty Ltd. The General Practice Statistics and Classification Unit is a collaborating unit of the Australian Institute of Health and Welfare.
|
PubMed Central
|
2024-06-05T03:55:51.798384
|
2004-12-22
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545648/",
"journal": "BMC Med Res Methodol. 2004 Dec 22; 4:30",
"authors": [
{
"first": "Stephanie A",
"last": "Knox"
},
{
"first": "Patty",
"last": "Chondros"
}
]
}
|
PMC545776
|
Background
==========
Development of eukaryotic cells towards particular cell fates is regulated by complex and dynamic changes in gene expression. These changes, when monitored on a genome-wide scale, provide a detailed framework for the analysis and modeling of cellular development. To monitor patterns of gene expression it is important to be able to isolate cells at precise stages along a developmental pathway. Well-developed procedures for cell culture and single-cell PCR techniques have allowed genome-wide changes in gene expression to be monitored during animal cell differentiation \[[@B1]-[@B4]\].
Transcriptomic studies of single cell types in plants have focused on diploid sporophytic cell types, including undifferentiated cell suspensions \[[@B5],[@B6]\], leaf epidermal and mesophyll cells \[[@B7]\], stomatal guard cells \[[@B8]\] and cultured mesophyll cells \[[@B9],[@B10]\]. These studies have provided valuable information about gene expression in single cell types; however, their coverage of the transcriptome has been limited and/or hampered by low RNA yields from individual cells, requiring mRNA preamplification steps that can bias the complementary RNA (cRNA) \[[@B11],[@B12]\]. Moreover, such studies have not involved the use of the most comprehensive tools for monitoring gene expression that are now available for *Arabidopsis*- which include the Affymetrix ATH1 gene arrays. Recently, a significant advance in transcriptome analysis of plant cell types has been achieved through fluorescence-activated cell sorting of cell-type marked and protoplasted root cells using Affymetrix ATH1 micorarrays \[[@B13]\]. This has provided a near-comprehensive transcriptomic view of cell-fate determination at three developmental stages in five different domains of the root apex.
In contrast to such enabling technologies and procedures developed for sporophytic cell types there have been no studies that provide a genome-wide perspective of cell fate determination and differentiation during haploid gametophyte development. The haploid male gametophyte generation of flowering plants has a simple and well-defined pathway of development and consists of two- or three-celled pollen grains that deliver two sperm cells via the pollen tube to the embryo sac at fertilization. The highly reduced cell lineage and functional specialization of the male gametophyte are thought to be key factors in the reproductive fitness and evolutionary success of flowering plants. Moreover, pollen ontogeny provides an attractive model of cellular development in which to dissect the regulation of cell growth and division, cellular differentiation and intercellular communication (for reviews see \[[@B14]-[@B17]\]).
Recent progress in understanding of molecular and cellular aspects of pollen development has emerged from genetic studies that have identified mutants in *Arabidopsis*that affect all phases of pollen development \[[@B18]-[@B29]\]. In parallel, cDNA libraries and databanks have been obtained for sperm cells in maize, lily, tobacco and *Plumbago zelanica*\[[@B30]-[@B33]\]. Despite such advances there is limited information about developmental changes in gene expression associated with particular phases of male gametophyte development. Our objective was to develop procedures to enable the isolation of populations of microspores and developing pollen grains at precise developmental stages in *Arabidopsis*and to analyze changes in gene expression from unicellular microspores to mature differentiated pollen grains. A particular advantage of the male gametophyte generation is that developing microspores and pollen grains are symplastically isolated. This facilitates access to viable cell populations at different stages of haploid development without contaminating sporophytic cells.
Some initial progress has been made towards the definition of the male gametophytic transcriptome of *Arabidopsis*using serial analysis of gene expression (SAGE) \[[@B34]\] and Affymetrix AG microarrays that harbor probes for approximately 8,000 different genes \[[@B35],[@B36]\]. These studies have provided valuable insight into the complexity of gene expression in mature pollen and the extent of overlap between male gametophytic and sporophytic gene expression (reviewed in \[[@B37]\]). However, these studies monitored the expression of only 30% of the annotated genes in *Arabidopsis*and analyzed mRNA populations only in mature differentiated pollen grains. These studies do not, therefore, provide a developmental perspective of gene expression during development and differentiation of the male gametophyte.
Here we describe spore isolation procedures for *Arabidopsis*and the use of Affymetrix ATH1 Genome Arrays to analyze transcript expression profiles throughout four successive stages of male gametophyte development in *Arabidopsis*. Isolated spore populations were large enough to enable RNA extraction for direct microarray hybridization without any preceding amplification step that could lead to bias in expression signals between stages or between genes within individual stages. Progression from proliferating microspores to terminally differentiated pollen was characterized by large-scale repression and the activation of a unique collection of late-program genes during pollen maturation. Putative male gametophyte-specific genes and distinct clusters of coexpressed genes are identified, including key groups of regulatory factors including cell cycle, transcription and translation factors. Bioinformatic and experimental data are used to address the importance of transcription and translation during pollen germination and tube growth
Results
=======
Isolation and characterization of developing spores
---------------------------------------------------
Transcriptome profiling throughout microgametogenesis in *Arabidopsis*required the introduction of a procedure for the isolation of homogeneous populations of viable spores at precisely defined stages of development. The method was based on centrifugation of isolated mixed spores in a Percoll step-gradient \[[@B38],[@B39]\]. Large homogeneous spore populations at three developmental stages were collected: uninucleate microspores (UNM), bicellular pollen (BCP) and immature tricellular pollen (TCP). In addition, a homogeneous mature pollen grain (MPG) population was isolated from open flowers according to Honys and Twell \[[@B35]\].
Microscopic examination of isolated spore populations revealed no contaminating sporophytic cells and little or no other cellular debris (Figure [1a](#F1){ref-type="fig"}). Vital staining revealed more than 90% viable spores at each stage (data not shown). The purity of spore populations was evaluated by DAPI staining (Figure [1b-e](#F1){ref-type="fig"}). The UNM population was the most homogeneous, containing 95% uninucleate microspores, 2.5% tetrads and 2.5% bicellular pollen. The BCP population was 77% pure, but also contained some tetrads (3.5%), microspores (12%) and tricellular grains (7.5%). The TCP population comprised 88% tricellular pollen and 12% bicellular pollen. The MPG population was 100% pure with approximately 2% aborted pollen.
Developmental changes in the male gametophytic transcriptome
------------------------------------------------------------
*Arabidopsis*ATH1 Genome Arrays were used to explore the dynamics of gene expression throughout male gametophyte development in comparison with sporophytic tissues. Microarrays were hybridized with cRNA probes made from total RNA purified from isolated spores. Hybridization data from two biological replicates derived from independently grown populations of plants were compared. Only genes with a positive hybridization signal and a detection call value of 1 in both experiments were scored as expressed. Microarray data from each pair of replicates were highly correlated, with correlation coefficients of 0.986 (UNM), 0.972 (BCP), 0.991 (TCP) and 0.971 (MPG). Complete microarray data are publicly available at the European *Arabidopsis*Stock Centre (NASC) microarray database \[[@B40]\]. Sporophytic ATH1 Genome Array datasets were downloaded from the NASC website \[[@B41]\]. This provided transcriptome data for seedlings at open cotyledon stage (COT, stage 0.7 \[[@B42]\]), leaves (LEF, stage 6.0), petiole (PET, stage 3.9), stems (STM, stage 6.1), roots (ROT), root hair zone (RHR, stage 1.02), and suspension cell cultures (SUS). Genes that were consistently expressed in replicate sporophytic datasets were identified using the same algorithm used for gametophytic data.
We have previously confirmed and validated the expression pattern of 15 putative pollen-specific genes identified using Affymetrix AG arrays by reverse transcription-PCR analysis \[[@B35]\]. Similarly we validated the current ATH1 datasets by RT-PCR analysis in two separate experiments that included analysis of 41 genes encoding predicted glycosylphosphotidylinositol-anchored proteins (GAPs) \[[@B21]\] and 16 cation/proton exchanger proteins \[[@B43]\]. In both experiments the expression patterns of all genes tested that were identified as pollen-expressed, or pollen-specific by ATH1 analysis were confirmed by RT-PCR.
The ATH1 Genome Array harbors oligonucleotide probes representing 22,591 genes based on the Arabidopsis Genome Initiative annotation. This represents 80.7% of the most recent estimate of 28,000 protein-coding genes in *Arabidopsis*\[[@B44]\]. Of these, 13,977 genes gave a consistently positive expression signal in at least one stage of male gametophyte development, representing 61.9% of the unigene targets on the microarray. The majority of these were expressed in the two earliest developmental stages; 11,565 in microspores and 11,909 in bicellular pollen (Figure [1f](#F1){ref-type="fig"}). After pollen mitosis II, there was a sharp decline in the number of diverse transcripts to 8,788 in tricellular pollen and 7,235 in mature pollen.
To identify genes expressed preferentially or specifically in developing male gametophytes, hybridization data was compared with sporophytic ATH1 datasets (COT, LEF, PET, STM, ROT and RHR; see Additional data file 1). Transcripts with a consistent positive expression signal in at least one stage of male gametophyte development and a zero signal in any sporophytic dataset were considered male gametophyte-specific. In total, 1,355 specific transcripts were identified, representing 9.7% of the male gametophytic transcriptome. The number of male gametophyte-specific transcripts ranged from 857 (BCP) to 625 (MPG). Thus, in contrast to the decline in the total number of diverse transcripts expressed, the representation of male gametophyte-specific transcripts increased, from 6.9% and 7.2% at UNM and BCP-stages to 8.0% and 8.6% at TCP and MPG-stages respectively.
Analysis of the distribution of transcripts among three abundance classes: high (up to 10-fold less than the maximum signal), medium (10- to 100- fold less) and low (more than 100-fold less) (Figure [1f](#F1){ref-type="fig"}), revealed a decrease in the proportion of transcripts forming the high-abundance class during development from 20% to 12%. On the contrary, there was sharp increase in the proportion of mRNAs forming the low-abundance class after pollen mitosis II from 4% (UNM) to 14% (MPG). Moreover, 55% of low-abundance transcripts at MPG stage represented repressed mRNAs expressed more abundantly at earlier stages. Thus, the dramatic decrease in the number of transcripts expressed between bicellular and tricellular stages is paralleled by redistribution of mRNA from the high to the low abundance classes. These changes may be associated with reduced cellular activities and cell differentiation processes together with preferential expression of certain classes of genes during pollen maturation. This finding is in accord with the over-representation of cytoskeleton, cell-wall and signaling-related genes that comprise 26% of the high-abundance transcripts at MPG stage. In particular, the average expression signals of cytoskeleton, cell-wall and signaling-related transcripts were increased by 3.1, 3.7 and 2.3-fold, respectively, compared with the UNM stage.
Scatter-plot analysis was used to examine in more detail the complexity of the mRNA decline after PMII. The expression levels of individual genes were normalized using a scale of 0 to 100. Genes coexpressed in pairs of datasets were plotted using a logarithmic scale and a correlation coefficient (*R*value) calculated (Figure [2](#F2){ref-type="fig"}). There was an extremely high correlation (*R*= 0.96) between the transcriptomes of UNM and BCP, the two earliest developmental stages (Figure [2a](#F2){ref-type="fig"}). These stages are closely related, with a moderate increase in the expression of a number of genes at BCP stage. The profiles of the two latest developmental stages, TCP and MPG, were also very similar (Figure [2c](#F2){ref-type="fig"}, *R*= 0.858), but with greater deviation than the early stages. The scatter plot of TCP and MPG revealed the shift between extreme mRNA abundance classes as described above. This was more evident when bicellular and tricellular stages were compared (Figure [2b](#F2){ref-type="fig"}). The scatter of gene expression values and the low correlation (*R*= 0.541), provide evidence that the major quantitative shift in transcriptome size between BCP and TCP stages is not simply the result of repression, but also involves the activation of new groups of genes associated with pollen maturation. The lack of correlation (*R*= 0.194) between gene expression profiles in uninucleate microspores and mature pollen (Figure [2d](#F2){ref-type="fig"}), also reflects the pronounced change in cell status from proliferating microspore to terminally differentiated pollen.
The relationship between cell proliferation activities and transcriptome profiles was examined by comparison of early UNM and late MPG stages with a publicly available suspension cell culture dataset. These comparisons demonstrated that the microspore transcriptome was significantly more similar to that of cell suspensions (*R*= 0.474) than to mature pollen (*R*= 0.194). This is also in accord with the lack of correlation between transcriptome profiles of mature pollen and cell suspensions (*R*= 0.13).
Co-regulated clusters of gametophytic genes
-------------------------------------------
To identify gametophytic genes that may form co-regulated clusters, all 13,977 male gametophyte-expressed genes were hierarchically clustered using EPCLUST clustering and analysis software. Application of a threshold value of 0.05 resulted in the definition of 39 gene clusters covering all phases of male gametophyte development (Figure [3](#F3){ref-type="fig"}; see also Additional data files 1 and 2). Cluster 37 contained 735 early genes (5.3% of all gametophytic genes) with positive expression signals only at UNM stage. Transcriptome data reflect steady-state mRNA profiles that result from the combination of transcription and mRNA turnover rates. In this regard, some transcripts grouped in early cluster 37 may be inherited through meiosis and/or from the tetrad stage.
The majority of male gametophyte-expressed genes (52%) were grouped into four clusters (25, 27, 29 and 35) comprising early expressed genes repressed after PMII. Several large gene clusters collectively containing 1,899 genes (13.6%) were associated with pollen maturation. These were activated or upregulated between BCP and TCP stages, forming clusters 5, 7, 11, 13, 18-24, 26, 28, 38 and 39. In contrast, a discrete set of 298 genes forming cluster 17 was upregulated only after TCP stage. In total, 3,342 late genes (24%), forming clusters 1-3, 6, 8 and in particular, cluster 17, encode proteins that are likely to function during post-pollination development.
Expression of regulatory genes throughout male gametophyte development
----------------------------------------------------------------------
We focused our further analysis on three key categories of genes with likely regulatory significance in male gametophyte development; core cell-cycle genes, transcription factors and core translation factors (Figure [4](#F4){ref-type="fig"}). Core cell-cycle genes \[[@B45]\] were defined according to TAIR \[[@B46]\]. Genes comprising *Arabidopsis*transcription factor families were derived by compilation of data available at The Ohio State University *Arabidopsis*Gene Regulatory Information Server \[[@B47]\], data published in \[[@B48]\] and databases homology searches. Recent annotations of the MADS-box and bHLH transcription factor gene families were defined according to \[[@B49],[@B50]\], respectively. Core translation factors \[[@B51]\] were defined according to TAIR \[[@B46]\].
### Core cell-cycle genes
Among 61 core cell-cycle genes, 55 genes were present on the ATH1 GeneChip and 45 (82%) of these were expressed in the male gametophyte (see Additional data file 1). Representative(s) of all families and subfamilies were expressed. The majority of gametophytic core cell-cycle genes showed similar expression profiles (Figure [4a](#F4){ref-type="fig"}), with a decline in mRNA abundance after UNM stage to zero (or low levels) at TCP and MPG stages. This pattern is consistent with the termination of proliferation of the microspore and generative cell before pollen maturation.
### Putative transcription factors
We identified 1,594 genes encoding putative transcription factors that were divided into 34 gene families (see Additional data file 1). Their representation on the ATH1 GeneChip was 1,350 (85%). Of these, 608 (45%) were expressed in the male gametophyte, including 54 (15.7%) that were male gametophyte-specific. There were distinct differences in the representation of large transcription factor families (with over 25 members) in the gametophyte. Among those over-represented were the p-coumarate 3-hydroxylase (C3H) family (67% of family members present on the ATH1 GeneChip), the CCAAT family (64%), C2H2 zinc finger proteins (57%), the WRKY family (53%), the bZIP family (51%), the TCP family (50%) and the GRAS family (50%). In contrast, the AUX/IAA (20%), HSF (33%), bHLH (34%), NAC (34%), AP2-EREBP (35%), HB (36%), R2R3-MYB (37%), MADS (37%) and C2C2 zinc finger (37%) gene families were all under-represented.
The dominant expression pattern of transcription factor genes reflected the general repression of mRNA diversity between BCP and TCP stages (Figure [4b](#F4){ref-type="fig"}). Besides a limited number of constitutively expressed genes, two major transcription factor gene groups could be distinguished. One contained a major group of early-expressed genes and the second a smaller group of genes that were more abundantly expressed late during pollen maturation. The same general tendency was apparent when the profiles of individual transcription factor families were analyzed (exemplified by the C3H family, Figure [4d](#F4){ref-type="fig"}). Several gene families comprised predominantly early-expressed genes. These were the NAC, WRKY, TCP, ARF, Aux/IAA, HMG-box and Alfin-like gene families (Figure [4c-e](#F4){ref-type="fig"}, Additional data file 3). Complete lists of transcription factor gene families and their expression profiles are presented in Additional data files 1 and 3.
### Core translation factors
Among 100 annotated core translation factor genes, 82 were present on the ATH1 GeneChip and 75 (91%) of these were expressed in the male gametophyte (see Additional data file 1). The vast majority of translation factor genes belonged to the early group and these were strongly expressed (Figure [4g](#F4){ref-type="fig"}). Reflecting the constitutive requirement for protein synthesis, only six genes showed male gametophyte-specific expression. These were: *AtPAB3*(At1g22760), *AtPAB6*(At3g16380), *AtPAB7*(At2g36660), *AteIF2*-*B3*(At3g07920), *AteIF4G*-like (At4g30680) and *AteIF6*-*2*(At2g39820). There was a striking over-representation of poly(A)-binding (PAB) proteins among the male gametophyte-specific genes; seven out of eight PAB genes were male gametophyte-expressed, three of which were specific. Moreover, two of these gametophyte-specific PAB genes were among the few late pollen genes encoding translation initiation factors (Figure [4h](#F4){ref-type="fig"}).
Integrating transcriptomic and experimental data
------------------------------------------------
The rapid decline of mRNAs encoding translation initiation factors after bicellular stage and the parallel *de novo*synthesis of a new set of late pollen transcription factors, suggested storage of translation factors and ongoing transcription after pollen germination. Therefore we investigated the dependence of *Arabidopsis*pollen germination and tube growth on transcription and translation. Pollen was cultured with increasing concentrations of actinomycin D and cyclohexmide to examine the importance of transcription and translation, respectively. Actinomycin D had only moderate effects on both pollen germination and tube growth even at high concentrations (Figure [5a](#F5){ref-type="fig"}). Similar results were observed when another transcription inhibitor, cordycepin, was used (data not shown). In contrast, cycloheximide had a dramatic effect on pollen tube growth (Figure [5b](#F5){ref-type="fig"}). The presence of 0.1 μg/ml cycloheximide only inhibited pollen germination by 40%, but pollen tube growth was inhibited by 90%. At higher concentrations, 40% of pollen was still able to germinate, but further pollen tube growth was blocked. We conclude that active pollen tube growth is strictly dependent upon protein synthesis, and that pollen germination and tube growth are relatively independent of transcription.
Discussion
==========
To identify patterns of gene expression involved in *Arabidopsis*male gametophyte development, we compared the transcriptomes of isolated spores at four discrete developmental stages using ATH1 microarrays. ATH1 microarrays harbor probe sets for 22,591 annotated genes \[[@B52]\]. Of these, 61.9% (13,977 genes) gave positive hybridization signals in at least one stage of male gametophyte development. A comparable proportion of active genes was reported for isolated root cells which expressed 10,492 genes (46%) on ATH1 microarrays \[[@B8]\]. Moreover, in similar studies of animal cell development, 53% of 13,179 arrayed genes were found to be expressed during early murine adipocyte differentiation \[[@B1]\].
As the proportion of known genes embedded on the ATH1 array is 80.7%, we estimate the total number of genes expressed throughout *Arabidopsis*male gametophyte development to be more than 17,000. Similarly, the total number of genes expressed at individual developmental stages is estimated to be 14,300 at UNM stage, 14,800 at BCP stage, 10,900 at TCP stage and 9,000 at MPG stage. Previous gene-by-gene approaches identified only 21 different genes expressed during *Arabidopsis*male gametophyte development (for a review see \[[@B16]\]). Moreover, only three of these genes were shown to be expressed at microspore stage \[[@B53]-[@B55]\]. The data sets reported here include more than 11,000 microspore-expressed genes, representing a 3,600-fold increase in knowledge of gene expression in haploid microspores.
Two recent studies of the *Arabidopsis*mature pollen transcriptome using Affymetrix 8K AG arrays led to the identification of 992 and 1,584 pollen-expressed mRNAs, respectively \[[@B35],[@B36]\]. Results obtained with ATH1 and AG arrays are considered comparable and largely independent of the different probe sets used \[[@B56]\]. However, there was a significant discrepancy in the number of incorrectly annotated genes between both arrays, with 6.3% of probe sets on the AG array being incorrectly annotated in comparison with only 0.4% on the ATH1 array \[[@B56]\]. Therefore, results from ATH1 arrays are more accurate as well as more comprehensive. Accordingly, the use of the more complete ATH1 array and more accurate microarray normalization protocols led to an increase in the estimated total number of genes expressed in mature pollen from around 3,500 \[[@B35]\] to around 9,000 (this study). The proportion of these genes that are considered male-gametophyte specific is strongly dependent on the choice of the set of reference sporophytic datasets. In the work reported here, the availability of more comprehensive sporophytic datasets and the application of more stringent criteria therefore led to a decrease in the estimated number of putative pollen-specific genes from around 1,400 \[[@B35]\] to around 800 (this study). This number could be reduced further if cell-type-specific expression within an organ limits detection of overlap with pollen expression. Our data highlight the extensive overlap between sporophytic and gametophyte gene expression and reveal the subset of the transcriptome that is strongly enhanced or specifically expressed during male gametophyte development. Considering all stages of microsporogenesis the total number of putative male-gametophyte-specific genes was 1,355 with the proportion of specific genes increasing from 6.9% at UNM-stage to 8.6% at MPG-stage. Among the male-gametophyte-specific genes identified there was an increase in the collective proportion of cell-wall, cytoskeleton, signaling and transport-related genes from 22% at UNM stage to 34% in MPG stage. This reflects the increasing functional specialization of mature pollen in preparation for a dramatic change in the pattern of cell growth during pollen germination and pollen tube growth.
Developmental analysis of transcriptome data revealed two striking features, a sharp reduction in transcript diversity after BCP stage and a major shift in mRNA populations between BCP and TCP stages. The decline in mRNA diversity after BCP stage is associated with terminal differentiation as well as the documented phenomenon of protein storage in pollen (see \[[@B57]\], and this study). Moreover, this large-scale repression associated with termination of cell proliferation after PMII is accompanied by the selective activation of new groups of genes that are likely to function during pollen maturation and post-pollination development.
It is interesting that the expression profiles of UNM stage and BCP stages are similar despite the presence of two different cell types in pollen grains at BCP stage - the larger vegetative cell and the smaller generative cell. Given the limited volume of cytoplasm associated with the generative cell, developmental changes in gene expression in the gametic or male germline cells are likely to be masked by the predominant contribution of the vegetative cell cytoplasm. Therefore, our male gametophytic gene expression profiles largely reflect the passage of the microspore through cell division and changes in gene expression associated with the differentiation of the vegetative cell. Large-scale changes in gene expression occur between BCP and TCP stages, and therefore do not coincide with asymmetric division of the microspore. UNM expression patterns persist into the bicellular stage, which is consistent with experiments that demonstrate that vegetative cell fate is specified independently of cell division at pollen mitosis I \[[@B58]\].
In contrast, generative cell fate appears to be dependent on asymmetric division at pollen mitosis I \[[@B25],[@B58]\]. Sperm-cell cDNAs and databanks recently established in maize, lily, tobacco and *Plumbago zelanica*\[[@B30]-[@B33]\] provide valuable gametic gene-expression data in other species. Although our data do not provide direct information about gametic gene expression in *Arabidopsis*, further development of cell gamete isolation sorting \[[@B36]\] would allow genome-wide identification of generative- and sperm-cell-specific genes in comparison with the datasets generated here.
Hierarchical cluster analysis provided detailed evidence for the dramatic switch between early and late developmental programs. We identified 39 gene clusters that could correspond to co-regulated genes. These included early clusters, several clusters of late genes, those with constitutive expression profiles and clusters showing transient expression with peaks at BCP or TCP stages. The large size of cluster 29 (4,464 genes) documents the homogeneity in expression profiles of most early genes. In contrast, late gene clusters included a significant number of genes with similar profiles between BCP and TCP stages, followed by expression profiles that deviated between TCP and MPG stages. Cluster 1, and in particular cluster 17, contained genes strongly upregulated in TCP and MPG, with likely functions in post-pollination events. The differential fate of certain late gene clusters is likely to be a feature of their requirement during pollen maturation or post-pollination events.
Our analysis revealed completely different expression profiles of transcription factors when compared to core translation factors. The majority of core translation factors belonged to the early-group genes with few that were male gametophyte-specific. This may be expected, given that many genes are involved in general cellular activities. However, genes encoding PAB proteins did not follow the general trend. Seven out of eight *Arabidopsis*PAB mRNAs were gametophytically expressed. Three PAB genes (*PAB3*, *PAB6*and *PAB7*) appeared to be male gametophyte-specific and *PAB5*was preferentially expressed in pollen. Moreover, *PAB3*and *PAB5*are the most abundant early and constitutive PAB mRNAs and *PAB6*and *PAB7*belong among the few late core translation-factor genes. Although these data suggest-specific expression, our data do not rule out expression in other sporophytic tissues, particularly in flowers. Indeed, previously published expression data confirmed the expression of these PABs in other reproductive tissues together with pollen \[[@B59]\].
Conversely, transcription factors showed more diverse spectra of expression profiles including early, constitutive and late. There was a considerable variation in the expression profiles of individual transcription factor families. The most over-represented was the C3H family, members of which are known to have roles in lignin and other phenylpropanoid pathways in plants \[[@B60]\]. Although sporopollenin synthesis is believed to be under strict sporophytic control (see \[[@B16]\]), the diversity of gametophytic C3H transcription factors might suggest a function for these genes in regulating chemical interactions between phenylpropanoid precursors secreted by the tapetum. One candidate is the At1g74990 gene encoding a putative RING finger protein, which is abundantly and preferentially expressed at UNM and BCP stages.
The majority of core translation factors belonged to the early gene clusters. In contrast, a significant number of transcription-factor genes were strongly expressed during pollen maturation. These data alone did not obviously support the fact that pollen germination and early tube growth in many species are largely independent of transcription, but vitally dependent on translation \[[@B61]\]. Similarly, we found that *Arabidopsis*pollen germination and tube growth were relatively independent of transcription, and that active pollen-tube growth, and to a lesser extent pollen germination, were dependent upon protein synthesis. It is known for some plant species that mRNAs and rRNAs accumulate during pollen maturation and are stored for use during pollen germination \[[@B62],[@B63]\]. Our results show that *Arabidopsis*pollen is charged with a diverse complement of stored mRNAs that could be used to support pollen germination and pollen tube growth. Moreover, the early synthesis of mRNAs encoding translation factors strongly suggests that these are preformed and stored in mature pollen grains to support rapid activation upon hydration and germination. We also suggest that some abundant late transcription factors could regulate maturation-associated genes or act as repressors of inappropriate transcription in growing pollen tubes.
Conclusions
===========
The key impact of this work is that it provides a genome-wide view of the complexity of gene expression during single cell development in plants. Analysis of the male gametophytic transcriptome provides comprehensive and unequivocal evidence for the unique state of differentiation that distinguishes the developing male gametophyte from the sporophyte. Male gametogenesis is accompanied by large-scale repression of gene expression that is associated with the termination of cell proliferation and the selective activation of new groups of genes involved in maturation and post-pollination events. Development is associated with major early and late transcriptional programs and the expression of about 600 putative transcription factors that are potential regulators of these developmental programs. This wealth of information lays the foundation for new genomic-led studies of cellular functions and the identification of regulatory networks that operate to specify male gametophyte development and functions.
Materials and methods
=====================
Plant material and spore isolation
----------------------------------
*Arabidopsis thaliana*ecotype Landsberg *erecta*plants were grown in controlled-environment cabinets at 21°C under illumination of 150 μmol/m^2^/sec with a 16-h photoperiod. Isolated spores from three stages of immature male gametophyte were obtained by modification of the protocol of Kyo and Harada \[[@B38],[@B39]\]. After removal of open flowers, inflorescences from 400 plants were collected and gently ground using a mortar and pestle in 0.3 M mannitol. The slurry was filtered through 100 μM and 53 μM nylon mesh. Mixed spores were concentrated by centrifugation (50 ml Falcon tubes, 450 *g*, 3 min, 4°C). Concentrated spores were loaded onto the top of 25%/45%/80% Percoll step gradient in a 10-ml centrifuge tube and centrifuged (450 *g*, 5 min, 4°C). Three fractions were obtained containing: (1) microspores mixed with tetrads; (2) microspores mixed with bicellular pollen; and (3) tricellular pollen (Figure [1](#F1){ref-type="fig"}). Fraction 2 was diluted with one volume of 0.3 M mannitol loaded onto the top of a 25%/30%/45% Percoll step gradient and centrifuged again under the same conditions. Three subfractions of immature pollen were obtained: (2.1) microspores; (2.2) microspores and bicellular pollen mixture; and (2.3) bicellular pollen. Spores in each fraction were concentrated by centrifugation (eppendorf tubes, 2,000 *g*, 1 min, 4°C) and stored at -80°C. The purity of isolated fractions was determined by light microscopy and 4\',6-diaminophenylindole (DAPI) staining according to \[[@B25]\]. Viability was assessed by fluorescein 3\',6\'-diacetate (FDA) treatment \[[@B58]\]. Mature pollen was isolated as described previously \[[@B35]\]. Pollen tubes were cultivated *in vitro*for 4 h according to \[[@B21]\]. Pollen was scored as germinated when pollen tubes were longer than half a pollen grain diameter. Pollen-tube growth was scored by counting those with tubes longer than two pollen grain diameters.
RNA extraction, probe preparation and DNA chip hybridization
------------------------------------------------------------
Total RNA was extracted from 50 mg of isolated spores at each developmental stage using the RNeasy Plant Kit (Qiagen) according to the manufacturer\'s instructions. The yield and RNA purity was determined spectrophotometrically and using an Agilent 2100 Bioanalyzer at the NASC.
Biotinylated target RNA was prepared from 20 μg of total RNA as described in the Affymetrix GeneChip expression analysis technical manual. Double-stranded cDNA was synthesized using SuperScript Choice System (Life Technologies) with oligo(dT)~24~primer fused to T7 RNA polymerase promoter. Biotin-labeled target cRNA was prepared by cDNA *in vitro*transcription using the BioArray High-Yield RNA Transcript Labeling Kit (Enzo Biochem) in the presence of biotinylated UTP and CTP.
*Arabidopsis*ATH1 Genome Arrays were hybridized with 15 μg labeled target cRNA for 16 h at 45°C. Microarrays were stained with streptavidin-phycoerythrin solution and scanned with an Agilent 2500A GeneArray Scanner.
Data analysis
-------------
Sporophytic data from public baseline GeneChip experiments used for comparison with the pollen transcriptome were downloaded from the NASC website \[[@B41],[@B64]\]. The list of dataset codes was as follows: COT (three replicates), Cornah\_A4-cornah-wsx\_SLD\_REP1-3; LEF (three replicates), A4-LLOYD-CON\_REP1-3; PET (three replicates), Millenaar\_A1-MILL-AIR-REP1-3; STM (two replicates), Turner\_A-7-Turne-WT-Base1-2\_SLD; ROT (two replicates), Sophie\_A1-Fille-WT-nodex\_SLD, Sophie\_A5-Fille-WT-nodex\_SLD; RHR (two replicates), Jones\_A1-jones-WT1, SLD, Jones\_A1-jones-WT2\_SLD; SUS (three replicates), A1-WILLA-CON-REP1-3.
All gametophytic and sporophytic datasets were normalized using freely available dChip 1.3 software \[[@B65]\]. The reliability and reproducibility of analyses was ensured by the use of duplicates or triplicates in each experiment, the normalization of all 26 arrays to the median probe intensity level and the use of normalized CEL intensities of all arrays for the calculation of model-based gene-expression values based on the Perfect Match-only model \[[@B66],[@B67]\]. A given gene was scored as \'expressed\' when it gave a reliable expression signal in all replicates. Expression signal value \'0\' means that the detection call value was not \'present\' in all replicates provided. All raw and dChip-normalized gametophytic datasets are available at the Institute of Experimental Botany AS CR website \[[@B68]\]. Although a RT-PCR validation of microarray data was not performed specifically for the purpose of this publication, our confidence in the quality of the data presented is based on our previously published RT-PCR validation of the expression of 70 genes \[[@B21],[@B35],[@B41]\].
Microsoft Excel was used to manage and filter the microarray data. For annotation of genes present on the ATH1 Array, the *Arabidopsis*Genome Annotation Release 3.0 published by The Institute for Genomic Research \[[@B52]\] was used. Genes were sorted into functional categories created according to data mined from the Munich Information Center for Protein Sequences *Arabidopsis thaliana*Database \[[@B69]\], Kyoto Encyclopedia of Genes and Genomes \[[@B70]\] and TAIR \[[@B46]\]. Hierarchical clustering of expressed genes was performed using expression-profile data clustering and analysis software EPCLUST \[[@B71]\], with correlation measure based distance and average linkage clustering methods.
Additional data files
=====================
The following additional data is available with the online version of this article: Additional data file [1](#s1){ref-type="supplementary-material"} is an Excel file containing the following items. The table Data contains the complete transcriptomic datasets used. Data were normalized using dChip 1.3 as described in Materials and methods. Expression signal value \'0\' means that the detection call value for particular gene was not \'present\' in all replicates provided. In the column \'Cluster\', the appropriate cluster for each male gametophyte-expressed gene is shown. The table Clusters gives the number of genes comprising all 37 clusters of genes coexpressed during male gametophyte development. The table Cell-cycle data lists core cell-cycle genes showing their expression values in male gametophytic datasets. Genes were defined according to \[[@B21]\]. The chart shows expression profiles of male gametophyte-expressed core cell-cycle genes. The table Transcription data lists transcription-factor genes, showing their expression values in male gametophytic datasets. Genes comprising *Arabidopsis*transcription factor families were derived by compilation of data available at the Ohio State University Arabidopsis Gene Regulatory Information Server \[[@B47]\], data published in \[[@B22]\] and database homology searches. MADS-box and bHLH gene families were defined according to \[[@B23]\] and \[[@B24]\], respectively. The table Translation data lists core translation-factor genes showing their expression values in male gametophytic datasets. Genes were defined according to the FIAT database \[[@B51]\]. The chart shows expression profiles of male gametophyte-expressed core translation-factor genes. The Transcription table summarizes transcription factor gene families showing the number of genes expressed during male gametophyte development. Additional data file [2](#s2){ref-type="supplementary-material"} lists a complete set of 39 clusters of genes coexpressed during male gametophyte development. Clusters were determined using EPCLUST software with a threshold value of 0.05. The list of genes comprising each cluster is given in Additional data file [1](#s1){ref-type="supplementary-material"}. Additional data file [3](#s3){ref-type="supplementary-material"} gives the expression profiles of male gametophyte-expressed transcription factors sorted into individual gene families. Expression data are given in Additional data file [1](#s1){ref-type="supplementary-material"}.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
The complete transcriptomic datasets used
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
The complete set of 39 clusters of genes coexpressed during male gametophyte development
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 3
The expression profiles of male gametophyte-expressed transcription factors sorted into individual gene families
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We gratefully acknowledge support from the BBSRC and the GARNet transcriptomic centre at NASC for performing pollen microarray hybridizations. We thank Andy Johnson for help with microspore extraction, John Okyere for advice on microarray normalization protocols and all members of the Twell laboratory for helpful comments on the manuscript. D.H. was supported through a Royal Society/NATO Fellowship, a Ministry of Education of the Czech Republic Grant 1K03018 and a Grant Agency of the ASCR grant KJB6038409.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Spore isolation and numbers of genes expressed throughout *Arabidopsis*male gametophyte development. **(a)**Purity of isolated spores in each developmental stage determined by microscopy: UNM, microspores; BCP, bicellular pollen; TCP, tricellular pollen; MPG, mature pollen. **(b-e)**DAPI-stained populations of developing spores: (b) microspores; (c) bicellular pollen; (d) tricellular pollen; and (e) mature pollen. **(f)**Total number of genes expressed in developing pollen and their distribution among three relative abundance classes. Gene-abundance classes were defined as follows: high (up to 10-fold less than the maximum signal), medium (10- to 100-fold less) and low (more than 100-fold less).
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Scatter-plots comparing relative gene expression in pairs of developmental stages. The expression levels of individual genes were normalized using a logarithmic scale of 0 to 100 and genes coexpressed in pairs of transcriptome datasets were plotted. **(a)**UNM versus BCP stage; **(b)**BCP versus TCP stage; **(c)**TCP versus MPG stage; **(d)**UNM versus MPG stage. *R*-value represents the correlation coefficient.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
A selection of clusters of genes coexpressed during male gametophyte development. A complete set of all 39 clusters determined using EPCLUST software with a threshold value of 0.05 is available as Additional data files 1 and 2. *n*, number of genes comprising a cluster.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Expression profiles of regulatory genes throughout male gametophyte development. **(a)**Core cell-cycle genes; **(b)**transcription factors; **(c-f)**selected transcription factor families. (c) R2R3-MYB; (d) WRKY; (e) NAC; (f) C3H. Expression profiles of all transcription factor families analyzed are available as Additional data files 1 and 3. **(g)**Core translation factors. **(h)**Poly(A)-binding (PAB) protein genes. Putative male gametophyte-specific (bold) or enhanced (bold dashed) genes are highlighted.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Pollen germination and pollen tube growth *in vitro*in the presence of inhibitors of **(a)**transcription and **(b)**translation. The percentage of germinated pollen and the percentage of pollen capable of extended tube growth were scored independently for each treatment. PT, pollen tube.
:::
:::
|
PubMed Central
|
2024-06-05T03:55:51.801951
|
2004-10-27
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545776/",
"journal": "Genome Biol. 2004 Oct 27; 5(11):R85",
"authors": [
{
"first": "David",
"last": "Honys"
},
{
"first": "David",
"last": "Twell"
}
]
}
|
PMC545777
|
Background
==========
Chromosomes have evolved to effectively retrieve and transmit genetic information stored in DNA. Significant progress has been made recently in our understanding of how DNA is packaged into chromosomes, particularly at low compaction levels determined by supercoiling and/or protein-dependent condensation \[[@B1]\]. Available structural information about bacterial chromosomes indicates that the chromosome is supercoiled *in vivo*\[[@B2]\] and organized in topologically constrained domains \[[@B3]\]; diffusion of supercoils over the chromosome is impeded in actively replicating cells \[[@B4]\]; the chromosome is mildly condensed *in vivo*\[[@B5],[@B6]\]; chromosomal loci inside the cell are specifically organized and arrayed in linear order according to the linear genetic map \[[@B7],[@B8]\]; and at least two chromosomal loci are actively moved and positioned inside the cell \[[@B7],[@B9]\].
Although the molecular bases of these structural features are not known, the bacterial chromosome can be viewed as a cellular organelle, whose dynamics may be coupled to the state of the cell. In turn, the state of the cell is reflected in the whole-genome transcriptional activity \[[@B10]\]. Therefore, genome-wide transcription can be used to probe chromosomal organization. Global transcriptional profiles have been successfully used to probe the organization of transcriptional units into operons \[[@B11]\] and regulons \[[@B12],[@B13]\]. However, such analysis is limited by assumptions about the nature of transcriptional units.
Covariation in transcriptional activity along the chromosome determine spatial transcriptional patterns \[[@B14],[@B15]\]. Such co-variation might result from differing DNA accessibility along the chromosome \[[@B16],[@B17]\]. The variation in accessibility, in turn, may be determined by chromosomal structural features. By analogy, chromosomal regions that do not reveal any spatial covariation could represent unstructured portions of the chromosome.
Using signal processing and statistical techniques, we systematically examined transcriptional activity of genes as a function of their position in the bacterial chromosome. Here we report the discovery of stable, short- and long-range patterns in genome-wide transcription in *Escherichia coli*K12. Moreover, we demonstrate that such patterns are affected by genetic and environmental factors, thereby offering the first biologically relevant insights into the nature of the spatial organization of transcription in bacteria.
Results
=======
Local structure in the spatial series of transcriptional activity
-----------------------------------------------------------------
We modeled transcriptional activity of the chromosome as a one-dimensional spatial series of transcript abundances. Transcript abundances were measured in cells grown in batch cultures to OD~600~= 0.5 in LB or M9 medium supplemented with 0.2% glucose (cultures reached stationary phase at OD~600~= 3.5 in LB medium and at 2.5 in M9 medium). Samples of total RNA extracted using hot-phenol method \[[@B18]\] were labeled with Cy-fluorophors and hybridized against genomic DNA as a reference. We carried out two types of hybridization: one using genomic DNA from the same cells from which we extracted the RNA and the other using genomic DNA from cultures with arrested initiation of DNA replication that completed ongoing rounds of replication. Two different types of genomic reference produced indistinguishable results in spectral analysis, and for the sake of simplicity we present here the analysis of results obtained in hybridizations against genomic DNA isolated from non-replicating cultures. The mRNA abundances have been recorded for almost every gene in the chromosome by two-color hybridization on whole-genome DNA microarrays. The data are publicly available at the Gene Expression Omnibus \[[@B19]\], accession numbers GSE1730 and GSE1735. To determine the degree of similarity in transcriptional activity of individual genes as a function of their position on the chromosome we calculated the autocorrelation function (ACF) as a function of the distance between genes, where the distance is measured as the number of intervening genes (Figure [1](#F1){ref-type="fig"}).
Independent of the growth conditions, the ACF could be characterized as a decaying function whose largest statistically significant portion assumes positive values and corresponds to relatively short gene-to-gene distances (fewer than 100 genes; Figure [1a](#F1){ref-type="fig"}, data not shown). By definition \[[@B20]\], the portion of the ACF that constitutes significant correlations may reflect the existing stable structure in the series. Thus, it is likely that the transcription of any two genes, separated on the chromosome by a distance less than or equal to the length of the structured portion of the series, is similarly affected on average across the entire chromosome. Scrambling the linear order of genes leads to the loss of structure in the series (Figure [1b](#F1){ref-type="fig"}), indicating that the relative positioning of genes determines the autocorrelation properties.
If the autocorrelation reflects dominant patterns of co-transcription, then by determining the properties of the ACF we should be able to describe these patterns. Any one-dimensional pattern could be characterized by the stability (the distance between two genes at which transcriptional similarity drops to 50% of the average correlation between immediate neighbors) and by the range of stable correlations (maximal length of an average stable co-transcribed chromosomal region). Following the two-parameter exponential fitting of at least eight independently measured ACFs, we determined that the transcriptional correlations decay by half in the rich medium over 7.5 kilobases (kb) (Table [1](#T1){ref-type="table"}, 0 minute entry).
The range of local transcriptional patterns was measured as the maximal length of a continuous, significantly correlated region of the chromosome that is not affected by experimental error. We determined that in rich medium up to 16 genes in a row could demonstrate apparently coherent transcriptional activity (Figure [1c](#F1){ref-type="fig"}). The short-range correlations of the transcriptional series obtained in minimal medium had comparable characteristics (data not shown). While it is expected that genes organized in operons would show significant autocorrelation, the stability as well as the significant range of autocorrelations observed here extend far beyond those expected if gene expression was only coordinated within operons, which have an average size of three genes \[[@B21]\].
Analysis of long-range correlations
-----------------------------------
In addition to continuous transcriptional correlations over short distances, we also observed individual, statistically significant spikes in correlations of transcriptional activities of genes located about 100 and 700 kilobases (kb) apart. To investigate such long-range transcriptional patterns in more detail we decomposed the original signal of transcript abundances into a series of harmonics. The frequency spectrum of transcript abundances is shown in Figure [2](#F2){ref-type="fig"}. Four frequencies appear to be significant (at 95% confidence level) in the spatial series obtained from the cultures grown in LB medium: 690^-1^kb^-1^; 129^-1^kb^-1^, 115^-1^kb^-1^and 103^-1^kb^-1^. The sequential signal recorded in the mid-exponential cells grown in M9 salts supplemented with glucose contained similar significant frequencies (Figure [2b](#F2){ref-type="fig"}): 690^-1^kb^-1^, a clump of frequencies around 115^-1^kb^-1^and a free-standing frequency of 414^-1^kb^-1^.
The Fourier transform provides signal average characteristics and does not determine the frequencies at a particular spatial locality. To localize significant frequencies determined by Fourier transform and to find potentially significant local spikes of transcriptional activity, we subjected the spatial series of transcript abundances to a wavelet transform. The wavelet analysis revealed significant spectral components at the scales very similar to significant periods in the Fourier spectrum: approximately 125 kb and 600-700 kb (Figure [3a](#F3){ref-type="fig"}). In addition, in the range of scales from 100 to 1,000 kb, the wavelet transform identified pronounced local patterns at frequencies corresponding to 235^-1^kb^-1^, 300^-1^kb^-1^, 365^-1^kb^-1^and 555^-1^kb^-1^. The wavelet spectrum also shows that patterns of transcriptional activity are not symmetrical with respect to the chromosome. The dominant patterns are localized largely in the left replichore (the half chromosome divided by the replication axis counter-clockwise from the *oriC*) and appear to be bounded by the origin of replication. The most dominant pattern wave in transcriptional series, represented by a period of about 600 kb, spreads for 2.3 megabases (Mb), from the origin of replication to the *terC*site. The second most pronounced pattern (around 125 kb) consistent with the Fourier spectrum results from the transcriptional activity between the origin of replication and *terG*, about 1.3 Mb away. While significant components of the wavelet spectrum consistent with the Fourier spectrum were largely distributed on the left replichore, the scales unique to the wavelet spectrum were narrowly distributed along the scattered parts of the right replichore.
Modulation of transcriptional spectra
-------------------------------------
Inhibition of transcription initiation by rifampicin completely eliminates significant frequencies from the spectra after 30 min of treatment (data not shown). We rationalized that if transcription is not only inhibited, but modulated, globally, we might be able to track changes in the transcriptional spectra. We used a topoisomerase inhibitor, norfloxacin, to try to modulate transcription in the cell by inhibiting topoisomerase activities. As supercoiling and transcription in the cell are expected to be tightly coupled \[[@B22]\], we anticipated that inhibition of DNA topoisomerization would affect spatial transcriptional patterns. Norfloxacin was used at a concentration that ensured 50% and 90% killing of a bacterial population after 10 minutes and 30 minutes treatment, respectively. We examined the local correlations in transcription following norfloxacin treatment and found that by 30 minutes the range of significant autocorrelations has been reduced by about 50% from 16.5 to 7.7 kb and the stability of local patterns was reduced from 7.5 to 4.3 kb (Table [1](#T1){ref-type="table"}). We have observed changes in the amplitude (the magnitude of squared amplitudes integrated over space - the power - is plotted along the vertical axis in Figure [3](#F3){ref-type="fig"}) as well as in spatial ranges of significant wavelet components (Figure [3b,c](#F3){ref-type="fig"}). The most obvious changes in the wavelet spectrum can be described as a recession of the largest-scale wavelets from the terminus of replication in the left replichore. At a higher resolution, as seen in the local wavelet power spectrum in Figure [4](#F4){ref-type="fig"}, the spatial range of the characteristic 100-125 kb wavelet narrowed by 25 to 90%. Similarly, the Fourier analysis revealed that the main periods in untreated cells, including 115 kb and 690 kb, were significantly (*p*\< 0.05) diminished by (in some case) 30 minutes with the drug (Figure [5](#F5){ref-type="fig"}).
Global modulation of transcription by a single point mutation in DNA gyrase
---------------------------------------------------------------------------
In addition to modulation of the transcriptional spectra by non-equilibrium perturbations, we were interested in steady-state alterations in the process of transcription that would not be associated with irreversible changes in bacterial physiology. Such alterations may result from a compensation to a partial reduction in gyrase activity. A transient increase in transcription of the gyrase genes has been observed following inhibition of gyrase activity \[[@B23]\]. We rationalized that a partial loss of function of the gyrase enzyme would be accompanied by a compensatory steady-state increase in transcription of the gyrase genes. We predicted these compensatory changes in transcription would be genome-wide rather than confined to the gyrase genes.
Using resistance to low concentrations of norfloxacin as a screening method, we selected mutants with increased levels of *gyrA*and/or *gyrB*transcripts. We characterized one spontaneous mutant of gyrase, resistant to 0.8 μg/ml of norfloxacin, which carries the D82G mutation in the *gyrA*allele and causes elevated levels of *gyrA*and *gyrB*mRNA *in vivo*and lowers supercoiling activity of DNA gyrase *in vitro*(K.S.J, Hiroshi Hiasa and A.B.K, unpublished work). This mutant had a normal rate of growth and cell density in stationary phase (less than 5% different from the isogenic wild-type strain in LB). As predicted, the compensatory transcriptional mechanism was not limited to transcriptional regulation of gyrase expression but had a global effect. Using microarrays we estimated that steady-state transcriptional activity of as many as 847 genes had changed in the mutant relative to the isogenic wild-type strain. The microarray results were in part confirmed by reverse transcription-PCR (RT-PCR) for 10 out of 10 randomly chosen differentially expressed genes with an observed change in transcript abundance of 50% or greater. The distribution of differentially expressed genes along the chromosome is shown in a histogram in Figure [6](#F6){ref-type="fig"}. Interestingly, transcription in the area of the chromosome spanning about 1.5 Mb from the vicinity of the *terE*site through the *terG*site appeared to be most affected in the mutant.
If transcriptional activity of the chromosome is differentially perturbed, that is, some regions are being more affected than others, then patterns that existed in unperturbed chromosome may no longer spread across regions with such an \'out of sync\' activity, resulting in a partial spatial confinement of originally wider patterns. Moreover, if transcriptional activity is inhibited in a spatial locality, it may cause reduction, or even complete elimination, of local patterns. The wavelet power spectrum in Figure [3d](#F3){ref-type="fig"} unambiguously confirmed such truncation of wider patterns and the disappearance of significant local patterns in the area of the chromosome where transcription was spatially differentially affected. No changes in transcription have been observed in the gyrase mutant carrying an S83L mutation, a naturally occurring mutation that is not accompanied by a compensatory increase in gyrase transcription (data not shown).
Spatial patterns of DNA gyrase distribution on the chromosome
-------------------------------------------------------------
While global and local patterns of transcription are likely to be due to multiple causes, our data suggest DNA gyrase as an important factor in pattern formation. It has been suggested that DNA gyrase is not randomly distributed along the chromosome of *E. coli*\[[@B24]\]. We examined the distribution of DNA gyrase along the chromosome in a chromatin immunoprecipitation (ChIP) chip assay with GyrA-specific antibodies. The averaged and de-noised spatial signal recorded in eight ChIP-chip microarray experiments were analyzed through a wavelet transform. The contours of significant scales calculated from the gyrase-binding signal and the transcriptional signal obtained under identical growth condition are overlaid in Figure [7a](#F7){ref-type="fig"}. Similarities between the two spectra were quantified as the dot product of power vectors at corresponding scales (Figure [7b](#F7){ref-type="fig"}). We observed high positive correlation between the wavelet power spectra of gyrase distribution and transcriptional signal across multiple scales. We found no correlations between wavelet spectra of transcriptional signal and chromosomal distributions of several sequence-specific or nonspecific proteins, including Lrp, Topo IV, FtsK and LexA (data not shown; the results of these ChIP-chip experiments will be summarized in a separate paper).
Discussion
==========
The development of new high-throughput technologies for parallel analysis of gene expression \[[@B25]\] and the completion of the full *E. coli*genome sequence \[[@B26]\] have enabled us to study the activity of the entire bacterial chromosome simultaneously. Owing to the physical limitations of some techniques and the invasiveness of others, the study of a system is limited to the analysis of signals coming from the system. The bacterial chromosome is a perfect example of a system that cannot be studied directly without interfering with its properties. Therefore, in order to obtain insights into the macroscopic properties of the bacterial chromosome, we chose to study the transcriptional signal that is generated by the chromosome and can be recorded on DNA microarrays. Transcript abundance along the chromosome can be represented as a one-dimensional signal in a spatial domain. The most pronounced feature of this transcriptional signal is a high degree of correlation between genes close together on the chromosome. While such a correlation is expected from genes organized in operons, the observed stability and range of correlations extends far beyond the expected size of the average operon. The stability of the short-range correlations can be significantly reduced by norfloxacin (Table [1](#T1){ref-type="table"}, Figure [8](#F8){ref-type="fig"}) suggesting that short-range correlations depend on negative supercoiling. Such dependence offers an intriguing hypothesis about the physical basis of the short-range transcriptional correlations: the transcription of the genes within a confined supercoiled domain is more similar to each other than to genes in other such domains, with the size of a domain being approximated by the linear stability of the short-range correlations. While this paper was in preparation, Postow *et al*. have shown that the size of a supercoiled domain in the *E. coli*chromosome is of the order of 10 kb \[[@B27]\]. The similarity between the dimensions of supercoiled domains and short-range transcriptional patterns makes our hypothesis even more plausible. Also consistent with our hypothesis is the observation that in a gyrase mutant with a transcriptionally compensated supercoiling function (plasmid DNA supercoiling is normal in the mutant, data not shown) the autocorrelation stability is statistically identical to that in the wild-type cells. By the same argument, however, other characteristics of transcriptional patterns do not appear to be strictly supercoiling-dependent.
The observed changes in the long-range correlations could not be explained by changes in supercoiling because they are observed in the mutant as well as after drug treatment. It is more likely that changes in the medium- and long-range transcriptional patterns are associated with a change in the distribution of gyrase binding to the chromosome. The similarity between patterns of gyrase binding and transcription provides the basis for such a conjecture. It also seems plausible that coherent transcription, or changes in transcription, among clusters of functionally related genes \[[@B28]\] could be associated, in part, with their apparent regular spacing. Co-regulation of such clusters could contribute to the formation of observed local and global transcriptional patterns. We also note that transcription within operons is not sufficient to account for observed short- and long-range patterns. Randomization of the order of operons, as well as of individual genes, completely abrogates both types of correlations. It remains to be seen whether short-range patterns or nonrandom distribution of any sequence features, including extreme secondary structures \[[@B29]\], can contribute to the modulation of long-range correlations.
Analysis of the structure in any signal can be complicated by instrument biases. Such biases in microarray measurements were originally pointed out by Speed and co-workers \[[@B30]\]. Although it has been argued that such systematic effects may preclude adequate spatial-temporal analysis of microarray data \[[@B31],[@B32]\], we offer several reasons why our results are not a microarray artifact. First, the documented patterns could only be observed following subtraction of intensities in the reference DNA channel from the abundance transcript channel and not in the reference channel alone, suggesting that the patterns are not a property of hybridization efficiency. Second, the patterns could be significantly changed. Third, the patterns become more pronounced following the removal of the systematic biases and do not depend on the array design-specific periodicities whose removal by low-pass filtering has no effect on transcriptional patterns; and fourth, modeling of promoter activities carried out by Allen *et al*. \[[@B33]\] revealed the existence of at least one frequency component in the corresponding spatial series that was identical to the lowest-frequency component identified in this study through a direct modeling of transcript abundances.
Conclusions
===========
This study demonstrates the existence of spatial patterns of transcription in the *E. coli*chromosome. These patterns can be classified on the basis of overall similarity in transcriptional activity of individual genes as well as on the basis of regional similarities. Three major spatial patterns have been identified: short-range correlations that are stable, on average, over 7 kb and could extend up to 15-16 kb; medium-range correlations over 100-125 kb; and long-range correlations over 600-800 kb. These patterns are experimentally stable and can be reproducibly detected in mid-exponential cells grown in batch culture. The growth rate and medium composition appear to have very minimal effects on pattern formation. However, these patterns could be modulated by perturbing DNA gyrase. The significant patterns of gyrase distribution on the chromosome match those of transcriptional activity. Among several proteins (see Results) whose distribution we mapped on the chromosome in mid-exponential culture, the pattern of gyrase binding was the only one coinciding with the patterns of transcription. Although it remains to be seen whether the observed patterns resulted from coherent transcription of functionally related genes regularly distributed along the chromosome and/or through chromosomal organization, the findings presented here are the first evidence of physiologically determined higher-order organization of transcription in any chromosome studied to date.
Materials and methods
=====================
Strains and culture conditions
------------------------------
All experiments in this study were carried out in *E. coli*K12 strain MG1655 obtained from the ATCC. Mutant *gyrA*^*R*^alleles used in this study carried D82G (this study) and S83L \[[@B34]\] mutations. Wild-type and mutant *gyrA*^*R*^alleles were linked to a Tn*10*marker and P1 transduced into *E. coli*K12 MG1655 as described previously \[[@B35]\]. MG1655 *dnaC2*\[[@B36]\] was used to obtain the DNA reference sample for transcript abundance measurements. Bacterial cultures were grown at 37°C in LB or M9 supplemented with 0.2% glucose.
Microarray analysis
-------------------
Whole-genome DNA microarrays of *E. coli*were designed, printed and probed as described previously \[[@B13],[@B18]\]. To ensure the success of PCR amplification and to minimize cross-hybridization we redesigned more than 700 primer pairs from the original set of primers supplied by Sigma-Genosys \[[@B37]\]. The relative transcript abundances were determined as described \[[@B38]\]. The RNA samples were extracted from the cultures grown to an OD~600~of 0.5-0.6 using the hot-phenol method \[[@B18]\]. The experimental error of the measurements of RNA abundances was assessed from at least three independent replicates, where one replicate corresponds to the RNA sample from a bacterial culture grown from a separate colony. Differentially expressed genes were identified using two-class comparisons of the adjusted relative expression values by SAM \[[@B39]\] at 1% false-discovery rate at the 90th percentile. RT-PCR was carried out on ABI Prism 7900 according to the Applied Biosystems protocol with SYBR Green dye as a fluorescent probe.
Spectral analysis of spatial series
-----------------------------------
Following the removal of the array-, pin- and dye-specific effects \[[@B40],[@B41]\], the estimated relative abundance values were ordered according to the position of a corresponding gene on the chromosome and subjected to spectral analysis. The spatial domain is defined as a function of the position of the center of mass of the open reading frame (ORF) or operon. In the search for significant frequencies, 2,071 positive frequencies were examined in a signal consisting of 4,143 samples corresponding to individual genes. The autocorrelation function of transcriptional spatial series was calculated as in \[[@B20]\]:
with *j*= 0,1,2,\...,*J*, where *y*~*x*~is the series value at the index corresponding to a given ORF location,  is the mean over all *N*observations and *J*is the number of genes in the genome minus 1. The standard error of autocorrelation estimates is determined as (*N*)^-1/2^, where *N*is the number of samples in the series.
Information about the process of transcription could be extracted by identifying a pattern in the observed variations in gene activities followed by a search for the cause or explanation of the pattern. For instance, the pattern may consist of a defined dependence of transcriptional activity as a function of a chromosomal position, or it may consist of the dependence of transcriptional activity as a function of time following a treatment or during the cell cycle. This functional dependence may express itself as a linear variation or harmonic oscillation partially hidden behind the noise. We considered a physical variable *Y*that corresponds to the relative abundance of mRNA. This variable could be measured as a function of the position, *x*, on the *E. coli*chromosome. Values of *Y*are discretely recorded following two-color hybridization on the whole-genome DNA microarrays. Thus there is a finite series of values of *x*, {*x~i~*}, *i*= 1, 2, 3,\...., *N*(*N*corresponds to the total number of genes) and corresponding values of *Y*, {*Y~i~*}, *i*= 1, 2, 3,\...., *N*. We call {*Y~i~*} a spatial series or sequential series. For the sake of convenience we approximated positions of the centers of mass of individual genes, {*x~i~*}, as evenly spaced. A process is a rule or procedure that generates a sequential series, that is, a prescription for determining the values *Y*for a given set of values of *x*. We defined a spatial series recorded in one whole-genome hybridization as one realization of the process.
The nature of the data that we register using microarrays as well as the nature of the process itself is likely to be such that the rule(s) generating a sequential series specify the probability distribution of {*Y*~*i*~} and not specific values that are the same at every realization. However, defining such a distribution did not seem feasible. Instead, we determined significant autocorrelation components in the spatial series by assessing the effect of experimental error on the significance of correlations. Following the recording of spatial series in at least three independent biological replicates, we simulated realizations of the process of transcription by resampling relative abundances in a gene-specific manner (see \[[@B42]\] for details of the bootstrap). For each of the realizations we calculated the ACF at all acceptable lags. We counted the number of times the ACF value appeared to be significant (α = 0.05) across all simulated realizations at each of the lags.
The Fourier spectrum was determined using the Lomb algorithm \[[@B43]\]. To ensure comparability between the wavelet scale and the Fourier period we used the Morlet wavelet as a basis \[[@B44]\] in Matlab 6.5.1 \[[@B45]\] or AutoSignal 1.6 \[[@B46]\]. The significance of the Fourier and wavelet peaks was estimated from peak-type critical limits. The critical limits were generated from Monte Carlo trials with uniformly spaced spatial domain coordinates. The null hypothesis was simulated using white noise as a background distribution.
Analysis of gyrase binding on the chromosome
--------------------------------------------
Chromatin immunoprecipitation of DNA sequences bound by DNA gyrase and their detection using whole-genome DNA microarrays were adapted from \[[@B47]\]. Briefly, cells were grown to an OD~600~of 0.5-0.6 in LB or M9 medium and DNA was cross-linked to proteins with formaldehyde at 1% (v/v) final concentration. Following incubation with monoclonal antibodies against the GyrA subunit of DNA gyrase (TopoGEN), the protein-DNA complexes were precipitated using Protein A-agarose beads (Sigma). Ligation-mediated PCR (LM-PCR; adaptor sequences: 5\'-GCGGTGACCCGGGAGATCTGAATTC-3\' and 5\'-GAATTCAGATC-3\') was used to amplify DNA following the reversal of cross-links. Sonicated genomic DNA served as a reference in two-color hybridizations, in which both the sample and the reference were labeled in the Klenow reaction following random amplification by LM-PCR.
Additional data files
=====================
Additional data file [1](#s1){ref-type="supplementary-material"}, available with the online version of this article, consists of a table of the intensities used in the analysis of spatial patterns. The file contains fluorescent intensities recorded in individual channels in two-color microarray hybridizations. The data from replicate arrays are included in the same worksheet, which also contains a brief description of an experiment. The data are also available at \[[@B19]\]; series accession numbers are GSE1730 and GSE1735.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
A table of the intensities used in the analysis of spatial patterns
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank Carol Gross, Stanley Cohen, Jon Bernstein and Todd Lowe for their help with genomics resources, Dennis Royzenfeld for implementing the expression database, Ed Boas and Paul Fawcett for suggestions on the Lomb algorithm, and Nancy Crisona and Melinda Hough for comments on the manuscript. We thank Boris Belotserkovskii, Nicholas Cozzarelli, Eric Siggia, Paul Spellman and anonymous reviewers for a constructive critique of the manuscript. A.K. thanks Patrick Brown for suggestions and support, and also Valentin Rybenkov and Hiroshi Hiasa for discussions and encouragement. This work was supported by Grant GM066098 from the National Institutes of Health to A.K.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
The autocorrelation function (ACF) of the spatial series of transcript abundances. **(a)**ACF of the series recorded in LB medium. **(b)**ACF of the series after scrambling the linear order of genes. The dotted line delimits the 95% confidence band. **(c)**The robustness of autocorrelations in the series recorded in LB as determined by bootstrapping. The robustness is expressed as a percentage of times the autocorrelation at the corresponding lag satisfied the critical value threshold (α = 0.05) following bootstrapping of the observed relative abundances. The sampling was done 1,000 times.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
The Lomb periodogram of transcriptional activity. Wild-type cells in **(a)**LB or **(b)**M9 medium. The significant frequencies at the 95% confidence level (dashed line) are indicated by their scales in kilobase-pairs.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Three-dimensional (3D) wavelet power spectrum of the transcriptional signal. The wavelet power spectrum from 100 to 1,000 kb in wild-type *E. coli*cells grown in LB and treated with norfloxacin at 10 μg/ml for **(a)**0 min, **(b)**10 min, **(c)**30 min and **(d)**in the *gyrA*(D82G) mutant grown in LB without a drug. The analysis was performed in AutoSignal 1.6 using the Morlet wavelet as the basis with the randomized set showing no significant pattern at 95% (magenta) confidence level. The significant wavelet components are represented as parts of a power spectral density (PSD) plot, where power is calculated as the space (time in standard notations)-integral squared amplitude (TISA). Our model assumes white background noise.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Two-dimensional (2D) wavelet power spectrum of transcriptional activity in the vicinity of the origin of replication. The wavelet power spectrum from 20 to 200 kb of transcriptional activity of wild-type *E. coli***(a)**untreated and **(b)**treated with norfloxacin for 30 min. The 2D wavelet spectrum was calculated as in Figure 3.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Comparison of significant frequencies from the Fourier spectra in the cultures treated with norfloxacin. The amplitudes of the most significant frequencies have been compared using the two-tailed Student\'s *t*-test; the asterisks mark the frequencies whose power was significantly different (*p*\< 0.05) from the power of the corresponding frequencies in the spectrum from the untreated cells (0 min).
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
A positional histogram of genes differentially affected in the *gyrA*D82G mutant relative to the wild-type strain. Positional bins of the 487 relatively downregulated genes are shown as filled bars and of the 360 upregulated genes as open bars. The number of genes in each bin are normalized to the bin\'s size. The sizes of the bins have been determined from the distribution of annotated open reading frames (ORFs) in the *E. coli*genome.
:::
:::
::: {#F7 .fig}
Figure 7
::: {.caption}
######
Comparison of the chromosomal spatial patterns of transcriptional activity and of gyrase binding. **(a)**The overlay of the contours of significant components of 2D wavelet power spectra (cyan, transcription activity; magenta, gyrase binding). **(b)** The dot products of power from the transcriptional and binding spectra at corresponding frequencies.
:::
:::
::: {#F8 .fig}
Figure 8
::: {.caption}
######
The exponential decay functions fitted by the least-square method into the corresponding autocorrelation functions. The graphs of fitted functions are shown. The goodness of fit was determined by the chi-squared method. Prob~*f*~(χ^2^≥ χ^2^~0~) for the ACF functions corresponding to transcriptional series obtained for 0, 10 and 30 min treatment with drug, and for a gyrase mutant were all above 95%.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Characteristic parameters of the short-range transcriptional patterns
:::
Strain and condition Stability (kb) Range (kb)
------------------------------- ---------------- ------------
Wild type, norfloxacin 0 min 7.5 ± 0.9 16.5
Wild type, norfloxacin 10 min 5.6 ± 1.3\* 11.0
Wild type, norfloxacin 30 min 4.3 ± 0.9^†^ 7.7
*gyrA*(D82G) 8.1 ± 1.1 16.5
\*Significantly and ^†^highly significantly different (*p*\< 0.05 and *p*\< 0.01, respectively) from the stability of the ACF in the untreated sample (0 min), based on the two-tailed *t*-test comparison of the mean fitted parameters for eight independent replicates.
:::
|
PubMed Central
|
2024-06-05T03:55:51.805615
|
2004-10-27
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545777/",
"journal": "Genome Biol. 2004 Oct 27; 5(11):R86",
"authors": [
{
"first": "Kyeong Soo",
"last": "Jeong"
},
{
"first": "Jaeyong",
"last": "Ahn"
},
{
"first": "Arkady B",
"last": "Khodursky"
}
]
}
|
PMC545778
|
Background
==========
The chromosome of *Escherichia coli*is a circular double-stranded DNA molecule that is maintained in a negatively supercoiled state. Supercoiling induces torsional tension in the DNA, and thus can influence processes that involve the opening of the double helix, such as replication initiation \[[@B1]\], DNA looping \[[@B2]\] and transcription \[[@B3]\]. A number of external stimuli, such as osmotic stress, oxygen tension, nutritional shifts, and temperature change affect supercoiling (for review see \[[@B4]\]), suggesting that supercoiling is a mechanism by which environmental changes could be communicated to the transcriptional machinery.
In *E. coli*, supercoiling is maintained at a precise range during log phase growth by the topoisomerases DNA gyrase, topoisomerase I (topo I), and topoisomerase IV (topo IV) \[[@B5]-[@B7]\]. DNA gyrase and topo IV are ATP-dependent type II topoisomerases that introduce negative supercoils and remove positive supercoils, respectively \[[@B8]-[@B10]\], whereas topo I is a type IA topoisomerase that removes negative supercoils \[[@B11]\]. Together, these activities remove the topological effects of translocating proteins, such as RNA polymerase, that create (+) supercoils in front and (-) supercoils behind the moving protein \[[@B12],[@B13]\]. The balanced activities of these enzymes result in a steady-state level of negative supercoiling. In turn, supercoiling modulates the expression of the genes for gyrase (*gyrA*and *gyrB*), and for topo I (*topA*). Relaxation of the chromosome upregulates *gyrA*and *gyrB*and downregulates *topA*as a form of feedback control \[[@B14]-[@B16]\]. This dual response also indicates that (-) supercoiling can promote, as well as inhibit, gene expression. It is perhaps not surprising that transcription of topoisomerase genes may be sensitive to supercoiling changes. Yet transcription of other genes, such as *fis*(a nucleoid-associated protein and transcriptional regulator), *ilvG*(an amino-acid synthase subunit) and *cydAB*(an oxidase involved in aerobic respiration), has been found to be sensitive to supercoiling \[[@B17]-[@B19]\], suggesting that a wider class of genes whose expression is sensitive to supercoiling may exist. Furthermore, a recent search for osmotic shock genes found a cluster of genes with enhanced sensitivity to supercoiling \[[@B20]\]. If supercoiling is used as a mechanism to sense environmental changes, we predict that genes from many functional classes would be affected by supercoiling, because environmental changes such as temperature and osmotic strength will affect many different reactions in the cell. Determining which genes are supercoiling sensitive may illuminate principles of promoter activation, such as common sequence characteristics in promoters and regulation of transcription initiation \[[@B14],[@B17],[@B18]\].
In this study, we used cDNA microarrays \[[@B21],[@B22]\] representing nearly the entire *E. coli*K-12 genome to systematically identify those genes that respond to relaxation of the chromosome during log-phase growth. We used antibiotics and mutations in the topoisomerase genes to change supercoiling levels by independent mechanisms and thus discerned the general effects of chromosome relaxation. We classify supercoiling-sensitive genes, or SSGs, according to their response to DNA relaxation. Therefore, we call \'relaxation-induced genes\' those genes whose expression is increased upon DNA relaxation, and \'relaxation-repressed genes\' those whose expression is repressed by DNA relaxation.
An extensive statistical analysis of our experimental results revealed 200 relaxation-repressed genes and 106 relaxation-induced genes; in total, around 7% of all genes in the genome were found to be significantly affected by supercoiling changes. Many of these genes are more sensitive to supercoiling than *gyrA*or *topA*, and their expression patterns correlated with the supercoiling level of a reporter plasmid in the cells. SSG transcripts have the same rates of RNA decay as non-SSG transcripts, and thus the changes in expression were due to a change in the rate of RNA synthesis, rather than RNA decay.
We discovered that the sequences of the relaxation-induced genes are significantly (*p*\< 0.0001) AT-rich in their upstream sequences, and also have AT-rich coding regions. Relaxation-repressed genes had a corresponding preference for GC-rich sequences. The SSGs are dispersed throughout the chromosome, and fall into many different functional classes. We propose that the large number and functional diversity of the SSGs reflects the role of supercoiling as a second messenger that responds to environmental changes and feeds into different regulatory networks.
Results
=======
Topoisomerase inhibition
------------------------
We sought to determine the genes that are activated or repressed by relaxation of the (-) supercoils in the chromosome. To isolate the expression changes due to the loss of supercoiling from those due to the experimental approach, we used three different methods to relax the chromosome. In two of the methods we inhibited gyrase and topoIV with either quinolone or coumarin antibiotics, and in the third we used a temperature-sensitive strain in which gyrase is inhibited at 42°C. Because it is technically difficult to quantify the supercoiling state of the bacterial chromosome, we used a plasmid, pBR322, in the strains as a reference. Co-transcriptional translation of the *tetA*gene in pBR322 anchors this plasmid to the membrane \[[@B23]\], and thus this plasmid has been used as a model for the chromosome \[[@B7]\]. The superhelical density, or σ, of plasmids can be readily measured. Plasmid σ values for all of the relaxation experiments are shown in Table [1](#T1){ref-type="table"}.
### Inhibition of topoisomerases by norfloxacin
The quinolone antibiotic norfloxacin selectively and immediately inhibits gyrase and topo IV \[[@B24]-[@B26]\]. We used isogenic strains with resistance mutations in the genes for gyrase (*gyrA*and *gyrB*) or topo IV (*parC*and *parE*) as controls, to separate expression changes due to undiscovered drug targets from those directly due to changes in supercoiling. When we inhibited gyrase by treating *gyrA*^+^*parC*^*R*^cells with 15 μg/ml norfloxacin, the reporter plasmid in the cells was rapidly relaxed (Table [1](#T1){ref-type="table"}). In a parallel experiment, plasmid DNA in a drug resistant *gyrA*^*R*^*parC*^*R*^strain remained (-) supercoiled. After 30 minutes, there was a 10^3^-fold drop in viability in the sensitive strain, but only a 17% drop in the resistant strain. A norfloxacin concentration of 50 μg/ml fully inhibited both gyrase and topoisomerase IV in the wild-type strain (data not shown), while the resistant strain retained wild type levels of (-) supercoiling and showed only a slight drop (15%) in viability, indicating that we did not overcome the drug resistance. At bacteriocidal concentrations similar to these, quinolones cause a decrease in the sedimentation coefficient of the nucleoid, indicating relaxation of the chromosomal supercoils \[[@B27]\]. The reference RNA sample was from cells removed immediately before addition of the drug (*t*= 0) and was labeled with Cy3 (green). RNA samples taken 2, 5, 10, 20 and 30 minutes after drug addition were labeled with Cy5 (red). The labeled experimental and reference samples were mixed in equal amounts before hybridization to a microarray.
Inhibition of topoisomerases by quinolones leads to double-strand breaks in the chromosome \[[@B28]\]. Thus, norfloxacin not only reduces supercoiling, but also induces the SOS response to DNA damage \[[@B29]\]. We found that the induction of the SOS response by norfloxacin was significantly slower and less extensive than either the responses of the SSGs (see below) or the SOS induction caused by UV treatment (see Additional data file 1). We conclude that the induction of SOS by norfloxacin is not a significant impediment to our search for SSGs.
### Inhibition of topoisomerases by a coumarin antibiotic
We also relaxed the chromosome using novobiocin, a coumarin antibiotic that inhibits gyrase, and at a higher concentration, topo IV \[[@B30],[@B31]\]. Novobiocin inhibits the ATPase activity of the enzyme \[[@B32],[@B33]\], and the mechanism of inhibition is completely different from that of norfloxacin \[[@B34]\]. We treated cells with 20, 50 and 200 μg/ml novobiocin for 5 minutes and measured the DNA relaxation by gel electrophoresis (Table [1](#T1){ref-type="table"}) and the gene-expression changes by microarray. In addition to changes due to a loss of topoisomerase activity, we saw changes in a set of non-overlapping genes between the norfloxacin and novobiocin experiments, indicating that there are also drug-specific transcriptional effects. Since we focused our analysis on the genes that responded to supercoiling independent of the relaxation method used, these drug-specific changes were removed from consideration.
### Inhibition of gyrase by mutation
We also used a mutant that is temperature-sensitive for gyrase activity \[[@B35]\], which results in relaxation of the chromosome at the restrictive temperature \[[@B36]\]. We measured expression changes in *gyrB234*cells upon shift to the restrictive temperature and subsequent relaxation of the DNA (Table [1](#T1){ref-type="table"}). To control for the effects of the temperature shift on gene expression, we compared the changes in the *gyrB*^*TS*^mutant to those in an identically treated isogenic wild-type strain. The *gyrB*^*TS*^data were combined with the norfloxacin and novobiocin data to make a body of experiments and controls where the transcriptional effects of relaxation were isolated from effects due to the method used to relax the chromosome.
Identification of supercoiling-sensitive genes by statistical analysis
----------------------------------------------------------------------
We obtained a dataset from a total of 35 arrays. Fourteen of the arrays were controls in which either drug was added to resistant cells or the temperature was shifted for wild-type cells. The supercoiling of the reporter plasmid did not change in these controls (Table [1](#T1){ref-type="table"}). The remaining 21 arrays represented experiments in which the DNA was relaxed by different methods and over various time courses. This rich dataset allowed us to use statistical methods to determine those genes whose expression significantly varied with supercoiling levels.
Using threshold ratio values (for example, requiring a twofold change in expression) to determine which genes change significantly during an experiment can bias expression analysis towards genes with very low or variable expression levels \[[@B37]\]. We used statistical methods to minimize the bias. To assess the significance of the difference in gene expression between supercoiled and relaxed samples we used the method described by Dudoit *et al*. \[[@B38]\]. Briefly, we performed a *t*-test for each gene and corrected the obtained *p*-values for multiple testing by a step-down procedure \[[@B39]\]. The corrected *p*-value represents the probability that the differences in gene expression between the controls and relaxation experiments could have arisen by chance, after taking the expression of all genes into consideration. We obtained *p*-values ranging from 0.000125 to 1.
As an independent metric of supercoiling sensitivity, we measured how closely gene expression followed the level of DNA supercoiling, by calculating the correlation of the expression of each gene across all of the experiments with the level of supercoiling in the reporter plasmid. Relaxation-induced genes showed a positive correlation with plasmid linking number (that is, as (-) supercoiling is lost, both linking number and gene expression increase), up to a maximum observed Pearson correlation coefficient of 0.91. Relaxation-repressed genes showed a negative correlation with plasmid linking number to a minimum Pearson coefficient of -0.88. The majority of genes (3,190, or 80%) showed no strong correlation with plasmid supercoiling, resulting in Pearson coefficients between 0.5 and -0.5. The *p*-value represents the robustness of the response to relaxation, whereas correlation with plasmid supercoiling may represent sensitivity to changes in supercoiling levels. For example, a gene that is always completely repressed in response to relaxation will have a low *p*-value, but may show little correlation with intermediate levels of supercoiling. Similarly, a gene with more variable expression may have a higher *p*-value, but may also have a higher sensitivity to intermediate supercoiling levels. Taken together, these metrics provide a detailed account of supercoiling sensitivity.
The *p*-values for all of the genes versus their correlation to plasmid supercoiling are plotted in Figure [1a](#F1){ref-type="fig"}. The great majority of the genes have both high *p*-values and little correlation with plasmid supercoiling. Those genes with the lowest *p*-values (and thus, the most significant expression change upon relaxation) tended to be more strongly correlated (or anticorrelated) to plasmid supercoiling. The data for all genes can be found in Additional data file 2. Among all genes there is a continuous variation in both *p*-value and correlation to plasmid supercoiling. We found a total of 306 genes at *p*\< 0.05, which we define as SSGs. Of these, 106 genes were induced by DNA relaxation and have a positive correlation with plasmid linking number, while 200 genes were repressed by relaxation and these have a negative correlation with plasmid linking number. The correlations of the SSGs with plasmid supercoiling are shown in Figure [1b](#F1){ref-type="fig"}, which is an expansion of the significant region of the plot in Figure [1a](#F1){ref-type="fig"}. All the SSGs have a correlation with plasmid supercoiling with an absolute value greater than 0.5, which validated our selection on the basis of *p*-value. Just over half of the SSGs have high significance, *p*\< 0.005. The high redundancy of our data (21 arrays measuring responses to DNA relaxation, and 14 control arrays with negatively supercoiled DNA) minimized the influence of any single array measurement. Thus we can be confident that the genes we classed as SSGs have a reproducible response to supercoiling changes.
Figure [2a](#F2){ref-type="fig"} shows the expression changes in the 200 relaxation-repressed genes across the 35 conditions tested, with each numbered column representing one array. Each row represents the expression of one gene across all experiments, ranked by *p*-value (from the top). Each colored entry in the diagram corresponds to one spot on one array (that is, expression of a gene for a point in a given experiment: red if expression increased during the experiment, green if it decreased). Conversely, these relaxation-repressed genes should have low ratios (and black squares) in the control experiments 1 to 14. The significant difference in SSG expression between the controls and relaxation experiments is reflected by the sharp contrast between the mostly black controls and the bright green relaxation experiments. At the top we have shown a model expression profile representing the supercoiling of the reporter plasmid in each experiment (Table [1](#T1){ref-type="table"}), with black indicating no change in plasmid supercoiling and bright green indicating complete relaxation of the plasmid. These plasmid relaxation data match very well the expression data of the SSGs. The names of the top 10% of genes (those with the lowest *p*-value) are listed, along with their correlations to plasmid supercoiling levels.
The 106 genes that are induced by relaxation are similarly shown in Figure [2b](#F2){ref-type="fig"}. Red squares indicate expression at a higher level when the DNA is relaxed. Once again there is a striking difference in color between the control and relaxation experiments, and the SSGs show a strong similarity to the model profile at the top (in this model profile, red color indicates relaxation of the reporter plasmid). Several of the relaxation-induced genes are marginally repressed (shown by green color) in some control experiments. This is due to the fact that our statistical selection did not require the SSGs to be unchanged in the controls, but only required a significant difference in expression between the controls and relaxation experiments. However, this trend highlights the large expression change (from repression to induction) caused by chromosomal relaxation. It is striking how many genes respond significantly to a loss of chromosomal supercoiling (7% of the total genes). The full list of SSGs, with their *p*-values, correlations to supercoiling, and expression levels in each experiment can be found in Additional data file 3.
Kinetic analysis of gene expression and supercoiling
----------------------------------------------------
We expected that changes in SSG expression that are a direct effect of supercoiling changes (rather than mediated through other genes) should respond quickly to relaxation. We used a finer time-course experiment to determine which genes had the fastest response to chromosomal relaxation. When 15 μg/ml norfloxacin was added to *gyrA*^+^*parC*^+^cells, plasmid supercoiling levels changed dramatically within the first minute (Figure [3](#F3){ref-type="fig"}). Significant changes in gene expression followed by 2 minutes (Figure [4](#F4){ref-type="fig"}). We ranked the SSGs according to their correlation to plasmid supercoiling levels in this experiment. Thus, genes with transcriptional changes that match the kinetics of plasmid relaxation have high correlations. About 90% of the SSGs had a correlation higher in absolute value than 0.5, and more than half had correlations better than 0.75. The expression profiles of all of the SSGs, ranked by their correlation to plasmid supercoiling, are shown in Figure [4](#F4){ref-type="fig"}. The correlation of the SSGs to plasmid relaxation kinetics shows the sensitivity of gene expression to changes in supercoiling, while the *p*-value is a good indicator of the reproducibility of the response to supercoiling across the different experimental conditions we tested.
The speed of the transcriptional response to relaxation, combined with the strong correlations to supercoiling of the reporter plasmid in the cells, is strong evidence that the SSGs are directly regulated by supercoiling changes. Furthermore, given that *E. coli*mRNAs have a mean half-life of 5.2 ± 0.3 minutes in LB media \[[@B40]\], RNA synthesis of the relaxation-repressed genes must have slowed almost immediately upon DNA relaxation, in order to produce the quick changes we recorded (Figure [4](#F4){ref-type="fig"}). More than half of the relaxation-repressed genes changed by twofold or more in the first 5 minutes of this experiment.
We found no correlation of *p*-value with the published values of RNA half-life \[[@B40]\] and in general the mRNA half-lives of the SSGs were not significantly different from those from the rest of the genome (data not shown). We conclude that the changes in SSG expression are direct effects on transcription, rather than an effect on RNA degradation.
Sequences surrounding the start codon of supercoiling sensitive genes
---------------------------------------------------------------------
We searched for a basis of supercoiling sensitivity at the nucleotide sequence level by examining the AT content in and around the SSGs. We considered only the first genes in an operon. Whereas relaxation-repressed genes have a slightly depleted AT content both upstream of their promoters and within the coding sequence, relaxation-induced genes have an emphatic elevation of AT content in similar regions. The AT content of relaxation-induced genes from 800 nucleotides before to 200 nucleotides after the start codon is 54.6%, compared with 51.7% for non-SSGs. To illustrate the very low probability of selecting by chance a set of genes with such an elevated AT content, we randomly selected groups of first-in-operon non-SSGs 50,000 times and calculated AT content within the same window. We never found a set with an AT content as high as the relaxation-induced genes (red circle, Figure [5a](#F5){ref-type="fig"}). The difference in AT content is highly statistically significant (*p*= 3E-4).
This is not the only region in which the AT content of SSGs deviates from the rest of the genome. Figure [5b](#F5){ref-type="fig"} shows the mean AT content in a 100-nucleotide window for relaxation-induced, relaxation-repressed, and non-SSGs from 2 kilobases (kb) upstream to 1.5 kb downstream of the start codon. Nearly all genes, including non-SSGs, have elevated AT content upstream and just downstream of the start codon. The relaxation-induced genes, however, have a higher maximum AT richness and the elevated AT content extends over a wider region. Also, the relaxation-repressed genes showed a highly statistically significant reduction in AT content from -400 to +1,000 relative to the start codon (*p*= 1E-6).
Striking as these differences in AT content are for SSGs as a group, they are not sufficient to distinguish an individual SSG from a non-SSG. That is, not all genes with high or low AT content were supercoiling sensitive in our experiments. Although such genes are rare in the non-SSG population, the greater size of the pool of non-SSGs results in many genes with wide variations in AT content. Also, supercoiling sensitivity cannot solely be due to differences in AT content, as a few SSGs were highly sensitive to supercoiling changes in spite of having an AT content similar to the rest of the genome.
Discussion
==========
In this analysis of supercoiling effects on transcription, we identified 306 genes that quickly and reproducibly respond to chromosomal relaxation. The comprehensive nature of our investigation, with responses of 93% of the genome (4,003 protein-coding genes) in 21 different relaxation experiments and 14 control experiments, allowed us to be more stringent than previous studies in our definition of SSGs, and to identify those genes that had statistically significant changes after the chromosome was relaxed by different methods. Genes that are sensitive to relaxation but are also affected by temperature shifts (including *topA*\[[@B41]\] and *gyrA*\[[@B42]\]) showed changes in our control experiments, and thus had less significant *p*-values. Accordingly, although the topoisomerase genes *topA*and *gyrA*both clearly respond to supercoiling (see Figure [1](#F1){ref-type="fig"} and \[[@B14]-[@B16]\]), they have *p*-values of 0.058 and 0.062, respectively (compared to the *p*-value of 0.001625 for *gyrB*). The omission of these topoisomerase genes from our list of SSGs reflects the conservative statistical approach we used to define the list. There are probably other genes that respond to supercoiling changes in different conditions from those we tested (log-phase growth in rich media). Also, we defined SSGs by focusing on the immediate effects of relaxation, and thus considered only primary transcriptional changes, rather than downstream effects mediated by other gene products (though we note that 14 of the SSGs encode known transcriptional regulators). When downstream effects are considered, changes in supercoiling are likely to affect transcription of an even greater proportion of the genome.
There have been several previous attempts to measure the effects of supercoiling on gene expression in *E. coli*. Two early studies used either nylon membranes or two-dimensional protein gels to compare topoisomerase mutants with slightly different homeostatic levels of supercoiling, and neither study found a large number of genes \[[@B43],[@B44]\]. This could be due to the lower sensitivity of these earlier studies and because they measured steady-state gene expression, generations after the initial mutations and subsequent adjustment to the new supercoiling levels.
A more recent analysis by Church and colleagues used microarrays to gauge the osmotic stress response of *E. coli*\[[@B20]\]. Surveying 2,146 genes that were above their threshold of detection, the authors scored a subset of 30 genes that should be significantly enriched for supercoiling-sensitive transcription. Four of the genes identified are on our list of SSGs (*ynhG*, *yrbL*, *otsB*and *yifE*). Seven others had *p*\< 0.1 in our relaxation experiments, and the rest had still higher *p*-values in this study. It is possible that these genes are only responsive to supercoiling changes in the context of osmotic stress.
Just as supercoiling is affected by many environmental changes, such as osmotic shock, oxygen tension, nutrient upshift and temperature change, so too do changes in supercoiling affect genes in a large number of classes. Not surprisingly, a substantial fraction (6.9%) of the SSGs encode genes involved in DNA replication, recombination, modification and repair. However, the SSGs span many other classes, and thus are well positioned to feed into many different regulatory networks. Thus, supercoiling can act as a second messenger that quickly translates environmental changes to transcriptional programs, inducing and repressing specific genes independently of protein synthesis.
Several of the SSGs warrant further inspection. For example, the repression of the *smtAmukBEF*operon on loss of supercoiling is intriguing, given the importance of *mukB*, *mukE*and *mukF*in chromosome supercoiling and segregation \[[@B45],[@B46]\]. Consistent with this, the XerC site-specific recombinase, which is needed for proper chromosome partitioning, is also repressed by relaxation. As (-) supercoiling promotes chromosome segregation in *E. coli*\[[@B47]\], these genes may represent part of a supercoiling \'checkpoint\' that senses whether supercoiling levels are sufficient for proper chromosome segregation. Thus, if there is insufficient (-) supercoiling to support chromosome segregation, transcription of these genes may be suppressed until supercoiling is re-established.
Another relaxation-repressed gene that may be involved in chromosomal maintenance is *yrdD*, a \'putative topoisomerase\'. *yrdD*encodes a 19 kilodalton (kDa) protein 30-40% identical to the carboxy-terminal domain of topoisomerase I from *Bacillus subtilis*, *Helicobacter pylori*and *Methanococcus jannaschii*. The function of YrdD is unknown, but the repression by chromosomal relaxation provides an intriguing lead.
Chromosomal relaxation leads to the repression of *cls*(cardiolipin synthase) and *ileS*(isoleucine tRNA synthetase), which is consistent with the earlier discovery that these genes were involved in sensitivity to gyrase inhibitors \[[@B48]\]. Also, we noted that the nucleotide salvage genes *deoA*and *deoC*are induced on relaxation. For these genes, DNA relaxation may be a signal of DNA damage, and their induction would allow the cell to recycle nucleotides necessary for DNA repair. Finally, the induction of *rpoD*, the σ^70^subunit of RNA polymerase, may help the cell compensate for the increased difficulty of melting the relaxed DNA template.
What is the basis of supercoiling sensitivity? Most of the well controlled analyses of supercoiling-sensitive promoters, notably of the *lacp*^*s*^and *ilv*~*G*~*P*\[[@B18],[@B49]-[@B51]\], were done on plasmids *in vitro*. The more relevant issue is promoter regulation on chromosomes *in vivo*, where other factors may dominate. The CRP protein increases *lac*operon transcription at the low to moderate superhelicities found *in vivo*, and the nucleoid-associated protein IHF is implicated in the supercoiling sensitivity of the *ilvGMEDA*operon \[[@B52]\]. Also, the relative levels of the nucleoid-associated proteins IHF, H-NS and, especially, Fis, can influence the local topology of DNA and accordingly affect transcription of nearby promoters \[[@B53]-[@B55]\]. We found no significant enrichment of genes regulated by IHF, H-NS, or Fis in our list of SSGs. However, we found that chromosomal relaxation affected different promoters to varying extents, and it is possible that the effect of changes in supercoiling may be amplified or attenuated at specific promoters by the actions of DNA-binding proteins. Finally, the proximity of genes to surrounding promoters and other barriers to supercoil diffusion may affect the response to supercoiling. For example, the modulation of the *Salmonella leu-500*promoter by supercoiling requires that the promoter is either on the chromosome or on a plasmid anchored to the cell membrane by transcription and translation of a gene such as *tetA*\[[@B23]\]. Further analysis of supercoiling-sensitive promoters will be more straightforward with the set of genes identified in this paper and our finding that relaxation-induced genes have an enriched AT content in the promoter and initially transcribed sequences.
It is striking that there are so many relaxation-induced genes that are relatively repressed when the chromosome is (-) supercoiled. This is surprising because (-) supercoiling should favor formation of an open promoter complex. The promoter regions of many of the genes induced by relaxation are AT rich, which will make it easier to form an open promoter complex even when the DNA is relaxed and the energy required is greater. Alternatively, the difference in AT content could reflect structural features such as curvature or flexibility. Curved sequences of DNA can influence the position of plectonemic supercoils, and thus could serve to localize a promoter sequence to the apex of a superhelical loop \[[@B56]\]. We note that the AT richness for the relaxation-induced genes extends on both sides of the transcriptional start site. It has been previously shown that promoter activity can be regulated by the initial transcribed sequence \[[@B57],[@B58]\]. Moreover, in their analysis of the *gyrA*and *gyrB*promoters, Menzel and Gellert \[[@B14]\] found that base-pairs downstream of the transcriptional start were important for the supercoiling sensitivity of these promoters. These authors proposed that promoter clearance may be the rate-limiting step during relaxation-induced transcription of *gyrA*and *gyrB*. Promoter clearance has also been invoked in the mechanism of supercoiling sensitivity of some promoters *in vitro*\[[@B51]\]. As our group of relaxation-induced genes is AT rich over this region, we can extend this hypothesis to transcription of many relaxation-induced genes *in vivo*, and propose that promoter clearance is generally a key regulatory step for supercoiling sensitive transcription.
The AT-rich regions of our relaxation-induced genes extend downstream of the translational start site, and thus may involve transcription elongation in addition to promoter clearance. There is growing appreciation of the regulation of transcription elongation \[[@B59]-[@B61]\]. The AT-rich regions deep within the coding sequence of relaxation-induced genes may reflect such regulation; easily melted regions of DNA may facilitate the continued movement of RNA polymerase along a relaxed, covalently closed template. At a given level of (-) supercoiling, there is likely to be an optimum AT content that facilitates both unwinding and subsequent closure of the transcription bubble. This hypothesis is strengthened by the fact that the genes with the opposite response, the relaxation-repressed genes, have a significantly depressed AT content over the same region.
The SSGs are useful as topological probes of the chromosome in living cells. While the SSGs are scattered throughout the chromosome, they are not evenly spread, but rather have regions of high and low density. The SSGs are plotted on a chromosomal map in Figure [6](#F6){ref-type="fig"}. The density of SSGs as a percentage of all genes in a 20-kb region varies from 2% to more than 20%. The regions with high SSG density may reflect spatial covariations in transcription which were recently described in the *E. coli*chromosome \[[@B62]\]. The distribution of the SSGs may also be influenced by the organization of the chromosome into topologically separate domains of supercoiling. We have already used the SSGs as local reporters of supercoiling to test the domains hypothesis. In recent work, we monitored expression from the SSGs after cleaving the chromosome with a restriction enzyme, and found that the SSGs accurately reported the resulting relaxation of the chromosome \[[@B63]\]. Relaxation diminished rapidly with distance from a restriction site, indicating that there are about 450 topologically separate domains in the chromosome. We also monitored transcription from the SSGs during replication in synchronized cells \[[@B64]\]. Here we found that the relaxation-induced and relaxation-repressed genes reported that supercoiling is re-established very quickly after the passage of the replication fork, again consistent with a large number of topological domains. Thus, the SSGs are not only a useful tool to study promoter regulation and the physiological effects of supercoiling changes, but also can lead to new findings about chromosome structure.
Conclusions
===========
We have shown that supercoiling acts as a transcription factor, with positive and negative effects on specific genes while leaving the majority of the genome unchanged. Like other transcription factors such as TrpR \[[@B65]\] and ArcA-P \[[@B66]\], supercoiling affects transcription from a wider array of genes than at first anticipated. The 306 genes that we identified as robust SSGs are classified into many different functional groups \[[@B67]\], including transcriptional regulators and genes in the SOS, PhoB and stringent-response regulons \[[@B68]\]. Transcriptional changes from the SSGs will affect a variety of transcriptional and regulatory networks, and thus supercoiling level is a global regulator that can affect a wide array of processes in the cell. As the topology of the chromosome is affected by anoxia, ionic strength and growth conditions, the cell can use supercoiling levels both to sense the environment and to effect appropriate transcriptional responses.
Materials and methods
=====================
PCR materials and conditions
----------------------------
Amino- and carboxy-terminal primers for protein-coding open reading frames (ORFs) of *E. coli*K-12, strain MG1655 (Sigma-Genosys), were generously supplied by Fred Blattner (University of Wisconsin) and Carol Gross (University of California San Francisco). ORFs were amplified from MG1655 genomic DNA using ExTaq polymerase (PanVera) and failed PCR reactions were attempted again using Platinum Taq (Invitrogen) or previous successful reactions as the DNA template. Ninety-six percent of the ORFs were successfully amplified. PCR conditions were set according to those supplied with the primers. DNA was precipitated with isopropanol and prepared for microarray printing as described in \[[@B69]\]. We did not include the RNA-coding genes on the arrays because primers for these genes were not initially available, though we note that some genes, such as *tyrT*, have been shown to respond to changes in supercoiling \[[@B70]\].
Microarray printing and processing
----------------------------------
Detailed instructions on slide preparation, microarray printing and processing microarrays can be found online \[[@B69]\]. 384-well plates were dried down between prints and resuspended in deionized water each time after the first print.
RNA preparation and microarray hybridization
--------------------------------------------
*E. coli*cells were grown with shaking in LB media to an OD~600~= 0.45-0.55 at 37°C, or at 30°C for temperature-sensitive strains. Samples of cells were withdrawn at intervals and added to a 1/10 volume of either 95% ethanol plus 5% phenol or 2 M NaN~3~to stop transcription. Cells were then quickly harvested by centrifugation in a microcentrifuge. The supernatant was aspirated and the pellets frozen in liquid N~2~. Total RNA was prepared using the Qiagen RNeasy mini kit, except that 4 mg/ml lysozyme and a 30 sec incubation was used in the first step. For each microarray, 20 μg total RNA was primed with 1-2 μg of random hexamers and labeled by reverse transcription in the presence of Cy3- and Cy5-conjugated dUTP (Amersham Biosciences). For each experiment or condition, a Cy5-labeled experimental sample was combined with a Cy3-labeled reference sample and hybridized to a processed microarray as described \[[@B69]\]. After 5-7 h hybridization, microarrays were washed and scanned at 10 μm resolution with a GenePix 4000A scanner (Axon Instruments).
Image processing
----------------
Scanned array images were visually inspected, and non-uniform spots were excluded from further analysis. The background was subtracted from the images that were then (median) normalized using the Scanalyze 2.0 program, v. 1.44 (Michael Eisen, Lawrence Berkeley National Laboratory) such that the total fluorescence in each channel was equal.
Data analysis
-------------
We tested several methods of imputation to estimate the values of spots missing due to hybridization defects (described in \[[@B71]\]), and after error analysis of the different methods we chose the weighted mean of K-nearest neighbors for K = 20. With this method we obtained a total of 4,003 genes, or 93% of the total number of *E. coli*genes, that could be considered for further study. Because we were interested in changes in expression levels due to variations in supercoiling rather than to drug or genetic effects, we used a two-sample comparison approach (comparing the mean over all relaxation experiments with that of the control experiments) rather than a factorial analysis approach. We tested two commonly used methods to determine differentially expressed genes in the comparison of two samples. We found that the method of Dudoit *et al*\[[@B38]\], which controls the family-wise error (that is, the probability of finding at least one false positive) was slightly more stringent for our data than that developed by Tusher *et al*\[[@B37]\].
Northern analysis
-----------------
Samples were run on formaldehyde-MOPS 1% agarose gels and blotted onto a nylon membrane \[[@B72]\]. ^32^P-labeled DNA probes for *gyrB*and *asnB*(as a loading control) were synthesized from their respective PCR products, and radioactivity was quantified by a phosphorimager.
Assays of DNA topology
----------------------
Plasmid DNA was extracted from cells by the alkaline lysis method \[[@B72]\] or the Qiagen spin miniprep kit. The norfloxacin-resistant mutants and the *gyrB234*mutant are in a C600 strain background, but all strains used have been described in greater detail elsewhere \[[@B26],[@B35],[@B73]\]. To increase the intracellular concentration of novobiocin, we used a *ΔacrA*strain that greatly slows drug efflux \[[@B74]\]. The superhelical density, σ, of pBR322 was determined by band counting \[[@B75]\] from the mean of the topoisomer distributions to a relaxed, covalently closed reference plasmid (σ = 0) which had been relaxed with calf thymus topoI. σ of pBR322 was calculated with the formula σ =Δ Lk/Lk~0~, where Lk~0~for pBR322 = 4,361 bp/10.5 bp/turn = 415. Samples were run on parallel 20-cm gels containing 0, 2.8 or 10 μg/ml chloroquine for 18-26 h at 48-52 V with constant buffer recirculation, which allowed visualization of the entire distribution of topoisomers. Gels were southern blotted \[[@B72]\], and hybridized with a ^32^P-labeled probe made from random priming of pBR322. Radioactive blots were quantified using a phosphorimager.
Microarray validation
---------------------
We tested the validity of our microarrays in three ways. First, we compared gene expression ratios measured with microarrays to values obtained by northern hybridization. We measured induction ratios for *gyrB*by both methods 5 min after addition of the gyrase inhibitor novobiocin to *ΔacrA*cells at 5, 20, 50 and 200 μg/ml. The microarray ratios for these concentrations were 2.3, 4.8, 4.9 and 6.3, respectively, while the ratios from northern hybridizations were 2.8, 4.7, 4.7 and 5.1. Second, as an internal control we compared the transcription of genes in 153 known polycistronic operons. We found no operons with genes that changed expression more than 1.5-fold in opposite directions (data not shown). Third, we compared two identically grown cultures with the same microarray (see Additional data file 4). We used two strains that were isogenic, except that one had point mutations conferring norfloxacin resistance on gyrase and topo IV. The correlation coefficient of the gene-expression levels was 0.982, indicating the negligible variation between the cultures. In contrast, when we treated cells with the gyrase inhibitor norfloxacin (see Additional data file 4), the correlation coefficient with respect to the untreated cells was only 0.391 and hundreds of genes showed large differences in expression. We conclude that gene-expression changes resulting from slight genotypic changes or experimental repeats were negligible compared with the changes resulting from topoisomerase inhibition, and that the *E. coli*microarrays are a reliable method for quantifying these changes.
Selection of supercoiling-sensitive genes
-----------------------------------------
We limited the list of SSGs to those whose expression difference between treatments and controls was statistically significant (*p*-values \< 0.05) over a total of 35 experiments, in which DNA gyrase was inhibited with novobiocin, norfloxacin or by a mutation that rendered gyrase temperature-sensitive. Next we determined the correlation of gene expression with the σ of a reference plasmid in the same cells. To calculate the correlation of gene expression to plasmid supercoiling, we created a model profile made up of the ratio of plasmid σ in each experiment to plasmid σ in the (supercoiled) reference for that experiment (Table [1](#T1){ref-type="table"}). The maximum ratio was scaled to 2.5, representing a σ of 0 (complete relaxation) and the minimum ratio was scaled to 1, representing native supercoiling levels (-0.06). The model repression profile is simply the inverse of the model induction profile. Changes of the arbitrary scaling values did not alter the results. Correlation coefficients in Figure [4](#F4){ref-type="fig"} were calculated with respect to those 13 arrays only.
Additional data files
=====================
The following additional data files are available with the online version of this paper: Additional data file [1](#s1){ref-type="supplementary-material"} contains data on the induction of the SOS response to DNA damage; Additional data file [2](#s2){ref-type="supplementary-material"} contains gene-expression ratios for all genes across all experiments; Additional data file [3](#s3){ref-type="supplementary-material"} contains gene-expression ratios for supercoiling-sensitive genes across all experiments; Additional data file [4](#s4){ref-type="supplementary-material"} contains data on the reproducibility of microarray measurement of RNA levels.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
Data on the induction of the SOS response to DNA damage
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
Gene-expression ratios for all genes across all experiments
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 3
Gene-expression ratios for supercoiling-sensitive genes across all experiments
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 4
Data on the reproducibility of microarray measurement of RNA levels
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank Carol Gross, Wonchul Suh, and Joe DeRisi for sharing PCR primers, technical assistance and useful discussion. We also thank Sydney Kustu and Dan Zimmer for assistance with array printing and databases. Finally we thank Lisa Postow for help with data processing and Figure [6](#F6){ref-type="fig"}, and S. Dudoit, T. Speed and members of the Cozzarelli lab for useful discussion. This work was supported by NIH grants to N.R.C. J.A. was partially supported by NSF grant DMS-9971169. A.M.B. is supported by a Howard Hughes Medical Institute Predoctoral Fellowship. P.O.B. is an associate investigator of the Howard Hughes Medical Institute.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Significance versus correlation of gene expression and plasmid supercoiling values for all genes over all experiments. For each gene we computed the correlation coefficient between its gene expression ratios (base 2 logarithm) over all experiments with the superhelical density (σ) of a reporter plasmid, as measured by gel electrophoresis. These values are plotted against the *p*-value, which represents the chance that the difference in expression between relaxation and control experiments could have arisen randomly. **(a)**Scatter plot for all genes. There is a general trend in which genes with low *p*-values showed very high correlation (absolute value) between expression and plasmid supercoiling. The points corresponding to the topoisomerase genes *gyrA*, *gyrB*, *topA*and *topB*are indicated. **(b)**Expanded portion of (a) highlighting those genes classified as significant (*p*\< 0.05). Genes with very low *p*-values show high positive or negative correlation between expression and plasmid supercoiling.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Expression profiles of relaxation-induced and repressed genes. The figure shows a cluster diagram ordered according to the *p*-value of each gene (from 0.000125 to 0.05). Each row represents a gene and each column an experiment. Therefore, each of the entries of the array shows the expression level for a gene in a given experiment. **(a)**Relaxation-repressed genes; **(b)**relaxation-induced genes. The set of experiments labeled 1 to 14, to the left of the vertical mark in (a and b), represents the control set in which plasmid supercoiling did not change. Experiments to the right of the vertical mark, labeled from 15 to 35, are experiments in which the chromosome is relaxed. As experiments were done in a time-dependent fashion, red color means that gene expression is higher at time points after relaxation of the chromosomes, while green means the opposite. Black indicates no change in expression during the experiment. Columns 1-5, gene expression measured after addition of 15 μg/ml norfloxacin to a norfloxacin-resistant strain at times *t*= 2, 5, 10, 20 or 30 min; columns 19-27, gene expression measured after addition of 15 μg/ml norfloxacin to an isogenic wild-type strain at times *t*= 2, 3, 4, 5, 7, 10, 15, 20 or 30 min; columns 6-10, gene expression at times *t*= 2, 5, 10, 20 or 30 min after addition of 50 μg/ml norfloxacin to a norfloxacin-resistant strain; columns 28-32, gene expression at these times after addition of the same concentration of norfloxacin to an isogenic wild-type strain; columns 15-18, gene expression at times *t*= 2, 5, 10 or 20 min after temperature shift in a temperature-sensitive mutant strain; columns 11-14, gene expression at times *t*= 2, 5, 10 or 20 min after temperature shift in an isogenic wild-type strain; columns 33-35, gene expression at fixed time *t*= 5 min and varying concentrations of novobiocin (Novo) = 20, 50 or 200 μg/ml on a wild-type strain. A total of 200 genes are repressed in response to DNA relaxation, while 106 genes are induced. The top row is a model expression profile of the supercoiling of the reporter plasmid in each experiment (Table 1). *p*-values and correlation coefficients with plasmid supercoiling levels for the top 10% of genes in each class are listed. The complete expression data for each gene can be found in Additional data file 2.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Plasmid relaxation kinetics. *gyrA*^+^*parC*^+^cells were treated with 15 μg/ml norfloxacin for the indicated times before samples were removed for DNA and microarray analysis. **(a)**pBR322 plasmid DNA was isolated and run on a 1% agarose gel + 2.8 μg/ml chloroquine to provide an indicator of topoisomerase activity in the cells. The positions of open circular (oc) and relaxed (rel) marker plasmids on the gel are shown. The distribution of native (-) supercoiled DNA is shown in lane 1. As the plasmid becomes relaxed, the center of the distribution first moves toward the open circular form and then moves down the gel to the relaxed position. The calculated superhelical density values for the plasmids (σ) are shown at bottom of each lane. **(b)**Graph of the average σ values from (a). Values of σ stabilized around 0 for times greater than 10 min and are not shown.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Kinetics of the expression changes of the supercoiling-sensitive genes. Norfloxacin was added to wild-type *E. coli*cells and RNA was extracted from cells removed from the culture at the time points shown (in minutes) above each column. This diagram illustrates the kinetics of the SSG responses, which are ranked by their correlation to plasmid supercoiling levels in this experiment (see Figure 3). *p*-values and correlation coefficients for each gene are listed (see Materials and methods for calculation). The model profiles shown at the top are colored representations of plasmid supercoiling levels, as in Figure 2. Red squares indicate that a gene is induced during the experiment, green squares that a gene is repressed.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Analysis of AT content in upstream regions of SSGs. **(a)**The average upstream AT content of 50,000 groups of 106 randomly selected genes. The actual average upstream AT content of the group of 106 relaxation-induced genes (red circle) lies well outside the distribution. **(b)**Average AT content in a 100-nucleotide window is plotted against distance from the start codon for relaxation-induced (red), relaxation-repressed (green) and all other (black) genes for 300 nucleotides to either side of the translation start site. The *y*-axis is drawn at the first nucleotide of the start codon, and a horizontal line indicates 50% AT content. The relaxation-induced genes show a significantly increased AT content relative to the other sets of genes both before and after the start codon. The relaxation-repressed genes show a milder depression of AT content over this region, which is still significantly different from the rest of the genome. We found no significant differences outside the region shown.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
Chromosomal map of SSGs. Supercoiling-sensitive genes were mapped across the *E. coli*genome. Relaxation-induced genes are colored red and relaxation-repressed genes are in green. Genes are dispersed through the entire chromosome, making them good sensors for local changes of supercoiling of the chromosome.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Plasmid supercoiling measurements from relaxation experiments
:::
Genotype Experimental treatment Plasmid σ Model ratio
------------------------ ------------------------- ----------- -------------
*gyrA*^*R*^*parC*^*R*^ **15 μg/ml Nor, 0 min** -0.057 1.0
*gyrA*^*R*^*parC*^*R*^ 15 μg/ml Nor, 2 min -0.057 1.0
*gyrA*^*R*^*parC*^*R*^ 15 μg/ml Nor, 5 min -0.057 1.0
*gyrA*^*R*^*parC*^*R*^ 15 μg/ml Nor, 10 min -0.057 1.0
*gyrA*^*R*^*parC*^*R*^ 15 μg/ml Nor, 20 min -0.060 1.0
*gyrA*^*R*^*parC*^*R*^ 15 μg/ml Nor, 30 min -0.060 0.9
*gyrA*^*R*^*parC*^*R*^ **50 μg/ml Nor, 0 min** -0.057 1.0
*gyrA*^*R*^*parC*^*R*^ 50 μg/ml Nor, 2 min -0.057 1.0
*gyrA*^*R*^*parC*^*R*^ 50 μg/ml Nor, 5 min -0.057 1.0
*gyrA*^*R*^*parC*^*R*^ 50 μg/ml Nor, 10 min -0.058 1.0
*gyrA*^*R*^*parC*^*R*^ 50 μg/ml Nor, 20 min -0.059 1.0
*gyrA*^*R*^*parC*^*R*^ 50 μg/ml Nor, 30 min -0.061 0.9
*gyrB*^+^ **37°C** -0.059 1.0
*gyrB*^+^ 42°C, 2 min -0.059 1.0
*gyrB*^+^ 42°C, 5 min -0.059 1.0
*gyrB*^+^ 42°C, 10 min -0.059 1.0
*gyrB*^+^ 42°C, 20 min -0.061 1.0
*gyrB*^*TS*^ **37°C** -0.044 1.0
*gyrB*^*TS*^ 42°C, 2 min -0.023 1.6
*gyrB*^*TS*^ 42°C, 5 min -0.016 1.8
*gyrB*^*TS*^ 42°C, 10 min ND ND
*gyrB*^*TS*^ 42°C, 20 min 0.000 2.5
*gyrA*^+^*parC*^*R*^ **15 μg/ml Nor, 0 min** -0.057 1.0
*gyrA*^+^*parC*^*R*^ 15 μg/ml Nor, 2 min -0.025 1.7
*gyrA*^+^*parC*^*R*^ 15 μg/ml Nor, 5 min -0.009 2.2
*gyrA*^+^*parC*^*R*^ 15 μg/ml Nor, 10 min -0.006 2.3
*gyrA*^+^*parC*^*R*^ 15 μg/ml Nor, 20 min -0.002 2.4
*gyrA*^+^*parC*^*R*^ 15 μg/ml Nor, 30 min 0.000 2.5
*gyrA*^+^*parC*^+^ **50 μg/ml Nor, 0 min** -0.057 1.0
*gyrA*^+^*parC*^+^ 50 μg/ml Nor, 2 min -0.016 1.9
*gyrA*^+^*parC*^+^ 50 μg/ml Nor, 5 min -0.002 2.4
*gyrA*^+^*parC*^+^ 50 μg/ml Nor, 10 min 0.000 2.5
*gyrA*^+^*parC*^+^ 50 μg/ml Nor, 20 min 0.007 2.8
*gyrA*^+^*parC*^+^ 50 μg/ml Nor, 30 min 0.016 3.2
*gyrA*^+^*parC*^+^ **15 μg/ml Nor, 0 min** -0.055 1.0
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 10 sec -0.048 1.1
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 25 sec -0.040 1.3
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 45 sec -0.032 1.5
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 1 min -0.027 1.6
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 1.5 min -0.022 1.7
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 2 min -0.017 1.9
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 3 min -0.012 2.1
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 4 min -0.008 2.2
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 5 min -0.003 2.4
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 7 min 0.000 2.5
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 10 min 0.003 2.6
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 15 min 0.002 2.6
*gyrA*^+^*parC*^+^ 15 μg/ml Nor, 20 min 0.000 2.5
*ΔacrA* **0 μg/ml Novo, 0 min** -0.057 1.0
*ΔacrA* 20 μg/ml Novo, 5 min -0.011 2.1
*ΔacrA* 50 μg/ml Novo, 5 min 0.005 2.7
*ΔacrA* 200 μg/ml Novo, 5 min 0.011 3.0
pBR322 plasmid DNA was isolated from cells and analyzed by electrophoresis. Experimental treatments in bold indicate samples taken immediately before addition of drug or temperature shift, which were used as a reference for the following time points. Model ratios represent values derived from plasmid σ by taking the ratio of σ in each time point and dividing by σ in the reference, and scaling that value such that a sigma of 0 corresponds to a model ratio of 2.5. Nor, norfloxacin; Novo, novobiocin; SSG, supercoiling-sensitive gene; TS, temperature-sensitive. ND, not determined.
:::
|
PubMed Central
|
2024-06-05T03:55:51.809329
|
2004-11-1
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545778/",
"journal": "Genome Biol. 2004 Nov 1; 5(11):R87",
"authors": [
{
"first": "Brian J",
"last": "Peter"
},
{
"first": "Javier",
"last": "Arsuaga"
},
{
"first": "Adam M",
"last": "Breier"
},
{
"first": "Arkady B",
"last": "Khodursky"
},
{
"first": "Patrick O",
"last": "Brown"
},
{
"first": "Nicholas R",
"last": "Cozzarelli"
}
]
}
|
PMC545779
|
Background
==========
*Cryptosporidium*is a member of the Apicomplexa, a eukaryotic phylum that includes several important parasitic pathogens such as *Plasmodium*, *Toxoplasma*, *Eimeria*and *Theileria*. As an emerging pathogen in humans and other animals, *Cryptosporidium*often causes fever, diarrhea, anorexia and other complications. Although cryptosporidial infection is often self-limiting, it can be persistent and fatal for immunocompromised individuals. So far, no effective treatment is available \[[@B1]\]. Furthermore, because of its resistance to standard chlorine disinfection of water, *Cryptosporidium*continues to be a security concern as a potential water-borne bioterrorism agent \[[@B2]\].
*Cryptosporidium*is phylogenetically quite distant from the hemosporidian and coccidian apicomplexans \[[@B3]\] and, depending on the molecule and method used, is either basal to all Apicomplexa examined thus far, or is the sister group to the gregarines \[[@B4],[@B5]\]. It is unusual in several respects, notably for the lack of the apicoplast organelle which is characteristic of all other apicomplexans that have been examined \[[@B6],[@B7]\]. The apicoplast is a relict plastid hypothesized to have been acquired by an ancient secondary endosymbiosis of a pre-alveolate eukaryotic cell with an algal cell \[[@B8]\]. All that remains of the endosymbiont in Coccidia and Haemosporidia is a plastid organelle surrounded by four membranes \[[@B9]\]. The apicoplast retains its own genome, but this is much reduced (27-35 kilobases (kb)), and contains genes primarily involved in the replication of the plastid genome \[[@B10],[@B11]\]. In apicomplexans that have a plastid, many of the original plastid genes appear to have been lost (for example, photosynthesis genes) and some genes have been transferred to the host nuclear genome; their proteins are reimported into the apicoplast where they function \[[@B12]\]. Plastids acquired by secondary endosymbiosis are scattered among eukaryotic lineages, including cryptomonads, haptophytes, alveolates, euglenids and chlorarachnions \[[@B13]-[@B17]\]. Among the alveolates, plastids are found in dinoflagellates and most examined apicomplexans but not in ciliates. Recent studies on the nuclear-encoded, plastid-targeted glyceraldehyde-3-phosphate dehydrogenase (GAPDH) gene suggest a common origin of the secondary plastids in apicomplexans, some dinoflagellates, heterokonts, haptophytes and cryptomonads \[[@B8],[@B18]\]. If true, this would indicate that the lineage that gave rise to *Cryptosporidium*contained a plastid, even though many of its descendants (for example, the ciliates) appear to lack a plastid. Although indirect evidence has been noted for the past existence of an apicoplast in *C. parvum*\[[@B19],[@B20]\], no rigorous phylogenomic survey for nuclear-encoded genes of plastid or algal origin has been reported.
Gene transfers, either intracellular (IGT) from an endosymbiont or organelle to the host nucleus or horizontal (HGT) between species, can dramatically alter the biochemical repertoire of host organisms and potentially create structural or functional novelties \[[@B21]-[@B23]\]. In parasites, genes transferred from prokaryotes or other sources are potential targets for chemotherapy due to their phylogenetic distance or lack of a homolog in the host \[[@B24],[@B25]\]. The detection of transferred genes in *Cryptosporidium*is thus of evolutionary and practical importance.
In this study, we use a phylogenomic approach to mine the recently sequenced genome of *C. parvum*(IOWA isolate; 9.1 megabases (Mb)) \[[@B7]\] for evidence of the past existence of an endosymbiont or apicoplast organelle and of other independent HGTs into this genome. We have detected genes of cyanobacterial/algal origin and genes acquired from other prokaryotic lineages in *C. parvum*. The fate of several of these transferred genes in *C. parvum*is explored by expression analyses. The significance of our findings and their impact on the genetic makeup of the parasite are discussed.
Results
=======
BLAST analyses
--------------
From BLAST analyses, the genome of *Cryptosporidium*, like that of *Plasmodium falciparum*\[[@B26]\], is more similar overall to those of the plants *Arabidopsis*and *Oryza*than to any other non-apicomplexan organism currently represented in GenBank. The program Glimmer predicted 5,519 protein-coding sequences in the *C. parvum*genome, 4,320 of which had similarity to other sequences deposited in the GenBank nonredundant protein database. A significant number of these sequences, 936 (E-value \< 10^-3^) or 783 (E-value \< 10^-7^), had their most significant, non-apicomplexan, similarity to a sequence isolated from plants, algae, eubacteria (including cyanobacteria) or archaea (Table [1](#T1){ref-type="table"}). To evaluate these observations further, phylogenetic analyses were performed, when possible, for each predicted protein in the entire genome.
Phylogenomic analyses
---------------------
The Glimmer-predicted protein-coding regions of the *C. parvum*genome (5,519 sequences) were used as input for phylogenetic analyses using the PyPhy program \[[@B27]\]. In this program, phylogenetic trees for each input sequence are analyzed to determine the taxonomic identity of the nearest neighbor relative to the input sequence at a variety of taxonomic levels, for example, genus, family, or phylum. Using stringent analysis criteria (see Materials and methods), 954 trees were constructed from the input set of 5,519 predicted protein sequences (Figure [1](#F1){ref-type="fig"}). Analysis of the nearest non-apicomplexan neighbor on the 954 trees revealed the following nearest neighbor relationships: eubacterial (115 trees), archaeal (30), green plant/algal (204), red algal (8), and glaucocystophyte (4); other alveolate (61) and other eukaryotes made up the remainder. As some input sequences may have more than one nearest neighbor of interest on a tree, a nonredundant total of 393 sequences were identified with nearest neighbors to the above lineages.
Searches of the *C. parvum*predicted gene set with the 551 *P. falciparum*predicted nuclear-encoded apicoplast-targeted proteins (NEAPs) yielded 40 significant hits (E-value \< 10^-5^), 23 of which were also identified in the phylogenomic analyses. A combination of these two approaches identified 410 candidates requiring further detailed analyses. Of these candidates, the majority were eliminated after stringent criteria were applied because of ambiguous tree topologies, insufficient taxonomic sampling, lack of bootstrap support or the presence of clear vertical eukaryotic ancestry (see Materials and methods). Thirty-one genes survived the screen and were deemed to be either strong or likely candidates for gene transfer (Table [2](#T2){ref-type="table"}).
Of the 31 recovered genes, several have been previously published or submitted to the GenBank \[[@B20]\], including those identified as having plant or eubacterial \'likeness\' on the basis of similarity searches when the genome sequence was published \[[@B7]\]. The remaining sequences were further tested to rule out the possibility that they were artifacts (*C. parvum*oocysts are purified from cow feces which contain plant and bacterial matter). Two experiments were performed. In the first, nearly complete genomic sequences (generated in a different laboratory) from the closely related species *C. hominis*were screened using BLASTN for the existence of the predicted genes. Twenty out of 21 *C. parvum*sequences were identified in *C. hominis*. The remaining sequence was represented by two independently isolated expressed sequence tag (EST) sequences in the GenBank and CryptoDB databases (data not shown). In the second experiment, genomic Southern analyses of the IOWA isolate were carried out (Figure [2](#F2){ref-type="fig"}) for several of the genes of bacterial or plant origin. In each case, a band of the predicted size was identified (see Additional data file 1). The genes are not contaminants.
Genes of cyanobacterial/algal origin
------------------------------------
Extant *Cryptosporidium*species do not contain an apicoplast genome or any physical structure thought to represent an algal endosymbiont or the plastid organelle it contained \[[@B6],[@B7]\]. The only possible remaining evidence of the past association of an endosymbiont or its cyanobacterially derived plastid organelle might be genes transferred from these genetic sources to the host genome prior to the physical loss of the endosymbiont or organelle itself. Several such genes were identified.
A leucine aminopeptidase gene of cyanobacterial origin was found in the *C. parvum*nuclear genome. This gene is also present in the nuclear genome of other apicomplexan species (*Plasmodium*, *Toxoplasma*and *Eimeria*), as confirmed by similarity searches against ApiDB (see Materials and methods). In *P. falciparum*, leucine aminopeptidase is a predicted NEAP and possesses an amino-terminal extension with a putative transit peptide. Consistent with the lack of an apicoplast, this gene in *Cryptosporidium*contains no evidence of a signal peptide and the amino-terminal extension is reduced. Similarity searches of the GenBank nonredundant protein database revealed top hits to *Plasmodium*, followed by *Arabidopsis thaliana*, and several cyanobacteria including *Prochlorococcus*, *Nostoc*and *Trichodesmium*, and plant chloroplast precursors in *Lycopersicon esculentum*and *Solanum tuberosum*(data not shown). A multiple sequence alignment of the predicted protein sequences of leucine aminopeptidase reveals overall similarity and a shared indel among apicomplexan, plant and cyanobacterial sequences (Figure [3](#F3){ref-type="fig"}). Phylogenetic analyses strongly support a monophyletic grouping of *C. parvum*and other apicomplexan leucine aminopeptidase proteins with cyanobacteria and plant chloroplast precursors (Figure [4a](#F4){ref-type="fig"}). So far, this gene has not been detected in ciliates.
Another *C. parvum*nuclear-encoded gene of putative cyanobacterial origin is a protein of unknown function belonging to the biopterine transporter family (BT-1) (Table [2](#T2){ref-type="table"}). Similarity searches with this protein revealed significant hits to other apicomplexans (for example, *P. falciparum*, *Theileria annulata*, *T. gondii*), plants (*Arabidopsis*, *Oryza*), cyanobacteria (*Trichodesmium*, *Nostoc*and *Synechocystis*), a ciliate (*Tetrahymena*) and the kinetoplastids (*Leishmania*and *Trypanosoma*). *Arabidopsis thaliana*apparently contains at least two copies of this gene; the protein of one (accession number NP\_565734) is predicted by ChloroP \[[@B28]\] to be chloroplast-targeted, suggestive of its plastid derivation. The taxonomic distribution and sequence similarity of this protein with cyanobacterial and chloroplast homologs are also indicative of its affinity to plastids.
Only one gene of algal nuclear origin, glucose-6-phosphate isomerase (G6PI), was identified by the screen described here. Several other algal-like genes are probable, but their support was weaker (Table [2](#T2){ref-type="table"}). A \'plant-like\' G6PI has been described in other apicomplexan species (*P. falciparum*, *T. gondii*\[[@B29]\]) and a \'cyanobacterial-like\' G6PI has been described in the diplomonads *Giardia intestinalis*and *Spironucleus*and the parabasalid *Trichomonas vaginalis*\[[@B30]\]. Figure [4b](#F4){ref-type="fig"} illustrates these observations nicely. At the base of the tree, the eukaryotic organisms *Giardia*, *Spironucleus*and *Trichomonas*group with the cyanobacterium *Nostoc*, as previously published. In the midsection of the tree, the G6PI of apicomplexans and ciliates forms a well-supported monophyletic group with the plants and the heterokont *Phytophthora*. The multiple protein sequence alignment of G6PI identifies several conserved positions shared exclusively by apicomplexans, *Tetrahymena*, plants and *Phytophthora*. This gene does not contain a signal or transit peptide and is not predicted to be targeted to the apicoplast in *P. falciparum*. The remainder of the tree shows a weakly supported branch including eubacteria, fungi and several eukaryotes. The eukaryotes are interrupted by the inclusion of G6PI from the eubacterial organisms *Escherichia coli*and *Cytophaga*. This relationship of *E. coli*G6PI and eukaryotic G6PI has been observed before and may represent yet another gene transfer \[[@B31]\].
Genes of eubacterial (non-cyanobacterial) origin
------------------------------------------------
Our study identified HGTs from several distinct sources, involving a variety of biochemical activities and metabolic pathways (Table [2](#T2){ref-type="table"}). Notably, the nucleotide biosynthesis pathway contains at least two previously published, independently transferred genes from eubacteria. Inosine 5\' monophosphate dehydrogenase (IMPDH), an enzyme for purine salvage, was transferred from ε-proteobacteria \[[@B32]\]. Another enzyme involved in pyrimidine salvage, thymidine kinase (TK), is of α or γ-proteobacterial ancestry \[[@B25]\].
Another gene of eubacterial origin identified in *C. parvum*is tryptophan synthetase β subunit (*trpB*). This gene has been identified in both *C. parvum*and *C. hominis*, but not in other apicomplexans. The relationship of *C. parvum trpB*to proteobacterial sequences is well-supported as a monophyletic group by two of the three methods used in our analyses (Figure [4c](#F4){ref-type="fig"}).
Other HGTs of eubacterial origin include the genes encoding α-amylase and glutamine synthetase and two copies of 1,4-α-glucan branching enzyme, all of which are overwhelmingly similar to eubacterial sequences. α-amylase shows no significant hit to any other apicomplexan or eukaryotic sequence, suggesting a unique HGT from eubacteria to *C. parvum*. Glutamine synthetase is a eubacterial gene found in *C. parvum*and all apicomplexans examined. The eubacterial affinity of the apicomplexan glutamine synthetase is also demonstrated by a well supported (80% with maximum parsimony) monophyletic grouping with eubacterial homologs (data not shown). The eubacterial origin of 1,4-α-glucan branching enzyme is shown in Figure [5](#F5){ref-type="fig"}. Each copy of the gene is found in a strongly supported monophyletic group of sequences derived only from prokaryotes (including cyanobacteria) and one other apicomplexan organism, *T. gondii*. It is possible that these genes are of plastidic origin and were transferred to the nuclear genome before the divergence of *C. parvum*and *T. gondii*; the phylogenetic analysis provides little direct support for this interpretation, however.
Mode of acquisition
-------------------
We examined the transferred genes for evidence of non-independent acquisition, for example, blocks of transferred genes or evidence that genes were acquired together from the same source. Examination of the chromosomal location of the genes listed in Table [2](#T2){ref-type="table"} demonstrates that the genes are currently located on different chromosomes and in most cases do not appear to have been transferred or retained in large blocks. There are two exceptions. The *trpB*gene and the gene for aspartate ammonia ligase are located 4,881 base-pairs (bp) apart on the same strand of a contig for chromosome V; there is no annotated gene between these two genes. Both genes are of eubacterial origin and are not found in other apicomplexan organisms. While it is possible that they have been acquired independently with this positioning, or later came to have this positioning via genome rearrangements, it is interesting to speculate that these genes were acquired together. The origin of *trpB*is proteobacterial. The origin of aspartate ammonia ligase is eubacterial, but not definitively of any particular lineage. In the absence of genome sequences for all organisms, throughout all of time, exact donors are extremely difficult to assess and inferences must be drawn from sequences that appear to be closely related to the actual donor.
In the second case, *C. parvum*encodes two genes for 1,4-α-glucan branching enzymes. Both are eubacterial in origin and both are located on chromosome VI, although not close together. They are approximately 110 kb apart and many intervening genes are present. The evidence that these genes were acquired together comes from the phylogenetic analysis presented in Figure [5](#F5){ref-type="fig"}. The duplication that gave rise to the two 1,4-α-glucan branching enzymes is old, and is well supported by the tree shown in Figure [5](#F5){ref-type="fig"}. A number of eubacteria (11), including cyanobacteria, contain this duplication. The 1,4-α-glucan branching enzymes of *C. parvum*and *T. gondii*represent one copy each of this ancient duplication. This suggests that the ancestor of *C. parvum*and *T. gondii*acquired the genes after they had duplicated and diverged in eubacteria.
Expression of transferred genes
-------------------------------
Each of the genes identified in the above analyses (Table [2](#T2){ref-type="table"}) appears to be an intact non-pseudogene, suggesting that these genes are functional. To verify the functional status of several of the transferred genes, semi-quantitative reverse transcription PCR (RT-PCR) was carried out to characterize their developmental expression profile. Each of the RNA samples from *C. parvum*-infected HCT-8 cells was shown to be free of contaminating *C. parvum*genomic DNA by the lack of amplification product from a reverse transcriptase reaction sham control. RT-PCR detected no signals in cDNA samples from mock-infected HCT-8 cells. On the other hand, RT-PCR product signals were detected in the *C. parvum*-infected cells of six independent time-course experiments for each of the genes examined (those for G6PI, leucine aminopeptidase, BT-1, a calcium-dependent protein kinase, tyrosyl-tRNA synthetase, dihydrofolate reductase- thymidine synthetase (DHFR-TS)). The expression profiles of the acquired genes show that they are regulated and differentially expressed throughout the life cycle of *C. parvum*in patterns characteristic of other non-transferred genes (Figure [6](#F6){ref-type="fig"}).
A small published collection of 567 EST sequences for *C. parvum*is also available. These ESTs were searched with each of the 31 candidate genes surviving the phylogenomic screen. Three genes - aspartate ammonia ligase, BT-1 and lactate dehydrogenase - are expressed, as confirmed by the presence of an EST (Table [2](#T2){ref-type="table"}).
Discussion
==========
A genome-wide search for intracellular and horizontal gene transfers in *C. parvum*was carried out. We systematically determined the evolutionary origins of genes in the genome using phylogenetic approaches, and further confirmed the existence and expression of putatively transferred genes with laboratory experiments. The methodology adopted in this study provides a broad picture of the extent and the importance of gene transfer in apicomplexan evolution.
The identification of gene transfers is often subject to errors introduced by methodology, data quality and taxonomic sampling. The phylogenetic approach adopted in this study is preferable to similarity searches \[[@B33],[@B34]\] but several factors, including long-branch attraction, mutational saturation, lineage-specific gene loss and acquisition, and incorrect identification of orthologs, can distort the topology of a gene tree \[[@B35],[@B36]\]. Incompleteness in the taxonomic record may also lead to false positives for IGT and HGT identification. In our study, we have attempted to alleviate these factors, as best as is possible, by sampling the GenBank nonredundant protein database, dbEST and organism-specific databases and by using several phylogenetic methods. Still, these issues remain a concern for this study as the taxonomic diversity of unicellular eukaryotes is vastly undersampled and studies are almost entirely skewed towards parasitic organisms.
The published analysis of the *C. parvum*genome sequence identified 14 bacteria-like and 15 plant-like genes based on similarity searches \[[@B7]\]. Six of these bacterial-like and three plant-like genes were also identified as probable transferred genes in the phylogenomic analyses presented here. We have examined the fate of genes identified by one analysis and not the other to uncover the origin of the discrepancy. First, methodology is the single largest contributing factor. Genes with bacterial-like or plant-like BLAST similarities which, from the phylogenetic analyses, do not appear to be transfers were caused by the fact that PyPhy was unable to generate trees due to an insufficient number of significant hits in the database, or because of the stringent coverage length and similarity requirements adopted in this analysis. Only seven of the previously identified 15 plant-like and 11 of 14 eubacterial-like genes survived the predefined criteria for tree construction. Second, subsequent phylogenetic analyses including additional sequences from non-GenBank databases failed to provide sufficient evidence or significant support for either plant or eubacterial ancestry. Third, searches of dbEST and other organism-specific databases yielded other non-plant or non-eubacterial organisms as nearest neighbors, thus removing the possibility of a transfer.
The limitations of similarity searches and incomplete taxonomic sampling are well evidenced in our phylogenomic analyses. From similarity searches, *C. parvum*, like *P. falciparum*\[[@B26]\], is more similar to the plants *Arabidopsis*and *Oryza*than to any other single organism. Almost 800 predicted genes have best non-apicomplexan BLAST hits of at least 10^-7^to plants and eubacteria (Table [1](#T1){ref-type="table"}). Yet only 31 can be inferred to be transferred genes at this time with the datasets and methodology available (Table [2](#T2){ref-type="table"}). In many cases (for example, phosphoglucomutase) the *C. parvum*gene groups phylogenetically with plant and bacterial homologs, but with only modest support. In other cases, such as pyruvate kinase and the bi-functional dehydrogenase enzyme (AdhE), gene trees obtained from automated PyPhy analyses indicate a strong monophyletic grouping of the *C. parvum*gene with plant or eubacterial homologs, but this topology disappears when sequences from other unicellular eukaryotes, such as *Dictyostelium*, *Entamoeba*and *Trichomonas*are included in the analysis (data not shown).
The list of genes in Table [2](#T2){ref-type="table"} should be considered a current best estimate of the IGTs and HGTs in *C. parvum*instead of a definitive list. As genomic data are obtained from a greater diversity of unicellular eukaryotes and eubacteria, phylogenetic analyses of nearest neighbors are likely to change.
Did *Cryptosporidium* contain an endosymbiont or plastid organelle?
-------------------------------------------------------------------
The *C. parvum*sequences of cyanobacterial and algal origin reported here had to enter the genome at some point during its evolution. Formal possibilities include vertical inheritance from a plastid-containing chromalveolate ancestor, HGT from the cyanobacterial and algal sources (or from a secondary source such as a plastid-containing apicomplexan), or IGT from an endosymbiont/plastid organelle during evolution, followed by loss of the source. *Cryptosporidium*does not harbor an apicoplast organelle or any trace of a plastid genome \[[@B7]\]; thus an IGT scenario would necessitate loss of the organelle in *Cryptosporidium*or the lineage giving rise to it. The exact position of *C. parvum*on the tree of life has been debated, with developmental and morphological considerations placing it within the Apicomplexa, and molecular analyses locating it in various positions, both within and outside the Apicomplexa \[[@B3]\], but primarily within. If we assume that *C. parvum*is an apicomplexan, and if the secondary endosymbiosis which is believed to have given rise to the apicoplast occurred before the formation of the Apicomplexa, as has been suggested \[[@B18]\], *C. parvum*would have evolved from a plastid-containing lineage and would be expected to harbor traces of this relationship in its nuclear genome. Genes of likely cyanobacterial and algal/plant origin are detected in the nuclear genome of *C. parvum*(Table [2](#T2){ref-type="table"}) and thus IGT followed by organelle loss cannot be ruled out.
What about other interpretations? While it is formally possible that these genes were acquired independently via HGT in *C. parvum*, their shared presence in other alveolates (including the non-plastidic ciliate *Tetrahymena*) provides the best evidence against this scenario as multiple independent transfers would be required and so far there is no evidence for intra-alveolate gene transfer. Vertical inheritance is more difficult to address as it involves distinguishing between genes acquired via IGT from a primary endosymbiotic event versus a secondary endosymbioic event. Our data, especially the analysis of G6PI and BT-1 are consistent with both primary and secondary endosymbioses, provided that the secondary endosymbiosis is pre-alveolate in origin. As more genome data become available and flanking genes can be examined for each gene in a larger context, positional information will be informative in distinguishing among the alternatives.
The plastidic nature of some genes is particularly apparent. There is a shared indel among leucine aminopeptidase protein sequences in apicomplexans, cyanobacteria and plant chloroplast precursors (Figure [3](#F3){ref-type="fig"}). The *C. parvum*leucine aminopeptidase does contain an amino-terminal extension of approximately 85-65 amino acids (depending on the alignment) relative to bacterial homologs, but this extension does not contain a signal sequence. The extension in *P. falciparum*is 85 amino acids and the protein is believed to be targeted to the apicoplast \[[@B26],[@B37]\]. No similarity is detected between the *C. parvum*and *P. falciparum*amino-terminal extensions (data not shown).
Other genes were less informative in this analysis. Among these, aldolase was reported in both *P. falciparum*\[[@B38]\] and the kinetoplastid parasite *Trypanosoma*\[[@B38]\] as a plant-like gene. The protein sequences of aldolase are similar in *C. parvum*and *P. falciparum*, with an identity of 60%. In our phylogenetic analyses, *C. parvum*clearly forms a monophyletic group with *Plasmodium*, *Toxoplasma*and *Eimeria*. This branch groups with *Dictyostelium*, Kinetoplastida and cyanobacterial lineages, but bootstrap support is not significant. The sister group to the above organisms are the plants and additional cyanobacteria, but again with no bootstrap support (see Additional data file 1 for phylogenetic tree). Another gene, enolase, contains two indels shared between land plants and apicomplexans (including *C. parvum*) and was suggested to be a plant-like gene \[[@B29]\], but alternative explanations exist \[[@B39]\].
The biochemical activity of the polyamine biosynthetic enzyme arginine decarboxylase (ADC), which is typically found in plants and bacteria, was previously reported in *C. parvum*\[[@B19]\]. However, we were unable to confirm its presence by similarity searches of the two *Cryptosporidium*genome sequences deposited in CryptoDB using plant (*Cucumis sativa*, GenBank accession number AAP36992), cyanobacterial (*Nostoc*sp., NP-487441; *Synechocystis*sp., NP-439907) and other bacterial (*Yersinia pestis*, NP-404547) homologs.
A plethora of prokaryotic genes
-------------------------------
Several HGTs from bacteria have been reported previously in *C. parvum*\[[@B25],[@B32],[@B40]\]. We detected many more in our screen of the completed *C. parvum*genome sequence (Table [2](#T2){ref-type="table"}). In most cases, the exact donors of these transferred genes were difficult to determine. However, for those genes whose donors could be more reliably inferred (Table [2](#T2){ref-type="table"}), several appear to be from different sources and hence represent independent transfer events. In one compelling case, both the *trpB*and aspartate ammonia ligase genes are located 4,881 bp apart on the same strand of a contig for chromosome V and there is no gene separating them. Both genes are of eubacterial origin and neither gene is detected in other apicomplexans. In addition, the aspartate ammonia ligase gene is expressed, as evidenced by an EST. In another case, copies of a 1,4-α-glucan branching enzyme gene duplication pair that is present in many eubacteria, were detected on the same chromosome in *C. parvum*. *C. parvum*also contains many transferred genes from distinct eubacterial sources that are not present in other apicomplexans (for example, IMPDH, TK (thymidine kinase), *trpB*and the gene for aspartate ammonia ligase).
The endosymbiotic event that gave rise to the mitochondrion occurred very early in eukaryotic evolution and is associated with significant IGT. However, most of these transfer events happened long before the evolutionary time window we explored in this study \[[@B41]\]. Many IGTs from the mitochondrial genome that have been retained are almost universally present in eukaryotes (including *C. parvum*which does not contain a typical mitochondrion \[[@B7],[@B42]-[@B44]\]) and thus would not be detected in a PyPhy screen since the \'nearest phylogenetic neighbor\' on the tree would be taxonomically correct and not appear as a relationship indicative of a gene transfer.
The impact of gene transfers on host evolution
----------------------------------------------
Gene transfer is an important evolutionary force \[[@B21],[@B22],[@B45],[@B46]\]. Several of the transferred genes identified in *C. parvum*are known to be expressed. IMPDH has been shown to be essential in *C. parvum*purine metabolism \[[@B32]\] and TK has been shown to be functional in pyrimidine salvage \[[@B25]\]. It is not yet clear whether these genes were acquired independently in this lineage, or have been lost from the rest of the apicomplexan lineage, or whether both these have happened. However, it is clear that their presence has facilitated the remodeling of nucleotide biosynthesis. *C. parvum*no longer possesses the ability to synthesize nucleotides; instead it relies entirely on salvage.
Many apicoplast and algal nuclear genes have been transferred to the host nuclear genome, where they were subsequently translated in the cytosol and their proteins targeted to the apicoplast organelle. However, as there is no apicoplast in *C. parvum*, acquired plastidic proteins are theoretically destined to go elsewhere. In the absence of an apicoplast, it is tempting to suspect that plastid-targeted proteins would have been lost, or would be detected as pseudogenes. No identifiable pseudogenes were detected and at least one gene is still viable. The *C. parvum*leucine aminopeptidase, which still contains an amino-terminal extension (without a signal peptide), is intact and is expressed, as shown in Figure [6](#F6){ref-type="fig"}. None of the cyanobacterial/algal genes identified in our study contains a canonical presequence for apicoplast targeting. One exception to this is phosphoglucomutase, a gene not present in Table [2](#T2){ref-type="table"} because of its poorly supported relationships in phylogenetic analyses. This gene exists in two copies as a tandem duplication in the *C. parvum*genome. One copy has a long amino-terminal extension (97 amino acids) beginning with a signal peptide. The extension does not contain characteristics of a transit peptide. Expression of a fluorescent reporter construct containing this extension in a related parasite, *T. gondii*, did not reveal apicoplast targeting but instead secretion via dense granules (see Additional data file 1). Exactly how and where intracellularly transferred genes (especially those that normally target the apicoplast) have become incorporated into other metabolic processes remains a fertile area for exploration.
Conclusions
===========
*Cryptosporidium*is the recipient of a large number (31) of transferred genes, many of which are not shared by other apicomplexan parasites. The genes have been acquired from several different sources including α-, β-, and ε-proteobacteria, cyanobacteria, algae/plants and possibly the Archaea. We have described two cases of two genes that appear to have been acquired together from a eubacterial source: *trpB*and the aspartate ammonia ligase gene are located within 5 kb of each other, while the two copies of 1,4-α-glucan branching enzyme represent copies of an ancient gene duplication also observed in cyanobacteria.
Once thought to be a relatively rare event, reports of gene transfers in eukaryotes are increasingly common. The abundance of available eukaryotic genome sequence is providing the material for analyses that were not possible only a few years ago. Analysis of the *Arabidopsis*genome \[[@B47]\] has revealed potentially thousands of genes that were transferred intracellularly. HGTs are still a relatively rare class of genes among multicellular eukaryotes, most probably because of the segregation of the germ line. By definition, unicellular eukaryotes do not have a separate germ line and are obligated to tolerate the acquisition of foreign genes if they are to survive. Among unicellular eukaryotes, there are now many reports of HGTs: *Giardia*\[[@B48],[@B49]\], *Trypanosoma*\[[@B38]\], *Entamoeba*\[[@B21],[@B49]\], *Euglena*\[[@B50]\], *Cryptosporidium*\[[@B25],[@B32],[@B40]\] and other apicomplexans \[[@B51]\].
As discussed earlier, genes transferred from distant phylogenetic sources such as eubacteria could be potential therapeutic targets. In apicomplexans, transferred genes are already some of the most promising targets of anti-parasitic drugs and vaccines \[[@B7],[@B25],[@B52]\]. We have shown that several transferred genes are differentially expressed in the *C. parvum*genome, and in two cases (IMPDH and TK), the transferred genes have been shown to be functional \[[@B25],[@B32]\]. The successful integration, expression and survival of transferred genes in the *Cryptosporidium*genome has changed the genetic and metabolic repertoire of the parasite.
Materials and methods
=====================
Cryptosporidium sequence sources
--------------------------------
Genomic sequences for *C. parvum*and *C. hominis*were downloaded from CryptoDB \[[@B53]\]. Genes were predicted for the completed *C. parvum*(IOWA) sequence as previously described using the Glimmer program \[[@B54]\] trained on *Cryptosporidium*coding sequences \[[@B52]\]. A few predicted genes that demonstrated apparent sequence incompleteness were reconstructed from genomic sequence by comparison with apicomplexan orthologs. The predicted protein encoding data set contained 5,519 sequences. A comparison of this gene set to the published annotation revealed that the Glimmer-predicted gene set contained all but 40 of the 3,396 annotated protein-encoding sequences deposited in GenBank. These 40 were added to our dataset and analyzed. Glimmer does not predict introns and some introns are present in the genome \[[@B7],[@B20]\]; thus our gene count is artificially inflated. Likewise, the official *C. parvum*annotation did not consider ORFs of less than 100 amino acids that did not have significant BLAST hits and thus may be a slight underestimate \[[@B7]\].
Database creation
-----------------
An internal database (ApiDB) containing all available apicomplexan sequence data was created \[[@B25]\]. A second BLAST-searchable database, PyPhynr, was constructed that included SwissProt, TrEMBL and TrEMBL\_new, as released in August 2003, predicted genes from *C. parvum*, ORFs of more than 120 amino acids from *Theileria annulata*, and more than 75 amino acids from consensus ESTs for several apicomplexan organisms. Genomic sequences for *T. gondii*(8x coverage) and clustered ESTs were downloaded from ToxoDB \[[@B55],[@B56]\]. Genomic data were provided by The Institute for Genomic Research (TIGR), and by the Sanger Institute. EST sequences were generated by Washington University. In addition, this study used sequence data from several general and species-specific databases. Specifically, the NCBI GenBank nr and dbEST were downloaded \[[@B57]\] and extensively searched. To provide taxonomic completeness, additional genes were obtained via searches of additional databases including: *Entamoeba histolytica*\[[@B58]\], *D. discoideum*\[[@B59]\], the kinetoplastids *Leishmania major*\[[@B59]\], *T. brucei*\[[@B59]\], *T. cruzi*\[[@B60]\], and a ciliate *Tetrahymena thermophila*\[[@B61]\]. Sequence data for *T. annulata*, *E. histolytica*, *D. discoideum*, *L. major*and *T. brucei*were produced by the Pathogen Sequencing Unit of the Sanger Institute and can be obtained from \[[@B62]\]. Preliminary sequence data for *T. thermophila*was obtained from TIGR and can be accessed at \[[@B63]\].
Phylogenomic analyses and similarity searches
---------------------------------------------
The source code of the phylogenomic software PyPhy \[[@B27]\] was kindly provided by Thomas Sicheritz-Ponten and modified to include analyses of eukaryotic groups, and changes to improve functionality \[[@B51]\]. For initial phylogenomic analyses, a BLAST cutoff of 60% sequence length coverage and 50% sequence similarity was adopted and the neighbor-joining program of PAUP 4.0b10 for Unix \[[@B64]\] was used. A detailed description of our phylogenomic pipeline and PyPhy implementation are described \[[@B51]\] and outlined in Figure [1](#F1){ref-type="fig"}.
Output gene trees with phylogenetic connections (that is, the nearest non-self neighbors at a distinct taxonomic rank) \[[@B27]\] to prokaryotes and algae-related groups were manually inspected. As the trees are unrooted, several factors were considered in the screen for candidate transferred genes. If the *C. parvum*gene does not form a monophyletic group with prokaryotic or plant-related taxa regardless of rooting, the subject gene was eliminated from further consideration. If the topology of the gene tree is consistent with a phylogenetic anomaly caused by gene transfer, but may also be interpreted differently if the tree is rooted otherwise, it was removed from consideration at this time. If the top hits of both nr and dbEST database searches are predominantly non-plant eukaryotes, and the topology of the tree was poor, the subject gene was considered an unlikely candidate. Finally, all 551 protein sequences predicted to be NEAPs in the malarial parasite *P. falciparum*\[[@B26]\] were used to search the *C. parvum*genome and the results were screened using a BLAST cutoff E-value of 10^-5^and a length coverage of 50%. Sequences identified by these searches were added to the candidate list (if not already present) for manual phylogenetic analyses to verify their likely origins. It should be noted that all trees were screened for the existence of a particular phylogenetic relationship. In some cases the proteins utilized to generate a particular tree are capable of resolving relationships among many branches of the tree of life, and in others they are not. Despite these differences in resolving power, the proteins which survive our phylogenetic screen and subsequent detailed analyses described below exhibit significant support for the branches of the tree in which we are interested. Similar procedures were used to characterize the complement of nuclear-encoded genes of plastid origin in the *Arabidopsis*genome \[[@B65]\]. BLAST searches were performed on GenBank releases 138-140 \[[@B57]\].
Detailed phylogenetic analyses of candidate genes identified by phylogenomic screening: candidate genes surviving the PyPhy phylogenomic screen were reanalyzed with careful attention to taxonomic completeness, including representative species from major prokaryotic and eukaryotic lineages when necessary and possible. New multiple sequence alignments were created with ClustalX \[[@B66]\], followed by manual refinement. Only unambiguously aligned sequence segments were used for subsequent analyses (see Additional data file 1). Phylogenetic analyses were performed with a maximum likelihood method using TREE-PUZZLE version 5.1 for Unix \[[@B67]\], a distance method using the program neighbor of PHYLIP version 3.6a package \[[@B68]\], and a maximum parsimony method with random stepwise addition using PAUP\* 4.0b10 \[[@B64]\]. Bootstrap support was estimated using 1,000 replicates for both parsimony and distance analyses and quartet puzzling values were obtained using 10,000 puzzling steps for maximum likelihood analyses. Distance calculation used the Jones-Taylor-Thornton (JTT) substitution matrix \[[@B69]\], and site-substitution variation was modeled with a gamma-distribution whose shape parameter was estimated from the data. For maximum likelihood analyses, a mixed model of eight gamma-distributed rates and one invariable rate was used to calculate the pairwise maximum likelihood distances. The unrooted trees presented in Figures [4](#F4){ref-type="fig"} and [5](#F5){ref-type="fig"} were drawn by supplying TREE-PUZZLE with the maximum parsimony tree and using TREE-PUZZLE distances as described above to calculate the branch lengths. The trees were visualized and prepared for publication with TreeView X Version 0.4.1 \[[@B70]\].
Genomic Southern analysis
-------------------------
*C. parvum*(IOWA) oocysts (10^8^) were obtained from the Sterling Parasitology Laboratory at the University of Arizona and were lysed using a freeze/thaw method. Genomic DNA was purified using the DNeasy Tissue Kit (Qiagen). Genomic DNA (5 μg) was restricted with *Bam*H1 and *Eco*R1 respectively and electrophoresed on a 0.8% gel in 1x TAE buffer, transferred to a positively charged nylon membrane (Bio-Rad), and fixed using a UVP crosslinker set at 125 mJ as described in \[[@B71]\]. *C. parvum*genomic DNA for the probes (700-1,500 bp) was amplified by PCR (see Additional data file 1).
Semi-quantitative reverse transcription-PCR
-------------------------------------------
Sterilized *C. parvum*(IOWA isolate) oocysts were used to infect confluent human adenocarcinoma cell monolayers at a concentration of one oocyst per cell as previously described \[[@B72]\]. Total RNA was prepared from mock-infected and *C. parvum*-infected HCT-8 cultures at 2, 6, 12, 24, 36, 48 and 72 h post-inoculation by directly lysing the cells with 4 ml TRIzol reagent (GIBCO-BRL/Life Technologies). Purified RNA was resuspended in RNAse-free water and the integrity of the samples was confirmed by gel electrophoresis.
Primers specific for several transferred genes identified in the study were designed (see Additional data file 1) and a semi-quantitative RT-PCR analysis was carried out as previously described \[[@B72]\]. Primers specific for *C. parvum*18S rRNA were used to normalize the amount of cDNA product of the candidate gene to that of *C. parvum*rRNA in the same sample. PCR products were separated on a 4% non-denaturing polyacrylamide gel and signals from specific products were captured and quantified using a phosphorimaging system (Molecular Dynamics). The expression level of each gene at each time point was calculated as the ratio of its RT-PCR product signal to that of the *C. parvum*18S rRNA. Six independent time-course experiments were used in the analysis.
Additional data files
=====================
Additional data is provided with the online version of this paper, consisting of a PDF file (Additional data file [1](#s1){ref-type="supplementary-material"}) containing: materials and methods for genomic Southern analysis; the amino-acid sequences of genes listed in Table [2](#T2){ref-type="table"}; accession numbers for sequences used in Figure [4](#F4){ref-type="fig"}; accession numbers for sequences used in Figure [5](#F5){ref-type="fig"}; expression of *C. parvum*phosphoglucomutase in *T. gondii*; table of primers used for RT-PCR experiments; phylogenetic tree of aldolase; alignment files for phylogenetic analyses in Figure [4](#F4){ref-type="fig"}; and the alignment of 1,4-α-glucan branching enzyme sequences used in Figure [5](#F5){ref-type="fig"}.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
This file contains the materials and methods for genomic Southern analysis; the amino-acid sequences of genes listed in Table 2; accession numbers for sequences used in Figure 4; accession numbers for sequences used in Figure 5; expression of *C. parvum*phosphoglucomutase in *T. gondii*; table of primers used for RT-PCR experiments; phylogenetic tree of aldolase; alignment files for phylogenetic analyses in Figure 4; and the alignment of 1,4-α-glucan branching enzyme sequences used in Figure 5
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank G. Buck (Virginia Commonwealth University), G. Widmer and S. Tzipori (Tufts University) for access to *C. hominis*genotype I genomic sequence data. Genomic sequence for *Toxoplasma gondii*(8X coverage) and clustered ESTs were downloaded from ToxoDB \[[@B56]\]. Genomic data were provided by The Institute for Genomic Research (supported by the NIH grant \#AI05093), and by the Sanger Institute (Wellcome Trust). Apicomplexan EST sequences were generated by Washington University (NIH grant \#1R01AI045806-01A1). Occasionally, genes were obtained via searches of several databases containing *Entamoeba histolytica*\[[@B58]\], *Dictyostelium discoideum*\[[@B59]\], kinetoplastids *Leishmania major*\[[@B59]\], *Trypanosoma brucei*\[[@B59]\], *T. cruzi*\[[@B60]\], and ciliate *Tetrahymena thermophila*\[[@B61]\]. Sequence data for *T. annulata*, *E. histolytica*, *D. discoideum*, *Leishmania major*and *T. brucei*were produced by the Pathogen Sequencing Unit of the Sanger Institute and can be obtained from \[[@B62]\]. Preliminary sequence data for *Tetrahymena thermophila*was obtained from The Institute for Genomic Research and can be accessed at \[[@B63]\]. We thank Fallon Hampton for her work on the phosphoglucomutase expression constructs. She was supported by a summer undergraduate fellowship in genetics (SUNFIG) award. This study was funded by a research grant from the University of Georgia Research Foundation to J.C.K. and NIH grant U01 AI 46397 to M.S.A. J.H. is supported by a postdoctoral fellowship from the American Heart Association. We thank Boris Striepen, Marc-Jan Gubbels and three anonymous reviewers for comments that greatly increased the clarity and precision of the analyses in the manuscript.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Phylogenomic analysis pipeline. The procedures used to analyze, assess and manipulate the protein-sequence data at each stage of the analysis are diagrammed.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
*Cryptosporidium parvum*genomic Southern blot. *C. parvum*genomic DNA, 5 μg per lane. Lanes were probed for the following genes: (1) aminopeptidase N; (2) glucose-6-phosphate isomerase; (3) leucine aminopeptidase; (4) pteridine transporter (BT-1); and (5) glutamine synthetase. Lanes (1-4) were restricted with *Bam*H1 and lane (5) with *Eco*R1. The ladder is shown in 1 kb increments. See Additional data file 1 for probes and methods.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Region of leucine aminopeptidase multiple sequence alignment that illustrates several characters uniting apicomplexan sequences with plant and cyanobacterial sequences. The red box denotes an indel shared between apicomplexans, plants and cyanobacteria. The number preceeding each sequence is the position in the individual sequence at which this stretch of similarity begins. GenBank GI numbers for each sequence are as indicated in Additional data file 1. Colored boxes preceeding the alignment indicate the taxonomic group for the organisms named to the left. Red, apicomplexan; green, plant and cyanobacterial; blue, eubacterial; lavender, other protists and eukaryotes.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Phylogenetic analyses. **(a)**Leucine aminopeptidase; **(b)**glucose-6-phosphate isomerase; **(c)**tryptophan synthetase β subunit. Numbers above the branches (where space permits) show the puzzle frequency (with TREE-PUZZLE) and bootstrap support for both maximum parsimony and neighbor-joining analyses respectively. Asterisks indicate that support for this branch is below 50%. The scale is as indicated. GI accession numbers and alignments are provided in Additional data file 1.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Phylogenetic analyses of 1,4-α-glucan branching enzyme. Numbers above the branches (where space permits) show the puzzle frequency (TREE-PUZZLE) and bootstrap support for both maximum parsimony and neighbor-joining analyses respectively; Asterisks indicate that support for this branch is below 50%. The scale is as indicated. GI accession numbers and alignment are provided in Additional data file 1.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
Expression profiles of select genes in *C. parvum*-infected HCT-8 cells. The expression level of each gene is calculated as the ratio of its RT-PCR product to that of *C. parvum*18s rRNA. **(a)**glucose-6-phospate isomerase; **(b)**leucine aminopeptidase; **(c)**pteridine transporter (BT-1); **(d)**tyrosyl-tRNA synthetase; **(e)**calcium-dependent protein kinase; **(f)**dihydrofolate reductase-thymidine synthetase (DHFR-TS). The genes examined in (a-c, e) represent transferred genes of different origins; (d, f) represent non-transferred references. Error bars show the standard deviation of the mean of six independent time-course experiments.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Distribution of best non-apicomplexan BLAST hits in searches of the GenBank non-redundant protein database
:::
Category E \< 10^-3^ E \< 10^-7^
------------------------------- ------------- -------------
Plants 670 588
Algae 30 21
Non-cyanobacterial eubacteria 188 117
Cyanobacteria 22 16
Archaea 26 11
Total 936 783
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Genes of algal or eubacterial origin in *C. parvum*
:::
Putative gene name Accession Location Expression Indel Putative origin Putative function
------------------------------------------------------------------ ----------- ---------- ------------- ------- --------------------- ----------------------------------
Lactate dehydrogenase\* AAG17668 VII EST \+ α-proteobacteria Oxidoreductase
Malate dehydrogenase\* AAP87358 VII \+ α-proteobacteria Oxidoreductase
Thymidine kinase AAS47699 V Assay \+ α/γ-proteobacteria Kinase; nucleotide metabolism
Hypothetical protein A^†^ EAK88787 II γ-proteobacteria Unknown
Inosine 5\' monophosphate dehydrogenase AAL83208 VI Assay \+ ε-proteobacteria Purine nucleotide biosynthesis
Tryptophan synthetase β chain EAK87294 V Proteobacteria Amino acid biosynthesis
1,4-α-glucan branching enzyme CAD98370 VI Eubacteria Carbohydrate metabolism
1,4-α-glucan branching enzyme CAD98416 VI Eubacteria Carbohydrate metabolism
Acetyltransferase EAK87438 VIII Eubacteria Unknown
α-amylase EAK88222 V Eubacteria Carbohydrate metabolism
DNA-3-methyladenine glycosylase EAK89739 VIII Eubacteria DNA repair
RNA methyltransferase AY599068 II Eubacteria RNA processing and modification
Peroxiredoxin AY599067 IV Eubacteria Oxidoreductase; antioxidant
Glycerophosphodiester phosphodiesterase AY599066 IV Eubacteria Phosphoric ester hydrolase
ATPase of the AAA class EAK88388 I Eubacteria Post-translational modification
Alcohol dehydrogenase EAK89684 VIII Eubacteria Energy production and conversion
Aminopeptidase N AAK53986 VIII Eubacteria Peptide hydrolase
Glutamine synthetase CAD98273 VI \+ Eubacteria Amino acid biosynthesis
Conserved hypothetical protein B CAD98502 VI Eubacteria Unknown
Aspartate-ammonia ligase^†^ EAK87293 V EST Eubacteria Amino acid biosynthesis
Asparaginyl tRNA synthetase^†^ EAK87485 VIII Eubacteria Translation
Glutamine cyclotransferase^†^ EAK88499 I Eubacteria Amido transferase
Leucine aminopeptidase EAK88215 V RT-PCR \+ Cyanobacteria Hydrolase
Biopteridine transporter (BT-1) CAD98492 VI RT-PCR /EST \+ Cyanobacteria Biopterine transport
Hypothetical protein C^†^(possible Zn-dependent metalloprotease) EAK89015 III Archaea Putative protease
Superoxide dismutase^†^ AY599065 V Eubacteria /archaea Oxidoreductase; antioxidant
Glucose-6-phosphate isomerase EAK88696 II RT-PCR \+ Algae/plants Carbohydrate metabolism
Uridine kinase/uracil phosphoribosyltransferase^†^ AAS47700 VIII Algae/plants Nucleotide salvage metabolism
Calcium-dependent protein kinases\* ^†^ AAS47705 II RT-PCR Algae/plants Kinase; cell signal transduction
AAS47706 II
AAS47707 VII
\*Genes that have been derived from a duplication following transfer; ^†^transferred genes that have less support. GenBank accession numbers are as indicated. Locations are given as chromosome number. The expression status for each gene is indicated by method: EST, RT-PCR or assay. Only 567 EST sequences exist for *C. parvum*. A + in the indel colum indicates the presence of a shared insertion/deletion between the *C. parvum*sequence and other sequences from organisms identified in the putative origin column.
:::
|
PubMed Central
|
2024-06-05T03:55:51.815322
|
2004-10-19
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545779/",
"journal": "Genome Biol. 2004 Oct 19; 5(11):R88",
"authors": [
{
"first": "Jinling",
"last": "Huang"
},
{
"first": "Nandita",
"last": "Mullapudi"
},
{
"first": "Cheryl A",
"last": "Lancto"
},
{
"first": "Marla",
"last": "Scott"
},
{
"first": "Mitchell S",
"last": "Abrahamsen"
},
{
"first": "Jessica C",
"last": "Kissinger"
}
]
}
|
PMC545780
|
Background
==========
In the post-sequencing phase of genome characterization, it is important to understand the contribution of non-coding sequences to higher-order genome structure and stability. Maintenance of genome integrity and the faithful transmission of genetic information in mitosis and meiosis are essential to organism survival and are critically dependent on two repetitive chromosomal elements. Telomeres protect against chromosomal truncation or fusion events \[[@B1]\], while centromeres ensure faithful chromosome segregation through cell division \[[@B2]-[@B4]\]. Failure in the function of these elements can lead to genomic instability, with often catastrophic consequences in humans such as miscarriage, congenital birth defects or cancer. In contrast to the telomere, whose properties have been well explored at the genomic and molecular levels \[[@B5]\], the human centromere remains relatively poorly characterized, and experimental systems for the genomic study of centromere formation and behavior are only just being developed and optimized \[[@B6]-[@B14]\].
Defining the minimal DNA sequences required for centromere function on a normal human chromosome has proved challenging, owing to the complex nature of inter- and intra-chromosomal homology and variability in genomic DNA content near the primary constriction. Common to all normal human centromeres are large amounts of alpha-satellite DNA, which is comprised of a family of diverged \'monomers\' of around 171 base-pairs (bp) that have been amplified in multimeric groups (higher-order repeats) on different chromosomes to form chromosome-specific arrays typically megabases in length \[[@B15]-[@B17]\]. In addition, the core of higher-order repeat alpha-satellite is, where examined in detail, surrounded by other alpha-satellite sequences that fail to form a recognizable higher-order structure (so-called \'monomeric\' alpha satellite) \[[@B10],[@B18]-[@B20]\]. Together, the two types of centromeric repeat span up to several megabases of genomic DNA at each centromere region and account for much of the largest remaining gaps in the human genome sequence assembly \[[@B21],[@B22]\]. Support for a critical role for alpha-satellite DNA in centromere function comes from recent studies on the human X chromosome, where the most abundant alpha-satellite sequence at this centromere, DXZ1, has been shown to be sufficient for centromere function \[[@B10],[@B23]\] and, more generally, from studies demonstrating the formation of *de novo*centromeres on human artificial chromosomes following transfection of some types of alpha-satellite sequences into human cells \[[@B6]-[@B14]\].
Paradoxically, despite conservation of the functional role of the centromere in every eukaryotic cell, DNA sequences at eukaryotic centromeres are quite divergent in sequence even between closely related species \[[@B24],[@B25]\]. Although primary genomic sequence has not been conserved at eukaryotic centromeres, they do, nonetheless, share features in common such as a structure based on tandem repeats, overall AT-rich composition, and packaging into specialized centromeric chromatin marked by the presence of centromere-specific histone H3 (CenH3) variants (reviewed in \[[@B4],[@B26],[@B27]\]). The ability of different genomic sequences to fulfill centromeric requirements in different species is in accord with data showing that the DNA normally associated with the genetically mapped centromere on normal human chromosomes is not always sufficient or necessary for centromere function. Rare chromosomal rearrangements can result in either dicentric chromosome formation, where one centromere is typically inactivated \[[@B28],[@B29]\], or in the formation of neocentromeres, where a centromere assembles on DNA that is not associated with the normal centromere genomic locus (reviewed in \[[@B3]\]). Together, these observations suggest that epigenetic factors are critical for centromere function \[[@B30]\] and point to the as-yet incompletely understood interplay of underlying genomic DNA sequences located in the centromeric region and their ability to package into specialized centromeric chromatin \[[@B2],[@B4],[@B27]\].
Recent evidence suggests that a complex system of epigenetic modifications based on histone variants and histone tail modifications is important for centromere activity (reviewed in \[[@B4],[@B31]\]), in much the same way as a histone code is involved in determining the transcriptional competence of DNA \[[@B32]\]. Although the epigenetic basis of centromere function is not yet fully defined, a strong candidate for specifying the site of the functional centromere (kinetochore-forming region) is the family of CenH3 variants, which are conserved from yeast to humans and are essential to viability of the organism (reviewed in \[[@B2]\]). In humans and flies, CenH3 is restricted to the centromere where CenH3- and typical H3-containing nucleosomes exist in an alternating arrangement, generating a unique chromatin structure that may be important for centromere function \[[@B33],[@B34]\].
The most completely studied complex eukaryotic centromere at the molecular level is that of the fission yeast *Schizosaccharomyces pombe*. Detailed analyses of a 40-kilobase (kb) *S. pombe*centromere revealed that it encompasses both the kinetochore, as defined by the exclusive association of Cnp1, the fission yeast CenH3, with the central core element \[[@B35]\] and adjacent repeats enriched for heterochromatin-associated factors \[[@B36]\] that are important for centromeric cohesion \[[@B37]-[@B40]\]. Within the heterochromatic domains, histone H3 is methylated at lysine 9 (H3MeK9), resulting in the recruitment of the heterochromatin protein HP1-homolog Swi6 \[[@B41]\].
There is substantial evidence that HP1 is involved in setting up and/or maintaining a repressed chromatin state in several epigenetic systems (reviewed in \[[@B42]\]). HP1 proteins are conserved and localize to centromere regions in human and mouse cells \[[@B43]-[@B45]\]. Human cells express three HP1 isoforms, HP1α, HP1β and HP1γ. HP1α and HP1β localize primarily to pericentromeric regions, while HP1γ is dispersed at sites along chromosome arms \[[@B43]\]. Furthermore, modified H3MeK9 nucleosomes, which create a binding site for HP1 (reviewed in \[[@B46]\]), have also been localized cytologically to centromere regions in flies and mice \[[@B44],[@B47]-[@B53]\]. These observations suggest a model in which local modifications of chromatin composition represent a crucial and highly conserved element necessary for the specification and/or maintenance of complex eukaryotic centromeres \[[@B2]\]. Consistent with these models, chromatin immunoprecipitation assays with highly specific antibodies have shown that both mouse minor and major satellite DNA sequences exhibit trimethylation of histone H3 at lysine 9 \[[@B51],[@B53]\]. However, while the association of histone modifications typical of repressive heterochromatin has been clearly demonstrated for sequences that flank the functional centromere, it is less certain what modifications, if any, may characterize the CenH3-containing chromatin of the functional centromere itself. Indeed, many of the characteristics historically assigned to pericentromeric DNA (that is, repressive heterochromatin and late-replication in S phase \[[@B54],[@B55]\]) may be features of the surrounding heterochromatin, more so than of the functional centromere *per se*.
One way to address the interacting and complementary role(s) of DNA sequence and *trans*-acting chromatin factors in human centromere function is through the construction of detailed genomic maps of human centromeric regions and evaluation of their associated proteins \[[@B10],[@B19],[@B56],[@B57]\]. An alternative empirical approach is to construct minimal human artificial chromosomes from defined alpha-satellite DNA sequences \[[@B6]-[@B14]\] as tools for evaluating the essential genomic requirements of centromere specification. Indeed, previous studies have shown that the human CenH3 - centromere protein A (CENP-A) - is deposited at the centromere on artificial chromosomes constructed from alpha-satellite DNA \[[@B12],[@B13],[@B58]\]. However, it is not known whether heterochromatin formation is required for centromere establishment and propagation and/or whether *de novo*centromeres on human artificial chromosomes without large amounts of adjacent heterochromatin demonstrate the same chromatin characteristics as either normal human centromeres or human artificial chromosomes with large amounts of heterochromatin.
In the present study, we have characterized the nature of heterochromatin and euchromatin formed on a series of human artificial chromosomes derived from higher-order repeat alpha satellite from chromosomes X or 17 \[[@B12],[@B14]\]. While large artificial chromosomes contain substantial amounts of heterochromatin (characterized by the presence of modified H3MeK9 nucleosomes and HP1α) and replicate later in S phase, small artificial chromosomes show features more consistent with the euchromatin of the chromosome arms, including the presence of histone variants typical of expressed euchromatin and replication earlier in S phase. These data suggest that the chromatin environment required for *de novo*centromere formation and function is likely to be generally conducive to gene expression, as will probably be required for either gene-transfer experiments and/or functional genomic applications of the artificial chromosome technology. Further, the data raise the possibility that functional centromeres may adopt a novel chromatin state that is, contrary to what has been long assumed, quite distinctive from that of conventional heterochromatin.
Results
=======
To examine the chromatin composition of human artificial chromosomes, we used a panel of artificial chromosomes formed after transfection with vectors containing either synthetic chromosome 17 (D17Z1) or cloned X chromosome (DXZ1) alpha-satellite sequences \[[@B12],[@B14]\]. Each of the artificial chromosomes tested contains a functional *de novo*centromere assembled from the transfected DNA, as well as at least one copy of a functioning gene used as a selectable marker. Together, this panel of artificial chromosomes provides an opportunity to examine the nature of heterochromatin and euchromatin assembled on the transfected DNA sequences. The high mitotic stability and *de novo*composition of artificial chromosomes generated from D17Z1 (17-E29, 17-D34 and 17-B12) or DXZ1 (X-4 and X-5) have been described \[[@B12],[@B14]\]. As a more direct measure of artificial chromosome segregation errors, we have used an assay that allows cells to undergo anaphase but cannot complete cytokinesis \[[@B14]\]. Using fluorescence *in situ*hybridization (FISH), artificial and host chromosome segregation products can be measured and nondisjunction or anaphase lag defects recorded.
In X-4 and X-5, artificial chromosomes mis-segregated in 1.8% and 2.4% of cells, respectively (\[[@B14]\] and Table [1](#T1){ref-type="table"}). Similar analyses of artificial chromosome segregation errors in 17-B12 revealed that they mis-segregated in 2.4% of the cells (Table [1](#T1){ref-type="table"}). This segregation error rate is comparable to that found for the majority of other human artificial chromosomes previously characterized \[[@B14]\]. Artificial chromosomes in 17-E29 and 17-D34 have segregation efficiencies corresponding to more than 99.9% per cell division, using metaphase analyses \[[@B12]\]. For comparison, we also examined an additional cell line, 17-C20, which contains highly mitotically unstable D17Z1-based artificial chromosomes. In 17-C20, artificial chromosome copy number was high (average 4.7 per cell) and artificial chromosomes were lost from the cell population by 30-40 days of culture without selection, despite containing both inner (CENP-A) and outer (CENP-E) kinetochore proteins (data not shown). In the anaphase assay, 12.2% of artificial chromosomes in 17-C20 were mis-segregating (at 12 days without selection) and the predominant defect was anaphase lag (Table [1](#T1){ref-type="table"}). Sizes of D17Z1-containing artificial chromosomes were based on comparison of the signal intensity on the approximately 3 Mb D17Z1 array on chromosome 17 to intensities on the artificial chromosomes using FISH analyses with a D17Z1 probe (Table [2](#T2){ref-type="table"}; see also Figures 2 and 3 in \[[@B12]\]). Artificial chromosomes that had signal intensities several-fold less than the endogenous D17Z1 signals were estimated to be 1-3 Mb in size, whereas artificial chromosomes that produced signals similar to or several-fold more intense than those of the endogenous D17Z1 arrays were estimated to be in the 3-10 Mb size range. Similar comparisons of the signal intensities on the DXZ1-based artificial chromosomes with those of the host DXZ1 signals were used to estimate the sizes of the DXZ1-based human artificial chromosomes (Table [2](#T2){ref-type="table"} and data not shown). Properties of artificial chromosomes used in the present study are summarized in Tables [1](#T1){ref-type="table"} and [2](#T2){ref-type="table"}.
Variation in levels of heterochromatin-associated factors correlates with artificial chromosome size
----------------------------------------------------------------------------------------------------
To test whether human artificial chromosomes were capable of forming heterochromatin, we first examined several established markers of heterochromatin on the artificial chromosome panel. Indirect immunofluorescence with an antibody recognizing histone H3 modified by trimethylation at lysine 9 and lysine 27 (H3TrimK9/K27) was applied to metaphase spreads. Methylation of lysines at these sites has been associated with formation of repressive chromatin, including pericentric heterochromatin in mouse cells \[[@B32],[@B51]-[@B53],[@B59],[@B60]\]. As shown in Figure [1a](#F1){ref-type="fig"} and [1b](#F1){ref-type="fig"}, small D17Z1-based artificial chromosomes, estimated to be in the 1-3 Mb size range (Table [2](#T2){ref-type="table"}), do not stain detectably with the H3TrimK9/K27 antibody, in contrast to the centromeric regions of the natural human chromosomes that stain, in some cases intensely, with this antibody. On the other hand, larger artificial chromosomes, estimated to be in the 3-20 Mb size range (Table [2](#T2){ref-type="table"}), stained strongly for H3TrimK9/K27 modifications (Figure [1c-g](#F1){ref-type="fig"}), often at levels greater than those of many endogenous centromeric regions (Figure [1g](#F1){ref-type="fig"}). It is clear that at least large amounts of transfected alpha satellite are capable of assembling into heterochromatin in the context of human artificial chromosomes. Whether small artificial chromosomes are truly negative for this marker of heterochromatin, or whether they assemble only small amounts of heterochromatin below the level of detection, cannot be assessed with this assay. Nonetheless, they clearly have assembled far less of this epigenetically modified heterochromatin than exists at the relevant endogenous 17 centromeric regions (Figure [1](#F1){ref-type="fig"}).
In a parallel approach, we examined the distribution of HP1α in four lines containing D17Z1-based artificial chromosomes. Each line was stably transfected with a Myc-epitope tagged form of HP1α (see Materials and methods) to permit detection of HP1α using an anti-Myc antibody. The smaller artificial chromosomes stained very weakly (at a level similar to that of the staining on the euchromatic chromosome arms), well below the levels of HP1α detected at the centromeric region of the endogenous chromosome 17s (Figure [2a,b](#F2){ref-type="fig"}). As seen with the H3TrimK9/K27 antibody, the larger artificial chromosomes stained strongly for HP1α (Figure [2c,d](#F2){ref-type="fig"}), at levels comparable to the endogenous chromosome 17s. The intensity of HP1α-Myc staining was variable at endogenous human centromere regions (Figure [2d](#F2){ref-type="fig"}); similar results were obtained using a primary anti-HP1α antibody (data not shown). This contrasts with the amount of CENP-A, which appears to be present at consistent levels at all normal human centromeres \[[@B61]\] and artificial chromosomes tested (Figure [2d](#F2){ref-type="fig"}) \[[@B12],[@B13],[@B58]\]. Notably, the CENP-A signal is localized to a discrete subdomain within the larger artificial chromosomes, whereas HP1α covers a much larger area of the artificial chromosome (Figure [2d](#F2){ref-type="fig"}). This suggests that HP1α may be a marker for generalized pericentromeric heterochromatin that flanks the kinetochore-associated alpha satellite of the functional centromere, rather than a marker of the functional centromere *per se*. Such a model \[[@B2],[@B3]\] is also consistent with the observation that small artificial chromosomes, which contain little if any of the flanking heterochromatin, do not contain elevated levels of HP1α (Figure [2a,b](#F2){ref-type="fig"}; Table [2](#T1){ref-type="table"}).
Euchromatin forms on artificial chromosomes
-------------------------------------------
For their potential use as gene-transfer vectors or as general vehicles suitable for interrogation of genome function, human artificial chromosomes must also be capable of forming euchromatin to support gene expression. Indeed, one would hypothesize that at least small amounts of transcriptionally active chromatin must form during artificial chromosome formation to permit expression of the selectable marker gene(s) contained on the transfected constructs \[[@B10],[@B12],[@B14]\]. It has previously been shown using immunocytochemical methods \[[@B62],[@B63]\] that methylation of histone H3 at lysine 4, an epigenetic modification associated with transcriptionally permissive chromatin \[[@B64]-[@B66]\], is generally enriched on autosomes and depleted at the repressed inactive X chromosome and human centromere regions.
As a test for formation of permissive chromatin on artificial chromosomes, we stained metaphase spreads with an antibody that recognizes histone H3 dimethylated at lysine 4 (H3DimK4). All artificial chromosomes tested stained positively for H3DimK4 modifications (Figure [3a-f](#F2){ref-type="fig"}; Table [2](#T2){ref-type="table"}). In contrast, the endogenous centromeric regions were depleted for H3DimK4 staining, although, as noted above for markers of heterochromatin formation, this depletion may reflect the state of the surrounding heterochromatin, rather than that of the functional centromere *per se*.
Previous structural analyses of artificial chromosomes indicate that they consist of input DNA multimers arranged as blocks of alpha-satellite DNA interspersed with vector sequences \[[@B7],[@B11],[@B12]\]. This structural organization is consistent with the presence of multiple selectable marker genes and differs from the large uninterrupted blocks of alpha-satellite DNA found at all human centromeres that are typically under-represented for this active chromatin mark (Figure [3](#F3){ref-type="fig"}). Because mitotically stable artificial chromosomes can have permissive as well as repressive chromatin present, these data suggest that this chromatin configuration does not significantly disturb mitotic centromere function.
Two modes of artificial chromosome replication timing
-----------------------------------------------------
While the genomic determinants of potential origins of DNA replication in the human genome, as well as of their timing of replication during S phase, are still not well understood, the generally accepted paradigm is that expressed sequences replicate in the first half of S phase, while non-expressed sequences replicate in the second half \[[@B67]\]. Consistent with this pattern, alpha-satellite DNA, as well as constitutive heterochromatin (such as that found on the Yq arm), replicate in the mid to late S phase period \[[@B54],[@B55],[@B68],[@B69]\]. In the present study, we have asked whether D17Z1-based artificial chromosomes replicate at a similar time to endogenous chromosome 17 alpha-satellite DNA. To determine the time of replication, unsynchronized cells were pulsed with bromodeoxyuridine (BrdU) for 2 hours, followed by a thymidine chase for varying lengths of time before harvesting cells in metaphase (see Materials and methods). Detection of BrdU incorporation at sites of DNA replication was performed using indirect immunofluorescence with an anti-BrdU antibody on metaphase spreads.
While there was overlap between artificial chromosome replication timing patterns and those of the host 17 centromere regions during mid S phase (Table [3](#T3){ref-type="table"}), we found two modes of artificial chromosome replication timing. The heterochromatin-enriched artificial chromosomes (17-B12 and 17-C20; see Table [2](#T2){ref-type="table"}) commenced replication in mid S phase (2-4 hours into S phase) and completed replication by 6 hours into S phase (Figures [4](#F4){ref-type="fig"} and [5c](#F5){ref-type="fig"}; Table [3](#T3){ref-type="table"}). In contrast, the heterochromatin-depleted artificial chromosomes (17-D34 and 17-E29; see Table [2](#T2){ref-type="table"}) started replicating within the first 2 hours of S phase (early S phase) and their replication was completed by 4 hours into S phase (Figure [5a,b](#F5){ref-type="fig"}; Table [3](#T3){ref-type="table"}). That these differences are characteristic of each particular artificial chromosome is suggested by the observation that, in all lines, when multiple artificial chromosomes were present in a given cell, they are frequently replicated synchronously (Figures [4c](#F4){ref-type="fig"} and [5a,c](#F5){ref-type="fig"}). From these data, it is tempting to propose that the presence of large amounts of heterochomatin in the larger artificial chromosomes may have influenced replication timing on these artificial chromosomes and promoted a shift towards later in S phase.
Discussion
==========
Human artificial chromosomes provide a novel system for analyzing *cis*- and *trans*-acting factors necessary for chromosome segregation and offer potential for both functional genomics and gene-transfer applications. The artificial chromosomes we used contain defined alpha-satellite DNA sequences \[[@B12],[@B14]\]. Studying how epigenetic components assemble with alpha satellite to form a *de novo*centromere on artificial chromosomes may reveal the critically important components and may help distinguish between those features that are characteristic of the functional centromere itself and those that are markers of the surrounding heterochromatin. Such a distinction is extremely difficult in normal human chromosomes but should be enhanced by the ability to generate a variety of different artificial chromosomes made with different input sequences.
Recent detailed molecular studies in the fission yeast have revealed that such epigenetic factors are critical for centromere function. The fission yeast CenH3, Cnp1, is deposited only at the central core domain, while heterochromatin (marked by methylation of histone H3 at lysine 9 and by binding of the HP1 homolog, Swi6) forms on the surrounding inverted repeats \[[@B35],[@B36],[@B41]\]. The yeast data, together with the observations that CenH3s are conserved and that H3K9-modified nucleosomes and HP1 proteins are often found close to the centromere in higher eukaryotes, have contributed to the development of models for centromere packaging in the larger chromosomes of multicellular eukaryotes, including mammals. In these models, a specific centromeric chromatin configuration, in which CenH3-containing chromatin is surrounded by pericentric heterochromatin, is conserved and may be an important determinant of centromere function \[[@B2]-[@B4]\].
While the data presented here are largely consistent with these models, they permit two important refinements. First, large amounts of heterochromatin (containing alpha satellite and marked by H3TrimK9/K27 staining, HP1α binding and late replication) are not required for effective chromosome segregation during mitosis; indeed, the small artificial chromosomes examined here do not contain detectable amounts of H3TrimK9/K27 (Table [2](#T2){ref-type="table"}). Second, the cytological characteristics of heterochromatin (repressive chromatin and later replication in S phase), classically attributed to the centromere \[[@B54],[@B55]\], may instead reflect features of the surrounding heterochromatin and do not appear to define critical properties of the functional centromere. Our own data would argue that the functional centromere - at least as assembled on the smaller D17Z1-based human artificial chromosomes - is instead characterized by a distinctive chromatin containing CenH3 (CENP-A) that can form within regions epigenetically modified with markers of euchromatin (Tables [1](#T1){ref-type="table"} and [2](#T2){ref-type="table"}). This conclusion is consistent with parallel work on the organization of centromeric chromatin of normal *Drosophila*and human chromosomes \[[@B34]\]. The finding that CENP-A-containing chromatin can be deposited within euchromatin-rich artificial chromosomes that are highly mitotically stable (more than 99.9 % segregation efficiency per cell division) yet depleted for heterochromatin modifications, suggests that only a very small amount of heterochromatin may be required on an artificial chromosome (from observations in yeast \[[@B37]-[@B40]\] and chicken DT40 cells \[[@B70]\] this is presumably for assembling the cohesin complex), and that this could also be true for human centromeres.
This study also addresses the question of timing of replication of D17Z1-based artificial chromosomes. The smaller artificial chromosomes that completely overlap with CENP-A \[[@B12]\] and euchromatic modifications (Figure [3](#F3){ref-type="fig"}) replicate early in S phase whereas the larger artificial chromosomes that have assembled heterochromatin (H3TrimK9/K27 and HP1α) in addition to euchromatin replicate later in S phase (Table [3](#T3){ref-type="table"}). The later onset of replication on the larger artificial chromosomes is similar to that of host chromosome 17 centromere regions that are also enriched for H3TrimK9/K27 and HP1α (Figures [1](#F1){ref-type="fig"} and [2](#F2){ref-type="fig"}, Tables [2](#T2){ref-type="table"} and [3](#T3){ref-type="table"}). With the caveats that higher-resolution methods will be required to determine the precise replication timing of the CENP-A domain on the artificial chromosomes, and that differences in vector DNA content may be influencing origin establishment and/or usage, our observations are consistent with local chromatin modification being an important factor influencing artificial chromosome replication.
Chromatin composition as a factor in determining replication timing has also been implicated in a study of a *Drosophila*minichromosome deletion series. In this study, replication timing was shifted to an earlier point in mid-S phase following deletion of large amounts of pericentromeric heterochromatin from the minichromosomes \[[@B71]\]. Support for a direct role of chromatin composition in replication timing comes from studies in budding yeast, where regions associated with acetylated histones (an epigenetic mark of active chromatin) replicate earlier than those depleted for this histone modification \[[@B72]\]. However, unexpected recent evidence from fission yeast has shown that centromeric heterochromatin replicates early in S phase, suggesting that chromatin composition is not a uniform determinant of replication timing in lower eukaryotes \[[@B73]\]. As the euchromatin-rich and highly mitotically stable artificial chromosomes replicate in the first half of S phase (in 17-E29, the majority of artificial chromosomes (75%, *n*= 20) replicated in the first 2 hours of S phase (Table [3](#T3){ref-type="table"})) these findings challenge the current dogma that replication later in S phase is an obligatory function of the centromere. The present findings are also supportive of earlier studies suggesting that replication timing of CenH3-containing chromatin is not a determinant of the functional centromere \[[@B69],[@B71]\].
Cytological data indicate that the amount of CENP-A modified chromatin (in addition to several other kinetochore-associated CENPs) is similar on endogenous human chromosomes and on all artificial chromosomes regardless of the amount of total alpha satellite present. This suggests that the amount of CENP-A chromatin and/or the size of the kinetochore is regulated and/or limited in some manner \[[@B6]-[@B14],[@B58],[@B61]\]. In contrast, the results of the present study indicate that the heterochromatic fraction of centromeric DNA (on both endogenous chromosomes and artificial chromosomes) is highly variable. In line with current models, we did detect elevated levels of H3TrimK9/K27 modifications and HP1α, diagnostic of heterochromatin on large artificial chromosomes generated from chromosome 17 (D17Z1) or X (DXZ1) alpha-satellite DNA. However, no immunocytochemically detectable heterochromatin (H3TrimK9/K27) was associated with the smaller artificial chromosomes.
To evaluate their potential for characterization of genome sequences and, eventually, for gene transfer or gene therapy applications, we sought to determine the extent of transcriptionally competent chromatin formation in artificial chromosomes. Epigenetic modification of histone H3 by dimethylation at lysine 4 (H3DimK4), a marker of transcriptionally competent chromatin, was present on all artificial chromosomes tested. This contrasts with the staining pattern associated with the centromere regions on human metaphase spreads, where this modification is largely undetectable, probably reflecting the general absence of genes mapping to centromere regions (Figure [3](#F3){ref-type="fig"}). As selectable marker genes are expressed on artificial chromosomes, it may be presumed that at least a portion of the artificial chromosome chromatin structure is transcriptionally permissive, consistent with the positive staining for H3DimK4. In line with these observations, large human transgenes have been expressed from artificial chromosomes \[[@B74]-[@B76]\] and selectable marker genes on artificial chromosomes assemble acetylated histones, another marker of euchromatin \[[@B77]\]. Furthermore, detection of transcription of genes within the CenH3 domain of a human neocentromere \[[@B78]\] and a rice centromere \[[@B79]\] suggests that CenH3-containing chromatin can be transcriptionally competent. The relationship between active and repressive chromatin and underlying genomic sequences on the larger artificial chromosomes is not known and will require more detailed follow-up analyses. As other detailed chromatin immunoprecipitation studies have shown that methylation of histone H3 at lysine 4 or lysine 9 seem to be mutually exclusive \[[@B64],[@B65]\], it will be interesting to find out how the two types of chromatin are assembled during artificial chromosome formation and to find out if there is a mechanism that prevents spreading of chromatin between the heterochromatic and euchromatic sub-domains. An advantage of the artificial chromosome system is the capacity to manipulate sequence content and to test directly the involvement of candidate sequences in gene expression, chromatin establishment or timing of DNA replication.
In this study we included one line, 17-C20, that contains *de novo*D17Z1-based artificial chromosomes that retain both inner and outer kinetochore components yet are highly mitotically unstable as a result of their rapid loss in the absence of selection and the very high segregation error rate (12.2%) detected in the anaphase assay (Table [1](#T1){ref-type="table"}). The artificial chromosomes in this line have a global chromatin composition indistinguishable to that of similar-sized D17Z1-based mitotically stable artificial chromosomes, as both H3DimK4- and H3TrimK9/K27-modified nucleosomes and HP1α are assembled (Table [2](#T2){ref-type="table"}). Our study has not revealed the cause of the segregation defect of artificial chromosomes in 17-C20, and so a more extensive examination of additional epigenetic markers or centromere-associated factors may be informative. Detailed anaphase segregation analyses of D17Z1- and DXZ1-based artificial chromosomes have revealed that there is a range of mitotic stability among artificial chromosomes \[[@B14]\]; future studies will aim to characterize the mechanistic basis of the segregation defects and the relative contribution of genomic and/or epigenetic factors to chromosome behavior.
Conclusions
===========
In summary, we have shown that artificial chromosomes assemble transcriptionally permissive chromatin and that there is a link between artificial chromosome size and the assembly of heterochromatin. Our results with the artificial chromosome panel are largely consistent with current models proposing that the formation of heterochromatin within the vicinity of CENP-A chromatin is functionally important, although the amount of heterochromatin assembled is quite variable, suggesting either that it is required only in small amounts or that it perhaps could even be dispensable. Strikingly, the studies here on the chromatin composition of artificial chromosomes, in combination with studies on normal human centromeres \[[@B34]\], strongly suggest that the chromatin state of the functional centromere region (as defined by CenH3 association) is quite distinct from pericentric heterochromatin. The artificial chromosome system provides a new set of reagents for investigating the role of both defined alpha-satellite DNA sequences and *trans*-acting epigenetic factors that cooperate to form a functional human centromere. A fuller understanding of the structure-function relationships of the chromatin and DNA composition of artificial chromosomes is important not only to further our understanding of the role of centromeres in genome stability, but also for the potential development of artificial chromosomes for gene transfer applications.
Materials and methods
=====================
Cell lines
----------
Characterization of cell lines containing mitotically stable human artificial chromosomes formed after transfection with either synthetic D17Z1 arrays (PAC17HT1.E29 (17-E29), PAC17HT1.D34 (17-D34), BAC17HT4.B12 (17-B12) or cloned DXZ1 sequences (X-4, X-5) have been described previously \[[@B12],[@B14]\]. The artificial chromosomes in 17-C20 were generated using VJ104-17α32 \[[@B12]\], hybridize with both D17Z1 and BAC vector probes, are *de novo*in composition and assemble CENP-A and CENP-E (data not shown). All artificial chromosomes were formed in human HT1080 cells. Cell lines were grown as described \[[@B12]\] and supplemented with either 100 μg/ml G418 (Gibco) (17-B12, 17-C20) or 2 μg/ml Blasticidin S HCl (ICN) (17-E29, 17-D34, X-5, X-6), as described \[[@B12]\].
Anaphase assays
---------------
Anaphase assays used to directly measure chromosome segregation defects in 17-B12 and 17-C20 (Table [1](#T1){ref-type="table"}) were carried out as previously described \[[@B14]\]. Assays were carried out at either 45 days (17-B12) or 12 days (17-C20) culture without selection. The spectrum orange-labeled D17Z1 probe (Vysis) hybridized with host 17 centromere regions and artificial chromosomes, whereas the spectrum green-labeled BAC vector probe VJ104 \[[@B6]\] hybridized exclusively with the artificial chromosomes. Co-localization of vector and D17Z1 probes produced yellow fluorescence on the artificial chromosomes, which allowed them to be distinguished from the host D17Z1 sequences (data not shown).
Generation of clonal lines expressing Myc-tagged HP1α
-----------------------------------------------------
The nucleotide sequence of human HP1α (NCBI Nucleotide database: S62077) was used in BLAST searches against entries in the human expressed sequence tag (EST) database using the NIH BLAST server \[[@B80]\]. A representative HP1α cDNA clone (IMAGE 627533) was obtained from Research Genetics. DNA was prepared with the Wizard-plus mini-prep DNA purification system (Promega), and the cDNA was sequenced on an ABI 373 (Perkin-Elmer) with a fluorescence labeled dye-terminator cycle sequencing kit according to the manufacturer\'s instructions (PRISM Ready DyeDeoxy Terminator Premix from Applied Biosystems). The full coding sequence of IMAGE 627533 was PCR-amplified with primers incorporating an *Eco*RI restriction enzyme recognition site (HP1α forward primer, 5\'-GGAATT CTGATGGGAAAGAAAACCAAGCG-3\'; reverse primer, 5\'-GGAATTCGCTCTTTGCTGTTT CTTTC-3\') and subcloned using standard techniques \[[@B81]\] into pcDNA3.1-CT-Myc-His (Invitrogen). Subclones were sequenced to verify sequence integrity and orientation as above. The HP1α-Myc tagged construct (pHP1α-Myc) was transfected into 17-C20, 17-B12, 17-E29 or 17-D34 cell lines using lipofectamine (Invitrogen), resulting in the formation of clonal lines (17-C20-1.B22, 17-B12-1.B10, 17-E29-1.C23 and 17-D34-1.A2, respectively) that stably express Myc-tagged HP1α. G418 selection at 400 μg/ml was applied to select clonal lines 17-E29-1.C23 and 17-D34-1.A2. Since 17-C20 and 17-B12 cells are G418-resistant, pHP1α-Myc was co-transfected in the presence of a second construct, pPAC4 \[[@B82]\] that carries a *bs*^r^marker gene. Clonal lines (17-C20-1.B22 and 17-B12-1.B10) resistant to 4 μg/ml Blasticidin S HCl (ICN) were selected and expanded. Confirmation of Myc-tagged HP1α expression was by immunofluorescence using a mouse monoclonal anti-Myc antibody (Invitrogen).
Immunofluorescence and fluorescence *in situ*hybridization (FISH)
-----------------------------------------------------------------
Metaphase spreads were prepared for immunofluorescence using previously described protocols \[[@B29]\]. Primary antibodies to the dimethylated form of histone H3 at lysine 4 (anti-H3DimK4) were purchased from Upstate Biotechnology (anti-dimethyl-histone H3 (Lys4)). Modification of histone H3 by trimethylation at lysine 9 (H3TrimK9) was detected using an antibody to the tri-methylated form of histone H3 at lysine 9 purchased from Abcam (anti histone H3-tri methyl K9). This antibody cross-reacts with lysine 27 on histone H3 and is termed anti-H3TrimK9/K27 in the present study. The CENP-A antibody was a generous gift from Manuel Valdivia (Cadiz University, Spain) \[[@B83]\]. Antibodies to H3DimK4, H3TrimK9/K27 and CENP-A were raised in rabbits. Primary and secondary antibody incubations were in 1× PBS supplemented with 1% BSA (Sigma). Secondary antibodies were purchased from Jackson ImmunoResearch. After immunofluorescence detection, 20-50 spreads were captured and their positions on the slide recorded. Slides were subsequently hybridized with an appropriate alpha-satellite probe to detect transfected and endogenous alpha-satellite sequences. FISH was carried out using standard protocols.
Replication timing assay
------------------------
Cells were pulsed with 10 μM BrdU (Roche) in T25 cm^2^flasks for 2 h intervals. Following three PBS washes, medium supplemented with 50 μM thymidine (Sigma) was added. Cells were left in thymidine-containing medium until chromosome harvest. At appropriate intervals, colcemid was added to block cells in mitosis. Cells were harvested and fixed in 3:1 methanol/acetic acid. Metaphase spreads on microscope slides were baked for 1 h at 60°C. Primary anti-BrdU antibody (Roche) was added for 1 h at room temperature. Rhodamine-donkey-anti-mouse secondary antibodies (Jackson ImmunoResearch) were used to visualize sites of BrdU incorporation. Typically, 25 metaphase spreads were captured following BrdU detection and their coordinates on the slide recorded. Subsequent analyses with a D17Z1 probe were used to confirm identity of artificial or endogenous chromosomes.
Acknowledgements
================
We thank Chris Yan for helpful discussions and Beth Sullivan for communicating results before publication. This work was supported by a Franklin Delano Roosevelt Award from the March of Dimes Birth Defects Foundation.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Heterochromatin forms on artificial chromosomes in the 3-20 Mb size range but is depleted on smaller artificial chromosomes that are approximately 1-3 Mb. Indirect immunofluorescence using an antibody that recognizes modification of histone H3 by trimethylation at lysine 9/lysine 27 (H3TrimK9/K27) (red signal) demonstrated that these heterochromatin markers are not detectable on the smaller D17Z1-based artificial chromosomes (arrowheads) in lines **(a)**17-D34 and **(b)**17-E29, but are readily detectable on the larger D17Z1- and DXZ1-based artificial chromosomes (arrowheads) as shown in lines **(c)**17-B12, **(d)**17-C20, **(e)**X-4 and **(f)**X-5. Arrows indicate chromosome 17 centromere regions (a-d) or host X centromere regions (e, f). Host D17Z1 sequences typically stained positive for H3TrimK9/K27 in most spreads (arrows in a-d). It was difficult to detect the X centromere signal (for example, arrow in (e)) but in about 30% of spreads there was a clearly positive signal as indicated by the arrow in (f). **(g)**Variation in H3TrimK9/K27 levels at host centromere regions is shown in a larger area of the spread shown in (c): artificial chromosomes are indicated by arrowheads; arrows point to the consistently strongly positive signals on the long arm of the Y chromosome (Yq). Artificial chromosome size estimates are listed in Table 2. Confirmation of artificial chromosomes and relevant host centromere regions were determined by FISH analyses with appropriate alpha-satellite probes (data not shown).
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Detection of HP1α on D17Z1-based artificial chromosomes. **(a-d)**Cell lines stably expressing a Myc-tagged form of HP1α. HP1α was detected using an anti-Myc antibody (red). The artificial chromosomes (about 1-3 Mb; indicated by small arrows) in lines **(a)**17-D34-1.A2 and **(b)**17-E29-1.C23 exhibit faint HP1α staining at a level similar to the general arm staining. Larger artificial chromosomes (3-10 Mb; small arrow) in lines **(c)**17-C20-1.B22 and **(d)**17-B12-1.B10 stain strongly for HP1α. Inserts in (a-c) show either DAPI (blue)-stained artificial chromosomes or HP1α (red). Host 17 centromere regions are indicated by the large arrows in (a-c). In (d), simultaneous staining for CENP-A (green) shows that CENP-A is restricted to a portion of the artificial chromosome (arrows) whereas the HP1α signal coats the entire artificial chromosome. In contrast to CENP-A, which is present at comparable levels on all artificial chromosomes tested \[12,13,58\] and host kinetochores \[61\], HP1α staining levels are more variable at host centromere regions (d).
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Transcriptionally competent chromatin is present on artificial chromosomes. Dimethylation of lysine 4 on histone H3 (H3DimK4) was visualized using an antibody against H3DimK4 (red). This euchromatin mark was detected on all artificial chromosomes (arrowheads) generated from either D17Z1 in lines **(a)**17-D34, **(b)**17-E29, **(c)**17-B12 and **(d)**17-C20, or DXZ1 in lines **(e)**X-4 or **(f)**X-5. Host centromere regions were generally depleted for H3DimK4 as indicated by arrows pointing to centromere regions of chromosome 17 (a-d) and the X chromosome (e, f).
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Replication timing of human artificial chromosomes in line 17-B12. BrdU detection (red) in cells that have been blocked with colcemid in mitosis following BrdU pulses during S phase (see Materials and methods). Artificial chromosome (small arrows; enlarged artificial chromosomes are shown in inserts) and chromosome 17 (large arrow) locations in each spread were confirmed by FISH analyses using a D17Z1 probe (data not shown). **(a-d)**Images from different periods in S phase. (a) Early in S phase, at 0-2 h, the two artificial chromosomes present in this spread are not replicating. Some incorporation of BrdU on chromosome 17 is detectable. (b) In the middle of S phase, at 2-4 h, two of four artificial chromosomes are replicating. (c) Later, at 4-6 h, all three artificial chromosomes are being coordinately replicated. Some BrdU incorporation within chromosome 17 arms is detectable. (d) Late in S phase, at 6-8 h, artificial chromosomes are not replicating. The centromere region on chromosome 17 is replicating (large arrow). Because of the A-rich sequence composition of satellite III on Yq, BrdU is preferentially incorporated into one strand, producing an asymmetrical staining pattern on Yq (arrowheads) \[84\].
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Replication timing in different human artificial chromosomes. **(a-c)**Detection of BrdU (red) on artificial chromosomes (small arrows; larger version in inserts). (a) In mid S phase, at 2-4 h, two artificial chromosomes in line 17-D34 are BrdU positive. (b) The artificial chromosome in line 17-E29 is replicating early in S phase, in the 0-2 h period. (c) In mid S phase (2-4 h), three artificial chromosomes are being coordinately replicated in this spread from line 17-C20. Images shown are from the first half of S phase, and, as expected, Yq (arrowhead) is not replicating at this time.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Artificial chromosome segregation errors
:::
Number (%) of chromosome mis-segregation events\*
-------- ----- --------------------------------------------------- ---- ---- -------------- ---------------- ---- ---
17-B12 86 5/208 (2.4) 3 2 3/295 (1.0) 2 1
17-C20 224 91/745 (12.2)^†^ 31 60 13/884 (1.5) 9 4
X-4^‡^ 400 16/866 (1.8) 10 6 25/1,596 (1.6) 20 5
X-5^‡^ 400 43/1,954 (2.2) 30 13 4/1,588 (0.2) 4 0
\*Chromosome segregation errors (either artificial chromosomes or host chromosomes 17 or X) were nondisjunction (NDJ) or anaphase lag (Lag) events. ^†^The predominant artificial chromosome segregation error in 17-C20 was due to anaphase lag (66%, *n*= 91). ^‡^Data for X-4 and X-5 have been published \[14\]. Segregation errors that could not be classified (for example, 1:0) were excluded from these analyses.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Chromatin formation on artificial chromosomes
:::
Line Artificial chromosomes Heterochromatin Euchromatin
--------------- ------------------------ ----------------- ------------- ----- ---- ----
17-E29 D17Z1 1-3 Mb \- (-) \+ \+
17-D34 D17Z1 1-3 Mb \- (-) \+ \+
17-B12 D17Z1 3-10 Mb \+ \+ \+ \+
17-C20^‡^ D17Z1 3-10 Mb \+ \+ \+ \+
X-4 DXZ1 10-20 Mb \+ ND \+ ND
X-5 DXZ1 10-20 Mb \+ ND \+ ND
Host controls
17 cen \+ \+ \- \+
X cen \+ ND \- \+
Summary of results obtained by immunofluorescence staining on metaphase chromosomes containing artificial chromosomes using antibodies to either heterochromatin (H3TrimK9/K27; HP1α) or euchromatin (H3DimK4) components (Figures 1-3). + positive staining; - signal not detectable; (-) weak staining comparable to general arm staining; ND, not done. \*Comparison of alpha satellite signal intensities (using FISH analyses) on the artificial chromosomes with those of the relevant host centromere regions was used to estimate artificial chromosome sizes. ^†^CENP-A stains uniformly on artificial chromosomes and at a level comparable to the host staining level \[12\]. Controls, staining pattern at either host 17 or X centromere regions overlapping with D17Z1 or DXZ1 probes (respectively). ^‡^17-C20 contains mitotically unstable artificial chromosomes (Table 1).
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Replication timing of artificial chromosomes
:::
Artificial chromosomes Replication timing
---------- ----------------------------- -------------------- ---- ----- ---- ----- ---- ----- ----- -----
17-E29 Euchromatin Early/mid 15 10 4 15 1 29 0 26
17-D34 Euchromatin Early/mid 16 11 26 13 0 31 0 32
17-B12 Euchromatin/heterochromatin Mid 2 24 14 15 9 19 0 30
17-C20 Euchromatin/heterochromatin Mid 0 56 20 57 0 57 0 57
Controls
17 cen Heterochromatin Mid 5 182 27 166 46 209 7 236
Yq Heterochromatin Mid/late 0 138 1 166 70 116 118 81
The number of either labeled (L) or unlabeled (U) artificial chromosomes in lines 17-E29, 17-D34, 17-B12 or 17-C20 or host control 17 centromere regions (17 cen) or Y long arm sequences (Yq) following BrdU detection at 2 h intervals in S phase is indicated in columns early, mid or late S phase. \*Chromatin composition of artificial chromosomes in the four lines indicated or control host 17 centromere or Yq regions (see Table 2). Euchromatin: euchromatin present; heterochromatin depleted. Euchromatin/heterochromatin: both euchromatin and heterochromatin present. Heterochromatin: predominantly heterochromatin; euchromatin depleted. ^†^Predominant phase in S phase during which replication occurs: early/mid: first half (0-4 h) of S phase; mid S phase (2-6 h into S phase); mid to late S (4-8 h into S phase). Pooled data from all experiments were used to generate the numbers for the controls.
:::
|
PubMed Central
|
2024-06-05T03:55:51.820719
|
2004-10-27
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545780/",
"journal": "Genome Biol. 2004 Oct 27; 5(11):R89",
"authors": [
{
"first": "Brenda R",
"last": "Grimes"
},
{
"first": "Jennifer",
"last": "Babcock"
},
{
"first": "M Katharine",
"last": "Rudd"
},
{
"first": "Brian",
"last": "Chadwick"
},
{
"first": "Huntington F",
"last": "Willard"
}
]
}
|
PMC545781
|
Background
==========
The delta subdivision of proteobacteria is a very diverse group of Gram-negative microorganisms that include aerobic genera *Myxococcus*with complex developmental lifestyles and *Bdellovibrio*, which prey on other bacteria \[[@B1]\]. In this study, we focus on anaerobic metal-reducing δ-proteobacteria, seven representatives of which have been sequenced recently, providing an opportunity for comparative genomic analysis. Within this group, sulfate-reducing bacteria, including *Desulfovibrio*and *Desulfotalea*species, are metabolically and ecologically versatile prokaryotes often characterized by their ability to reduce sulfate to sulfide \[[@B2]\]. They can be found in aquatic habitats or waterlogged soils containing abundant organic material and sufficient levels of sulfate, and play a key role in the global sulfur and carbon cycles \[[@B1]\]. Industrial interest in sulfate reducers has focused on their role in corrosion of metal equipment and the souring of petroleum reservoirs, while their ability to reduce toxic heavy metals has drawn attention from researchers interested in exploiting this ability for bioremediation. Psychrophilic sulfate-reducing *Desulfotalea psychrophila*has been isolated from permanently cold arctic marine sediments \[[@B3]\]. In contrast to sulfate-reducing bacteria, the genera *Geobacter*and *Desulfuromonas*comprise dissimilative metal-reducing bacteria, which cannot reduce sulfate, but include representatives that require sulfur as a respiratory electron acceptor for oxidation of acetate to carbon dioxide \[[@B4]\]. These bacteria are an important component of the subsurface biota that oxidizes organic compounds, hydrogen or sulfur with the reduction of insoluble Fe(III) oxides \[[@B5]\], and have also been implicated in corrosion and toxic metal reduction.
Knowledge of transcriptional regulatory networks is essential for understanding cellular processes in bacteria. However, experimental data about regulation of gene expression in δ-proteobacteria are very limited. Different approaches could be used for identification of co-regulated genes (regulons). Transcriptional profiling using DNA microarrays allows one to compare the expression levels of thousands of genes in different experimental conditions, and is a valuable tool for dissecting bacterial adaptation to various environments. Computational approaches, on the other hand, provide an opportunity to describe regulons in poorly characterized genomes. Comparison of upstream sequences of genes can, in principle, identify co-regulated genes. From large-scale studies \[[@B6]-[@B9]\] and analyses of individual regulatory systems \[[@B10]-[@B14]\] it is clear that the comparative analysis of binding sites for transcriptional regulators is a powerful approach to the functional annotation of bacterial genomes. Additional techniques used in genome context analysis, such as chromosomal gene clustering, protein fusions and co-occurrence profiles, in combination with metabolic reconstruction, allow the inference of functional coupling between genes and the prediction of gene function \[[@B15]\].
Recent completion of finished and draft quality genome sequences for δ-proteobacteria provides an opportunity for comparative analysis of transcriptional regulation and metabolic pathways in these bacteria. The finished genomes include sulfate-reducing *Desulfovibrio vulgaris*\[[@B16]\], *D. desulfuricans G20*, and *Desulfotalea psychrophila*, as well as the sulfur-reducing *G. sulfurreducens*\[[@B17]\], while the *G. metallireducens*genome has been completed to draft quality. A mixture of *Desulfuromonas acetoxidans*and *Desulfuromonas palmitatis*has been sequenced, resulting in a large number of small scaffolds, the identity of which (*acetoxidans*or *palmitatis*) has not been determined, and we refer to this sequence set simply as *Desulfuromonas*. Though draft-quality sequence can make it difficult to assert with confidence the absence of any particular gene, we have included these genomes in our study because they do provide insight as to the presence or absence of entire pathways, they can be compared to the related finished genome of *G. sulfurreducens*, and because complete genome sequence is not necessary for the methodology we use to detect regulatory sequences.
In this comprehensive study, we identify a large number of regulatory elements in these δ-proteobacteria. Some of the corresponding regulons are highly conserved among various bacteria (for example, riboswitches, BirA, CIRCE), whereas others are specific only for δ-proteobacteria. We also present the reconstruction of a number of biosynthetic pathways and systems for metal-ion homeostasis and stress response in these bacteria. The most important result of this study is identification of a novel regulon involved in sulfate reduction and energy metabolism in sulfate-reducing bacteria, which is most probably controlled by a regulator from the CRP/FNR family.
Results
=======
The results are organized under four main headings for convenience. In the first, we analyze a number of specific regulons for biosynthesis of various amino acids and cofactors in δ-proteobacteria. Most of them are controlled by RNA regulatory elements, or riboswitches, that are highly conserved across bacteria \[[@B18]\]. In the next section we describe several regulons for the uptake and homeostasis of transition metal ions that are necessary for growth. These regulons operate by transcription factors that are homologous to factors in *Escherichia coli*, but are predicted to recognize entirely different DNA signals. We then describe two stress-response regulons: heat-shock regulons (σ^32^and HrcA/CIRCE), which operate by regulatory elements conserved in diverse bacteria, and newly identified peroxide stress response regulons that are quite diverse and conserved only in closely related species. Finally, we present a completely new global regulon in metal-reducing δ-proteobacteria, which includes various genes involved in energy metabolism and sulfate reduction.
Biosynthesis and transport of vitamins and amino acids
------------------------------------------------------
### Biotin
Biotin (vitamin H) is an essential cofactor for numerous biotin-dependent carboxylases in a variety of microorganisms \[[@B19]\]. The strict control of biotin biosynthesis is mediated by the bifunctional BirA protein, which acts both as a biotin-protein ligase and a transcriptional repressor of the biotin operon. The consensus binding signal of BirA is a palindromic sequence TTGTAAACC-\[N~14/15~\]-GGTTTACAA \[[@B20]\]. Consistent with the presence of the biotin repressor BirA, all bacteria in this study have one or two candidate BirA-binding sites per genome, depending on the operon organization of the biotin genes (Table [1](#T1){ref-type="table"}). In the *Desulfovibrio*species, the predicted BirA site is located between the divergently transcribed biotin operon and the *birA*gene. In other genomes, candidate binding sites for BirA precede one or two separate biotin biosynthetic loci, whereas the *birA*gene stands apart and is not regulated.
All δ-proteobacteria studied possess genes for *de novo*biotin synthesis from pimeloyl-CoA precursor (*bioF*, *bioA*, *bioD*, *bioB*) and the bifunctional gene *birA*, but the initial steps of the biotin pathway are variable in these species (Figure [1](#F1){ref-type="fig"}). The *Geobacter*species have the *bioC-bioH*gene pair, which is required for the synthesis of pimeloyl-CoA in *Escherichia coli*. The *Desulfuromonas*species contain both *bioC-bioH*and *bioW*genes, representing two different pathways of pimeloyl-CoA synthesis. In contrast, *D. psychrophila*is predicted to synthesize a biotin precursor using the *bioC-bioG*gene pair, where the latter gene was only recently predicted to belong to the biotin pathway \[[@B20]\]. Both *Desulfovibrio*species have an extended biotin operon with five new genes related to the fatty-acid biosynthetic pathway. Among these new biotin-regulated genes not present in other δ-proteobacteria studied, there are homologs of acyl carrier protein (ACP), 3-oxoacyl-(ACP) synthase, 3-oxoacyl-(ACP) reductase and hydroxymyristol-(ACP) dehydratase. From positional and regulatory characteristics we conclude that these genes are functionally related to the biotin pathway. The most plausible hypothesis is that they encode a novel pathway for pimeloyl-CoA synthesis, as the known genes for this pathway, *bioC*, *bioH*, *bioG*and *bioW*, are missing in the *Desulfovibrio*species.
### Riboflavin
Riboflavin (vitamin B~2~) is an essential component of basic metabolism, being a precursor to the coenzymes flavin adenine dinucleotide (FAD) and flavin mononucleotide (FMN). The only known mechanism of regulation of riboflavin biosynthesis is mediated by a conserved RNA structure, the *RFN*-element, which is widely distributed in diverse bacterial species \[[@B21]\]. The δ-proteobacteria in this study possess a conserved gene cluster containing all genes required for the *de novo*synthesis of riboflavin (*ribD-ribE-ribBA-ribH*), but lack this regulatory element. The only exception is *D. psychrophila*, which has an additional gene for 3,4-dihydroxy-2-butanone-4-phosphate synthase (*ribB2*) with an upstream regulatory *RFN*element.
### Thiamine
Vitamin B~1~in its active form, thiamine pyrophosphate, is an essential coenzyme synthesized by the coupling of pyrimidine (HMP) and thiazole (HET) moieties in bacteria. The only known mechanism of regulation of thiamine biosynthesis in bacteria is mediated by a conserved RNA structure, the *THI*-element \[[@B22]\]. Search for thiamine-specific regulatory elements in the genomes of δ-proteobacteria identified one or two *THI*-elements per genome that are located upstream of thiamine biosynthetic operons (Figure 1 in Additional data file 1). The δ-proteobacteria possess all the genes required for the *de novo*synthesis of thiamine (Figure [2](#F2){ref-type="fig"}) with the exception of *Geobacter*species, which lack some genes for the synthesis and salvage of the HET moiety (*thiF*, *thiH*and *thiM*), and *D. psychrophila*, which has no *thiF*. In most δ-proteobacteria there are two paralogs of the thiamine phosphate synthase *thiE*, and *Geobacter*and *Desulfuromonas*species have fused genes *thiED*. In *D. psychrophila*, the only *THI*-regulated operon includes HET kinase *thiM*and previously predicted HMP transporter *thiXYZ*\[[@B22]\], whereas other thiamine biosynthetic genes are not regulated by the *THI*-element (Figure [2](#F2){ref-type="fig"}).
In most cases, downstream of a *THI*-element there is a candidate terminator hairpin, yielding regulation by the transcription termination/antitermination mechanism. The two exceptions predicted to be involved in translational attenuation are *THI*-elements upstream of genes *thiED*in *Desulfuromonas*and *thiM*in *D. psychrophila*. In the *Desulfovibrio*species, the *thiSGHFE*operon is preceded by two tandem *THI*-elements, each followed by a transcriptional terminator. This is the first example of possible gene regulation by tandem riboswitches.
### Cobalamin
Adenosylcobalamin (Ado-CBL), a derivative of vitamin B~12~, is an essential cofactor for several important enzymes. The studied genomes of δ-proteobacteria possess nearly complete sets of genes required for the *de novo*synthesis of Ado-CBL (Figure [3](#F3){ref-type="fig"}). The only exception is the precorrin-6x reductase, *cbiJ*, which was found only in *Desulfuromonas*but not in other species. The occurrence of CbiD/CbiG enzymes instead of the oxygen-dependent CobG/CobF ones suggests that these bacteria, consistent with their anaerobic lifestyle, use the anaerobic pathway for B~12~synthesis similar to that used by *Salmonella typhimurium*\[[@B23]\].
Ado-CBL is known to repress expression of genes for vitamin B~12~biosynthesis and transport via a co- or post-transcriptional regulatory mechanism, which involves direct binding of Ado-CBL to the riboswitch called the *B12*-element \[[@B24],[@B25]\]. A search for *B12*-elements in the genomes of δ-proteobacteria produced one *B12*-element in *D. desulfuricans*, *D. psychrophila*and *G. metallireducens*, two in *D. vulgaris*and *G. sulfurreducens*, and four in *Desulfuromonas*(Figure 2 in Additional data file 1). In *Geobacter*species these riboswitches regulate a large locus containing almost all the genes for the synthesis of Ado-CBL (Figure [3](#F3){ref-type="fig"}). One *B12*-element in the *Desulfovibrio*species regulates both the cobalamin-synthesis genes *cbiK-cbiL*and the vitamin B~12~transport system *btuCDF*, whereas three such regulatory elements in *Desulfuromonas*precede different vitamin B~12~transport loci. In *D. psychrophila*, a *B12*-element occurs within a large B~12~synthesis gene cluster and precedes the *cbiK-cbiL*genes.
The most interesting observation is that genes encoding the B~12~-independent ribonucleotide reductase NrdDG are preceded by *B12*-elements in *D. vulgaris*and *Desulfuromonas*. Notably, all δ-proteobacteria have another type of ribonucleotide reductase, NrdJ, which is a vitamin B~12~-dependent enzyme. We propose that when vitamin B~12~is present in the cell, expression of the B~12~-independent isozyme is inhibited, and a relatively more efficient B~12~-dependent isozyme is used. This phenomenon has been previously observed in other bacterial genomes \[[@B26]\].
### Methionine
The sulfur-containing amino acid methionine and its derivative *S*-adenosylmethionine (SAM) are important in protein synthesis and cellular metabolism. There are two alternative pathways for methionine synthesis in microorganisms, which differ in the source of sulfur. The *trans*-sulfuration pathway (*metI-metC*) utilizes cysteine, whereas the direct sulfhydrylation pathway (*metY*) uses inorganic sulfur instead. All δ-proteobacteria in this study except the *Desulfovibrio*species possess a complete set of genes required for the *de novo*synthesis of methionine (Figure [4](#F4){ref-type="fig"}). The *Geobacter*species and possibly *Desulfuromonas*have some redundancy in the pathway. First, these genomes contain the genes for both alternative pathways of the methionine synthesis. Second, they possess two different SAM synthase isozymes, classical bacterial-type MetK and an additional archaeal-type enzyme \[[@B27]\]. Moreover, it should be noted that the B~12~-dependent methionine synthase MetH in these bacteria lacks the carboxy-terminal domain, which is involved in reactivation of spontaneously oxidized coenzyme B~12~.
In Gram-positive bacteria, SAM is known to repress expression of genes for methionine biosynthesis and transport via direct binding to the S-box riboswitch \[[@B28]\]. In contrast, Gram-negative enterobacteria control methionine metabolism using the SAM-responsive transcriptional repressor MetJ. The δ-proteobacteria in this study have no orthologs of MetJ, but instead, we identified S-box regulatory elements upstream of the *metIC*and *metX*genes in the genomes of the *Geobacter*species and *Desulfuromonas*(see Figure 3 in Additional data file 1). A strong hairpin with a poly(T) region follows all these S-boxes, implying involvement of these S-boxes in a transcriptional termination/antitermination mechanism.
Both *Desulfovibrio*species have genes involved in the conversion of homocysteine into methionine (*metE*, *metH*and *metF*), which could be involved in the SAM recycling pathway, but not those genes required for *de novo*methionine biosynthesis. The ABC-type methionine transport system (*metNIQ*), which is widely distributed among bacteria, was also not found in these δ-proteobacteria. The *Desulfovibrio*species appear to have the single-component methionine transporter *metT*\[[@B28]\].
### Lysine
The amino acid lysine is produced from aspartate through the diaminopimelate (DAP) pathway in most bacteria. The first two stages of the DAP pathway, catalyzed by aspartokinase and aspartate semialdehyde dehydrogenase, are common for the biosynthesis of lysine, threonine, and methionine. The corresponding genes were found in δ-proteobacteria where they form parts of different metabolic operons. Four genes for the conserved stages of the lysine synthesis pathway (*dapA*, *dapB*, *dapF*and *lysA*) were further identified in δ-proteobacteria, whereas we did not find orthologs for three other genes (*dapC*, *dapE*and *dapD*), which vary in bacteria using different meso-DAP synthesis pathways. The lysine synthesis genes are mostly scattered along the chromosome, and in only some cases are *dapA*and either *dapB*, *dapF*or *lysA*clustered. All δ-proteobacteria studied lack the previously known lysine transporter LysP. However, in *D. desulfuricans*and *D. psychrophila*we found a gene for another candidate lysine transporter, named *lysW*, which was predicted in our previous genomic survey \[[@B29]\].
In various bacterial species, lysine is known to repress expression of genes for lysine biosynthesis and transport via the L-box riboswitch \[[@B30]\]. In addition, Gram-negative enterobacteria use the lysine-responsive transcriptional factor LysR for control of the *lysA*gene. Among the δ-proteobacteria studied, we found neither orthologs of LysR, nor representatives of the L-box RNA regulatory element. In an attempt to analyze potential lysine regulons in this phylogenetic group, we collected upstream regions of all lysine biosythesis genes and applied SignalX as a signal detection procedure \[[@B31]\]. The strongest signal, a 20-bp palindrome with consensus GTGGTACTNNNNAGTACCAC, was observed upstream of the *lysX-lysA*operons in both *Desulfovibrio*genomes and the candidate lysine transporter gene *lysW*in *D. desulfuricans*(Table [2](#T2){ref-type="table"}). The first gene in this operon, named *lysX*, encodes a hypothetical transcriptional regulator with a helix-turn-helix motif (COG1378) and is the most likely candidate for the lysine-specific regulator role in *Desulfovibrio*. To find new members of the regulon, the derived profile (named LYS-box) was used to scan the *Desulfovibrio*genomes. The lysine regulon in these genomes appears to include an additional gene (206613 in *D. vulgaris*, and 394397 in *D. desulfuricans*), which encodes an uncharacterized membrane protein with 14 predicted transmembrane segments. We predict that this new member of the lysine regulon might be involved in the uptake of lysine or some lysine precursor.
Metal ion homeostasis
---------------------
### Iron
Iron is necessary for the growth of most bacteria as it participates in many major biological processes \[[@B32]\]. In aerobic environments, iron is mainly insoluble, and microorganisms acquire it by secretion and active transport of high-affinity Fe(III) chelators. Under anaerobic conditions, Fe(II) predominates over ferric iron, and can be transported by the ATP-dependent ferrous iron transport system FeoAB. Genomes of anaerobic δ-proteobacteria contain multiple copies of the *feoAB*genes, and lack ABC transporters for siderophores. Regulation of iron metabolism in bacteria is mediated by the ferric-uptake regulator protein (FUR), which represses transcription upon interaction with ferrous ions. FUR can be divided into two domains, an amino-terminal DNA-binding domain and a carboxy-terminal Fe(II)-binding domain. The consensus binding site of *E. coli*FUR is a palindromic sequence GATAATGATNATCATTATC \[[@B33]\].
In all δ-proteobacteria studied except *D. psychrophila*, we identified one to three FUR orthologs that form a distinct branch (FUR\_Delta) in the phylogenetic tree of the FUR/ZUR/PerR protein family (see below). One protein, FUR2 in *D. desulfuricans*, lacks an amino-terminal DNA-binding domain and is either non-functional or is involved in indirect regulation by forming inactive heterodimers with two other FUR proteins. Scanning the genomes with the FUR-box profile of *E. coli*did not result in identification of candidate FUR-boxes in δ-proteobacteria. In an attempt to analyze potential iron regulons in this phylogenetic group, we collected upstream regions of the iron-transporter genes *feoAB*and applied SignalX to detect regulatory signals. The strongest signal, a 17-bp palindrome with consensus WTGAAAATNATTTTCAW (where W indicates A or T), was observed upstream of the multiple *feoAB*operons and *fur*genes in all δ-proteobacteria except *D. psychrophila*(Table [3](#T3){ref-type="table"}). The constructed search profile (dFUR-box) was applied to detect new candidate FUR-binding sites in these five genomes (Figure [5](#F5){ref-type="fig"} and Table [3](#T3){ref-type="table"}).
The smallest FUR regulons were observed in the *Geobacter*and *Desulfuromonas*species, where they include the ferrous iron transporters *feoAB*(one to four copies per genome), the *fur*genes themselves (one copy in the *Geobacter*species and two copies in *Desulfuromonas*), and two hypothetical porins. The first one, named *psp*, was found only in *G. metallireducens*and *Desulfuromonas*genomes, where it is preceded by two tandem FUR-boxes. The *psp*gene has homologs only in *Aquifex aeolicus*and in various uncultured bacteria, and in one of them (a β-proteobacterium) it is also preceded by two FUR-boxes (GenBank entry AAR38161.1). This gene is weakly similar to the family of phosphate-selective porins (PFAM: PF07396) from various Gram-negative bacteria. The second hypothetical porin was found only in *G. sulfurreducens*(383590), where it is preceded by a FUR-box and followed by *feoAB*transporter. This gene, absent in other δ-proteobacteria, has only weak homologs in some Gram-negative bacteria and belongs to the carbohydrate-selective porin OprB family (PFAM: PF04966). Thus, two novel genes predicted to fall under FUR control encode hypothetical porins that could be involved in ferrous iron transport.
Another strong FUR-box in the *G. sulfurreducens*genome precedes a cluster of two hypothetical genes located immediately upstream of the *feoAB*-containing operon. The first gene in this operon, named *genX*(383594), has no orthologs in other bacteria and the encoded protein has a heme-binding site signature of the cytochrome *c*family (PFAM: PF00034). The second gene, named *genY*(383592), encodes a two-domain protein that is not similar to any known protein. In *Desulfuromonas*, an ortholog of the *genY*amino-terminal domain (391875) is divergently transcribed from a predicted ferric reductase (391874), and their common upstream region contains a strong FUR-box. Moreover, orthologs of the *genY*C-terminal domain were identified in *Desulfovibrio*species, where they are again preceded by two tandem FUR-boxes and form a cluster with the hypothetical gene, *genZ*, encoding a protein of 100 amino acids with two tetratricopeptide repeat domains that are usually involved in protein-protein interactions (PFAM: PF00515). From genomic analysis alone it is difficult to predict possible functions of these new members of the FUR regulon in δ-proteobacteria.
Two *Desulfovibrio*species have significantly extended FUR regulons that are largely conserved in these genomes and include ferrous iron transporter genes *feoAB*and many hypothetical genes. Another distinctive feature of the FUR regulon in *Desulfovibrio*species is a structure of two partially overlapping FUR-boxes shifted by 6 bp. Interestingly, the flavodoxin gene, *fld*, is predicted to be regulated by FUR in both *Desulfovibrio*species. In addition to this iron-repressed flavodoxin (a flavin-containing electron carrier), the *Desulfovibrio*species have numerous ferredoxins (an iron-sulfur-containing electron carrier). One possible explanation is that in iron-restricted conditions these microorganisms can replace ferredoxins with less-efficient, but iron-independent alternatives. A similar regulatory strategy has been previously described for superoxide dismutases in *E. coli*, *Bordetella pertusis*and *Pseudomonas aeruginosa*\[[@B34]-[@B36]\] and predicted, in a different metabolic context, for B~12~-dependent and B~12~-independent enzymes \[[@B26]\]; see the discussion above.
Other predicted regulon members with conserved FUR-boxes in both *Desulfovibrio*species are the hypothetical genes *pep*(Zn-dependent peptidase), *gdp*(GGDEF domain protein, PF00990), *hdd*(metal dependent HD-domain protein, PF01966), and a hypothetical P-type ATPase (392971) that could be involved in cation transport, and a long gene cluster starting from the *pqqL*gene (Zn-dependent peptidase). The latter cluster contains at least 10 hypothetical genes encoding components of ABC transporters and biopolymer transport proteins (*exbB*, *exbD*and *tonB*). In *D. vulgaris*, the first gene in this FUR-regulated cluster is an AraC-type regulator named *foxR*, since it is homologous to numerous FUR-controlled regulators from other genomes (*foxR*from *Salmonella typhi*, *alcR*from *Bordetella pertussis*, *ybtA*from *Yersinia*species, *pchR*from *Pseudomonas aeruginosa*), which usually regulate iron-siderophore biosynthesis/transport operons \[[@B33]\]. An ortholog of *foxR*, a single FUR-regulated gene, was identified in *D. desulfuricans*located about 30 kb away from the FUR-regulated *pqqL*gene cluster. Given these observations, we propose that this gene cluster is involved in siderophore transport and is regulated by FoxR.
A hypothetical gene in *D. vulgaris*(*209207*) has the strongest FUR-box in this genome; however, its orthologs in *D. desulfuricans*are not predicted to belong to the FUR regulon. Another operon in *D. desulfuricans*(392971-392970-392969), encoding three hypothetical proteins, is preceded by two candidate FUR-boxes, but these genes have no orthologs in other δ-proteobacteria. Thus, FUR-dependent regulation of these hypothetical genes is not confirmed in other species, and their possible role in the iron homeostasis is not clear.
### Nickel
The transition metal nickel (Ni) is an essential cofactor for a number of prokaryotic enzymes, such as \[NiFe\]-hydrogenase, urease, and carbon monoxide dehydrogenase (CODH). Two major types of nickel-specific bacterial transporters are represented by the NikABCD system of *E. coli*(the nickel/peptide ABC transporter family) and the HoxN of *Ralstonia eutropha*(the NiCoT family of nickel/cobalt permeases). Nickel uptake must be tightly regulated because excessive nickel is toxic. In *E. coli*and some other proteobacteria, nickel concentrations are controlled by transcriptional repression of the *nikABCD*operon by the Ni-dependent regulator NikR \[[@B37]\].
The genomes of δ-proteobacteria studied so far contain multiple operons encoding \[NiFe\] and \[Fe\] hydrogenases and Ni-dependent CODH, but lack urease genes. Both known types of nickel-specific transporters are absent in δ-proteobacteria, but these genomes contain orthologs of the nickel repressor *nikR*. In an attempt to identify potential nickel transporters in this taxonomic group, we analyzed the genome context of the *nikR*genes. The *nikR*gene in *Desulfuromonas*is co-localized with a hypothetical ABC transport system, which is weakly homologous to the cobalt ABC-transporter *cbiMNQO*from various bacteria. Orthologs of this system, named here *nikMNQO*, are often localized in proximity to Ni-dependent hydrogenase or urease gene clusters in various proteobacteria (data not shown). Among δ-proteobacteria, the *Geobacter*species have a complete *nikMNQO*operon, whereas operons in *D. desulfuricans*and *D. psychrophila*lack the *nikN*component but include two additional genes, named *nikK*and *nikL*, which both encode hypothetical proteins with amino-terminal transmembrane segments (Figure [6](#F6){ref-type="fig"}). *Desulfovibrio vulgaris*has a *nikMQO*cluster and separately located *nikK*and *nikL*genes. Since various other proteobacteria also have the same clusters including *nikK*and *nikL*, but not *nikN*(data not shown), we propose that these two genes encode additional periplasmic components of the NikMQO ABC transporter, possibly involved in the nickel binding.
By applying SignalX to a set of upstream regions of the *nikMQO*operons, we identified *de novo*the NikR binding signal in all δ-proteobacteria except *D. psychrophila*(Table [4](#T4){ref-type="table"}). This signal has the same structure as in enterobacteria (an inverted repeat of 27-28 bp), but its consensus (GTGTTAC-\[N~13/14~\]-GTAACAC) differs significantly from the consensus of NikR binding signal of enterobacteria (GTATGAT-\[N~13/14~\]-ATCATAC) \[[@B37]\]. Using the derived profile to scan the genomes of δ-proteobacteria we identified one more candidate NikR-binding site in *D. desulfuricans*. Thus the nickel regulon in this bacterium includes the *hydAB2*operon, encoding periplasmic iron-only hydrogenase. Altogether, *D. desulfuricas*has three paralogs of \[NiFe\] hydrogenase and two paralogs of \[Fe\] hydrogenase. We predict that an excess of nickel represses a nickel-independent hydrogenase isozyme using the Ni-responsive repressor NikR. Regulation of hydrogenase enzymes by NikR has not been described previously. A closer look at the upstream region of the putative nickel transport operon in *D. psychrophila*revealed similar NikR consensus half-sites but in the opposite orientation to each other (GTAACAC-\[N~13/14~\]-GTGTTAC). Searching the genomes with this reversed NikR signal, we observed one more hypothetical gene cluster in *D. psychrophila*which has two high-scoring NikR-sites in the upstream region, and a NikR-site upstream of the single *nikK*gene in *D. vulgaris*(Figure [6](#F6){ref-type="fig"}).
### Zinc
Zinc is an important component of many proteins, but in large concentrations it is toxic to the cell. Thus zinc repressors ZUR regulate high-affinity zinc transporters *znuABC*in various bacteria \[[@B38]\]. An orthologous zinc transporter was found in δ-proteobacteria (Figure [7](#F7){ref-type="fig"}). In *G. sulfurreducens*and the *Desulfovibrio*species, this cluster also includes a hypothetical regulatory gene from the FUR/ZUR/PerR family, named *zur\_Gs*and *zur\_D*, respectively. Phylogenetic analysis of this protein family demonstrated that ZUR\_Gs and ZUR\_D are not close relatives and are only weakly similar to known FUR, ZUR, and PerR regulators from other bacteria (see below). The predicted ZUR-binding site located just upstream of the *zur-znuABC*operon in *G. sulfurreducens*is highly similar to the ZUR consensus of Gram-positive bacteria (TAAATCGTAATNATTACGATTTA). Another strong signal, a 17-bp palindrome with consensus ATGCAACNNNGTTGCAT, was identified upstream of the *znuABC-zur*operons in two *Desulfovibrio*genomes (Table [5](#T5){ref-type="table"}). Although *znuABC*genes are present in all δ-proteobacteria, we observed neither candidate ZUR regulators, nor ZUR-binding sites in *G. metallireducens*, *Desulfuromonas*and *D. psychrophila*, suggesting either the absence of zinc-specific regulation or presence of another regulatory mechanism for these genes.
### Cobalt
The previously described cobalt transport system CbiMNQO was found only in the *Geobacter*species, where it is located within the B~12~-regulated *cbi*gene cluster close to the cobaltochelatase gene *cbiX*, responsible for incorporation of cobalt ions into the corrin ring (see the \'Cobalamin\' section above). In contrast, other δ-proteobacteria, possessing a different cobaltochelatase (*cbiK*), lack homologs of any known cobalt transporter. It was previously suggested by global analysis of the B~12~metabolism that different types of cobalt transporters are interchangeable in various bacterial species \[[@B26]\]. From genome context analysis and positional clustering with the *cbiK*gene, we predicted a novel candidate cobalt transporter in δ-proteobacteria, named *cbtX*(Figure [3](#F3){ref-type="fig"}), which was previously annotated as a hypothetical transmembrane protein conserved only in some species of archaea (COG3366).
### Molybdenum
Molybdenum (Mo) is another transition metal essential for bacterial metabolism. Bacteria take up molybdate ions via a specific ABC transport system encoded by *modABC*genes. Mo homeostasis is regulated by the molybdate-responsive transcription factor ModE, containing an amino-terminal DNA-binding domain and two tandem molybdate-binding domains. Orthologs of ModE are widespread among prokaryotes, but not ubiquitous \[[@B39]\]. All δ-proteobacteria have one or more homologs of the *modABC*transporter (Figure [8](#F8){ref-type="fig"}). However, full-length *modE*genes containing both DNA- and molybdate-binding domains were observed only in *G. sulfurreducens*and *Desulfuromonas*. In *G. sulfurreducens*, the molybdate transport operon is co-localized with *modE*and is preceded by a putative ModE-binding site (Table [6](#T6){ref-type="table"}), which is similar to the *E. coli*consensus of ModE (ATCGNTATATA-\[N~6~\]-TATATANCGAT). In contrast, we could not identify *E. coli*-type ModE-binding sites upstream of the *mod*operons in *Desulfuromonas*, indicating that these operons may be regulated by a different, unidentified signal.
Three other δ-proteobacteria (two *Desulfovibrio*species and *D. psychrophila*) have genes encoding a single DNA-binding domain of ModE (Figure [8](#F8){ref-type="fig"}). Searching with the *E. coli*-type profile did not reveal candidate binding sites of ModE in these species. To predict potential ModE sites *de novo*, we collected upstream regions of all molybdate transport operons and applied SignalX. In both *Desulfovibrio*genomes, we identified a common inverted repeat with consensus CGGTCACG-\[N~14~\]-CGTGACCG, which is considerably different from the *E. coli*consensus of ModE (Table [6](#T6){ref-type="table"} and Figure [8](#F8){ref-type="fig"}). The *modABC*gene cluster in these species includes an additional chimeric gene encoding a fusion of phage integrase family domain (PF00589) and one or two molybdate-binding domains (MOP). The functions of these chimeric molybdate-binding proteins, and the mechanism of Mo-sensing by DNA-binding ModE domains in the *Desulfovibrio*species, are not clear.
Stress response regulons
------------------------
### Oxidative stress
Under aerobic conditions, generation of highly toxic and reactive oxygen species such as superoxide anion, hydrogen peroxide and the hydroxyl radical leads to oxidative stress with deleterious effects \[[@B40]\]. Strictly anaerobic sulfate-reducing bacteria are adapted to survive in transient oxygen-containing environments by intracellular reduction of oxygen to water using rubredoxin:oxygen oxidoreductase (Roo) as the terminal oxidase \[[@B41]\]. The main detoxification system for reactive oxygen species in aerobic and anaerobic bacteria involves superoxide dismutase (Sod), catalase (KatA, KatG) and nonspecific peroxidases (for example, AhpC). In addition to these enzymes, *Desulfovibrio*species have an alternative mechanism for protecting against oxidative stress, which includes rubredoxin oxidoreductase (Rbo), which has superoxide reductase activity, rubrerythrin (Rbr) with NADH peroxidase activity, and rubredoxin-like proteins (Rub, Rdl), which are used as common intermediary electron donors \[[@B42]\].
Searching for orthologs of the oxidative stress-related genes in the genomes in this study revealed great variability in content and genomic organization (Figure [9](#F9){ref-type="fig"}). We also searched for homologs of transcription factors known to be involved in regulation of the peroxide and superoxide stress responses. Lacking orthologs of the *E. coli*OxyR and SoxR/SoxS regulators, the δ-proteobacteria studied have instead multiple homologs of the peroxide-sensing regulator PerR of *B. subtilis*\[[@B43]\]. The PerR-specific branch on the phylogenetic tree of the FUR/ZUR/PerR family contains at least three distinct sub-branches with representatives from δ-proteobacteria (Figure [10](#F10){ref-type="fig"}). In all cases except *D. psychrophila*, the *perR*genes are co-localized on the chromosome with various peroxide stress-responsive genes (Figure [9](#F9){ref-type="fig"}). However, the upstream regions of these genes contain no candidate PerR-binding sites conforming to the *B. subtilis*PerR consensus TTATAATNATTATAA. Applying the SignalX program to various subsets of upstream regions of peroxide stress-responsive genes resulted in identification of candidate PerR operators in δ-proteobacteria (Table [7](#T7){ref-type="table"}).
In the *Desulfovibrio*species, a common palindromic signal was found upstream of the *perR*and *rbr2*genes. In *D. vulgaris*, *perR*forms an operon with *rbr*and *rdl*genes \[[@B42]\]. Searching for genes with the derived profile identified additional candidate members of the PerR regulon, alkyl hydroperoxide reductase *ahpC*in *D. vulgaris*(*D. desulfuricans*has no ortholog of *ahpC*), and a hypothetical gene of unknown function in both *Desulfovibrio*species (206199 in *D. vulgaris*and 395549 in *D. desulfuricans*).
The *perR-rbr-roo*operon in both *Geobacter*species is preceded by a conserved palindromic region (Table [7](#T7){ref-type="table"}) which overlaps a candidate -10 promoter element (Figure [11](#F11){ref-type="fig"}). The second *perR*paralog in *G. sulfurreducens*(named *perR2*), which is followed by a gene cluster containing two cytochrome peroxidase homologs (*hsc*and *ccpA*), glutaredoxin (*grx*) and rubrerythrin (*rbr*), has a close ortholog in the *Desulfuromonas*species, where it precedes the *rbr*gene (Figures [9](#F9){ref-type="fig"}, [10](#F10){ref-type="fig"}). For these gene clusters we found a common palindromic signal, which is not similar to other predicted PerR signals in δ-proteobacteria (Table [7](#T7){ref-type="table"}). Two other *perR*paralogs in *Desulfuromonas*(*perR2*and *perR3*) probably result from a recent gene duplication (Figure [10](#F10){ref-type="fig"}), and both are co-localized on the chromosome with the peroxide stress-responsive genes *katG*and *rbr2*, respectively (Figure [9](#F9){ref-type="fig"}). A common new signal identified upstream of the *katG*and *perR3*genes is probably recognized by both PerR2 and PerR3 regulators in this organism (Table [7](#T7){ref-type="table"}).
The PerR regulons in δ-proteobacteria are predicted to include only a small subset of all peroxide stress-related genes identified in these genomes. In addition to the mainly local character of the predicted regulation, these regulons seem to be highly variable between different species, both in their content and DNA signals.
### Heat shock
In bacteria, two major mechanisms regulating expression of heat-shock proteins are positive control by alternative sigma factor σ^32^, encoded by the *rpoH*gene, and negative control by binding of the repressor protein HrcA to palindromic operators with a consensus TTAGCACTC-\[N~9~\]-GAGTGCTAA called CIRCE \[[@B44]\]. The *rpoH*gene was identified in the genomes of all δ-proteobacteria studied. Though the HrcA/CIRCE system is conserved in very diverse taxonomic groups of bacteria, it is not universal, as some γ-proteobacteria lack it \[[@B45]\]. We detected the *hrcA*genes and CIRCE sites in all genomes studied except *D. psychrophila*(Table [8](#T8){ref-type="table"}).
We then searched the genomes of δ-proteobacteria with previously constructed profiles for σ^32^promoters and CIRCE \[[@B45]\]. As was observed previously for other bacteria, the only constant member of the HrcA regulon in δ-proteobacteria is the *groESL*operon. In addition, CIRCE sites are present upstream of the *hrcA-grpE-dnaKJ*operons in the *Geobacter*and *Desulfuromonas*species and upstream of the *rpoH*gene in *G. sulfurreducens*. In contrast to the highly conserved CIRCE signal, the σ^32^promoters identified in multiple copies in various proteobacteria are less conserved \[[@B45],[@B46]\]. Among δ-proteobacteria, we identified σ^32^-like promoters upstream of some heat-shock-related genes encoding chaperons (GroE, DnaJ, DnaK, GrpE) and proteases (ClpA, ClpP, ClpX, Lon) (Table [9](#T9){ref-type="table"}). Thus, in δ-proteobacteria, as in most proteobacteria, σ^32^plays a central part in the regulation of the heat-shock response, although detailed regulatory strategies seem to vary in different species. The alternative HrcA/CIRCE system controls expression of *groE*and other major chaperons.
Central energy metabolism
-------------------------
### The CooA regulon for carbon monoxide utilization in *Desulfovibrio*species
Growth using carbon monoxide (CO) as the sole energy source involves two key enzymes in the γ-proteobacterium *Rhodospirillum rubrum*- CO dehydrogenase (CODH) and an associated hydrogenase - which are encoded in the *coo*operons and induced by the CO-sensing transcriptional activator CooA \[[@B47]\]. Among the sequenced δ-proteobacteria, only *Desulfovibrio*species have *coo*operons and the CooA regulator. *D. vulgaris*has two separate operons encoding CODH and the associated hydrogenase, whereas *D. desulfuricans*has only one operon encoding CODH (Figure [12](#F12){ref-type="fig"}). The strongest identified signal, a 16-bp palindrome with consensus TGTCGGCNNGCCGACA, was identified upstream of the *coo*operons from both *Desulfovibrio*species and *R. rubrum*(Table [10a](#T10){ref-type="table"}). This consensus conforms to the experimentally known CooA-binding region at the *R. rubrum cooFSCTJ*operon \[[@B48]\].
### New CRP/FNR-like regulon for sulfate reduction and prismane genes
Sulfate-reducing bacteria are characterized by their ability to utilize sulfate as a terminal electron acceptor. To try to identify the regulatory signals responsible for this metabolism, we applied the signal detection procedure SignalX to a set of upstream regions of genes involved in the sulfate-reduction pathway in *Desulfovibrio*species. A conserved palindromic signal with consensus sequence TTGTGANNNNNNTCACAA was detected upstream of the *sat*and *apsAB*operons, which encode ATP sulfurylase and APS reductase, respectively. This novel signal is identical to the *E. coli*CRP consensus, and we hypothesized that a CRP-like regulator might control the sulfate-reduction regulon in *Desulfovibrio*. Scanning the *Desulfovibrio*genomes resulted in identification of similar sites upstream of many hypothetical genes encoding various enzymes and regulatory systems (Table [10b](#T10){ref-type="table"} and Figure [12](#F12){ref-type="fig"}). One of them, the *hcp*gene in *D. vulgaris*, encodes a hybrid-cluster protein (previously called the prismane-containing protein) of unknown function \[[@B49]\], which is coexpressed with a hypothetical ferredoxin gene, named *frdX\**: new gene names introduced in this study are marked by asterisk. In both *Desulfovibrio*species, the *hcp-frdX\**genes are co-localized with a hypothetical regulatory gene from the CRP/FNR family of transcriptional regulators, named HcpR\* for the Hcp regulator (Figure [12](#F12){ref-type="fig"}).
Close HcpR\* orthologs were detected in two other δ-proteobacteria, *D. psychrophila*and *Desulfuromonas*; however, the same CRP-like signals were not present in their genomes. Examination of a multiple alignment of the CRP/FNR-like proteins revealed one specific amino acid (Arg 180) in the helix-turn-helix motif involved in DNA recognition, which is changed from arginine (for example, in *E. coli*CRP and *Desulfovibrio*HcpR\*) to serine and proline in these two δ-proteobacteria (data not shown). As both these species have multiple *hcp*and *frdX*paralogs, we applied SignalX to a set of corresponding upstream regions and obtained another FNR-like palindromic signal with consensus at ATTTGACCNNGGTCAAAT, which is notably distinct from the CRP-like signal in the third position (which has T instead of G). Such candidate sites were observed upstream of all *hcp*and *frdX*paralogs identified in *D. psychrophila*and *Desulfuromonas*, as well as upstream of some additional genes in *Desulfuromonas*, for example those encoding polyferredoxin and cytochrome *c*heme-binding protein (Table [10](#T10){ref-type="table"} and Figure [12](#F12){ref-type="fig"}).
The HcpR regulon was also identified in other taxonomic groups, including *Clostridium*, *Thermotoga*, *Bacteroides*, *Treponema*and *Acidothiobacillus*species, and in all cases candidate HcpR sites precede *hcp*orthologs (data not shown). Moreover, the *hcpR*gene is often co-localized with *hcp*on the chromosome. In clostridia, *frdX*orthologs are also preceded by candidate HcpR sites. These data indicate that the main role of HcpR is control of expression of two hypothetical proteins - hybrid-cluster protein and ferredoxin - which are most probably involved in electron transport. However, the HcpR regulon is significantly extended in some organisms. Additional members of this regulon that are conserved between the two *Desulfovibrio*species include two operons involved in sulfate reduction (*apsAB*and *sat*), a hypothetical cluster of genes (206515-206516) with similarity to dissimilative sulfite and nitrite reductases, polyferredoxin, a hypothetical gene conserved in Archaea (209119), and the putative thiosulfate reductase operon *phcAB*(209106-209105). Notably, both CooA and HcpR candidate sites precede the *cooMKLXUHF*operon for CODH-associated hydrogenase, which is present only in *D. vulgaris*.
Because regulators from the CRP/FNR family are able to both repress and activate gene expression, it was interesting to predict the mode of regulation of the HcpR regulon members. To this end, we investigated the positions of candidate HcpR sites in pairwise alignments of orthologous regulatory regions from the two *Desulfovibrio*species. These two closely related genomes are diverse enough to identify regulatory elements as conserved islands in alignments of intergenic regions. For the *sat*and *apsAB*operons, the HcpR sites were found within highly conserved parts of alignments and in both cases the site overlaps the -10 box of a site strongly resembling a promoter (Figure [13a,b](#F13){ref-type="fig"}), suggesting repression of the genes by HcpR. In contrast, positive regulation by HcpR could be proposed for the *hcp-frdX*, 206515-206516 and 209119 operons, which have HcpR sites upstream or slightly overlapping the -35 box of predicted promoters (Figure [13c](#F13){ref-type="fig"}). In the case of the *cooMKLXUHF*operon in *D. vulgaris*, the HcpR site is located upstream of the candidate site of the known positive regulator CooA; thus it is also predicted to be an activator site.
By analysis of the functions of genes co-regulated by HcpR, it is difficult to predict the effector for this novel regulon. The physiological role of the hybrid iron-sulfur cluster protein Hcp, the most conserved member of the HcpR regulon, is not yet characterized despite its known three-dimensional structure and expression profiling in various organisms. In two facultative anaerobic bacteria, *E. coli*and *Shewanella oneidensis*, the *hcp*gene is expressed only under anaerobic conditions in the presence of either nitrate or nitrite as terminal electron acceptors \[[@B50],[@B51]\]. More recent expression data obtained for anaerobic *D. vulgaris*have showed strong upregulation of the *hcp-frdX\**and 206515-206516 operons by nitrite stress (J. Zhou, personal communication). While HcpR is predicted to activate these two hypothetical operons, as well as the CODH-associated hydrogenase operon, it most probably represses two enzymes from the sulfate reduction pathway, APS reductase and ATP sulfurylase. We hypothesize that HcpR is a key regulator of the energy metabolism in anaerobic bacteria, possibly controlling the transition between utilization of alternative electron acceptors, such as sulfate and nitrate. The absence of the dissimilatory sulfite reductase DsrAB in the predicted HcpR regulon of *Desulfovibrio*could be explained by its experimentally defined ability to reduce both sulfite and nitrite \[[@B52]\].
Discussion
==========
Regulation of biosynthesis pathways
-----------------------------------
Because the organisms considered in this study are commonly identified on the basis of their catabolic capabilities, comparatively little is known about the regulation of their biosynthetic pathways. In this study, we identified a number of previously characterized regulatory mechanisms (involved in biotin, thiamine, cobalamin and methionine synthesis), all of which, excluding the biotin regulon, are mediated by direct interaction of a metabolic product with a riboswitch control element (summarized in Table [11](#T11){ref-type="table"}). Of particular interest in this set was observation of a dual tandem *THI*-element riboswitch in *Desulfovibrio*species. Multiple protein-binding sites are a common regulatory feature and often imply cooperative binding of multiple protein factors. Although true riboswitch units do not interact with *trans*-acting factors, it is theoretically possible for independently acting sites to yield a cooperative effect when ligand binding derepresses transcription. For switches that are repressed by ligand binding, however, tandem sites would simply lower the concentration threshold at which a response is seen, but not affect cooperativity unless some more complicated interaction of the sites were allowed. On the one hand, independently acting sites is a simpler mechanism to explain, while on the other hand, it seems unusual that duplicate sites would have evolved to adjust the concentration response instead of simply changing the binding affinity for the ligand at the sequence level. Moreover, it seems unlikely that a tandem switch would be preserved across a large evolutionary distance without offering some other advantage such as cooperativity. It would be interesting to investigate the biochemical behavior of these tandem *THI*-elements in the laboratory to resolve whether their genomic organization reflects a more sophisticated mode of regulation, or is simply an evolutionarily convenient way to adjust the concentration response, or is perhaps just a recombination remnant that has persisted in these genomes by chance.
Another interesting finding was the absence of complete machinery for the *de novo*synthesis of methionine in the *Desulfovibrio*species. These organisms have the necessary genes to form methionine from homocysteine, but no apparent process by which to produce homocysteine. Although the enzymatic pathway of cysteine synthesis has been studied in *Desulfovibrio vulgaris*\[[@B53]\], its ability to synthesize methionine has not been characterized. Growth in minimal medium using sulfate as the only source of sulfur is routine, however, and suggests that these bacteria use a previously uncharacterized mechanism for assimilation of sulfur into methionine. On the basis of genomic context analysis we also predicted that the *Desulfovibrio*species contain a novel set of genes involved in biotin synthesis.
Regulation of metal-ion homeostasis
-----------------------------------
A number of regulators believed to be involved in metal-ion homeostasis were identified on the basis of orthology with known factors from *E. coli*or *B. subtilis*. However, in almost all cases, with the possible exception of ZUR and ModE in *G. sulfurreducens*, which appear to have signals similar to the *B. subtilis*and *E. coli*consensus respectively, similarity to known binding signals was not observed (Table [11](#T11){ref-type="table"}). The presence of similar sets of target genes in the δ-proteobacteria studied allowed us to apply the signal detection procedure to elucidate novel regulatory signals, to expand core regulons, and to observe species-specific differences in regulation. Interestingly, the FUR/ZUR/PerR family of transcriptional regulators was found to be ubiquitous in these bacteria and responsible for a broad range of functions including iron and zinc homeostasis as well as oxidative stress response. In some cases, multiple paralogous factors were found, perhaps indicating previously uncharacterized functions for this versatile gene family.
The large number of iron-containing proteins predicted from the genome sequence of these organisms, and their ability to use ferric iron anaerobically as a terminal electron acceptor, makes iron homeostasis a key target for analysis. A number of new genes were identified that may belong to the FUR regulon of these organisms. First, uncharacterized porins with upstream FUR boxes were identified in the *Geobacter*and *Desulfuromonas*genomes, which we speculate might be involved in iron transport. Additionally, a two-domain protein with no homologs of known function was identified in all species except *D. psychrophila*. In *G. sulfurreducens*, this gene occurred downstream of another gene with a cytochrome-type heme-binding motif, while in *Desulfuromonas*it was divergently transcribed with a ferric reductase, and was associated with a tetratricopeptide repeat protein in the *Desulfovibrio*genomes. In both *Desulfovibrio*species, we identified an additional regulon, possibly under FoxR control, which might be involved in siderophore transport. This finding was particularly surprising because we did not identify any known siderophore biosynthetic pathway. A possible explanation is that these bacteria use a novel siderophore biosynthesis pathway, or alternatively, take up siderophores released by other bacteria in the environment.
Stress response
---------------
Oxidative stress is one of the most common environmental stressors for these organisms, especially in the metal-contaminated sites of interest for bioremediation. The bacteria in this study are unusual in that they contain both the aerobic superoxide dismutase (Sod)/catalase-type oxidative response as well as the anaerobic Sor/rubrerythrin-type response as previously noted for *D. vulgaris*\[[@B54]\]. Analysis of the signal peptides in these proteins indicates that the Sod/catalase system acts periplasmically, whereas the Sor/rubrerythrin system acts cytoplasmically \[[@B54]\]. While these organisms have no homologs of the OxyR or SoxRS regulators known to respond to changes in oxygen levels in *E. coli*, they do contain homologs of the PerR regulator of *B. subtilis*, known for its involvement in peroxide stress (Table [11](#T11){ref-type="table"}). Clustering of PerR homologs with oxidative stress genes, as well as their grouping with known *Bacillus*PerR genes in a phylogenetic analysis of the FUR/ZUR/PerR family of transcription factors, allowed the inference that they may, in part, be responsible for the control of the oxidative stress response of these organisms. Although we did not identify conserved regulatory elements for some known oxidative stress genes such as the Rbo/Rub/Roo operon in *Desulfovibrio*species, it has been observed that the Rub/Roo operon of *Desulfovibrio gigas*shows strong constituitive expression from a previously identified σ^70^promoter, indicating that additional factors may not be involved \[[@B55]\].
The heat-shock response of these bacteria was found to be mediated by two regulons previously described in other species (Table [11](#T11){ref-type="table"}). First, the σ^32^regulon was identified, with a consensus signal similar to that characterized for *E. coli*. The second observed regulon was the HrcA/CIRCE regulon known in *B. subtilis*and other bacteria, but not present in *E. coli*. These two regulons include a partially overlapping set of genes. Notably, CIRCE elements were identified in all of the genomes used in this study with the exception of *D. psychrophila*. It is tempting to speculate that the constant and cold temperatures encountered by this species in its environmental niche have removed the need for this particular heat-shock response.
Similarity of regulatory signals with those in other bacteria
-------------------------------------------------------------
Comparison with well studied bacterial model organisms has shown that δ-proteobacteria share regulatory components with both Gram-positive and Gram-negative microorganisms (Table [11](#T11){ref-type="table"}). For example, the use of NikR and ModE for the regulation of, respectively, nickel and molybdenum uptake and utilization is consistent with *E. coli*-like regulation. However, the presence of PerR, CIRCE elements and S-box motifs is reminiscent of *B. subtilis*-like regulation. Moreover, in the case of FUR, although the regulon structure showed overlap with known downstream targets in model organisms, the sequence of the FUR box, which is conserved in both *E. coli*and *B. subtilis*, was observed to be different in the metal-reducing δ-proteobacteria.
We recognize that this is one of the first direct studies comparing entire regulons in δ-proteobacteria. Two recent computational works, considering either a single *D. vulgaris*or two *Geobacter*species, used the AlignACE signal detection program, which is based on a Gibbs-sampling algorithm, to derive large sets of conserved DNA motifs without linking them to specific regulatory systems \[[@B56],[@B57]\]. Unfortunately, the predicted regulatory signals based on single genomes turned out not to be conserved across genomes, and could not be used for functional gene annotation. In this comparative work, we tried to extensively describe a set of biologically reasonable regulons in δ-proteobacteria. The regulatory sites predicted here were not detected in the other two computational studies by Hemme and Wall and by Yan *et al*. \[[@B56],[@B57]\]. Previously published experimental studies of sulfate-reducing δ-proteobacteria have focused mostly on the biochemistry unique to these organisms, and little is known about the regulation of gene expression. In part, this has been due to difficulties in genetically manipulating these strictly anaerobic bacteria. Recent advances in microarray technologies provide genome-scale expression data for *D. vulgaris*under various conditions. In support of our findings, all operons predicted to be co-regulated by the peroxide-responsive regulator PerR in *D. vulgaris*are significantly downregulated by oxygen stress (J. Zhou, personal communication). Furthermore, recent microarray data obtained for *G. sulfurreducens*in iron-limiting conditions confirm our prediction of the FUR regulon in this genome (R. O\'Neil, personal communication).
It is interesting to observe the extent to which regulatory motifs are conserved between δ-proteobacteria. Although riboswitches and some DNA signals (that is, CIRCE, σ^32^and BirA) seem to be conserved across vast spans of evolutionary time, in many cases we observe divergence in binding signals even when the core components of a regulon are conserved (NikR, FUR, PerR, ModE). These findings raise, but do not answer, questions such as what circumstances cause transcription factor binding specificities to change or remain conserved, and whether those changes reflect genetic drift, or active selection to alter the regulatory action of the factor.
Energy metabolism
-----------------
We identified two regulons involved in the control of energy metabolism (Table [11](#T11){ref-type="table"}). The first, controlled by the CooA protein, was present only in the *Desulfovibrio*genomes. It is orthologous to a known regulon in *R. rubrum*, and regulates genes involved in the oxidation of CO. The second regulon is novel and distributed widely among anaerobic and facultatively anaerobic bacteria. The primary downstream target of this newly identified regulator, which we called HcpR\*, is the hybrid-cluster protein Hcp. Upregulation of the *hcp*gene in response to growth on nitrate or nitrite in *Shewanella oneidensis*, *E. coli*and *D. vulgaris*indicates that Hcp is likely to be involved in the utilization of alternative electron acceptors.
Consistent with this hypothesis, we predicted positive regulation of Hcp and the associated ferredoxin FrdX by HcpR, and negative regulation of the sulfate-reduction genes by HcpR in the *Desulfovibrio*genomes, based on the position of the candidate HcpR-binding sites relative to the predicted promoters. Thus, HcpR is predicted to be responsible for switching between alternative electron acceptors during anaerobic respiration in these species. Interestingly, we found an HcpR site upstream of the CO-dependent hydrogenase that was also predicted to be under the control of CooA. This hydrogenase was recently proposed to play a key role in sulfate reduction \[[@B16]\], and it is tempting to speculate that its inclusion in a common regulon with known sulfate-reduction genes supports this hypothesis. The position of the binding site, however, suggests that it activates rather than represses transcription, contrary to predictions for other known sulfate-reduction genes, so its regulation is likely to be complex, and further experiments will be needed to determine whether it plays the role of the cytoplasmic hydrogenase necessary for the proposed \'hydrogen cycling\' of sulfate reduction \[[@B58]\]. The ubiquitous phylogenetic distribution of the HcpR regulon indicates that it has a central role in facilitating an anaerobic life style, yet very little is known about its specific function. We hope our elucidation of the core components and regulator of this important regulon will inspire future experimental studies to determine its cellular role.
Regulatory motifs for alternative cofactor adaptation
-----------------------------------------------------
In the course of this study we identified several cases in which different variants of genes were predicted to be regulated according to the availability of required cofactors or nutrients. Three examples were observed in which an alternative enzyme, not requiring a given cofactor, was repressed by the availability of that cofactor: B~12~-independent ribonucleotide reductase was repressed by the availability of B~12~; \[Fe\] hydrogenase was repressed by the availability of nickel (and presumably replaced by \[NiFe\] hydrogenase); and Fe(II) was predicted to repress a flavodoxin gene which we suspect may be used as an alternative to ferredoxins present in the genome. This mode of regulation for B~12~-independent isozymes of ribonucleotide reductase and methionine synthetase has been previously described \[[@B26]\]. Moreover, a similar regulatory strategy has been reported for one of the alternative superoxide dismutases and for paralogs of ribosomal proteins \[[@B34]-[@B36],[@B38],[@B59]\]. Taken together, these data suggest that this flexible strategy may represent a common theme in the adaptation of bacteria to their environment. Indeed, similar mechanisms may, in part, explain some of the apparent genetic redundancy in many genomes.
Materials and methods
=====================
The genomes of δ-proteobacteria that were analyzed in this study are *Desulfovibrio vulgaris*Hildenborough (DV); *Desulfovibrio desulfuricans G20*(DD); *Geobacter metallireducens*(GM); *Geobacter sulfurreducens PCA*(GS); *Desulfuromonas*species (DA); and *Desulfotalea psychrophila*(DP). Complete genomic sequences of DV and GS were downloaded from GenBank \[[@B60]\]. Draft sequences of DD, GM and DA genomes were produced by the US Department of Energy Joint Genome Institute and obtained from \[[@B61]\]. Draft sequence of the DP genome was provided by the Max Planck Institute for Marine Microbiology in Bremen, Germany \[[@B62]\]. Numerical gene identifiers from the Virtual Institute for Microbial Stress and Survival (VIMSS) Comparative Genomics database \[[@B63]\] are used for hypothetical genes without common names. New gene names introduced in this study are marked by an asterisk.
For *de novo*definition of a common transcription factor-binding signal in a set of upstream gene fragments, a simple iterative procedure implemented in the program SignalX was used \[[@B31]\]. Weak palindromes were selected in each region, and each palindrome was compared to all others. The palindromes most similar to the initial one were used to make a profile. The positional nucleotide weights in this profile were defined as
*W*(*b,k*) = log\[*N*(*b*,*k*) + 0.5\] - 0.25Σ~i\ =\ A,C,G,T~log\[*N*(*i*,*k*) + 0.5\],
where *N*(*b*,*k*) is the count of nucleotide *b*in position *k*\[[@B10]\]. The candidate site score *Z*is defined as the sum of the respective positional nucleotide weights
*Z*(*b*~1~\...*b*~*L*~) = Σ~*k*=\ 1\...*L*~*W*(*b*~*k*~,*k*),
where *k*is the length of the site.
These profiles were used to scan the set of palindromes again, and the procedure was iterated until convergence. Thus a set of profiles was constructed. The profile with the greatest information content \[[@B64]\] was selected as the recognition rule.
Each genome was scanned with the profile using the GenomeExplorer software \[[@B65]\], and genes with candidate regulatory sites in the 300-bp upstream regions were selected. The upstream regions of genes that are orthologous to genes containing regulatory sites were examined for candidate sites even if these were not detected automatically. The threshold for the site search was defined as the lowest score observed in the training set. Sets of potentially co-regulated genes contained genes that had candidate regulatory sites in their upstream regions and genes that could form operons with such genes (that is, located downstream on the same strand with intergenic distances of less than about 100 bp). A complete description of the GenomeExplorer software, including the SignalX program, is given at \[[@B65]\].
The RNApattern program \[[@B66]\] was used to search for conserved RNA regulatory elements (riboswitches) in bacterial genomes. The input RNA pattern for this program describes an RNA secondary structure and sequence consensus motifs as a set of the following parameters: the number of helices, the length of each helix, the loop lengths, and a description of the topology of helix pairs. The latter is defined by the coordinates of helices. For instance, two helices may be either independent or embedded helices, or they could form a pseudoknot structure. This definition is similar to the approach implemented in the Palingol algorithm \[[@B67]\].
Orthologous proteins were identified as bidirectional best hits \[[@B68]\] by comparing the complete sets of protein sequences from the two species using the Smith-Waterman algorithm implemented in the GenomeExplorerprogram \[[@B65]\]. When necessary, orthologs were confirmed by construction of phylogenetic trees for the corresponding protein families. Phylogenetic analysis was carried out using the maximum likelihood method implemented in PHYLIP \[[@B69]\]. Large-scale gene cluster comparisons were carried out using the VIMSS Comparative Genomics database \[[@B63]\]. Multiple sequence alignments were done using CLUSTALX \[[@B70]\]. The COG \[[@B68]\], InterPro \[[@B71]\], and PFAM \[[@B72]\] databases were used to verify the protein functional and structural annotation.
Note added in proof
-------------------
Recently it has been demonstrated by *in vitro*experiment that the glycine-specific riboswitch consists of two tandem aptamer sequences that appear to bind target molecules cooperatively \[[@B73]\]. This indirectly confirms our hypothesis of a cooperative effect of ligand binding to tandem *THI*-elements in *Desulfovibrio*spp. Also we have recently shown that *Geobacter*spp. have a modified HcpR regulon, which uses a signal similar to that found in DA and DP, but contains multiple nitrate/nitrite reductase genes.
Additional data files
=====================
An additional data file (Additional data [1](#s1){ref-type="supplementary-material"}) containing three figures with detailed description of DNA- and RNA-type regulatory sites is available with the online version of this paper and on our website \[[@B74]\].
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
Three figures with detailed description of DNA- and RNA-type regulatory sites
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We are grateful to Elizaveta Permina for the CIRCE and σ^32^-promoter recognition profiles and to Sergey Stolyar and Morgan Price for helpful discussions. This study was partially supported by grants from the Howard Hughes Medical Institute (55000309) (to M.G.), the Russian Fund of Basic Research (04-04-49361) (to D.R.), the Programs Molecular and Cellular Biology and Origin and Evolution of the Biosphere of the Russian Academy of Sciences (to M.G.), and by the US Department of Energy\'s Genomics: GTL program (DE-AC03-76SF00098, to A.P.A.). This study has been done in part during the visit by D.R. to the Lawrence Berkeley National Laboratory, Berkeley, CA, USA.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Genomic organization of the biotin biosynthetic genes and regulatory elements. DV (*Desulfovibrio vulgaris*); DD (*Desulfovibrio desulfuricans G20*); GM (*Geobacter metallireducens*); GS (*Geobacter sulfurreducens PCA*); DA (*Desulfuromonas*species); DP (*Desulfotalea psychrophila*).
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Genomic organization of the thiamin biosynthetic genes and regulatory *THI*-elements (yellow structures). See Figure 1 legend for abbreviations.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Genomic organization of the cobalamin biosynthetic genes and regulatory *B*12-elements (yellow cloverleaf-type structures). Genes of the first part of the pathway, involved in the corrin ring synthesis are shown as yellow arrows, the genes required for the attachment of the aminopropanol arm and assembly of the nucleotide loop in vitamin B~12~are in green. Cobalt transporters and chelatases used for the insertion of cobalt ions into the corrin ring are shown in pink and orange, respectively. ABC transport systems for vitamin B~12~are shown in blue. See Figure 1 legend for abbreviations.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Genomic organization of the methionine biosynthetic genes and regulatory S-boxes (yellow cloverleaf-type structures). See Figure 1 legend for abbreviations.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Genomic organization of the predicted iron-regulated genes and FUR-binding sites (small black rectangles). \*Name introduced in this study. See Figure 1 legend for abbreviations.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
Genomic organization of the nickel-regulated genes and NikR-binding sites (small blue arrows). See Figure 1 legend for abbreviations.
:::
:::
::: {#F7 .fig}
Figure 7
::: {.caption}
######
Genomic organization of predicted zinc ABC transporters and ZUR-binding sites. The black oval and blue box represent two different types of ZUR-binding site. See Figure 1 legend for abbreviations.
:::
:::
::: {#F8 .fig}
Figure 8
::: {.caption}
######
Genomic organization of predicted molybdate ABC transporters and ModE-binding sites (small ovals). The black and blue ovals represent two different types of ModE-binding site. See Figure 1 legend for abbreviations.
:::
:::
::: {#F9 .fig}
Figure 9
::: {.caption}
######
Genomic organization of genes involved in oxidative stress response. Dots of various colors represent predicted PerR-binding sites with different consensus sequences. See Figure 1 legend for abbreviations.
:::
:::
::: {#F10 .fig}
Figure 10
::: {.caption}
######
Maximum-likelihood phylogenetic tree of the FUR/ZUR/PerR family of transcriptional regulators. Consensus sequences of binding sites predicted in this study are underlined. See Figure 1 legend for abbreviations.
:::
:::
::: {#F11 .fig}
Figure 11
::: {.caption}
######
Pairwise sequence alignment of upstream regions of the *perR-rbr-roo*operons from *Geobacter*species. Conserved palindromic signal, that is the candidate PerR-box, is highlighted in gray. Predicted SD-boxes and start codons of the *perR*genes are in bold. Predicted -10 and -35 promoter boxes are underlined. \*Conserved position of alignment. See Figure 1 legend for abbreviations.
:::
:::
::: {#F12 .fig}
Figure 12
::: {.caption}
######
Genomic organization of genes predicted to be regulated by two transcription factors from the CRP/FNR-family. Black circles denote operators for the CO-responsive regulator CooA. Blue circles and squares denote predicted sites of the hypothetical transcriptional factor HcpR with two different consensus sequences, respectively. w, HcpR site with a weak score; \..., a set of gene names that are not shown. See Figure 1 legend for abbreviations.
:::
:::
::: {#F13 .fig}
Figure 13
::: {.caption}
######
Pairwise sequence alignment of upstream regions of the predicted HcpR-regulated operons from *Desulfovibrio*species. **(a)***sat*; **(b)***apsAB*; **(c)***206515-206516*. Candidate HcpR sites are highlighted in gray. Predicted SD-boxes and start codons of the first genes in the operons are in bold. Predicted \'-10\' and \'-35\' promoter boxes are underlined. \*Conserved position of alignment. See Figure 1 legend for abbreviations.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Candidate binding sites for the biotin repressor BirA
:::
Gene Site Position\* Score
---------------------------------- -------- ----------------------------- ------------ -------
*Desulfuromonas*sp.
387978 *bioW* `aTGTcAACC-[N14]-GGTTgACAg` -63 8.61
390011 *bioB* `acGTcAACC-[N14]-GGTTgACAA` -94 8.13
*Geobacter sulfurreducens*PCA
381880 *bioB* `TTGTcAACC-[N14]-aGTTgACAA` -78 8.50
382941 *bioF* `TTGTcAACC-[N14]-GGTTgACgA` -182 8.29
*Geobacter metallireducens*
377241 *bioB* `TTGTtAACC-[N14]-aGTTgACAA` -76 7.81
377542 *bioF* `TTGTcAACC-[N14]-GGTTgACgA` -64 8.29
*Desulfovibrio vulgaris*
208055 *bioB* `TTGTAAACC-[N15]-cGTTgACAg` 6 8.39
*Desulfovibrio desulfuricans*G20
394249 *bioB* `TTGTAAACC-[N15]-aGTTgACAA` -119 8.60
*Desulfotalea psychrophila*
425025 *bioB* `TTGTAAAtt-[N15]-ccaTTACAg` 233 6.19
\*Position relative to the start of translation. Lower case letters represent positions that do not conform to the consensus sequence.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Candidate binding sites for the predicted lysine-specific regulator LysX\*
:::
Gene Site Position^†^ Score
---------------------------------- --------------- ------------------------ ------------- -------
*Desulfovibrio vulgaris*
208064 *lysX\*-lysA* `GTGGTACTAATcAGTACCAC` -277 6.82
206613 *\~mviN\** `GTGGTtCTttgTAGTACtAC` -135 5.45
*Desulfovibrio desulfuricans*G20
394240 *lysX\*-lysA* `GTaGTACTAAaTAGTACCAC` -43 6.70
393213 *lysW\** `GgcGTtCTAAagAGTACCAC` -145 5.88
394397 *\~mviN\** `GTaGTtgTgATaAGaAaCAC` -275 4.70
^†^Position relative to the start of translation. \*New name introduced in this study. Lower case letters represent positions that do not conform to the consensus sequence.
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Candidate binding sites for the ferric uptake regulator FUR
:::
Gene Operon Function Site Position\* Score
---------------------------------- --------------------------- ----------------------------------------------- --------------------- ------------ -------
*Geobacter sulfurreducens*PCA
381665 *Fur* Ferric uptake regulator `ATGAtAtTCAcTTTCAg` -31 5.25
381666 *feoB1 - R* Fe^2+^transporter `cTGAAAgTGATTTTCAc` -192 5.18
383594 *genX\*-genY\** Cytochrome c family protein, putative `gTGAAAAaCATTTTCAa` -65 5.08
383590 *X-feoA-feoA-feoB2* Porin, Fe^2+^transporter `tTGAAAATGgaaTTCAT` -82 5.07
*Geobacter metallireducens*
379927 *Fur* Ferric uptake regulator `tTGAAAATCAcTTTCAg` -30 5.54
379928 *feoB1 - R* Fe^2+^transporter `tTGAAAgTGAaTaTCAa` -48 5.33
378774 *psp\** Porin? `tTGAAAAaGAcTTTCAT` **-259** 5.28
`ATGAAtATGAaTTTCAa` **-160** 5.35
*Desulfuromonas*species
392427 *fur2-feoB1 - R* Fe regulator, Fe^2+^transporter `tTGAAAATCATTTTCAg` -34 5.72
390939 *psp\** Porin? `tTGAtAATGgcTTTCAT` **-139** 5.22
`cTGAAAAcGATTTTCAT` **-86** 5.46
391943 *fur1* Ferric uptake regulator `tTGAAcATCATTTTCAT` -37 5.44
387887 *feoA-feoB4* Fe^2+^transporter `ATGAAAAcGAaTTTCAT` 93 5.43
`tTGAtAAaGAcTTTCAT` 39 5.12
391875 *genY\*(N)* `tTGAAAAcGgTTTTCAT` -105 5.28
389803 *feoA-feoB2* Fe^2+^transporter `cTGAAAAcCgTTTTCAa` -39 5.16
392265 *feoA-feoB3* Fe^2+^transporter `ATGAAAtaCAcTTTCAa` -54 5.13
*Desulfovibrio vulgaris*
209207 ? `tTGAAAATtATTTTCAa` **-35** 5.42
`ATtAtttTCAaTaTCAg` **-29** 4.06
206189 *gdp\** GGDEF domain protein `tTGActtTGAaaaTCAT` **-36** 4.04
`tTGAAAATCATaaTCAa` **-30** 5.32
208071 *feoA-feoA-feoB* Fe^2+^transporter `ATaAActTGAcaaTCAT` **-99** 3.91
`tTGAcAATCATTTTCAT` **-93** 5.18
207866 *foxR-pqqL\*-atpX\*-\...* Regulator, Zn-dependent peptidase, ABC operon `tTGActtTGATTTTCAc` **-195** 4.31
`tTGAtttTCAcTTTCAT` **-189** 5.01
209238 *genY\*(C)-genZ\** ? `tTGAcAtTGATTTTCgT` **-55** 4.31
`tTGAtttTCgTTTTCAa` **-49** 4.89
208179 *fld\** Flavodoxin `tTGAAAAcaAaaaTCAa` **-182** 4.49
`AcaAAAATCAaTTTCAa` **-176** 4.25
208641 *hdd\** HD-domain protein `tTGAcAATGATTTTCtT` **-93** 4.46
`ATGAtttTCtTTTTCAa` **-87** 4.81
208856 Has P-type ATPase/hydrolase domains `tTGAtttaGATTTTCAa` **-87** 4.79
`taGAtttTCAaTTTCAg` **-81** 4.20
`tTcAAttTCAgTaTCAa` **-75** 3.82
*Desulfovibrio desulfuricans*G20
395878 *fur3* Ferric uptake regulator `ATGAAAATaATTTTCAT` -77 5.46
393004 *pqqL\*-atpX\*-\...* Zn-dependent peptidase, ABC operon `ATGAAAATaAaTTTCAT` **-54** 5.31
`ATaAAttTCATTTTCAT` **-48** 4.65
392971 *392971-70-69* MoxR-like ATPase, CoxE-like protein `cTGAAAtTGgTTTTCAa` **-99** 5.29
`tTGgtttTCAaTaTCAg` **-93** 4.24
`tTGAAAATGAaaTTtAT` **-30** 4.63
`ATGAAAtTtATagTCAg` **-24** 4.19
393146 *genY\*(C)-genZ\** ? `tTGAcAtTGATTTTCAT` **-84** 5.03
`tTGAtttTCATTTTCAc` **-78** 4.81
393462 *fld\** Flavodoxin `tTGAcAATGAaTTTCAT` **-263** 5.03
`ATGAAttTCATTTTCAc` **-257** 4.99
394236 *feoA-feoB* Fe^2+^transporter `ATGAgAAgGATTTTCAa` **-83** 5.00
`AgGAtttTCAaTTTCAc` **-77** 3.96
394235 *feoA3* Fe^2+^transporter `AgGAActTGAcaaTCAT` **-60** 3.91
`tTGAcAATCATTcTCAT` **-54** 4.72
393956 *gdp\** GGDEF domain protein `tTGAtttTGAgTTTCAT` **-122** 4.56
`tTGAgttTCATaTTCAT` **-116** 4.55
395154 *FoxR* AraC-type regulator `tTGAcAtTGAaaaTCAT` **-189** 4.38
`tTGAAAATCATTTTCgc` **-183** 4.74
394231 *pep\*-fur1* Zn-dependent peptidase, Fe regulator `tTcAgAcTGgTTTTCAT` **-281** 3.75
`cTGgtttTCATTaTCAT` **-275** 4.41
395541 *hdd\** HD-domain protein `gTGAtAtTGAaaTTCtT` **-105** 3.96
`tTGAAAtTCtTTaTCgc` **-99** 4.05
395164 *fepA-feoA2-feoB2* Outer membrane receptor, Fe-transporter `cTGAtAAaGAaacTCAc` **105** 3.87
`AaGAAAcTCAcTaTCAg` **111** 4.05
\*Position relative to the start of translation. Lower case letters represent positions that do not conform to the consensus sequence. Multiple tandem sites in one regulatory region are shown in bold.
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Candidate binding sites for the nickel regulator NikR
:::
Gene Operon Function Site Orientation Position\* Score
------------------------------- --------------- ---------------------------------------- ------------------------- ------------- ------------ -------
*Geobacter sulfurreducens*PCA
381565 *nik(MN)QO\** Nickel transporter `GTGTTAC-[N14]-GTgACAC` →← -183 5.00
*Geobacter metallireducens*
379930 *nik(MN)QO\** Nickel transporter `GTGTTAC-[N13]-GTAACAC` →← -63 5.22
*Desulfuromonas*species
387207 *nikQO\** Nickel transporter `GTGccAC-[N13]-GTAACAC` →← -41 4.67
*Desulfovibrio vulgaris*
206492 *nikMQO\** Nickel transporter `GTGTTAt-[N13]-GTAACAC` →← -120 5.00
208275 *nikK\** Additional component of Ni transporter `GTgACAC-[N13]-GTGTaAC` ←→ -84 4.49
*Desulfovibrio desulfuricans*
395510 *nikKMLQO\** Nickel transporter `GTGTTAt-[N13]-GTAACAC` →← -104 5.00
394565 *hydAB* Periplasmic Fe-only hydrogenase `GTaTTAC-[N13]-GTAACAC` →← -83 4.67
*Desulfotalea psychrophila*
422915 *nikMLKQO\** Nickel transporter `GTAACAC-[N13]-GTGTTAC` ←→ -20 5.22
422176 422176-177 ? `GTAACAC-[N13]-GTGTTAC` ←→ -197 5.22
`GTAACAC-[N13]-GTGTTAC` ←→ -124 5.22
\*Position relative to the start of translation.
:::
::: {#T5 .table-wrap}
Table 5
::: {.caption}
######
Candidate binding sites for the zinc regulator ZUR
:::
Gene Operon Function Site Position\* Score
------------------------------- ------------------ --------------------------------- --------------------------- ------------ -------
*Geobacter sulfurreducens*PCA
383303 *zur\_Gs-znuABC* Zinc ABC transporter, regulator `TAAAtgGAAATgATTTCtgTTTA` -40 5.32
*Desulfovibrio vulgaris*
206785 *znuABC-zur\_D* Zinc ABC transporter, regulator `ATGCAACagtGTTGCAT` -216 6.65
*Desulfovibrio desulfuricans*
394629 *znuABC-zur\_D* Zinc ABC transporter, regulator `ATGCAACtgaGTTGCAT` -47 6.65
\*Position relative to the start of translation. Lower case letters represent positions that do not conform to the consensus sequence.
:::
::: {#T6 .table-wrap}
Table 6
::: {.caption}
######
Candidate binding sites for the molybdate regulator ModE
:::
Gene Operon Function Site Position\* Score
-------------------------------- ----------------- ------------------------ -------------------------------- ------------ -------
*Geobacter sulfurreducens PCA*
383279 *modDABC* Molybdate transport `ATCGTTATgTcaTgAAggtTATAGCGtT` -158 5.16
*Desulfovibrio vulgaris*
209110 *modA* Molybdate transport `CGGTCACG-[N14]-gGTGACCG` -131 5.56
209114 *modBC* Molybdate transport `CGGTCACc-[N14]-CGTGACCa` -218 5.38
*Desulfovibrio desulfuricans*
393254 *modAB2-393256* Molybdate transport, ? `CtGTCACG-[N14]-CGTGACCG` -183 5.56
393587 *modAB1-modC* Molybdate transport `ttGTCACG-[N14]-CGTGACCG` -119 5.38
\*Positionrelative to the start of translation. Lower case letters represent positions that do not conform to the consensus sequence.
:::
::: {#T7 .table-wrap}
Table 7
::: {.caption}
######
Candidate binding sites for the peroxide-responsive regulators PerR
:::
Gene Operon Function Site Position\* Score
------------------------------- -------------------- --------------------------------------------------- ----------------------- ------------ -------
*Desulfovibrio vulgaris*
207805 *rbr2* Rubrerythrin `AATAGGAATCGTTCCTGTT` -46 5.97
208612 *perR-rbr-rdl* PerR-like repressor, rubrerythrin, rubredoxin `AtCAGTAATtGTTACTGgT` -36 5.50
207732 *ahpC* Alkyl hydroperoxide reductase C `cACAGGAATGATTCCTGTT` -116 5.40
206199 ? `AtCAGTAATaGTTAtTGTT` -124 5.39
*Desulfovibrio desulfuricans*
395420 *rbr2* Rubrerythrin `AATAGGAATCGTTACTGaT` -76 5.91
395549 ? `AATAaGAATtGTTACTATT` -134 5.45
393457 *perR* PerR-like repressor `ttTAGGAATGGTTAtTATT` -41 5.23
*Desulfotalea psychrophila*
423938 *roo1-roo2* Rubredoxin-oxygen oxidoreductase `GTTAATGATAATCATTAct` -203 6.25
425393 *perR* PerR-like repressor `GaTAATttTTATtATTAAC` -74 5.97
*Geobacter sulfurreducens*
383613 *perR-rbr\*-roo* Rubredoxin-oxygen oxidoreductase `AaTGCAATAAAATACCAAT` -99 6
*Geobacter metallireducens*
378323 *perR2-rbr\*-roo* Rubredoxin-oxygen oxidoreductase `ATTGCAATAAAgTACCAAc` -99 5.79
*Desulfuromonas*species
387528 *katG1* Catalase ` GGTcTTGACAATtCC` -75 5.55
387530 *perR31* PerR-like repressor ` GaTATTGACAAacCC` -96 5.29
*Geobacter sulfurreducens*
383124 *hsc-grx-ccpA-rbr* Cytochrome peroxidase, glutaredoxin, rubrerythrin `TTGCGCATTCcATtCGTAA` -32 5.84
*Desulfuromonas*species
390120 *perR1-rbr* PerR-like repressor, rubrerythrin `TTGCGCgTTAAAacaGTAA` -91 5.54
\*Position relative to the start of translation. Lower case letters represent positions that do not conform to the consensus sequence.
:::
::: {#T8 .table-wrap}
Table 8
::: {.caption}
######
Candidate CIRCE sites for the heat shock-responsive regulator HrcA
:::
Gene Operon Site Position\* Score
------------------------------- ----------------------- ---------------------------- ------------ -------
*Desulfovibrio vulgaris*
207448 *groESL* `cTgGCACTC-[N9]-GAGTGCcAA` -68 6.53
*Desulfovibrio desulfuricans*
394393 *groESL* `TTgGCACTC-[N9]-GAGTGCTAA` -70 7.15
*Geobacter sulfurreducens*
380317 *hrcA-grpE-dnaK-dnaJ* `TTAGCACTC-[N9]-GAGTGCTAA` -49 7.50
380945 *rpoH* `TTAGCACTC-[N9]-GAGTGCTAA` -51 7.28
383663 *groESL* `TTAGCACTC-[N9]-GAGTGCTAA` -81 7.45
*Geobacter metallireducens*
379288 *groESL* `TTAGCACTC-[N9]-GAGTGCTAA` -80 7.41
379629 *hrcA-grpE-dnaK-dnaJ* `TTAGCACTC-[N9]-GAGTGCTAA` -45 7.29
*Desulfuromonas*species
387711 *hrcA-grpE-dnaK-dnaJ* `TTAGCACTC-[N9]-GAGTGCTAA` -85 7.06
389722 *groESL* `TTAGCACTC-[N9]-GAGTGCTAA` -99 7.20
\*Position relative to the start of translation. Lower case letters represent positions that do not conform to the consensus sequence.
:::
::: {#T9 .table-wrap}
Table 9
::: {.caption}
######
Candidate σ^32^-dependent promoters upstream of heat-shock genes
:::
Gene Operon Site Position\* Score
------------------------------- ------------------- ------------------------ ------------ -------
*Desulfovibrio vulgaris*
206437 *dnaJ-?-clpA* `gaTGAAt-[N15]-CCCCtT` -114 5.43
206776 *?-clp* `gTTGttg-[N15]-CCCCgT` -196 5.28
207035 *rpoH* `aTTGAAA-[N12]-aaCtAT` -110 5.71
207448 *groESL* `CaTaAAA-[N12]-CCCCtT` -239 5.23
*Desulfovibrio desulfuricans*
394616 *clpP-clpX-lon* `CTTGAAc-[N12]-CCCgAT` -82 6.45
394617 *clpX* `CTTGAAA-[N14]-aCCgAT` -136 6.94
394712 *rpoH* `aTTGAAA-[N12]-aaCtAT` -122 5.71
395109 *dnaJ-?-clpA* `CTTGAAA-[N13]-gaCggT` -81 5.16
`gTTGcAg-[N12]-CCgCAT` -57 5.28
395651 *dnaK* `CTcGAAA-[N14]-CCgCAg` -71 5.17
*Desulfotalea psychrophila*
422219 *groESL* `aTTGAAA-[N13]-CCCCtT` -201 6.33
`CTTGAtt-[N13]-aCCtAT` -134 5.98
423932 *grpE-dnaK* `CaTGAAc-[N12]-CtCCAT` -232 5.34
`CTTGAcA-[N13]-aCttAT` -135 5.67
424328 *dnaJ* `gTTtAcA-[N14]-gCCCAT` -113 5.62
`CTTGAct-[N14]-CCCtAa` -40 5.67
425016 *?-clpP-clpX-lon* `tTTGAtA-[N11]-CCCaAg` -123 5.33
*Geobacter sulfurreducens*
380319 *dnaK-dnaJ* `gTTGAgg-[N14]-CCCaAT` -208 6.05
382089 *?-clpP-clpX-lon* `gTTcAAA-[N12]-CCCCAT` -283 6.65
382697 *htpG* `CTTGAAA-[N11]-CatgAT` -75 5.85
*Geobacter metallireducens*
379288 *groESL* `gaTGAAA-[N12]-aCtCAT` -45 5.79
379647 *clpA* `CTTGAct-[N14]-gCCtAT` -58 5.72
379699 *?-clpP-clpX-lon* `gTTcAAA-[N13]-CCCaAT` -280 5.96
*Desulfuromonas*species
388073 *clpP-clpX-lon* `CTTGAAg-[N14]-gCCaAT` -203 6.41
`aTTGAAg-[N14]-aCCtAT` -110 6.20
389722 *groESL* `gTTGAgA-[N14]-CCCCtT` -163 5.91
\*Position relative to the start of translation. Lower case letters represent positions that do not conform to the consensus sequence.
:::
::: {#T10 .table-wrap}
Table 10
::: {.caption}
######
Candidate binding sites for the CO-responsive regulator CooA and the FNR/CRP-like HcpR factor
:::
Gene Operon Function Site Position\* Score
------------------------------- ----------------------------- ---------------------------------------------------------------------------------------- ---------------------- ------------ -------
**(a)**CooA regulon
*Desulfovibrio vulgaris*
207573 *cooSC* CO dehydrogenase (CODH) `TGTCGGCTAGCCGACA` -187 6.04
207772 *cooMKLXUHXF* CODH-associated hydrogenase `gGTCGGtcAaCCaACt` -64 4.43
*Desulfovibrio desulfuricans*
393975 *cooSC* CO dehydrogenase (CODH) `TGTCaGCcAGCCGACA` -111 5.78
**(b)**HcpR regulon
*Desulfovibrio vulgaris*
208467 Two-component response regulator `TTGTGAcATgTaTaACAA` -74 5.61
206736 *sat* ATP sulfurylase `TTGTaAAtTtTTTCACAA` -148 5.53
206272 *apsAB* APS reductase `TTGTtAAtTccaTCACAA` -168 5.29
209106 *phcAB* Putative thiosulfate reductase `aTGTGAcgcATTTCgCAA` -194 5.06
207772 *cooMKLXUHXF* CODH-associated hydrogenase `TTGgGAAtcgaTTCACAA` -116 4.97
208738 208738-208737 Two-component regulatory system `cTGTGAAAcATgTCgCAt` -104 4.88
206515 206515-206516 Putative sulfite/nitrite reductase, polyferredoxin `gTGTGAcccgcgTCACAg` -52 4.79
209119 Hypothetical protein conserved in Archaea `TTGTtcAcaAaaTCACAA` -218 4.61
208040 *hcp-frdX-adhE*-208043 Hybrid cluster-containing protein, ferredoxin, alcohol dehydrogenase, histidine kinase `aTtTGAcgcAcgTCACAA` -179 4.55
*Desulfovibrio desulfuricans*
392869 209119 Hypothetical protein conserved in Archaea `TTGTtAAATAaTTCACAA` -118 5.93
395578 *apsAB* APS reductase `TTGTtAAATATcTCACAA` -186 5.77
394579 *sat* ATP sulfurylase `TTGctAAAaATTTCACAA` -147 5.43
`TTGTtAcAatTaTCACAt` -328 4.93
393955 Two-component response regulator `TTGTGAcAgcTgTCACAA` -80 5.36
393201 Two-component response regulator `TTGTGAAggAaaTaACAA` -18 5.29
392939 \~ 6-aminohexanoate-cyclic-dimer hydrolase `TTGTtAAtTATTTaAaAA` -61 5.00
395499 395499-395498-395497-395496 Arylsulfatase, thioredoxin, thioredoxin reductase, sulfate transporter homolog `aTGTGAAAaAcaTCACAt` -129 4.98
393758 393758-..-393776 Large gene cluster encoding carboxysome shell proteins, aldehyde dehydrogeanses, \... `TTGTtAtATtTTTCtCAA` -148 4.97
394469 394469-394470 Putative sulfite/nitrite reductase, polyferredoxin `aTGTGAccTgcaTCACAg` -81 4.86
394261 *hcp-frdX-uspA* Hybrid cluster-containing protein, ferredoxin, universal stress protein UshA `TTGTGActccggTCACAt` -152 4.81
395604 *phcAB* Putative thiosulfate reductase `TTGTGcttTtTTgCACAA` -114 4.25
*Desulfotalea psychrophila*
425344 *frdX* Ferredoxin `ATTTGAtCTAGGTCAAAg` -103 5.81
423439 *hcp3/hcp2* Hybrid cluster-containing proteins `ccTTGACCTgGGTCAAtT` -200 5.47
422894 *hcp1* Hybrid cluster-containing protein `tcTTGACtTAGGTCAAAg` -117 5.44
*Desulfuromonas*species
389812 *hcp1/?-frdX2-?* Hybrid cluster-containing protein/ferredoxin `ATTTGACCTcGGTCAAga` -155 5.66
`AcaTGACgcAGaTCAAAa` -200 4.87
389024 *hcp3* Hybrid cluster-containing protein `tcTTGAtCTgGaTCAAAT` -85 5.45
391271 *dnrA* \~ Regulator of NO signaling `cTTTGACCcgGGTCAAtT` -109 5.44
390920 *hcp2* Hybrid cluster-containing protein `ATTTGACCTgGGTCAtgT` -127 5.40
390344 *galE* \~ Nucleoside-diphosphate-sugar epimerase `ATTTGACCccGGTCAAta` -117 5.39
392163 *yccM* Polyferredoxin `AaaTGACCcAGGTCAAAg` -80 5.14
392663 Two-component response regulator `AaTTGAttcAGGTCAAgg` -85 5.06
390999 Cytochrome c (heme-binding protein) `ATTTGACggccGTCAAAg` -83 5.02
390998 *frdX1* Ferredoxin `tTTTGAtgccGGTCAAgg` -96 5.00
388470 *hcp4* Hybrid cluster-containing protein `tTTTGAttTgtaTCAAtT` -126 4.66
\*Position relative to the start of translation. **(a)**Candidate sites of the CO-responsive regulator CooA in *Desulfovibrio*species; **(b)**candidate sites of the FNR/CRP-like HcpR factor regulating energy metabolism. Lower case letters represent positions that do not conform to the consensus sequence.
:::
::: {#T11 .table-wrap}
Table 11
::: {.caption}
######
Summary of predicted regulatory sites in δ-proteobacteria
:::
Regulator Regulon Consensus Genomes
------------------ ----------------------------------------------------- -------------------------------- ------------------------
BirA Biotin biosynthesis `TTGTAAACC-[N14/15]-GGTTTACAA` DD, DV, GM, GS, DA, DP
*RFN*riboswitch Riboflavin biosynthesis see Additional data files DP
*THI*riboswitch Thiamin biosynthesis see Additional data files DD, DV, GM, GS, DA, DP
*B*12 riboswitch Cobalamin biosynthsis and transport see Additional data files DD, DV, GM, GS, DA, DP
S-box riboswitch Methionine biosynthesis see Additional data files GM, GS, DA
LysX Lysine biosynthesis and transport `GTgGTaCTnnnnAGTACCAC` DD, DV
Fur Iron uptake and metabolism `GATAATGATnATCATTATC` DD, DV, GM, GS, DA
NikR Nickel uptake and metabolism `GTGTTAC-[N13/14]-GTAACAC` DD, DV, GM, GS, DA, DP
Zur Zinc uptake `ATGCAACnnnGTTGCAT` DD, DV
`TAAATCGTAATnATTACGATTTA` GS
ModE Molybdate uptake and metabolism `cgGTCACg-[N14]-cGTGACCg` DD, DV
`atCGnTATATA-[N6]-TATATAnCGat` GS
PerR Peroxide stress response `AwnAGnAAtngTTnCTnwT` DD, DV
`TtnCgnnTTnAAnncGnAA` DA, GS
`AatTGnnATnnnATnnCAatt` GM, GS-2
`GtTAATgATnATcATTAaC` DP
`GgnnTTGnCAAnncC` DA-2
HrcA Heat-shock response `TTAGCACTC-[N9]-GAGTGCTAA` DD, DV, GM, GS, DA
Sigma-32 Heat-shock response `CTTGAAA-[N11/16]-CCCCAT` DD, DV, GM, GS, DA, DP
CooA CO dehydrogenase `TGTCGGCnnGCCGACA` DD, DV
HcpR Sulfate reduction and energy metabolism (prismanes) `TTGTGAnnnnnnTCACAA` DD, DV
`atTTGAccnnggTCAAat` DA, DP
DV (*Desulfovibrio vulgaris*); DD (*Desulfovibrio desulfuricans G20*); GM (*Geobacter metallireducens*); GS (*Geobacter sulfurreducens PCA*); DA (*Desulfuromonas*species); DP (*Desulfotalea psychrophila*). Lower case letters represent positions that do not conform to the consensus sequence.
:::
|
PubMed Central
|
2024-06-05T03:55:51.825607
|
2004-10-22
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545781/",
"journal": "Genome Biol. 2004 Oct 22; 5(11):R90",
"authors": [
{
"first": "Dmitry A",
"last": "Rodionov"
},
{
"first": "Inna",
"last": "Dubchak"
},
{
"first": "Adam",
"last": "Arkin"
},
{
"first": "Eric",
"last": "Alm"
},
{
"first": "Mikhail S",
"last": "Gelfand"
}
]
}
|
PMC545782
|
Background
==========
Carbon and nitrogen are two major macronutrients required for plant growth and development. Specific carbon and nitrogen metabolites act as signals to regulate the transcription of genes encoding enzymes involved in many essential processes, including photosynthesis, carbon metabolism, nitrogen metabolism, and resource allocation \[[@B1]-[@B5]\]. For example, studies have shown that carbon sources (for example, glucose or sucrose) affect the expression of genes involved in nitrogen metabolism, including genes encoding nitrate transporters and nitrate reductase \[[@B6],[@B7]\]. Conversely, nitrogen sources (such as nitrate) have been shown to affect the expression of genes involved in carbon metabolism, including genes encoding PEP carboxylase and ADP-glucose synthase \[[@B8]\]. Responses to carbon and nitrogen result in important changes at the growth/phenotypic level as well. For example, carbon and nitrogen treatments have antagonistic effects on lateral root growth \[[@B9]\], while their effect on cotyledon size, chlorophyll content and endogenous sugar levels appear to be synergistic \[[@B10]\].
In plants, there are multiple carbon-responsive signaling pathways \[[@B11]-[@B13]\], and progress has been made in uncovering parts of the sugar-sensing mechanisms in plants, including the identification of a putative glucose sensor, hexokinase \[[@B14]\]. However, our current knowledge of the mechanisms by which genes and biological processes are regulated by carbon signaling in plants and how they are regulated at the level of transcription is still limited. For example, a search of the PlantCare \[[@B15],[@B16]\] and TRANSFAC \[[@B17]\] databases revealed only seven plant *cis*elements that have been shown to be carbon-responsive *cis*elements (C-elements) and none has been identified from studies in *Arabidopsis thaliana*. Although much less is known concerning the mechanisms controlling nitrogen signaling, microarray analysis has been used to identify nitrogen-responsive genes \[[@B8],[@B18]\]. It has recently been proposed that glutamate receptor 1.1 (AtGLR1.1) functions as a regulator of carbon and nitrogen metabolism in *A. thaliana*\[[@B19]\], but a global understanding of the genes and processes that are regulated by carbon and nitrogen signaling in plants and the mechanism by which this occurs is still lacking.
Previously, microarrays were used to identify genes and biological processes regulated by interactions between carbon and light signaling in *A. thaliana*, including the identification of a putative *cis*regulatory element that is responsive to either light or carbon signals \[[@B13]\]. In this study, we present a genome-wide analysis of the effects of transient carbon and/or nitrogen treatments on mRNA levels, with a particular focus on genes whose mRNA levels are affected by the carbon and nitrogen (CN) treatment. This study has enabled us to evaluate a number of models for intersections between carbon and nitrogen signaling (Figure [1](#F1){ref-type="fig"}) and to identify genes and biological processes that are regulated by the interactions between carbon and nitrogen signaling pathways. In addition, we have identified putative *cis*elements that may be responsible for coordinating a gene\'s responses to both these signaling pathways.
Results
=======
Testing models of carbon and nitrogen regulation
------------------------------------------------
The goal of this study was to use a genomic approach to test the hypothesis that carbon and nitrogen signaling pathways interact to regulate the expression of genes in *Arabidopsis*. We predicted six general models that could describe the possible modes of gene regulation due to carbon, nitrogen and CN together. Three of these models do not involve interactions between carbon and nitrogen signaling. The \'No effect\' model includes genes not regulated by carbon, nitrogen and/or CN. The \'C-only\' model includes genes regulated only by carbon. Finally, the \'N-only\' model includes genes regulated only by nitrogen. Three additional models are needed to describe the regulation of genes affected by interactions between carbon and nitrogen signaling (Figure [1a](#F1){ref-type="fig"}). Model 1 (CN independent) depicts a gene *W*, for which carbon and nitrogen signals act as independent pathways, so that the effects of carbon and nitrogen are additive. Model 2 (CN dependent) depicts a gene *X*, for which regulation requires carbon and nitrogen, and neither carbon alone nor nitrogen alone has an effect. Model 3 (CN dependent/independent) incorporates both an independent and a dependent component to the interactions of carbon and nitrogen signaling. For gene *Y*, carbon alone has an independent inductive effect, while nitrogen has a carbon-dependent effect as it can enhance the effect of carbon, but has no effect on its own (Model 3 CN-enhanced). For gene *Z*, nitrogen alone has an independent inductive effect, while carbon has a nitrogen-dependent effect. These general models can be broken down into more descriptive sub-models. For example, Model 2 can be broken into two sub-models for which CN results in either an inductive or repressive effect.
To test the *in vivo*significance of the above models, a microarray analysis of RNA from plants treated transiently with distinct carbon and nitrogen treatments was carried out, and the results were analyzed to determine the carbon and nitrogen regulation of different genes. For this study, we analyzed RNA isolated from *Arabidopsis*seedlings exposed to four different transient carbon and/or nitrogen treatments (-C/-N, +C/-N, -C/+N, and +C/+N) (Figure [2](#F2){ref-type="fig"}) using Affymetrix whole-genome microarray chips. Analysis of gene expression across these treatments was performed on the whole genome using InterAct Class \[[@B13],[@B20]\], an informatic tool that enabled us to classify genes into each of the above models based on their relative responses to carbon and/or nitrogen treatments. The analysis of the microarray data with InterAct Class enabled us to group genes whose relative responses to carbon, nitrogen and CN were similar to each other. In this case, each InterAct class is made up of four values listed in the following order: value 1 = the expression due to carbon; value 2 = the expression due to nitrogen, value 3 = the expression due to carbon and nitrogen supplied as a combined treatment (CN); and value 4 = the synthetic expression of C+N calculated by adding the expression due to carbon plus the expression due to nitrogen, which is a \'virtual\' treatment.
InterAct Class is a ranking system used to qualitatively compare gene-expression profiles across multiple treatments. For each gene, each treatment is assigned a value representing the effect of the treatment on the expression of that gene. Treatments that result in repression of a gene are assigned a negative number, treatments that do not significantly affect a gene are assigned zero, and treatments that cause induction are assigned a positive number. If more than one treatment causes induction or repression, the treatments are ranked so that the treatment that causes the most induction or repression will be assigned the number furthest from zero. The four hypothetical genes in Figure [1a](#F1){ref-type="fig"} (*W*, *X*, *Y*and *Z*) were classified by InterAct Class (Figure [1b](#F1){ref-type="fig"}), demonstrating that, with this program it becomes easy to determine whether the regulation of a gene is due to a complex (non-additive) interaction between carbon and nitrogen signaling. For such genes, the value assigned to CN (the third InterAct Class number) will be higher or lower than the value assigned to C+N (the fourth InterAct Class number). These genes will fall into Models 2 and 3 (Figure [1b](#F1){ref-type="fig"}, genes *X*, *Y*and *Z*).
Out of 23,000 genes on the Affymetrix chip, 3,652 passed our stringent filtering criteria for reproducibility among treatment replicates and were assigned an InterAct class. Our subsequent analysis of the expression patterns of these 3,652 genes validated the existence of 60 different InterAct classes (Table [1](#T1){ref-type="table"} and Additional data file 1). These 60 InterAct classes represent a broad spectrum of expression patterns that validate each of the six general models for gene regulation. This analysis shows that of the 3,652 genes in the analysis, the vast majority (2,485) is responsive to carbon and/or nitrogen treatment. Moreover, almost half of these genes (1,175 genes) are regulated by an interaction between carbon and nitrogen signaling (Table [1](#T1){ref-type="table"}). For example, there are 175 genes that are in Model 3 CN-enhanced, for which expression due to CN is greater than expression due to C+N (Table [1](#T1){ref-type="table"} and Additional data file 1). This suggests that an interaction between carbon and nitrogen signaling affects the expression of this set of genes.
MIPS funcat analysis uncovers biological processes that are regulated by carbon and/or nitrogen
-----------------------------------------------------------------------------------------------
The InterAct classes were assigned to one of the six general models. To identify biological processes that contain a significant number of genes regulated by carbon, nitrogen and/or CN, we determined which Munich Information Center for Protein Sequences (MIPS) functional categories (funcats) \[[@B21],[@B22]\] were statistically under-represented in the No effect model (InterAct class 0000), compared to all the genes assigned an InterAct class (Table [2](#T2){ref-type="table"}) (not to all the genes in the genome; this takes into account any bias that may have occurred as a result of the filtering process before InterAct class analysis). Under-representation of a biological process in the No effect model means that for that particular funcat, there are fewer genes in the No effect model than expected on the basis of how all the genes assigned to an InterAct class behave. This means that processes under-represented in the 0000 InterAct class contain a significant number of genes that respond to carbon and/or nitrogen treatments compared to the general population of genes in the analysis.
For example, 31.6% (1,089/3,447) of the genes assigned to an InterAct class and a funcat are assigned to the No effect model (Table [2](#T2){ref-type="table"}). This percentage was used as a basis of comparison to determine if genes in any specific funcat varied significantly from the general population. For example, if genes in the metabolism funcat are not regulated by carbon and/or nitrogen in a significant fashion, the number of genes expected to be in the No effect model would be equal to the total number of genes in the metabolism funcat that are assigned an InterAct class (496) times 0.316, which would equal 156.7 genes. However, the actual number of metabolism genes in the No effect model is 120, which is significantly less than 156.7 (*p*-value = 6.0 × 10^-4^). Therefore, the metabolism funcat is under-represented in the No effect model, showing that metabolism displays significant regulation by carbon and/or nitrogen. This analysis revealed several primary funcats (01 = metabolism, 02 = energy and 05 = protein synthesis) that are significantly under-represented in the No effect model (Table [2](#T2){ref-type="table"}). Thus, a significant number of genes involved in metabolism, protein synthesis and energy respond to carbon, nitrogen and/or CN.
For the funcats that are under-represented in the No effect model, this type of analysis was extended to examine the regulation of these funcats in all of the sub-models. This analysis enabled us to determine into which sub-models the genes from these funcats fell and to determine whether the genes in these funcats are under- and over-represented (-S and +S respectively) in these sub-models (Table [3](#T3){ref-type="table"}) (see Additional data file 1 for the *p*-value, and the funcat analysis extended to every sub-model and every funcat).
Identification of *cis*elements associated with CN-regulated genes
------------------------------------------------------------------
To begin to elucidate the mechanisms that control gene regulation in response to carbon and nitrogen treatments, we sought to identify putative *cis*elements that might be responsible for regulating genes in Model 3 CN-enhanced (Table [1](#T1){ref-type="table"}). These genes are likely to contain *cis*elements involved in interactions between carbon and nitrogen signaling because the expression due to CN is greater than that due to C+N. Previously, genes that are biologically related and similarly expressed were used to find putative *cis*-regulatory elements involved in carbon and/or light regulation \[[@B13]\]. For this study, to identify related genes in metabolism, we added a new statistical functionality to the informatic tool PathExplore \[[@B23]\], which enabled us to identify metabolic pathways that contain more genes than expected in a list of genes \[[@B24]\]. As used here, PathExplore is useful to find functionally related genes from analyses that combine data from multiple microarray chips (for example, InterAct Class and clustering). In this case, we searched for pathways that contained more than the expected number of genes in Model 3 CN-enhanced, compared to the general population. Three genes involved in ferredoxin metabolism were found to be over-represented in Model 3 CN-enhanced (*p*-value = 0.022) (Table [4a](#T4){ref-type="table"}). These genes were also found to be induced in roots and shoots of nitrate-treated plants \[[@B18]\], and the protein products of these genes are all predicted to be localized to the chloroplast \[[@B25]\], further suggesting that they are biologically related and co-regulated.
As we found that genes in the funcat protein synthesis are over-represented in Model 3 CN-enhanced (Table [3](#T3){ref-type="table"}), we selected a set of genes in protein synthesis that are in Model 3 CN-enhanced for additional *cis*search analysis. Four nuclear genes encoding ribosomal proteins predicted to be localized to the mitochondria \[[@B25]\] were assigned to InterAct class 1021 (Table [4b](#T4){ref-type="table"}). These four genes meet the criteria of being biologically related and having similar expression patterns and were also analyzed for potential *cis*-regulatory elements. Over-represented motifs in the promoters of the four protein synthesis genes or the three ferredoxin metabolism genes were identified using AlignAce \[[@B26],[@B27]\] (AlignAce motifs).
We predicted two general mechanisms for which we might be able to identify *cis*-regulatory elements by which carbon and nitrogen can have a non-additive effect (for example, Model 3 CN-enhanced) on the transcription of a gene (Figure [3](#F3){ref-type="fig"}). These models predict that because the genes used for *cis*discovery are induced by carbon alone, there must be a transcription factor (and cognate *cis*element) that responds to carbon alone. Such carbon-responsive *cis*elements (C-elements) can be identified because they should also be over-represented in the promoters of genes that are induced by carbon alone (the C-only inductive model). From this analysis, a number of the AlignAce motifs identified from the ferredoxin metabolism and protein synthesis genes in the Model 3 CN-enhanced were also shown to be associated with C-only inductive model genes (Table [5](#T5){ref-type="table"}; C1-C11). The simplest model that could result in the expression due to CN being greater than C+N is depicted in Figure [3a](#F3){ref-type="fig"}. In this model, the promoters that contain a C-element are also regulated by a completely independent transcription factor (and cognate *cis*element) that responds specifically to a CN-signaling pathway (Figure [3a](#F3){ref-type="fig"}). If such a CN-responsive *cis*element (CN-element) exists, it would be predicted to be over-represented in the promoters of genes in Model 3 CN-enhanced, but would not be over-represented in the C-only inductive model. Two of the AlignAce motifs fit this pattern (motifs CN1 and CN2, Table [5](#T5){ref-type="table"}), suggesting that they are CN-elements.
If CN1 and CN2 regulate gene expression, they might be expected to be evolutionarily conserved. Unfortunately, *A. thaliana*and/or *Oryza sativa*have multiple genes encoding ferredoxin and ferrodoxin reductase, and as such, the true orthologs of the genes used for this analysis can not be conclusively identified for a promoter analysis (the same is true for the ribosomal genes used for analysis). Another prediction is that if CN1 and CN2 regulate gene expression, biologically related genes might also contain CN1 and CN2. Interestingly, ferredoxin-dependent nitrite reductase (At2g15620) contains three copies of CN1 and one copy of CN2 in its promoter. This gene is in Model 3 CN-enhanced (InterAct class 1021), its protein product is localized to the chloroplast \[[@B25]\] and its expression is induced in shoot and roots of nitrate-treated plants \[[@B8]\], suggesting that the gene is biologically related to and co-regulated with the ferredoxin and ferredoxin reductase genes used for this analysis. We next tested if finding three copies of CN1 and one copy of CN2 in the promoter of ferredoxin-dependent nitrite reductase was statistically likely by testing randomized versions of the promoter. We found that three copies of CN1 were unlikely (*p*-value = 0.0364), but it would not be unlikely to find one copy of CN2 (*p*-value = 0.200). In addition, a total of four copies of CN1 and CN2 was very unlikely (*p*-value = 0.018) in any combination (for example, three CN1 and one CN2, two CN1 and two CN2, or one CN1 and two CN2, and so on).
As *A. thaliana*has only one copy of ferredoxin-dependent nitrite reductase, we searched the *O. sativa*genome sequence for ferredoxin-dependent nitrite reductase genes. Again, we found only one gene \[[@B28]\]. BLAST \[[@B29]\] did not find enough similarity between the promoters of the *A. thaliana*ferredoxin-dependent nitrite reductase gene and the *O. sativa*gene for an alignment. Despite this lack of similarity, we tested for the presence of CN1 and CN2 in the promoter of this gene; three copies of CN1 (*p*-value = 0.052) and one copy of CN2 (*p*-value = 0.389) were found. Again, it was very unlikely that a total of four copies of CN1 and CN2 (*p*-value = 0.045) would occur in the promoter sequence.
Identification of nitrogen-dependent enhancers of carbon regulation (NDEs)
--------------------------------------------------------------------------
A second mechanism by which the expression due to CN could be greater than C+N could involve a nitrogen-responsive *cis*element that alone has little or no effect on gene regulation, but when present in combination with a C-element, enhances the induction caused by carbon and is dependent on a carbon-responsive transcription factor (Figure [3b](#F3){ref-type="fig"}). Other regulatory modules in plants have been identified in which the regulation due to one *cis*element requires the presence of another \[[@B30]\]. In the example examined here, the nitrogen-dependent *cis*element enhances the induction caused by the C-element, making it a nitrogen-dependent enhancer of carbon regulation (NDE). To identify NDEs, our strategy for *cis*element identification was modified. NDEs would be expected to be over-represented in the promoters of Model 3 CN-enhanced genes, but only when present in combination with a separate C-element, as both elements are required to give the enhanced expression due to CN. However, some of the AlignAce motifs are potentially involved in regulating expression due to the carbon treatment in cooperation with the already identified C-elements. These *cis*elements would be similar to NDEs as they would be over-represented in genes induced by carbon in combination with the already identified C-elements. As these motifs are not NDEs, we sought to identify them and remove them from the analysis. AlignAce motifs were tested to determine whether they are over-represented in the promoters of genes whose promoters contain any of the C-elements and are in the C-only inductive model. Those that were found to be over-represented were eliminated from further analysis because these motifs are potentially involved in carbon regulation and are not NDEs. Next, the remaining 33 AlignAce motifs were tested to determine if any are NDEs by determining whether they are over-represented in combination with a C-element within the promoters of the Model 3 CN-enhanced genes. Seven of the potential NDEs are over-represented (*p*-value \< 0.05) with at least one C-element in the promoters of the Model 3 CN-enhanced genes, resulting in 12 significant combinations between putative NDEs and C-elements (that is, some of the potential NDEs are over-represented with more than one C-element; data not shown).
To determine if this approach resulted in an enrichment of NDEs, the promoter sequence of each gene was randomized, and the same test was performed. This enabled us to determine whether the remaining 33 AlignAce motifs were over-represented in combination with each C-element in the randomized promoters of the Model 3 CN-enhanced genes. Sets of the randomized promoters (200 sets) were tested, and none of them had as many significant pairs of potential nitrogen-dependent enhancers of carbon regulation and C-elements than the 12 found in the actual promoters. This randomization proves that our approach successfully enriched for NDEs in the actual promoters of the Model 3 CN-enhanced genes and that all the observed significant combinations cannot be due to false positives (*p*-value \< 0.005).
Not surprisingly, each of the seven potential NDEs was found to be over-represented with C-elements using the randomized promoters. This shows that false positives can occur in testing for NDEs. The results from the randomized promoters were used to identify which potential NDEs are over-represented with more C-elements than expected (that is, all the combinations for that NDE cannot be explained by false positives). Two NDEs (N1 and N2) were found to be associated with C-elements (Table [5](#T5){ref-type="table"}; C3, C6, C7 and C10) in six (N1C6, N1C7, N2C3, N2C6, N2C7 and N2C10) of the 12 significant combinations between the 33 remaining AlignAce motifs and the C-elements. N1 and N2 are involved in more significant combinations than expected on the basis of the randomization study (Table [6](#T6){ref-type="table"}; last column).
If N1 or N2 work with the C-elements (C3, C6, C7 and C10) to regulate gene expression in response to CN, then genes that contain both motifs and are in Model 3 CN-enhanced should be misrepresented in certain functional groups as these genes are truly co-regulated. This misrepresentation should occur not only with respect to the genome, but also with respect to the genes in Model 3 CN-enhanced. This result is expected because these genes are more closely related to each other than to the other genes in Model 3 CN-enhanced, and because their CN regulation is the result of the action of the same transcription factor(s). Funcat analysis was used to determine if any functional categories were misrepresented in the genes whose promoters contain N1C6, N1C7, N2C3, N2C6, N2C7 or N2C10 and are in Model 3 CN-enhanced. As the genes used to derive most of the pertinent *cis*motifs encode proteins that are localized to mitochondria, we also tested to see if these genes were misrepresented in the predicted localization of the proteins they encode with respect to the genes in Model 3 CN-enhanced. For the genes whose promoters contain N1C6, N1C7, N2C3, N2C6, N2C7, or N2C10 and are in Model 3 CN-enhanced, only the \'protein synthesis\' funcat was found to be misrepresented amongst the primary funcats as compared to all the genes in Model 3 CN-enhanced (Table [7](#T7){ref-type="table"}). The genes predicted to encode mitochondria-localized proteins are over-represented for some combinations, but genes localized to the cytoplasm or chloroplast are never misrepresented (Table [7](#T7){ref-type="table"}). Two combinations (N2C3 and N2C8) do not show over-representation in protein synthesis and/or genes encoding mitochondria-localized proteins, suggesting they are false positives. All the others show over-representation in some category, further suggesting the potential biological relevance of these *cis*elements (Table [7](#T7){ref-type="table"}).
Discussion
==========
This report contains the one of the first genome-wide investigations of carbon- and nitrogen-signaling interactions in *A. thaliana*\[[@B31]\]. While the focus of our analysis is related to genes controlled by carbon and nitrogen interactions, information from this study can also be used to globally identify genes and processes responsive to regulation by carbon or nitrogen alone. This type of analysis reveals that carbon is a more ubiquitous regulator of the genome compared to nitrogen. The most obvious manifestation of this is the number of genes assigned an InterAct class that are regulated by C-only (1,310) versus N-only (4) (Table [1](#T1){ref-type="table"}). This result is not surprising, because carbon plays a major part in many biological processes and is therefore a major regulator of those processes. However, our studies show that nitrogen has a significant role in modifying the effect of carbon on gene expression. In particular, it is noteworthy that many genes show a response to CN (208 genes) treatment that is different from plants treated with carbon alone (Table [1](#T1){ref-type="table"} and Additional data file 1). This analysis demonstrates that nitrogen does have an effect on gene expression, but that in the vast majority of cases, the nitrogen effect is largely carbon-dependent. The carbon dependence of nitrogen regulation may reflect the metabolic interdependence of carbon and nitrogen. For example, carbon skeletons are required on which to assimilate nitrogen into amino acids.
Biological processes containing genes that respond significantly to carbon, nitrogen and/or CN were initially identified by finding MIPS funcats \[[@B21],[@B22]\] that contained genes that were under-represented in InterAct class 0000 (the No effect model) (Table [2](#T2){ref-type="table"}). Funcats under-represented in the No effect model have a significant number of genes regulated by carbon and/or nitrogen. It is not surprising that processes like metabolism, protein synthesis, and energy are under-represented in the No effect model. These processes control metabolism or require energy generated by metabolism, and therefore expression of genes involved in these processes are likely to change in response to changes in levels of carbon, nitrogen and/or CN caused by external feeding or depletion after starvation. Protein synthesis regulation might be because it is a downstream process responding to an increase of amino acids as a result of feeding carbon, nitrogen and/or CN.
To gain a better understanding of how the metabolism, energy and protein synthesis funcats are regulated by carbon and/or nitrogen, the sub-models in which they are misrepresented were identified (Table [3](#T3){ref-type="table"}). This analysis revealed that the energy funcat is over-represented in InterAct classes that correspond to repression by carbon. It has been shown that carbon sources repress the expression of genes involved in photosynthesis \[[@B32]\]. As photosynthesis genes are part of the energy funcat, the photosynthesis sub-funcat (02.40) was tested and found to be over-represented in the C-only repressive model, in agreement with the previously observed repression of photosynthesis genes by carbon \[[@B32]\].
Surprisingly, metabolism is over-represented in Model 3 CN-suppressed, indicating that many of the genes involved in metabolism show less expression due to CN than expected. The majority of the genes (28 out of 34) were repressed by carbon, induced by nitrogen and repressed by CN, and were assigned to InterAct classes such as -21-2-1 (see Additional data file 1). Several of these genes encode enzymes involved in the catabolism of complex carbohydrates, including β-fructofuranosidase (At1g12240), β-amylase (At3g23920) and β-glucosidase (At3g60130 and At3g60140). *ASN1*(At3g47340), which has been proposed to be involved in producing asparagine for the transport of nitrogen when carbon levels are low and has been shown to be repressed by carbon \[[@B32]\], was assigned Model 3 CN-suppressed (-21-2-1). In addition, *GDH1*(At5g18170), which has been proposed to be involved in ammonia assimilation when ammonia levels are high, is repressed by carbon, and induced by nitrogen \[[@B33]\], and was assigned InterAct class -21-2-1, again a Model 3 CN-suppressed class. These genes therefore seem to be regulated as a result of decreased levels of carbon, increased levels of nitrogen or an imbalance between carbon and nitrogen. For example, when carbon sources are limiting (nitrogen is in excess), *ASN1*is induced because it is involved in shifting the excess nitrogen to asparagine, as asparagine is an efficient way to store and transport nitrogen with respect to carbon \[[@B34]\]. However, when carbon is in excess or carbon and nitrogen are balanced, *ASN1*is repressed. The regulation of these genes demonstrates the exquisite control of metabolic genes required to balance carbon and nitrogen availability.
Our studies also showed that protein synthesis is one of the processes most affected by the interactions between carbon and nitrogen signaling (Table [3](#T3){ref-type="table"}). In addition, the funcat entitled \'protein with binding function or cofactor requirement\' (structural or catalytic) is also over-represented in Model 3 CN-enhanced (see Additional data file 1), partly due to genes that encode proteins involved in translation, including At4g10450 (putative ribosomal protein L9 cytosolic; InterAct class 2132) and At4g25740 (putative ribosomal protein S10; InterAct class 1021) (see Additional data file 1). This suggests that protein synthesis is regulated by carbon (see above), but also by complex interactions between carbon and nitrogen signaling.
Little work has been done on the transcriptional control of protein synthesis by carbon and/or nitrogen signaling in plants. However, it has been shown in yeast that genes encoding ribosomal proteins are induced by nitrogen in the presence of carbon; whether this induction by nitrogen requires carbon to be present was not addressed in the yeast study \[[@B35]\]. Furthermore, in the fungus *Trichoderma hamatum*, the gene for ribosomal protein L36 is regulated by interactions between carbon and nitrogen, as it is induced only by CN, and not by carbon or nitrogen alone \[[@B36]\]. Our studies of carbon and nitrogen regulation of gene expression in plants, combined with the studies in fungi, suggest that transcriptional regulation of genes involved in protein synthesis by carbon and nitrogen signaling interactions is evolutionarily conserved.
Finally, we sought to identify the *cis*-regulatory mechanisms involved in carbon and nitrogen signaling interactions. We hypothesized that there could be two general transcriptional mechanisms that would result in the expression due to CN being greater than that due to C+N (Figure [3](#F3){ref-type="fig"}). In one case, the regulation due to carbon and the regulation due to CN are completely independent (Figure [3a](#F3){ref-type="fig"}), and in the other case, the regulation due to nitrogen is dependent on a carbon-responsive transcription factor and *cis*element (Figure 4b). Since CN1 and CN2 (Table [5](#T5){ref-type="table"}) are over-represented in Model 3 CN-enhanced genes (for example, InterAct class 1021) independently of a C-element, we propose that CN1 and CN2 regulate gene expression due to CN that is independent of a C-element (Figure [3a](#F3){ref-type="fig"}). This hypothesis is supported because CN1 and CN2 were found in the ferredoxin-related genes, which contain no C-elements that are over-represented in Model 3 CN-enhanced. However, we cannot rule out the possibility that CN1 and CN2 are promiscuous NDEs (Figure [3b](#F3){ref-type="fig"}) that interact with many C-elements, which might result in over-representation of CN1 and CN2 in Model 3 CN-enhanced genes, but not in over-representation of a specific C-element.
Further analysis suggests that CN1 is involved in regulating the expression of ferredoxin-dependent nitrite reductase. Finding three copies of CN1 in the promoter of the *A. thaliana*ferredoxin-dependent nitrite reductase gene is statistically unlikely (*p*-value = 0.0364), and while three copies in the promoter of the *O. sativa*gene did not reach the 0.05 cutoff, this might represent some small change in the specificity of the regulating factor between *O. sativa*and *A. thaliana*. The failure of BLAST to detect any similarity between the promoters of these two genes suggests that their transcriptional regulators share very little sequence specificity, so a slight change in specificity is not unexpected. The same analysis suggests that CN2 is a false positive because it is not over-represented in the promoters of ferredoxin-dependent nitrite reductase genes. However, we cannot rule out the possibility that the combination of CN1 and CN2 is what is important in regulating these genes, as having a total of four copies of CN1 and CN2 is unlikely in the promoters of both genes. One possibility is that there is a positional relationship between the copies of CN1 and CN2 that is important. From a quick visual inspection, there does not appear to be a conserved relationship between the three copies of CN1 and one copy of CN2 in the *A. thaliana*and *O. sativa*promoters. These issues will have to be resolved by further experimental work; however, these results do suggest that ferredoxin, ferredoxin reductase and ferredoxin-dependent nitrite reductase are co-regulated by carbon and nitrogen due to CN1 and/or CN2. CN1 and/or CN2 therefore might act to link nitrogen reduction and energy metabolism.
Our analysis found CN-elements in the promoters of the ferredoxin-related genes (Table [4a](#T4){ref-type="table"}), but not in those of the nuclear-encoded ribosomal mitochondrial protein genes (Table [4b](#T4){ref-type="table"}). Also none of the C-elements found in the ferredoxin-related genes (C1 through C5) is over-represented in the Model 3 CN-enhanced genes, suggesting that these elements have no role in CN regulation and that the CN and carbon signaling are independent (Table [5](#T5){ref-type="table"}). In contrast, most of the C-elements in the promoters of the ribosomal protein genes are also over-represented in the promoters of the Model 3 CN-enhanced genes (C6 through C9), suggesting that they have a role in carbon and CN regulation. In addition, the majority of the C-elements (C6, C7 and C10) found to be over-represented in combination with NDEs (N1 and N2), and the most statistically significant of these enhancers (N2), was found in the promoters of the ribosomal proteins (Table [6](#T6){ref-type="table"}). This suggests that the CN transcriptional regulation of genes for ribosomal proteins is primarily due to NDEs (Figure [3b](#F3){ref-type="fig"}). Thus, it is not surprising that many of the genes potentially regulated by the combination of C-elements and NDEs are involved in protein synthesis (Table [7](#T7){ref-type="table"}). However, the putative NDEs most probably regulate genes involved in a number of different biological processes. For example, genes that contain the combination N1C7 and are in Model 3 CN-enhanced include metabolic genes (for example, At3g25900 (homocysteine S-methyltransferase), At2g30970 (aspartate aminotransferase) and At3g52940 (C-14 sterol reductase)), histone-related proteins (for example, At1g54690 (histone H2A) and At2g27840 (histone deacetylase-related)), and putative signaling/regulatory proteins (for example, At4g39990 (Ras-related GTP-binding protein BG3), At5g38480 (14-3-3 protein) and At3g18130 (guanine nucleotide-binding protein)).
This analysis represents a first step in understanding how carbon and nitrogen signaling interact to control gene expression and has identified genes and putative *cis*elements that are responsive to carbon and nitrogen signaling interactions. It is noteworthy that the putative CN-elements and NDEs represent *cis*elements that have not been previously identified and as such may represent novel components of the CN regulatory circuit. Further study of the identified genes and *cis*elements is required to bring about a complete understanding of interactions between carbon and nitrogen signaling.
Materials and methods
=====================
Plant growth and treatment for analysis
---------------------------------------
*Arabidopsis thaliana*seeds of the Columbia ecotype were surface-sterilized and plated on designated media and vernalized for 48 h at 8°C. Plants were grown semi-hydroponically under 16-h-light (70 E/m^2^/sec)/8-h-dark cycles at a constant temperature of 23°C on basal Murashige and Skoog (MS) medium (Life Technologies) supplemented with 2 mM KNO~3~, 2 mM NH~4~NO~3~and 30 mM sucrose \[[@B37]\]. Two-week-old seedlings were transferred to fresh MS media without nitrogen (KNO~3~and NH~4~NO~3~) or carbon (sucrose) and dark-adapted for 48 h. To perform specific metabolic treatments, 25 dark-adapted seedlings were transferred to fresh MS medium containing 0% or 1% (w/v) sucrose and/or 2 mM KNO~3~and 2 mM NH~4~NO~3~or no nitrogen, and illuminated with white light for an additional 8 h (70 E/m^2^/s^1^). Following these transient carbon and nitrogen treatments, whole seedlings were harvested, immediately frozen in liquid nitrogen, and stored at -80°C before RNA extraction.
RNA isolation and microarray analysis
-------------------------------------
RNA was isolated from whole seedlings using a phenol extraction protocol as previously described \[[@B38]\]. Double-stranded cDNA was synthesized from 8 μg total RNA using a T7-Oligo (dT) promoter primer and reagents recommended by Affymetrix. Biotin-labeled cRNA was synthesized using the Enzo BioArray High Yield RNA Transcript Labeling Kit. The concentration and quality of cRNA was estimated through an A260/280 nm reading and running 1:40 of a sample on a 1% (w/v) agarose gel. cRNA (15 μg) was used for hybridization (16 h at 42°C) to the *Arabidopsis*ATH1 Target (Affymetrix). Washing, staining and scanning were carried out as recommended by the Affymetrix instruction manual. Expression analysis was performed with the Affymetrix Microarray Suite software (version 5.0) set at default values with a target intensity set to 150. Three biological replicates for each treatment were carried out.
Using Affymetrix probes to assign genes to InterAct classes
-----------------------------------------------------------
Only Affymetrix probes representing genes that were deemed to be expressed in all treatments and replicates were used for subsequent analysis by InterAct Class \[[@B13],[@B20]\]. For a gene to be considered expressed, the absolute call made by Affymetrix Microarray Suite 5.0 must be \'present\' (P) for each of three replicates for each of four treatments (12 chips total). These genes have reliable values assigned to them that can be used for further analysis, while the proper InterAct Class assignment of a gene with an A (\'absent\') call would not be ensured. It should also be noted that the always P genes are less noisy than the genes that have an A call (data not shown). In the InterAct Class analysis, four values were assigned to each gene on the basis of its response to carbon and/or nitrogen. The first three values are the expression due to carbon (the expression in treatment 2 minus the expression in treatment 1; see Figure [2](#F2){ref-type="fig"}), the expression due to nitrogen (the expression in treatment 2 minus the expression in treatment 1; see Figure [2](#F2){ref-type="fig"}), and the expression due to CN (the expression in treatment 4 minus the expression in treatment 1; see Figure [2](#F2){ref-type="fig"}). The fourth InterAct Class value represents the expected expression due to C+N, which was calculated by adding the expression due to carbon to the expression due to nitrogen. The expression due to carbon, the expression due to nitrogen, the expression due to CN and the C+N values were calculated for each replicate and then analyzed with InterAct Class without binning \[[@B20]\].
Statistical analysis of InterAct Classes and functional categories
------------------------------------------------------------------
*p*-values were calculated for the MIPS functional categories (funcats) \[[@B21],[@B22]\] analysis as described previously \[[@B13]\]. Briefly, the number of genes assigned to the funcat being analyzed and any InterAct class was used as *n*; *p*was the number of genes assigned to the specific model being analyzed divided by the number of genes assigned to an InterAct class and funcat; *k*was the number of genes in the funcat being analyzed and assigned to the model being analyzed. This analysis, with the baseline being all the genes assigned an InterAct class, accounts for any biases that may have been caused by discarding all the absent genes. The one-tailed *p*-value was considered when the Poisson approximation of binomial probabilities was used. For the binomial-ratio and the exact binomial probability test, the *p*-value for *k*or more out of *n*was used.
Identification of putative *cis*-regulatory elements in promoters of CN-regulated genes
---------------------------------------------------------------------------------------
Pathways whose genes are over-represented in Model 3 CN-enhanced were identified using the informatic tool PathExplore \[[@B23]\] function 13 \[[@B24]\]; the methodology is described in pages at these websites. Briefly, a binomial test is used, and the genes assigned an InterAct class were used as the parent list, *n*was the number of genes in Model 3 CN-enhanced (the child list), *k*was the number of genes in the pathway being analyzed and in the child list, and *p*was the number of genes in the pathway being analyzed and in the parent list divided by the number of genes in the parent list. We limited our search to pathways that contained more than two genes in the Model 3 CN-enhanced list. To identify *cis*-regulatory elements involved in regulating genes in Model 3 CN-enhanced and protein synthesis, we used genes involved in protein synthesis that were assigned Model 3 CN-enhanced, to drive the *cis*search: At1g07070 (60S ribosomal protein L35a), At2g36620 (60S ribosomal protein L24), At5g07090 (ribosomal protein S4), and At5g58420 (ribosomal protein S4 like).
The methodology used to identify putative carbon and CN regulatory elements was carried out as described previously \[[@B11]\]. RSA tools was used to extract the *A. thaliana*promoters for every gene \[[@B39],[@B40]\]; AlignAce was then used to identify over-represented motifs in the promoters of the genes being analyzed (AlignAce motifs) \[[@B24]\]. To determine if a motif is over-represented in the promoters of genes in a particular sub-model, the sequence extracted from RSA tools and its reverse complement were searched to determine how many promoters contained the AlignAce motif and in what copy number. Then a binomial test was used to determine if the number promoters that contain the motif in the proper number of copies are over-represented in a particular sub-model. For this analysis, the number of genes with the AlignAce motif being analyzed in their promoter is *n*, *p*is the number of genes in the sub-model (for example, Model 3 CN-enhanced) divided by the total number of genes assigned an InterAct class, and *k*is the number of genes whose promoters contain the AlignAce motif being analyzed (in a specific copy number) and that is in the particular sub-model being tested. A *p*-value was only calculated if *k*is greater than nine. In each case, the lowest *p*-value is given. *Cis*elements over-represented in the C-only inductive model are considered to be putative C-elements, and *cis*elements that are over-represented in the promoters of Model 3 CN-enhanced genes and are not over-represented in the promoters of C-only inductive genes, are considered to be putative CN-elements (Table [5](#T5){ref-type="table"}).
To identify interacting elements, a similar analysis was used. For example, to identify motifs interacting with a C-element (Table [5](#T5){ref-type="table"}) in regulating induction due to carbon (C-associated elements), genes whose promoters contain the C-element were identified. The promoters of these genes were then checked for a second motif. The number of genes that contained the C-element being analyzed and the second motif was used as *n*. The number of genes in the C-only inductive model that contained the C-element being analyzed divided by the number of genes assigned an InterAct class and that contained the C-element being analyzed was used as *p*. The number of genes whose promoters contain that C-element and the second motif being analyzed (in a specified copy number) and that are in the C-only inductive model was used as *k*. In this example, the analysis will determine if the genes that contain the second motif and the C-element being analyzed are over-represented in the C-only inductive model compared to the genes that just contain the C-element. The same approach was used to identify NDEs as described below.
Further analysis for NDEs
-------------------------
The 33 motifs (13 motifs from ribosomal proteins plus 20 motifs from ferredoxin-related proteins (data not shown)) that are not N-, CN- or C-associated elements were tested to determine whether they are potential NDEs. They were tested to see whether genes whose promoters contained these motifs plus a C-element (Table [5](#T5){ref-type="table"}) are over-represented in Model 3 CN-enhanced, as compared to all the genes whose promoters contain the C-element as described above. If a *p*-value less than 0.05 is obtained, the C-element and potential NDE are a significant combination and are likely to regulate carbon and nitrogen interactions. As each motif is tested with each of the 11 C-elements, two steps were taken to control for the multiple tests. First, single strands of the promoter sequences of the *A. thaliana*genes were randomized 200 times, the reverse complement of the randomized strand was determined, and the number of times the 33 remaining AlignAce motifs were found to be over-represented (*p*-value \< 0.05) with the C-elements was determined and compared to the number of significant combinations (*p*-value \< 0.05) between the 33 remaining motifs and the C-elements when the actual promoters were used. In no set of the randomized promoters were the potential NDEs found to form more significant combinations with the 11 C-elements than the actual promoter sequences (*p*-value \< 1/200 = 0.005). In the second control step, the number of significant combinations that each of the 33 remaining AlignAce motifs was involved in was determined and compared to the number of significant combinations found with the 200 sets of randomized promoters. For one motif, if one random set is significant with as many C-elements as the real promoters the *p*-value would be 0.005 (1/200).
Further analysis of CN1 and CN2
-------------------------------
The promoter for At2g15620 was extracted from RSA tools \[[@B39],[@B40]\]. The reverse complement of the strand from RSA tools was determined to identify the occurrence of CN1 and CN2 in either strand of the promoter as described above to determine over-representation of the AlignAce motifs in the promoters of the genes in Model 3 CN-enhanced. To determine whether CN1 and CN2 occur more times than expected in the promoter, the sequence from RSA tools \[[@B39],[@B40]\] was randomized 5,000 times and the above procedure was repeated. The number of times CN1 and/or CN2 were found in the randomized versions as many or more times than the actual promoter was determined and used to calculate a *p*-value (that is, if 50 random cases do as well as or better than the actual case *p*-value = 50/5,000 (0.05))
The sequence database was searched using BLAST \[[@B29]\] for a gene similar to At2g15620 in the *O. sativa*sequence. Only one hit was found. This gene is annotated as a ferredoxin-dependent nitrate reductase \[[@B28]\]. The 1,000 base-pairs upstream of this gene were taken and \'BLAST align two sequences\' was used to determine whether this sequence is similar to the promoter of At2g15620. BLAST did not find enough similarity to create an alignment. The sequence was then subjected to the same test described above for the promoter of At2g15620.
Funcat analysis of the NDEs
---------------------------
Funcat analysis of the genes whose promoters contain specific *cis*elements was performed similarly to the approach described above. Briefly, the number of genes assigned to the funcat being analyzed and Model 3 CN-enhanced was used as *n*; *p*was the number of genes assigned to Model 3 CN-enhanced and the funcat being analyzed divided by the number of genes assigned to Model 3 CN-enhanced and a funcat; *k*was the number of genes in the funcat being analyzed that was assigned to the Model 3 CN-enhanced category and containing the combination of C- and N-element being analyzed.
Statistical significance of localization was calculated similarly. The only difference being that instead of genes assigned a funcat, genes whose protein products are predicted to be localized in the compartment being analyzed were used. Predicted protein localizations were extracted from the TAIR web page \[[@B25]\].
Additional data files
=====================
The following additional data are available with the online version of this paper: Additional data file [1](#s1){ref-type="supplementary-material"} containing a table listing the Affymetrix probe ID, gene, and InterAct class for all the Affymetrix probes assigned an InterAct class; Additional data file [2](#s2){ref-type="supplementary-material"} listing the data from 12 Affymetrix microarray chips used in this study.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
A table listing the Affymetrix probe ID, gene, and InterAct class for all the Affymetrix probes assigned an InterAct class
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
The data from 12 Affymetrix microarray chips used in this study
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
This work was supported by the National Institutes of Health (grant number GM32877 to G.M.C) and a National Research Service Reward (grant number GM65690 to P.M.P).
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Transcriptional regulation by carbon and nitrogen interactions. **(a)**Interactions between carbon (C) and nitrogen (N) signaling can be explained by three models, and an example(s) of each is given. Model 1, carbon and nitrogen regulation are independent and therefore are additive. Model 2, carbon and nitrogen are dependent, as both are required for an effect. Model 3, there is a dependent and independent component to carbon and nitrogen regulation. Two examples of Model 3 are shown (genes *Y*and *Z*). For gene *Y*, nitrogen only has an effect in the presence of carbon, while for gene *Z*, carbon only has an effect in the presence of nitrogen. **(b)**The assignment of genes *W*, *X*, *Y*, and *Z*to InterAct classes.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Treatments for carbon and nitrogen interaction studies. +C, -C, with and without carbon, respectively. +N, -N, with and without nitrogen, respectively.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Two general mechanisms that would result in CN expression being greater than C+N. **(a)**Carbon (C) and CN regulatory elements are independent and do not interact. The data do not allow us to rule out the possibility that the C-element is inactive in the presence of CN and that the CN-element alone results in more expression than the C-element. **(b)**CN and carbon regulation are dependent. The increase in expression due to CN requires two interacting *cis*elements, one of which is a C-element and the other a nitrogen-dependent enhancer of carbon regulation (NDE).
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
InterAct classes that contain more than one gene
:::
CN model Number of genes InterAct class
---------------------------------- --------------- ----------------- ---------------- ---- ---- ----
No effect 1,167 0 0 0 0
C-only Inductive 1,011 1 0 1 1
Repressive 596 -1 0 -1 -1
N-only Repressive 4 0 -1 -1 -1
C dominates 187 1 -1 1 1
156 -1 1 -1 -1
142 2 1 2 2
140 -2 -1 -2 -2
Model 1 (independent) N dominates 3 -1 1 1 1
Equal effect 145 -1 -1 -1 -1
63 1 1 1 1
Antagonistic 6 2 -1 1 1
3 -2 1 -1 -1
Inductive 4 2 1 3 3
3 1 1 2 2
Repressive 3 -1 -1 -2 -2
2 -2 -1 -3 -3
Model 2 (dependent) Inductive 7 0 0 1 0
Repressive 2 0 0 -1 0
Model 3 (dependent/ independent) CN-enhanced 92 1 0 2 1
25 2 1 3 2
13 -1 -1 -1 -2
11 -2 -1 -2 -3
8 2 -1 2 1
8 1 -1 2 1
4 0 -1 0 -1
3 0 -1 1 -1
2 2 1 4 3
2 -3 -1 -2 -3
CN-suppressed 46 -2 1 -2 -1
17 -1 1 -1 0
9 -1 2 -1 1
9 -1 0 -2 -1
9 1 1 1 2
8 -1 1 -1 1
6 0 1 0 1
6 2 0 1 2
3 2 1 2 3
2 1 1 -1 1
2 -1 1 0 1
2 1 2 1 3
2 1 0 0 1
2 1 -1 -1 1
2 0 1 -1 1
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Funcats that are statistically under-represented in InterAct class 0000 (the No effect model)
:::
Funcats Number of genes assigned an InterAct class Number of InterAct class 0000 genes *p*-value
------------------- -------------------------------------------- ------------------------------------- ---------------
All funcats 3,447 1,089 \-
Metabolism 496 120 6.0 × 10^-4^
Protein synthesis 218 55 2.6 × 10^-2^
Energy 125 21 1.34 × 10^-4^
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Sub-models that are misrepresented in the metabolism, protein synthesis and energy funcats
:::
Funcats No effect (1,089) C-only CN interactions
------------------- ------------------- -------- ----------------- ------- ------- -------
Metabolism 120 -S 141 +S 62 -S 19 -S 34 +S 20
Protein synthesis 55 -S 81 +S 13 -S 4 -S 2 -S 32 +S
Energy 21 -S 32 29 +S 3 8 5
+S, sub-model over-represented; -S, sub-model under-represented. See text for details.
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Genes used to drive *cis*analysis
:::
-------------------------------------------------------------------------------------- ----------------------------- ------- --- ---- -----
Gene Enzyme Class
C N CN C+N
**(a)**Genes from pathways that are over-represented in Model 3 CN enhanced
At2g27510 Ferredoxin 1 0 2 1
At1g30510 Ferredoxin-NADP^+^reductase 1 0 2 1
At4g05390 Ferredoxin-NADP^+^reductase 2 1 3 2
**(b)**Genes involved in protein synthesis were also used to drive the *cis*analysis
At1g07070 60S Ribosomal protein L35a 1 0 2 1
At2g36620 60S Ribosomal protein L24 1 0 2 1
At5g07090 40S Ribosomal protein S4 1 0 2 1
At5g58420 40S Ribosomal protein S4 1 0 2 1
-------------------------------------------------------------------------------------- ----------------------------- ------- --- ---- -----
:::
::: {#T5 .table-wrap}
Table 5
::: {.caption}
######
Motifs that are over-represented in Model 3 CN-enhanced or in the C-only inductive model
:::
Motif C-only inductive *p*-value Model 3 CN-enhanced *p*-value Element name
---------------------------------- ---------------------------- ------------------------------- --------------
Ferredoxin-related motifs
RGAAVMANA NS 0.0262 CN1
GNAANVMGAHNM NS 0.0089 CN2
GAWYTGA 0.0073 NS C1
ARNNGANNCAA 0.00049 NS C2
KMSAGAG 0.0322 NS C3
WMNCHGAANC 0.0091 NS C4
GAGARRDDG 0.0375 NS C5
Protein-synthesis related motifs
WKGGGCC \<0.0001 \<0.0001 C6
GGCCSAW \<0.0001 \<0.0001 C7
AAACYCNA 0.0375 0.0038 C8
WTBGGCY 0.0022 0.011 C9
GDNTTGKAM 0.0359 NS C10
AAGAAAA 0.0344 NS C11
Nucleotide abbreviations: R; A or G, Y; C or T, W; A or T, S; G or C, M; A or C, K; G or T, H; A, C or T, B; G, C or T, V; G, A or C, D; G, A or TC, N; G, A, C or T.
:::
::: {#T6 .table-wrap}
Table 6
::: {.caption}
######
Potential NDEs
:::
NDEs Element name Ferredoxin C-elements Protein synthesis C-elements Total *p*-value
---------------------------------- -------------- ----------------------- ------------------------------ ----------------- -------- ---------
Ferredoxin-related motifs
CHHNAACHRA N1 NS 0.0222 0.0344 NS 0.046
N1C6 N1C7
Protein synthesis related motifs
TNNDNVNACAACA N2 0.0281 0.0207 0.0268 0.0037 \<0.005
N2C3 N2C6 N2C7 N2C8
For nucleotide abbreviations see the foonote for Table 5.
:::
::: {#T7 .table-wrap}
Table 7
::: {.caption}
######
Misrepresentation of genes that are potentially regulated by a combination of a C-element and N1 or N2
:::
Gene set Protein synthesis funcat Genes predicted to be localized to the mitochondria
------------------------------ -------------------------- -----------------------------------------------------
InterAct Class genes (3,652) 370 393
Model 3 CN enhanced (127) 32 +S 21
N1C6 (45) 14 +S 9
N1C7 (49) 15 +S 9
N2C3 (27) 5 4
N2C6 (16) 8 +S 7 +S
N2C7 (17) 9 +S 7 +S
N2C10 (15) 3 4
+S, sub-model over-represented; -S, sub-model under-represented. See text for details.
:::
|
PubMed Central
|
2024-06-05T03:55:51.839170
|
2004-10-29
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545782/",
"journal": "Genome Biol. 2004 Oct 29; 5(11):R91",
"authors": [
{
"first": "Peter M",
"last": "Palenchar"
},
{
"first": "Andrei",
"last": "Kouranov"
},
{
"first": "Laurence V",
"last": "Lejay"
},
{
"first": "Gloria M",
"last": "Coruzzi"
}
]
}
|
PMC545783
|
Background
==========
The analysis of genetic regulatory networks has received a major impetus from the huge amounts of data made available by high-throughput technologies such as DNA microarrays. The genome-wide, massively parallel monitoring of gene activity will increase the understanding of the molecular basis of disease and facilitate the identification of therapeutic targets.
To fully uncover regulatory structures, different analysis tools for transcriptomic and other high-throughput data will have to be used in an integrative or iterative fashion. In simple eukaryotes or prokaryotes, gene-expression data has been combined with two-hybrid data \[[@B1]\] and phenotypic data \[[@B2]\] to successfully predict protein-protein interaction and transcriptional regulation on a large scale. If the principal organization of a gene network has been established, differential equations may be used to study its quantitative behavior \[[@B3],[@B4]\].
In higher organisms, however, little is known about regulatory control mechanisms. As a first step in reverse engineering of genetic regulatory networks, structural relationships between genes can be explored on the basis of their expression profiles. Here, we focus on graphical models \[[@B5],[@B6]\] as a probabilistic tool to analyze and visualize conditional dependencies between genes. Genes are represented by the vertices of a graph and conditional dependencies between their expression profiles are encoded by edges. Graphical modeling can be carried out with directed and undirected edges, with discretized and continuous data. Over the past few years, graphical models, in particular Bayesian networks, have become increasingly popular in reverse engineering of genetic regulatory networks \[[@B7]-[@B10]\].
Graphical models are powerful for a small number of genes. As the number of genes increases, however, reliable estimates of conditional dependencies require many more observations than are usually available from gene-expression profiling. Furthermore, because the number of models grows super-exponentially with the number of genes, only a small subset of models can be tested \[[@B10]\]. Most important, a large number of genes often entails a large number of spurious edges in the model \[[@B11]\]. The interpretation of the graph within a conditional-independence framework is then rendered difficult \[[@B12]\]. Even a search for local dependence structures and subnetworks with high statistical support \[[@B7]\] provides no guarantee against the detection of numerous spurious features.
Some of these problems may be circumvented by restricting the number of possible models or edges \[[@B10],[@B13]\] or by exploiting prior knowledge on the network structure. So far, however, this prior knowledge is difficult to obtain.
As an alternative approach to modeling genetic networks with many genes, we propose not to condition on all genes at a time. Instead, we apply graphical modeling to small subnetworks of three genes to explore the dependence between two of the genes conditional on the third. These subnetworks are then combined for making inferences on the complete network. This modified graphical modeling approach makes it possible to include many genes in the network while studying dependence patterns in a more complex and exhaustive way than with only pairwise correlation-based relationships.
For an independent validation of our method, we compare our modified graphical Gaussian modeling (GGM) approach with conventional graphical modeling in a simulation study. We show at the end of the Results section that our approach outperforms the standard method in simulation settings with many genes and few observations. For a further evaluation with real data, we apply our approach to the galactose-utilization data from \[[@B14]\] to detect galactose-regulated genes in *Saccharomyces cerevisiae*.
The main aim of this methodological work, however, was to elucidate the regulatory network of the two isoprenoid biosynthesis pathways in *Arabidopsis thaliana*(reviewed in \[[@B15]\]). The greater part of this paper is therefore devoted to the inference and biological interpretation of a genetic regulatory network for these two pathways. To motivate our novel modeling strategy, we first describe the problems that we encountered with standard GGMs before presenting the results of our modified GGM approach.
Results
=======
Isoprenoids serve numerous biochemical functions in plants: for example, as components of membranes (sterols), as photosynthetic pigments (carotenoids and chlorophylls) and as hormones (gibberellins). Isoprenoids are synthesized through condensation of the five-carbon intermediates isopentenyl diphosphate (IPP) and dimethylallyl diphosphate (DMAPP). In higher plants, two distinct pathways for the formation of IPP and DMAPP exist, one in the cytosol and the other in the chloroplast. The cytosolic pathway, often described as the mevalonate or MVA pathway, starts from acetyl-CoA to form IPP via several steps, including the intermediate mevalonate (MVA). In contrast, the plastidial (non-mevalonate or MEP) pathway involves condensation of pyruvate and glyceraldehyde 3-phosphate via several intermediates to form IPP and DMAPP. Whereas the MVA pathway is responsible for the synthesis of sterols, sesquiterpenes and the side chain of ubiquinone, the MEP pathway is used for the synthesis of isoprenes, carotenoids and the side chains of chlorophyll and plastoquinone. Although both pathways operate independently under normal conditions, interaction between them has been repeatedly reported \[[@B16],[@B17]\].
Reduced flux through the MVA pathway after treatment with lovastatin can be partially compensated for by the MEP pathway. However, inhibition of the MEP pathway in seedlings leads to reduced levels in carotenoids and chlorophylls, indicating a predominantly unidirectional transport of isoprenoid intermediates from the chloroplast to the cytosol \[[@B16],[@B18]\], although some reports indicate that an import of isoprenoid intermediates into the chloroplast also takes place \[[@B19]-[@B21]\].
Application of standard GGM to isoprenoid pathways in *Arabidopsis thaliana*
----------------------------------------------------------------------------
To gain more insight into the cross-talk between both pathways at the transcriptional level, gene-expression patterns were monitored under various experimental conditions using 118 GeneChip (Affymetrix) microarrays (see Additional data files 1 and 2). To construct the genetic regulatory network, we focused on 40 genes, 16 of which were assigned to the cytosolic pathway, 19 to the plastidal pathway and five encode proteins located in the mitochondrion. These 40 genes comprise not only genes of known function but also genes whose encoded proteins displayed considerable homology to proteins of known function. For reference, we adopt the notation from \[[@B22]\] (see Table [1](#T1){ref-type="table"}).
The genetic-interaction network among these genes was first constructed using GGM with backward selection under the Bayesian information criterion (BIC) \[[@B23]\]. This was carried out with the program MIM 3.1 \[[@B24]\] (see Materials and methods for further details). The network obtained had 178 (out of 780) edges - too many to single out biologically relevant structures. Therefore, bootstrap resampling was applied to determine the statistical confidence of the edges in the model (Figure [1b](#F1){ref-type="fig"}). For the bootstrap edge probabilities, only a cutoff level as high as 0.8 led to a reasonably low number of selected edges (31 edges, Figure [2](#F2){ref-type="fig"}). However, a comparison between bootstrap-edge probabilities and the pairwise correlation coefficients suggested that for such a high cutoff level, many true edges may be missed. For example, the gene *AACT2*appears to be completely independent from all genes in the model although it is strongly correlated with *MK*, *MPDC1*and *FPPS2*(see Additional data file 4 for the correlation patterns).
This phenomenon had already been observed in a simulation study by Friedman *et al*. \[[@B25]\] and may be related to the surprisingly frequent appearance of edges with a low absolute pairwise correlation coefficient but a high bootstrap estimate (Figure [1c](#F1){ref-type="fig"}). Although there is no concise explanation for this pattern, one conjecture would be that the simultaneous conditioning on many variables introduces many spurious edges with little absolute pairwise correlation but high absolute partial correlation into the model. Our modification for GGMs is to improve upon this drawback.
Application of our modified GGM approaches
------------------------------------------
As described in more detail in Materials and methods, our approach aims at modeling dependencies between two genes by taking the effect of other genes separately into account. In the hope of identifying direct co-regulation between genes, an edge is drawn between two genes *i*and *j*when their pairwise correlation is not the effect of a third gene. Each edge has therefore a clear interpretation.
We have developed two versions of our method: a frequentist approach in which each edge is tested for presence or absence; and a likelihood approach with parameters *θ*~*ij*~, which describe the probability for an edge between *i*and *j*in a latent random graph. One main benefit of the second version over full graphical models is that one can easily test on a large scale how well additional genes can be incorporated into the network. This allows the selection of additional candidate genes for the network in a fast and efficient way.
We have applied and tested our modified GGM approaches by constructing a regulatory network of the 40 genes in the isoprenoid pathways in *A. thaliana*and by attaching 795 additional genes from 56 other metabolic pathways to it. Figure [3](#F3){ref-type="fig"} shows the network model obtained from the frequentist modified GGM approach. Because we find a module with strongly interconnected genes in each of the two pathways, we split the graph into two subgraphs, each displaying the subnetwork of one module and its neighbors. Our finding provides a further example that within a pathway many consecutive or closely positioned genes are potentially jointly regulated \[[@B26]\].
In the MEP pathway, the genes *DXR*, *MCT*, *CMK*and *MECPS*are nearly fully connected (upper panel of Figure [3](#F3){ref-type="fig"}). From this group of genes, there are a few edges to genes in the MVA pathway. Among these genes, *AACT1*and *HMGR1*form candidates for cross-talk between the MEP and the MVA pathway because they have no further connection to the MVA pathway. Their correlation to *DXR*, *MCT*, *CMK*and *MECPS*is always negative.
Similarly, the genes *AACT2*, *HMGS*, *HMGR2*, *MK*, *MPDC1*, *FPPS1*and *FPPS2*share many edges in the MVA pathway (lower panel of Figure [3](#F3){ref-type="fig"}). The subgroup *AACT2*, *MK*, *MPDC1*and *FPPS2*is completely interconnected. From these genes, we find edges to *IPPI1*and *GGPPS12*in the MEP pathway. Whereas *IPPI1*is positively correlated with *AACT2*, *MK*, *MPDC1*and *FPPS2*, *GGPPS12*displays negative correlation to the four genes.
In contrast to the conventional graphical model, we could now identify the connection between *AACT2*and *MK*, *MPDC1*and *FPPS2*. In general, we found a better agreement between the absolute pairwise correlation and the selected edges (frequentist approach) or the probability parameters *θ*(latent random graph approach). Figures [4a](#F4){ref-type="fig"} and [4b](#F4){ref-type="fig"} show the selected edges and *θ*-values as a function of the absolute pairwise correlation.
Attaching additional pathway genes to the network
-------------------------------------------------
Following construction of the isoprenoid genetic network, 795 additional genes from 56 metabolic pathways were incorporated. Among these were genes from pathways downstream of the two isoprenoid biosynthesis pathways, such as phytosterol biosynthesis, mono- and diterpene metabolism, porphyrin/chlorophyll metabolism, carotenoid biosynthesis, plastoquinone biosynthesis for example. Using the second version of our method, that is, the latent random graph approach, we compared *θ*-values for all gene pairs in the network with and without attaching these additional genes (Figure [4b](#F4){ref-type="fig"} and [4c](#F4){ref-type="fig"}). As expected, the parameters *θ*for the edge probabilities decreased if additional genes were included in the isoprenoid network (see Materials and methods). After addition, if for a gene pair *i*, *j*, *θ*~*ij*~dropped by more than 0.3, it was assumed that the dependence between *i*and *j*could be \'explained\' by some of the additional genes.
To find these genes out of all additionally tested candidates *k*, GGMs with genes *i*, *j*and *k*were formed. A gene *k*was considered to explain the dependency between *i*and *j*when an edge between *i*and *j*was not supported in the GGM, that is, when the null hypothesis *ρ*~*ij*\|*k*~= 0 was accepted in the corresponding likelihood ratio test. *k*was then taken to \'attach well\' to the gene pair *i*, *j*.
Thus, for each gene pair *i*, *j*whose parameter *θ*~*ij*~dropped by more than 0.3, we obtained a list of well-attaching genes. Genes appearing significantly frequently in these lists of well-attaching genes were assumed to connect well to the complete genetic network. We tested for significance by randomization: For each gene pair *i*, *j*, a randomized list of well-attaching genes was formed with the same size as the original gene list. To explore which pathways attach significantly well to the MVA and MEP pathways, the portion of genes from each of the 56 pathways was summed over all gene pairs *i*, *j*. These sums were then compared for the originally attached genes and the sums of randomly attached genes in 100 datasets.
Table [2](#T2){ref-type="table"} shows the pathways whose genes were found to attach significantly frequently to the MVA pathway, the MEP pathway, or both pathways. Interestingly, from all 56 metabolic pathways considered, we predominantly find that genes from downstream pathways fit well into the isoprenoid network. These results suggest a close regulatory connection between isoprenoid biosynthesis genes and groups of downstream genes. On the one hand, we find strong connections between the MEP pathway and the plastoquinone, the carotenoid and the chlorophyll pathways (experimentally supported by \[[@B15],[@B16],[@B27]\]). On the other hand, the plastoquinone and phytosterol biosynthesis pathways appear to be closely related to the genetic network of the MVA pathway.
On a metabolic level, our results are substantiated by earlier labeling experiments using \[1-^13^C\] glucose, which revealed that sterols were formed via the MVA pathway, while plastidic isoprenoids (β-carotene, lutein, phytol and plastoquinone-9) were synthesized using intermediates from the MEP pathway \[[@B27]\]. Moreover, incorporation of \[1-^13^C\]- and \[2,3,4,5-^13^C~4~\]1-deoxy-D-xylulose into β-carotene, lutein and phytol indicated that the carotenoid and chlorophyll biosynthesis pathways proceed from intermediates obtained via the MEP pathway \[[@B28]\].
In contrast, a close connection between the MVA and the MEP pathways could not be detected. This suggests that cross-talk on the transcriptional level may be restricted to single genes in both pathways.
In a further analysis step, we examined which gene pairs the four identified pathways (plastoquinone, carotenoid, chlorophyll, and phytosterols) attached to. Genes from the plastoquinone pathway were predominantly linked to the genes *DXR*, *MCT*, *CMK*, *GGPPS11*, *GGPPS12*, *AACT1*, *HMGR1*and *FPPS1*, supporting the hypothesis that *AACT1*and *HMGR1*are involved in communication between the MEP and MVA pathways.
Genes from the carotenoid pathway attached to *DXPS2*, *HDS*, *HDR*, *GGPPS11*, *DPPS2*and *PPDS2*, whereas the chlorophyll biosynthesis appears to be related to *DXPS2*, *DXPS3*, *DXR*, *CMK*, *MCT*, *HDS*, *HDR*, *GGPPS11*and *GGPPS12*. Genes from the phytosterol pathway attach to *FPPS1*, *HMGS*, *DPPS2*, *PPDS1*and *PPDS2*.
Incorporating 795 additional genes into the isoprenoid genetic network would not have been feasible with standard GGMs as the graphical model would have had to be newly fitted for each additional gene. Also, hierarchical clustering would not have been an appropriate tool for detecting the similarities in the correlation patterns between the two isoprenoid metabolisms and their downstream pathways. Figure [5](#F5){ref-type="fig"} shows the hierarchical clustering of the 40 isoprenoid genes and 795 additional pathway genes based on the distance measure 1 - \|*σ*~*ij*~\|, where *σ*~*ij*~denotes the pairwise correlation between genes *i*and *j*.
The positions of the MVA pathway genes (labeled \'m\') and the non-mevalonate pathway genes (labeled \'n\'), respectively, are shown to the right of the figure. The symbol + represents the positions of genes from the downstream pathways identified in Table [2](#T2){ref-type="table"}, whereby the vertical line is drawn to distinguish between genes downstream of the mevalonate and the non-mevalonate pathway. From Figure [5](#F5){ref-type="fig"} it can be easily seen that there is no clear pattern of (positional) association between genes of the isoprenoid biosynthesis and downstream pathways in the hierarchical clustering.
Simulation study
----------------
For an independent comparison between the modified and the conventional GGM approaches, we simulated gene-expression data with 40 genes and 100 observations. This simulation framework corresponds to the data for isoprenoid biosynthesis and is thought to be only exemplary at this point. An extensive simulation study is currently underway and will be presented elsewhere.
Following recent findings on the topology of metabolic and protein networks \[[@B29],[@B30]\], we simulated scale-free networks in which the fraction of nodes with *k*edges decays as a power law ∝ *k*^-*γ*^. For metabolic and protein networks, *γ*is usually estimated to range between 2 and 3, which would result in very sparse networks with fewer edges than nodes in our simulation settings. To allow for denser networks, we generated 100 graphs each for *γ*= 0.5, 1.5 and 2.5. With 40 nodes, these graphs then comprised 88.3, 49.7 and 30.5 edges on average. For each edge, the conditional dependence of the corresponding gene pairs was modeled with a latent random variable in a structural equation model as described in \[[@B31]\]. Further details are of technical nature and are omitted here. The use of latent random variables enabled us to model partial correlation coefficients according to the previously defined network structure while ensuring positive definiteness of the complete partial correlation matrix. This matrix was then transformed into a covariance matrix Σ, from which synthetic gene expression data with 100 observations were sampled according to a multivariate normal distribution *N*(0,Σ).
The performance of the graphical modeling approaches was monitored using the rate of true and false positives in receiver operator characteristics (ROC) curves (see \[[@B11]\] for a short introduction). For the standard graphical model, bootstrapping would have been too time-consuming, so we ranked all edges according to their sequential removal in the backward selection process. Figure [6a](#F6){ref-type="fig"} shows the ROC curves for the graphical modeling with backward selection and the modified graphical modeling approaches (frequentist and latent random graph approach). We also included the ROC curve for network inference with pairwise correlation coefficients. It can be seen that the modified GGM approaches outperform the conventional graphical modeling. Both the frequentist and the latent random graph method show a similar performance. Also, it should be noted that a simple measure such as the pairwise correlation can be quite powerful in detecting conditional dependencies between genes.
ROC curves depict the true-positive rate as a function of the false-negative rate. However, in our setting where the false-positive edges by far outnumber the true-positive ones, the proportion of true positives among the selected edges is also of interest (Figure [6b](#F6){ref-type="fig"}). Note that this proportion is the complementary false-discovery rate 1-FDR \[[@B32]\]. Figure [6b](#F6){ref-type="fig"} provides further evidence that the modified GGM approaches have a better performance than standard GGM.
Application to galactose utilization in *Saccharomyces cerevisiae*
------------------------------------------------------------------
For further evaluation, we applied our approach to the galactose-utilization dataset from \[[@B14]\] to detect galactose-regulated genes in *Saccharomyces cerevisiae*. Ideker *et al*. \[[@B14]\] used self-organizing maps to cluster 997 genes with significant expression changes in 20 systematic perturbation experiments of the galactose pathway. From the nine galactose genes under investigation, two subgroups with three and four genes, respectively, were found in two of the 16 clusters. Nine of the 87 genes in these two clusters carried GAL4p-binding sites and are thus candidate genes for regulation by the transcription factor GAL4p. Among these candidate genes, *GCY1*and *PCL10*are known to be targets of GAL4p \[[@B33]\], and *YMR318C*has been implicated in another binding-site study \[[@B34]\].
After incorporating all yeast genes into our network of the nine galactose genes, 13 genes were found to attach significantly well. Among these, *GCY1*and *PCL10*were also detected. Furthermore, three out of the remaining 11 candidate genes (*MLF3*, *YEL057C*and *YPL066W*) had GAL4p-binding sites. These genes were also identified in \[[@B14]\]. This result shows once more that with our approach we are not only able to model the dependence between genes but also find genes whose expression profiles fit well to the original genes in the model. In contrast to \[[@B14]\], we did not have to rely on gene clusters with a high occurrence of galactose genes to find these genes.
Discussion
==========
Analysis of gene expression patterns, for example cluster analysis, often focuses on coexpression and pairwise correlation between genes. Graphical models are based on a more sophisticated measure of conditional dependence among genes. However, with this measure, modeling is restricted to a small number of genes. With a larger set of genes, it is rather difficult to interpret the model and to generate hypotheses on the regulation of genetic networks.
In our approaches, in the search for significant co-regulation between two genes all other genes in the model are also taken into account. However, the effect of these genes is examined separately, one gene at a time. Because of this simplification, modeling can include a larger number of genes. Also, each edge has a clear interpretation, representing a pair of significantly correlated genes whose dependence cannot be explained by a third gene in the model. Our frequentist method has a resemblance to the first two steps in the SGS and PC algorithms \[[@B31]\]. By restricting the modeling to subnetworks with three genes, we avoid the statistically unreliable and computationally costly search for conditional independence in large subsets, as in the SGS algorithm. Also, we avoid having to remove edges in a stepwise fashion, as in the PC algorithm. Therefore, we do not run the risk of mistakenly removing an edge at an early stage, which leads to improved stability in the modeling process.
By using a Gaussian model, we can only reveal linear dependencies between genes. For handling nonlinearities, gene-expression profiles should be discretized and analyzed in a multinomial framework. In principle, it should be straightforward to adopt our approach to a multinomial model. Because we focused on linear dependencies, we have not addressed this problem so far.
For the isoprenoid biosynthesis pathways in *A. thaliana*, we constructed a genetic network and identified candidate genes for cross-talk between both pathways. Interestingly, both positive and negative correlations were found between the identified candidate genes and the corresponding pathways. *AACT1*and *HMGR1*, key genes of the MVA pathway, were found to be negatively correlated to the module of connected genes in the MEP pathway. This suggests that in the experimental conditions tested, *AACT1*and *HMGR1*may respond differently (than the MEP pathway genes) to environmental conditions, or that they possess a different organ-specific expression profile. In either case, expression within both groups seems to be mutually exclusive. On the other hand, a positive correlation was identified between *IPPI1*and members of the MVA pathway, suggesting that this enzyme controls the steady-state levels of IPP and DMAPP in the plastid when a high level of transfer of intermediates between plastid and cytosol takes place.
Although we have considered only metabolic genes in this analysis, the method can be extended to identify genes encoding other types of proteins belonging to the same transcription module. In fact, transcription factors and other regulator proteins, as well as structural proteins such as transporters, are often found in the same expression module \[[@B26]\]. Our results suggest that the expression of genes belonging to the chlorophyll and carotenoid biosynthesis pathways is controlled by a module that possibly includes genes from the MEP pathway.
Similarly, the expression of genes in the phytosterol pathway appears to be influenced by genes from the MVA pathway. For the downstream regulation of plastoquinone biosynthesis, however, genes from both pathways seem to be involved. This finding is in agreement with the dual localization of enzymes from the plastoquinone pathway in either the plastid or the cytosol. The regulation of this pathway may therefore depend on processes happening on the metabolic and regulatory level in both compartments.
We have shown in a simulation study that for gene-expression data with many genes and few observations, the modified GGM approaches have performed better in recovering conditional dependence structures than conventional GGM. However, a final evaluation of our inferred network for the isoprenoid biosynthesis pathways in *A. thaliana*can only be made on the basis of additional knowledge and biological experiments. At this stage, the use of domain knowledge has provided some means of network validation. As genes from the respective downstream pathways were significantly more often attached to the isoprenoid network than were candidate genes from other pathways, we are quite confident that our method can grasp the modularity in the dependence structure within groups of genes and also between groups of genes. Such modularity would have been difficult to detect by standard graphical modeling or clustering.
Materials and methods
=====================
Graphical Gaussian models (GGMs)
--------------------------------
Let *q*be the number of genes in the network, and *n*be the number of observations for each gene. The vector of log-scaled gene-expression values, *Y*= (*Y*~1~,\...,*Y*~*q*~) is assumed to follow a multivariate normal distribution *N*(*μ*,Σ) with mean *μ*= (*μ*~1~,\...,*μ*~*q*~) and covariance matrix Σ. The partial correlation coefficients *ρ*~*ij*\|*rest*~, which measure the correlation between genes *i*and *j*conditional on all other genes in the model are calculated as
where *ω*~*ij*~, *1*, *j*= 1,\...,*q*are the elements of the precision matrix Ω = Σ^-1^.
Using likelihood methods, each partial correlation coefficients *ρ*~*ij*\|*rest*~can be estimated and tested against the null hypothesis *ρ*~*ij*\|*rest*~= 0 \[[@B5]\]. An edge between genes *i*and *j*is drawn if the null hypothesis is rejected. Since the estimation of the partial correlation coefficients involves matrix inversion, estimators are very sensitive to the rank of the matrix. If the model comprises many genes, estimates are only reliable for a large number of observations.
Commonly, the modeling of the graph is carried out in a stepwise backward manner starting from the full model from which edges are removed consecutively. The process stops when no further improvement can be achieved by removal of an additional edge. The final model is usually evaluated by bootstrapping to exclude spurious edges in the model.
Modified GGM approaches
-----------------------
Let *i*, *j*be a pair of genes. The sample Pearson\'s correlation coefficient *σ*~*ij*~is the commonly used measure for coexpression. For examining possible effects of other genes *k*on *σ*~*ij*~, we consider GGMs for all triples of genes *i*, *j*, *k*with *k*≠ *i*, *j*. For each *k*, the partial correlation coefficient *ρ*~*ij*\|*k*~is computed and compared to *σ*~*ij*~. If the expression level of *k*is independent of *i*and *j*, the partial correlation coefficient would not differ from *σ*~*ij*~. If on the other hand, the correlation between *i*and *j*is caused by *k*since *k*co-regulates both genes, one would expect *ρ*~*ij*\|*k*~to be close to 0. Here, we use the terminology, that *k*\'explains\' the correlation between *i*and *j*.
In order to combine the different *ρ*~*ij*\|*k*~values in a biologically and statistically meaningful way, we define an edge between *i*and *j*if *ρ*~*ij*\|*k*~≠ 0 for all remaining genes *k*. In particular, if there is at least one *k*with *ρ*~*ij*\|*k*~= 0, no edge between *i*and *j*is drawn since the correlation between *i*and *j*may be the effect of *k*. Our approach can be implemented as a frequentist approach in which each edge is tested for presence or absence or alternatively, as a likelihood approach with parameters *θ*~*ij*~, which describe the probability for an edge between *i*and *j*in a latent random graph.
Frequentist approach
--------------------
For the gene pair *i*, *j*and all remaining genes *k*, p-values *ρ*~*ij*\|*k*~are obtained from the likelihood ratio test of the null hypothesis *ρ*~*ij*\|*k*~= 0. In order to combine the different *p*-values *ρ*~*ij*\|*k*~, we simply test whether a third gene *k*exists that \'explains\' the correlation between *i*and *j*. For this purpose, we apply the following procedure:
\(1) For each pair *i*, *j*form the maximum *p*-value
*p*~*ij*,max~= max{*p*~*ij*\|*k*~, *k*≠ *i, j*}.
\(2) Adjust each *p*~*ij*,max~according to standard multiple testing procedures such as FDR \[[@B32]\].
\(3) If the adjusted *p*~*ij*,max~value is smaller than 0.05, draw an edge between the genes *i*and *j*; otherwise omit it.
The correction for multiple testing in step 2 is carried out with respect to the possible number of edges (*q*(*q*- 1))/2 in the model. Implicitly, multiple testing over all genes *k*is also involved in step 1. However, because the maximum over all *p*~*ij*\|*k*~is considered, a multiple testing correction is not necessary.
Latent random graph approach
----------------------------
The frequentist approach has the disadvantage that a connection between two genes *i*and *j*is either considered to be present or absent. Also, it is not taken into account whether an edge between *i*and *k*respectively *j*and *k*is truly present when we test for *ρ*~*ij*\|*k*~= 0. In our second method, we introduce a parameter *θ*~*ij*~as the probability for an edge between two genes *i*and *j*in a latent random graph model. Let *θ*be the parameter vector of *θ*~*ij*~for all 1 ≤ *i*\<*j*≤ *q*and *y*= (*y*^1^,\...,*y*^*n*^) be a sample of *n*observations. For estimating *θ*, we maximize the log-likelihood *L*(*θ*) = log*P*~*θ*~(*y*) via the EM-algorithm \[[@B35]\].
Let *θ*^*t*^be a current estimate of *θ*. Further, let *g*be the unobserved graph encoded as an adjacency matrix with *g*~*ij*~∈ {0,1} depending on whether there is an edge between genes *i*and *j*or not. In the E-step of the EM-algorithm, the conditional expectation of the complete data log-likelihood is determined with respect to the conditional distribution *p*(*g*\|*y*,*θ*^*t*^),
By assuming independence between edges, Equation (1) becomes
and further, after replacing
and summing out Equation 2 we find
*P*(*g*~*ij*~= 1\|*y*,*θ*^*t*^) and *P*(*g*~*ij*~= 0\|*y*,*θ*^*t*^) at the right side of Equation (3) are approximated by the statistical evidence of edge *i*, *j*in GGMs with genes *i*, *j*and *k*. As we only want to estimate the effect of *k*on the correlation between *i*and *j*, we distinguish only the two cases whether *k*is a common neighbor of *i*and *j*, for example, *g*~*ik*~= 1 and *g*~*jk*~= 1 or not. When *k*is a common neighbor, we test *ρ*~*ij*\|*k*~≠ 0 versus *ρ*~*ij*\|*k*~= 0. When *k*is not a common neighbor of *i*and *j*, we test *σ*~*ij*~≠ 0 versus *σ*~*ij*~= 0 for the pairwise correlation coefficients instead. Thus, we obtain
where  and  are *p*-values of the corresponding likelihood ratio tests. After replacing Equation (4) in Equation (3), the M-step of the EM-algorithm, that is the maximization of *E*~*θ*~(log*P*~*θ*~(*g*)\|*y*,*θ*^*t*^) with respect to *θ*, leads to an iterative updating scheme *θ*^*t*^→ *θ*^*t*+1^with
In summary, we determine the probability parameters *θ*as follows
\(1) For gene pairs *i*, *j*, compute *P*(*ρ*~*ij*~\|*k*≠ 0) and *P*(*σ*~*ij*~≠ 0) for all genes *k*≠ *i*, *j*.
\(2) Starting with *θ*^0^, apply iteratively Equation (5) until the error \|*θ*^*t*+1^- *θ*^*t*^\| drops below a prespecified value, for example 10^-6^.
Our latent random graph approach also enables us to fit a large number of additional genes into a constructed genetic network. In this case, for a gene pair *i*, *j*in step 1 of the analysis, the partial correlation coefficients *ρ*~*ij*\|*k*~are not only computed and tested for genes *k*in the model but also for the additional candidate genes. However, the iteration in step 2 is not extended to these candidate genes. In other words, *θ*~*ij*~is only iteratively updated in Equation (5) if both genes *i*, *j*are in the original model. For candidate genes *k*, *θ*~*ik*~and *θ*~*jk*~are kept fixed at a prespecified value, for example 1, and are not re-estimated in the EM-iteration process.
This outline introduces a second level into the modeling process. At the first level, the network between the original genes is constructed. At the second level, we test how additional candidate genes influence the parameters *θ*. If these candidates have an effect on the correlation between *i*and *j*, *θ*~*ij*~will decrease. Thus, by comparing the original network with the network inferred from allowing for additional genes in step 1, we can determine which candidate genes lower the *θ*-values and, accordingly, fit well into the network.
Additional data files
=====================
Additional data is available with the online version of this paper. Additional data files [1](#s1){ref-type="supplementary-material"} and [2](#s2){ref-type="supplementary-material"} contain the gene expression values of the isoprenoid genes (Additional data file [1](#s1){ref-type="supplementary-material"}) and the 795 genes from other pathways (Additional data file [2](#s2){ref-type="supplementary-material"}). Additional data file [3](#s3){ref-type="supplementary-material"} contains a more detailed description of the microarray data (such as experimental conditions, hybridization and standardization). Additional data file [4](#s4){ref-type="supplementary-material"} describes the correlation pattern of the 40 isoprenoid genes.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
The gene expression values of the isoprenoid genes
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
The gene expression values of the 795 genes from other pathways
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 3
A more detailed description of the microarray data (such as experimental conditions, hybridization and standardization)
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 4
The correlation pattern of the 40 isoprenoid genes.
:::
::: {.caption}
######
Click here for additional data file
:::
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Bootstrapped GGM of the isoprenoid pathway. **(a)**Comparison between absolute pairwise correlation coefficients and presence of edges. Dots at 0 and 1 denote absent and present edges respectively. **(b)**Histogram of the bootstrap edge probabilities. **(c)**Comparison between absolute pairwise correlation coefficients and bootstrap edge probabilities for all 780 possible edges.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Bootstrapped GGM of the isoprenoid pathway with a cutoff at 0.8. The solid undirected edges connecting individual genes (in boxes) represent the GGM. Dotted directed edges mark the metabolic network, and are not part of the GGM. The grey shading indicates metabolic links to downstream pathways.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Dependencies between genes of the isoprenoid pathways according to the frequentist modified GGM method. **(a)**Subgraph of the gene module in the MEP pathway; **(b)**subgraph of the gene module in the MVA pathway. For an explanation of what the edges and shading indicate see legend to Figure 2.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Comparison of the absolute pairwise correlation coefficients and the modified GGM approaches. **(a)**Selected edges in the frequentist modified GGM approach (0 and 1 denote absent and present edges respectively). **(b)***θ*-values in the latent random graph approach. **(c)***θ*-values after attaching 795 genes from other pathways.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Hierarchical clustering of 40 genes involved in the isoprenoid pathway and 795 genes from other pathways. Clustering is depicted as a heatmap, in which red and green represent high and low expression values, respectively. Rows depict genes and columns depict hybridizations. Positions of the genes from the MEV pathway (m) and the plastoquinone and phytosterol pathways (+) are indicated in the left-hand column of the heatmap axis on the right side of the figure. Positions of the genes from the MEP pathway (n) and the plastoquinone, carotenoid and chlorophyll pathways (+) are indicated in the right column of the axis.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
Performance of different GGM approaches. **(a)**ROC curves and **(b)**the proportion of true-positive edges as a function of the number of selected edges for the different graphical modeling strategies. Black line, the standard GGM; red line, frequentist modified GGM approach; blue line, latent random graph modified GGM approach; green line, pairwise correlation. Sparse networks with fewer edges as nodes (*γ*= 2.5) are represented in the left column, networks with approximately as many edges as nodes (*γ*= 1.5) are represented in the middle column, and networks with approximately twice as many edges as nodes (*γ*= 0.5) are in the right column.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Genes coding for enzymes in the two isoprenoid pathways
:::
Name AGI number Subcellular location
--------- ------------ ----------------------
AACT1 At5g47720 C
AACT2 At5g48230 C
CMK At2g26930 P
DPPS1 At2g23410 C/ER
DPPS2 At5g58770 M
DPPS3 At5g58780 ER
DXPS1 At3g21500 P
DXPS2 At4g15560 P\*
DXPS3 At5g11380 P
DXR At5g62790 P\*
FPPS1 At4g17190 C
FPPS2 At5g47770 C/M\*
GGPPS1 At1g49530 M\*
GGPPS2 At2g18620 P
GGPPS3 At2g18640 C/ER\*
GGPPS4 At2g23800 C/ER\*
GGPPS5 At3g14510 M
GGPPS6 At3g14530 P
GGPPS7 At3g14550 P\*
GGPPS8 At3g20160 C/ER
GGPPS9 At3g29430 M
GGPPS10 At3g32040 P
GGPPS11 At4g36810 P\*
GGPPS12 At4g38460 P
GPPS At2g34630 P\*
HDR At4g34350 P
HDS At5g60600 P\*
HMGR1 At1g76490 C/ER\*
HMGR2 At2g17370 C/ER\*
HMGS At4g11820 C
IPPI1 At3g02780 P
IPPI2 At5g16440 C
MCT At2g02500 P\*
MECPS At1g63970 P
MK At5g27450 C
MPDC1 At2g38700 C
MPDC2 At3g54250 C
PPDS1 At1g17050 P
PPDS2 At1g78510 P
UPPS1 At2g17570 M
Subcellular locations are pooled from experimental data, the TargetP data base \[36\] and \[22\]. C, cytoplasm; ER, endoplasmic reticulum; M, mitochondrion; P, chloroplast. Experimentally verified subcellular locations are marked with an asterisk (\*).
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Pathways whose genes attach significantly well to the isoprenoid pathways
:::
Both isoprenoid pathways MEP pathway MVA pathway
-------------------------- ------------------------- -----------------
Plastoquinone\* Plastoquinone\* Plastoquinone\*
Carotenoid\* Carotenoid\* Phytosterol\*
Calvin cycle Porphyrin/chlorophyll\*
Histidine One carbon pool
One carbon pool Calvin cycle
Tocopherol\*
Porphyrin/chlorophyll\*
Downstream pathways are marked with an asterisk (\*). The Calvin cycle is also metabolically linked to the isoprenoid pathways.
:::
|
PubMed Central
|
2024-06-05T03:55:51.845187
|
2004-10-25
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545783/",
"journal": "Genome Biol. 2004 Oct 25; 5(11):R92",
"authors": [
{
"first": "Anja",
"last": "Wille"
},
{
"first": "Philip",
"last": "Zimmermann"
},
{
"first": "Eva",
"last": "Vranová"
},
{
"first": "Andreas",
"last": "Fürholz"
},
{
"first": "Oliver",
"last": "Laule"
},
{
"first": "Stefan",
"last": "Bleuler"
},
{
"first": "Lars",
"last": "Hennig"
},
{
"first": "Amela",
"last": "Prelić"
},
{
"first": "Peter",
"last": "von Rohr"
},
{
"first": "Lothar",
"last": "Thiele"
},
{
"first": "Eckart",
"last": "Zitzler"
},
{
"first": "Wilhelm",
"last": "Gruissem"
},
{
"first": "Peter",
"last": "Bühlmann"
}
]
}
|
PMC545784
|
Background
==========
There is a strong belief underpinning systems biology that between the individual molecules and an organism\'s phenotype there exist intermediary levels of organization \[[@B1]\]. The lowest level, and one that can be objectively defined, is that of the motif, for example a feedforward loop \[[@B2]-[@B5]\]. At the next level there exist putative modules within networks \[[@B6]-[@B16]\]. However, unlike motifs, modules are not objectively defined and are hence rather fuzzy. Moreover, even if a stringent definition or sophisticated algorithms could be envisaged, the data used to identify such modules are typically very noisy, for example, protein-protein interaction data. The central problem \[[@B17]\] with the notion of modules, therefore, is not identifying putative candidates but verifying which of them really reflect an important level of biological organization, rather than artifacts of the data or module-defining protocol. In addition, it would be of interest to determine the minimal information needed to identify such candidates, so that this level of organization can be readily probed, even in relatively poorly characterized systems.
Given that we could define such modules for a particular data source, for example, protein-protein interactions, there exists the further problem of understanding how modules relate to other forms of organization. Do for example, the proteins in a given module within a protein-protein interaction network show evidence of being coexpressed? Are they regulated by the same transcription factors and do they have the same level of dispensability?
Whether we can define modules in a stringent biologically relevant fashion is not just important for our understanding of the organization of biological systems. Many authors have conjectured that if modules are real they may also be more likely to contain proteins that are essential to viability. Hence, a network approach could be imagined to hone down potential drug targets such as, for instance, candidate targets for antimicrobials.
Here we ask whether phylogenetic information could be used to verify putative interaction-based modules. The assumption we make is that if a set of proteins belongs to the same module and that module has some biological relevance, then such a set should be generally conserved to act as an integrated functional unit \[[@B18],[@B19]\]. Hence we should expect a genome to contain roughly all the set components or none. The extent to which we find the module components present or absent together we define as the \'phylogenetic correlation\' of the module. We show that this correlation can be used to verify putative modules in a network context and that the modules identified in this way have important biological properties.
Results and discussion
======================
Extracting modules in protein networks
--------------------------------------
Several network-clustering algorithms have been developed recently that make use of the local and global properties of networks \[[@B9]-[@B11]\]. To this end, it is helpful to represent networks as graphs, with proteins playing the role of nodes and protein-protein interactions playing the role of edges between nodes. In such graphs, the presence of modular topology could be manifested in the fact that the shortest distance, *L*, between any given node and the rest of the nodes in the graph would exhibit a similar pattern for those nodes belonging to the same module. Alternatively, modularity could also imply that proteins within a module would interact more frequently with each other than with proteins of different modules, a property characterized by high values of a generalized clustering coefficient, *C*(see Materials and methods).
We introduce here a simple algorithm that makes use of both sources of information. The basic steps of the so-called overlap algorithm are as follows (see also Materials and methods and Figure [1a](#F1){ref-type="fig"}).
### Selection of the number of modules
*C*-based and *L*-based matrices were obtained from the interaction matrix. These matrices are the input data of a standard hierarchical agglomerative average-linkage clustering algorithm with a Pearson-based distance metric \[[@B20]\]. We obtained as an output of the clustering different sets of modules associated to each matrix by delimiting clusters according to a given number of branches present in the clustering tree () (discarding those ones containing just a single protein). In the next step we calculated an average overlap between both modular structures. A -value with significantly high maximal overlap was then chosen.
### Extraction of a particular modular structure
Having obtained *C*-based and *L*-based modules with a -value selected as previously described, we calculated the overlap of each *C*-based module with all those obtained with the *L*-based method. An *L*-based method less efficiently discriminates modular structures in small-world networks \[[@B21]\], collapsing some of the modules extracted with the *C*-based technique into a unique module. The *C*-based method is more robust but is weak at discriminating modules when organization levels are high. Therefore we used the *C*-based results as a template and the *L*-based method as a filter in the extraction of modular structure. In the *C*-based modular structure we kept in each module only those components which also appeared in the corresponding *L*-based module with which the selected *C*-module had the greatest overlap. In those cases with more than one module with maximal overlap, we selected one of them at random. Although finding the optimal classification choice is a common problem of clustering analysis, this simple algorithm allows one to select a -value with a high average maximal overlap and low overlap ratios between both methods, a measure of the reliability of the obtained modules (see Materials and methods and Additional data file 1 for more details).
The overlap method was applied to the yeast protein-interaction network; that is, yeast would act as an imaginary \'poorly\' characterized system where we can, however, check the relevance of our findings. This was derived from two public databases (see Materials and methods) and would be, more generally, the result of high-throughput experiments. In any case, these data are probably incomplete and no doubt contain false interactions \[[@B22]\]. Should the analysis be done on the whole network? Certainly this could be done - and many similar analyses have been done. However, one of the novelties of the current analysis is that we perform the analysis on sub-parts. This is because we are interested in knowing whether different functional categories differ in the extent to which they might be modular \[[@B1]\], not least because we also want to know whether this modularity might be reflected in such things as coexpression of the genes involved. This tendency is likely to vary by functional class. For example, cell-cycle genes should in principle show a strong coexpression signal if the modules are real. In contrast, one might imagine that all cell-signaling components need to be present under all circumstances and so coexpression need not be detectable. Analyzing the network as a whole, one might come to conclude that there exists no or just a weak correspondence between modules and coexpressed genes, when in reality there might be a very strong relationship for some categories while none for others.
We therefore opted to analyze networks consisting of proteins belonging to different Munich Information Center for Protein Sequences (MIPS) protein functional categories \[[@B23]\]. This also has some methodological advantages. First, as methods for detecting protein-protein interactions may vary systematically according to functional grouping - for example, cytoplasmic complexes tend to be under-reported - it can be helpful to isolate each grouping alone. Second, it is probably desirable to filter out highly connected proteins to avoid big hubs and star-like clusters with low statistical significance \[[@B9]\]. Projecting the networks onto functional categories is a possible way of achieving such a filter. In every functional network, we found a regime of -values with significantly high average maximal overlap, that is, overlap equal to or greater than 0.8, and low ratios, characterizing the reliability of the proposed modular organizations. For an analysis of the performance of the algorithm as a function of -see Additional datafile 1. Note that these results extend the presence of modularity found previously in some yeast networks \[[@B9],[@B10],[@B24]\] to the functional networks introduced here. Explicit -values in the regime described above were chosen such that the average module size is around  equal to 5 to 25 proteins, the so-called meso scale of biological networks \[[@B9]\] (Table [1](#T1){ref-type="table"}).
Modular phylogenetic profiles
-----------------------------
To ask whether the degree of phylogenetic correlation of the modules is higher than expected, we made use of the idea of phylogenetic profiles \[[@B18]\]; that is, patterns of presence or absence of homologs of a given protein across different genomes. We then adapted the underlying general assumption of phylogenetic profiles, that proteins belonging to a particular functional class should display a similar pattern of homologs in a set of organisms, to a more restricted hypothesis. We considered that modules within functional networks could indeed reflect a stronger functional link among their components than with the rest of the proteins. This stronger functional link, even when all proteins in the networks are part of the same functional classification, could consequently be reflected in the correlated presence or absence of module components across different organisms - that is, their phylogenetic profiles.
To verify this initial suggestion, we examined the corresponding null hypothesis, that there is no phylogenetic correlation of the proposed structures, which is based on a completely uncorrelated distribution of phylogenetic profiles with respect to the modular organization. We made use of a class of statistical methods termed multi-response permutation procedure (MRPP). MRPPs are commonly used in ecological and environmental studies to compare an *a priori*group classification of a population in which measurements of *r*responses (*r*≥ 1) are obtained from each member of the population \[[@B25]\]. In contrast to well-known parametric statistical techniques such as the univariate and multivariate analysis of variance, MRPPs do not require any assumption with respect to the distribution of the response measurements. In the present case, proteins are the members of the population, modules are the group classification, and the phylogenetic profiles play the role of response measurements. A further difference from standard statistical techniques is that similarity measures, or normed distances, and not individual object measurements, are the primary units of analysis.
We compared the within-module scores to the between-module scores. For each pair of modules we calculated each between-module protein pairwise similarity and took the average of these. To examine overall between-module similarity we calculated a weighted mean correlation of all between-module similarities. We then asked about the size of the difference between the mean within-module score and the mean between-module score, that is, *ξ*=  -  (see Materials and methods). Significance was tested by randomization; that is, we randomly permute the proteins within the modules while keeping the global modular organization fixed (Figure [1b](#F1){ref-type="fig"}). Not all putative network modular organizations, according to different s, are shown to be biologically significant. However, we find for all networks a strong signal of phylogenetic correlation between genes in a module for some -values within the regime of high reliability of the algorithm (Table [1](#T1){ref-type="table"} and Additional datafile 1).
We can extend the analysis to identify those modules showing the strongest signal. We used a method based on the analysis of each within-module similarity and the use of mean similarity dendrograms. For every module, we subtracted from the mean within-module similarity *W*~*m*~, the mean of all between-module similarity , a sort of representative of all pairs of between-module similarities: that is, *ξ*~*m*~= *W*~*m*~- . We estimated the significance of the values observed with such a modular test by performing again an approximate permutation procedure with a Holm\'s correction to multiple testing (Figure [1b](#F1){ref-type="fig"} and Materials and methods). This gives a significance measure of which module similarities reflect correlated evolution of their components in a particular functional network.
Statistical significance does not supply any information on the magnitude of the respective similarities. To this end, we constructed a graphical representation, a mean similarity dendrogram \[[@B26]\], with branches for each module joined at a node plotted at . Branches terminate at *W*~*m*~, giving branch lengths of *ξ*~*m*~in similarity units (Figure [2](#F2){ref-type="fig"}). Those branches with considerable positive length, for example, *ξ*~*m*~equal to or greater than 0.1, indicate correlated evolution of the respective module components according to the phylogenetic profiles of the whole functional network, even though some of them could not be shown to be statistically significant because of the conservative nature of Holm\'s test. Thus, this combined approach provides both statistical significance and a clear quantitative picture of the compactness and isolation of the proposed modules. Figure [2](#F2){ref-type="fig"} shows two examples of the application of this approach to evaluate modular network structures with the use of mean similarity dendrograms and phylogenetic profiles (we have chosen two small networks as examples to show a full picture of the modular characterization). Network phylogenetic profiles can be easily visualized as a matrix whose columns display the presence or absence of network nodes in a given organism and whose rows show the presence or absence of a given node in all the organism set. It then presents a full view of the degree of conservation of network modules for a collection of organisms. The arrangement of species in taxonomic groups is a convenient representation of the relative conservation of modules across the different lineages.
Module cores
------------
Previous studies suggest that any given module may have a module core and a periphery \[[@B10]\]. In addition, in an evolutionary context, it is not clear to what extent full modules should be present or absent in different species, considering the tinkering aspect of most evolutionary processes. Can we use the network method to discriminate a core and does the core have a stronger phylogenetic correlation? To examine this hypothesis, we selected the most connected components of each module that was part of a given network, according to their intra-modular connectivity, and applied again the overall and modular tests to these cores (see Materials and methods). We found a substantial increase in the validation of the evolutionary significance of the modules revealed, for example, by the presence of a bigger number of significant modules (Table [1](#T1){ref-type="table"}, \'core\' column group). Such statistically significant cores are mainly characterized by two distinct phylogenetic profiles; either their components had profiles with homologs present in all three kingdoms, or they had homologs present only in Eukarya (Table [2](#T2){ref-type="table"}). This agrees with previous results and seems to support a picture of network assembly with a combination of ancient and modern modules \[[@B12],[@B24],[@B27]\].
The phylogenetic correlation suggests that this core architecture is biologically meaningful. Such extracted structures could then be used to probe this intermediate level of organization even in the case of uncharacterized biological systems. Owing to the extensive biochemical knowledge about yeast we are ready to validate such hypothesis. We have made use of the MIPs yeast complexes database \[[@B12],[@B24]\] to characterize the biological relevance of the cores (see Additional data file 1 for a full list of phylogenetically distinct module cores and their biological characterization). As suggested, many, but not all, of the cores describe a significant part of relevant protein complexes, for example, anaphase-promoting complex, prenyltransferases (Ftase, GGTase I and GGTase II), some cytoplasmic translation initiation complexes such as eIF2 and eIF2B, Kel1p/Kel2p complex and Gim complexes (Table [3](#T3){ref-type="table"}). Other module cores are not identified as parts of known protein complexes. This could mean either that some of the cores correspond to uncharacterized complexes or that these cores represent dynamic modules. Dynamic modules control a particular cellular activity by means of interactions of different proteins at different times or places instead of by the assembly of a macromolecular machine \[[@B1]\]. Thus, the combination of modular analysis and phylogenetic correlation is useful to find relevant components of biological systems.
Do we also find that the significantly phylogenetically correlated cores have other properties of biologically relevant cores, that is, show a high degree of coexpression? We examined both the extent of coexpression \[[@B28]\] and degree of similarity in 5\' motifs \[[@B29]\], the latter being an indirect method of assaying possible expression parameters. As regards coexpression, most functional groups have cores with more similar coexpression than expected by chance, but the significance levels tend to be low and hence the effect, while widespread, is relatively weak. This is probably a consequence of the dynamic organization of modularity \[[@B15]\], a phenomenon previously observed in protein complexes \[[@B28]\] (Table [4](#T4){ref-type="table"} and Materials and methods). This weakness is similarly reflected in the extent of sharing of 5\' motifs. This latter result is probably as expected, given a lack of certainty over the relevance of many motifs and the fact that two genes of similar expression profile can have different motifs.
Do the modules also represent units of homogeneity of dispensability? That is, if one protein in the core is lethal are all lethal, if one is dispensable are all dispensable? This can be quantified by the absolute distance of the ratio of lethal proteins in the core (0 ≤ ratio ≤1) to 1/2. We then sum these distances for the relevant cores in each network and estimate statistical significance by randomization (Figure [1b](#F1){ref-type="fig"}). We find some cases where there is indeed higher homogeneity than expected (Table [4](#T4){ref-type="table"}). But does this also mean that the modules all contain more lethals than expected? We find that for some functional groups this is indeed very profoundly the case. However, for other functional groups this is not so (Table [4](#T4){ref-type="table"}).
Assuming that the putative functional group of a protein can be assigned blind to genes, this method then has the potential to narrow down the possible drug targets in poorly described species. Perhaps as expected, cell-cycle, protein synthesis and transcription-related modules have the most significant tendency to amass lethal genes. Could we apply the knowledge of validated network structures in a therapeutical context, for instance to identify targets for antimicrobials? In principle, identifying candidate proteins as antimicrobial targets is straightforward: the protein needs to be in the microbe and not the host and to be essential to the microbe. To this end, we calculated the probability of finding lethal genes in the set of proteins without human homolog belonging to the significant cores. We compared this with the probability of finding lethal genes in those yeast proteins not found in humans which are part of the full network. While the data on which genes are essential is questionable, owing to condition-dependent lethality \[[@B30]\], the ratio of these two measures should give an indication of the extent to which our method improves the search strategy. Crucially, the method greatly increases the probability of finding such essential genes (Table [4](#T4){ref-type="table"}). Some of these targets in yeast could be, for instance, the proteins APC4, ORC6 or POP5, which are part of complexes involved in the functional categories mentioned earlier (see Additional data file 1 for a detailed list).
Conclusions
===========
We have shown that by combining protein-protein data and phylogenetic information it is possible to systematically describe biologically relevant modules in protein networks which partially correlate with other types of organization. The analysis also suggests, however, that not all core modules within the functional network are equally vital for the organism\'s survival. This may just reflect condition-dependent lethality \[[@B30]\]. Indeed, the fact that fewer than half of the core metabolic modules show significant enrichment for lethal genes is possibly due to such condition-dependency. Given this result, in the development of antimicrobials it seems wiser to attack modules related to transcription, protein synthesis and the cell cycle than it is to attack metabolic pathways. This simple example hints at the relevance of knowledge about the modular organization of networks in other therapeutic settings, such as that in cancer, to home in on which modules and which parts of modules within these systems should be selected in a putative list of potential drug candidates. Overall, our results contribute to validate the relevance of the modular level of organization of biochemical networks.
Materials and methods
=====================
Data
----
We used two databases as of July 2003: MIPS \[[@B23]\], contributing 9,036 protein interactions; and DIP \[[@B31]\], contributing 15,116 interactions. Networks were assembled using a joint set of interactions after filtering common pairs. Protein information for the fully sequenced organisms selected is available at the website of the European Bioinformatics Institute \[[@B32]\]. A dataset on the presence of 5\' regulatory motifs was downloaded from the Church Laboratory \[[@B33]\]. Expression data was obtained from a whole-genome mRNA expression data compiled by the Eisen laboratory \[[@B34]\].
Network clustering matrices
---------------------------
Network clustering can be based on a global property, that is, *L*-based clustering, where *L*is referred to the shortest path length between two nodes in the network. From the interaction network, a matrix of distances is computed and transformed into an \'association\' matrix by taking 1/*L*^2^\[[@B10]\]. A second approach to network clustering is based on a local property, *C*-based clustering, where *C*is a generalized local connectivity coefficient measuring common interactors of any two proteins in the interaction graph \[[@B8],[@B9],[@B11]\] given by
Here \|\...\| denotes the size of the set, ∩ the intersection and *Adj*(*i*) the adjacency matrix, that is, the set of proteins interacting with protein *i*. Local properties tend to be more robust \[[@B11]\].
Module overlap
--------------
Given two different modules, *M~i~*, *M~j~*, we considered the following overlap \[[@B13]\]:
with \|\...\| denoting the size of the set and ∩ the intersection. The average overlap used to determine the number of branches present in the clustering tree () is given by:
In this case, \|*C*\| and \|*L*\| denote the number of *C*-based and *L*-based modules extracted in a given functional network.
Network small-worldness
-----------------------
To characterize the small-world property of the networks, we first calculated the clustering coefficient, , and characteristic path length, *L*, for all assembled networks.  = 2*j*/*m*(*m*- 1), the ratio between the number of interactions found among the *m*proteins connected to a given one, say *j*, and the maximal potential number of such interactions, which equals *m*(*m*- 1)/2 for a undirected graph. We obtained high values of such clustering coefficient and small characteristic path length for all cases, reflecting the small-worldness of the networks. To assess the statistical significance of these values, we generated 100 randomly rewired graphs for each functional network with the algorithm described in \[[@B21]\]. All cases were shown to be highly significant (*P*= 0.01), that is, , and *L*≥ *L~random~*(we obtained *P*\< 0.05 for *L*in the case of the energy network).
Phylogenetic profiles
---------------------
We calculated binary and continuous phylogenetic profiles \[[@B18]\] for different threshold values, obtaining robust results for all discussed tests in both cases. For each yeast protein of interest, BLAST searches were done against 70 proteomes of species from the Archaea (14), Bacteria (47), and Eukarya (9) (see organism list in Additional data file 1). BLAST hits with Karlin-Altschul *E*-values bigger than a given threshold, *E~th~*, were considered absent \[[@B35]\]. A particular value is then assigned to each homolog present, characterizing in this way every protein by means of a phylogenetic vector. For continuous profiles, homologs receive a score of -1/log*E*and the absent ones receive a score of -1/log*E~th~*. For the binary case, profiles take the value 1 or 0 when the *E*-values are below or above the threshold, respectively. Finally, note that *E*-values were corrected to account for the different database sizes. Results in the main text are for the case of binary phylogenetic profiles and a threshold value of *E~th~*= 1*e*^-6^.
Multi-response permutation procedures
-------------------------------------
Non-parametric randomization methods, such as MRPP, have several advantages compared to more well-known parametric procedures. In particular, if the assumption of normally distributed populations is not reasonable, the datasets have multiple measurements and if multivariate comparisons are desired \[[@B25]\].
### Similarity measure
Given two binary phylogenetic profiles corresponding to proteins *i*, and *j*, we considered the following matching coefficient as a simple similarity measure: *S*~*ij*~= (*x*+ *w*)/(*x*+ *y*+ *z*+ *w*), where *x*is the number of homologs present in both phylogenetic profiles, *y*is the number present in profile *i*only and *z*is the number present in profile *j*only. Finally, *w*is the number of absent homologs in both profiles.
### Mean within and between similarities
Within similarity
Here, *c~m~*is the ratio between the number of components of module *m*, *n~m~*, and the number of components of all modules, *N~M~*, that is, *c~m~*= *n~m~*/*N~M~*, *W~m~*is the mean of similarities between proteins belonging to module *m*, and *M*is the total number of modules.
Between similarity:
Here, *c~m,s~*is the ratio between the product of the number of components of modules *m*and *s*, *n~m-~n~s~*, and the total number of components squared, *N*^2^~*M*~, that is *c~m,s~*= *n~m~n~s~*/*N*^2^~*M*~. *W~m,s~*is the mean of similarities between proteins of modules *m*and *s*, and *M*is the total number of modules. Results for all discussed tests were robust to the use of Euclidean distances with continuous profiles instead of similarities with binary profiles, as it is argued in the main text.
Holm\'s test
------------
The Holm test \[[@B36]\] is a method that gets round the problem of the Bonferroni procedure being too conservative, by means of the added power of sequential stepping versions of the traditional Bonferroni tests. The procedure behind the Holm test is to find all the *P*-values for a set of *k*individual tests that are being performed and then rank them from smallest to largest. While Bonferroni would compare all null hypothesis to the same value *α*, the Holm test compares the smallest to *α*/*k*and, in case of rejection of the null case, to decreasing values *α*/(*k*- 1),\... until failing to reject the null.
To perform the MRPP Holm test, we computed the branch length, that is, *W~m~*-  (see above) and determined the unadjusted *P*-value for each module by means of a permutation test with 10,000 randomizations. Suppose that we have *M*modules. We assemble an ordered vector of size *M*whose components are the uncorrected *P*-values in increasing order, that is, *P*~1~is the smallest uncorrected *P*-value and *P*~*M*~is the largest. To adjust a particular vector component *P~i~*we multiply this component by *A*~i~= (*M*- *i*+ 1), thus generating a vector *P*for adjusted *P*-values. The added power of the Holm test can then be seen in a simple example. Imagine the case of three modules, that is, *M*= 3. The uncorrected *P*-values of the corresponding MRPP tests are: *P~ν~*= (0.01, 0.02, 0.03). A Bonferroni procedure for multiple testing would consider only the first test as significant according to a 0.05 significancy threshold. However, the adjusted *P*-values obtained with the Holm test would imply that all tests are significant, that is,  = *P~ν~*× (3,2,1) = (0.03, 0.04, 0.04).
Core components
---------------
To obtain the core component of the modules, we selected for each module those components with more than two interactions, for the case of a module whose component with maximal number of interactions (MNI) is less than ten, or those components with more than four interactions for the case of a module whose component with MNI is equal to or greater than 10. Slight modifications to these rules produced similar results.
5\' regulatory motifs, coexpression and lethality of module cores
-----------------------------------------------------------------
For each of the significant module cores, *ξ*~*m*~≥ 0.1, we calculated the mean of pairwise Euclidean distances between expression vectors of proteins belonging to a given module core. In the case of the 5\' motifs, the statistic measures the number of regulatory motifs common to at least more than half of the core size. Finally, for each significant core, we simply measured the number of components that are lethal. The overall statistic for all cases is the sum of each corresponding measure in each core weighted by the ratio of the core size vs network size. *P*values are obtained with 10,000 randomizations.
Additional data files
=====================
Additional data file [1](#s1){ref-type="supplementary-material"}, available with the online versin of this article, includes a discussion on the network clustering algorithm, the list of species and lineages for the phylogenetic profiles, and a list of phylogenetically distinct module core components and their biological characterization.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
A discussion on the network clustering algorithm, the list of species and lineages for the phylogenetic profiles, and a list of phylogenetically distinct module core components and their biological characterization
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
J.F.P thanks H.J. Dopazo, R. Díaz-Uriarte and, especially, J. Van Sickle for fruitful discussions, and the Evolutionary Systems Biology Initiative at CNIO and M. Baena for valuable comments. This research has been supported by the Spanish MCyT (Ministry of Science and Technology) Ramón y Cajal Program (J.F.P) and the UK Biotechnology and Biological Sciences Research Council (L.D.H.).
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Overlap algorithm and multi-response randomization test method. **(a)**Overlap algorithm. *C*-based and *L*-based matrices are obtained from the interaction matrix. These matrices are then the input data of a standard hierarchical agglomerative average-linkage clustering algorithm \[20\] which extracts modules according to a given number of branches present in the clustering tree () (see text). Finally, in the *C*-based modular structure, we kept in each module only those components which also appeared in the corresponding *L*-based module with which the selected *C*-module had the greatest overlap. The organization thus obtained is the putative modular organization of the network under consideration. **(b)**Multi-response permutation procedure. We validate the previous modular organization with the use of the phylogenetic conservation of module protein constituents across species. We calculate a matrix of mean pairwise similarities (or distances) among those phylogenetic profiles \[18\] of proteins belonging to the same module, *W~i~*, or every two pairs of modules, *W~ij~*, and computed a representative statistic *ξ*~*observed*~. *P*-values are obtained by randomly permuting the data and recomputing the statistic. This step is repeated a large number of times, 10,000 in our case. The resulting values form a randomized distribution. The observed value from the original data can then be compared with this distribution to compute the *P*-value.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Modular organization, mean similarity dendrogram and phylogenetic profile. Modular organization, mean similarity dendrogram and phylogenetic profile of **(a-c)**cellular rescue, and **(d-f)**cellular environment functional networks. (a-d) Modular organization extracted with the network clustering algorithm. Protein interactions are plotted in brown. Modules are highlighted in white. Proteins within each module have been reorganized to show those with the greatest intra-modular connectivity - the core proteins - in the center of the module. (b,e) Mean similarity dendrograms. Branches for each corresponding module in (a) and (d) are joined at a node plotted at . Branches terminate at the mean similarity of each module, *W~m~*, giving branch lengths of *W~m~*-  in similarity units. Dendrograms related to full modules are in black and those corresponding to the core components are in red. Those branches statistically significant (*P*\< 0.05) end in a circle. (c,f) Continuous phylogenetic profiles color-coded from dark blue (maximal homology) to brown (no homology). Columns show the presence or absence of network nodes in a given organism and rows show the presence or absence of a given node in all the organism set. Species are arranged in taxonomic groups separated by white dashed vertical lines: Bacteria (left), Archaea (center), and Eukarya (right) (see Additional data file 1). The horizontal white dashed lines represent the localization of modules. A quick look at these figures provides evidence that proteins that are part of the same module exhibit a loosely correlated degree of conservation, as should be the case if modules represent some sort of discrete functional unit. This argument is quantitatively estimated by the branch length in the mean similarity dendrogram and the corresponding statistical significance.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Global and follow-up analysis of the network modular organizations
:::
Function Full Core
------------------------ ----- ----- ---- ------- --------- ----- ------ ------- ---------- ------ ------
Cellular fate 34 323 14 0.012 \<0.001 2/5 16.7 0.035 \<0.001 3/6 6.5
Energy 25 84 5 0.066 \<0.001 1/1 12.4 0.156 \<0.001 1/4 4.4
Metabolism 102 420 15 0.067 \<0.001 2/8 15.7 0.177 \<0.001 4/9 4.7
Cellular transport 32 336 15 0.014 \<0.001 2/5 18.7 0.021 \< 0.001 -/2 10.8
Cell cycle 26 514 13 0.012 \<0.001 2/3 26.6 0.05 \<0.001 2/7 8.5
Protein fate 48 352 18 0.014 0.004 -/9 15.3 0.03 0.001 -/10 8.7
Transport facilitation 20 63 4 0.034 0.047 1/1 10.7 0.372 0.097 1/1 6.5
Cellular environment 18 87 8 0.037 0.007 2/3 8.5 0.072 0.002 3/4 5.6
Protein synthesis 16 137 7 0.038 0.002 1/1 17.3 0.194 \<0.001 2/5 4.8
Cell rescue 26 88 8 0.08 \<0.001 1/2 7.7 0.108 \<0.001 1/3 4.2
Signaling 14 67 6 0.017 0.082 -/2 9.3 0.018 0.157 -/2 6.2
Cellular organization 36 258 15 0.032 \<0.001 1/7 12.3 0.097 \<0.001 3/9 5.3
Transcription 40 654 21 0.019 \<0.001 2/7 25.1 0.037 \<0.001 4/9 12.3
For every functional network of size *n*, we applied the network clustering algorithm with a given number of branches in the clustering tree, . These -values were chosen to be among those with significantly high average maximal overlap, that is, overlap equal to or greater than 0.8, low overlap ratios, and meso-scale average module size, that is, \~5-25. The outcome of this algorithm is a modular organization with *M*modules. For the follow-up analysis of both full and core components of the modules, third and fourth column groups, the following quantities are shown: *ξ*, the overall statistic, *P*, statistical significance of global test, *P~m~*^†^, number of modules whose branch length in the similarity dendrogram (see text for details) is bigger than 0.1 in similarity units and *P~m~*, number of modules whose within-similarity is statistically significant (*P*\< 0.05) in the modular test. All *P*-values were obtained by means of an approximate permutation test with 10,000 randomizations and the use of binary phylogenetic profiles with a threshold of *E~th~*= 1*e*^-6^in the BLAST *E*-value \[35\].
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Conservation properties of module core components for those functional networks with more than one statistically significant module core
:::
Conservation
----------------------- -------------- ------
Cell fate 0(0) 6(3)
Metabolism 3(1) 6(3)
Cellular organization 3(0) 6(3)
Cellular environment 3(2) 1(1)
Protein synthesis 3(0) 2(2)
Transcription 1(1) 8(3)
Cell cycle 0(0) 7(2)
Conservation of components follows two distinct patterns: module core components are conserved in all three kingdoms: (B,A,E) Bacteria, Archaea and Eukarya, or are only present in eukaryotes, (-,-,E). The table shows the number of module cores, with branch length *ξ*~*m*~≥ 0.1, whose components have a representative phylogenetic profile of either type. Conservation profiles of statistically significant core components is shown in parenthesis. See also Table 1.
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
List of complexes significantly represented in the phylogenetically distinct module cores
:::
Function Cores (r~cc~≥ 5) Complexes
------------------------ ------------------ -----------------------------------------------------------------
Cell fate 6 (2) Actin-associated motor protein, 431
Energy 4 (2) 47, 346, Serine/threonine phosphoprotein phosphatase
Metabolism 9 (3) 521, GGTase II, OT
Cellular transport 2 (2) Class C Vps, 239, 77, AP-3, AP-2
Cell cycle 7 (4) Tubulins, CA, AP, 3, OR, SCF-GRR1, SCF-CDC4, RI
Protein fate 10 (5) Vps, Class C Vps, 71, 77, FT, GGTase I, 168, 651, OT, AP, 23
Transport facilitation 1 (1) TOM
Cell environment 4 (3) STE5-MAPK, Kel1p/Kel2p, 521
Protein synthesis 5 (2) elF3, elF2B, elF2, 340, 339, 613
Cell rescue 3 (3) No complexes
Signaling 2 (1) 167, 308, 521
Cell organization 9 (6) 272, 5, 71, 289, casein kinase II, 181, 167, Gim
Transcription 9 (6) 154, RM, RP, Ma, Cbf, Mb, 126, NSP1, TF, 178, CPK, 634, 160, CF
Numbers correspond to those complexes found by systematic analysis as described in MIPS \[23\]. Abbreviations: AP, anaphase-promoting complex; CA, chromatin-assembly complex; Cbf, Cbf1/Met4/Met28; CF, core factor; CPK, cAMP-dependent protein kinase; FT, farnesyltransferase; GGTase I, geranylgeranyltransferase I; GGTase II, geranylgeranyltransferase II; Ma, Met4/Met28/Met32; Mb, Met4/Met28/Met31; OR, origin-recognition complex; OT, oligosaccharyltransferase; RI, replication initiation complex; RM, RNase MRP; RP, RNase P; TF, TFIIIC; TOM, transport across the outer membrane complex; Vps, Vps35/Vps29/Vps2. Here, *r~cc~*is the ratio between the number of complex components being part of a core and the total number of complex constituents.
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Statistical significance of the overall analysis of coexpression, common 5\' regulatory motifs, homogeneity in dispensability and lethality for the phylogenetically distinct module cores
:::
Function *P*-exp *P*-mot *P*-hom *P*-let *p*-core *p*-net
------------------------ ---------- --------- ----------- --------- ---------- ---------
Cell fate \<0.05 \- \- \<0.05 0.28 0.08
Energy \- \<.005 \- \- 0 0.05
Metabolism \<0.0005 \<0.05 \- \<0.01 0.14 0.08
Cellular transport \- \- \< 0.01 \- None 0.28
Cell cycle \<0.05 \- \< 0.05 0.0001 0.35 0.29
Protein fate \<0.0005 \- \- \- 0.41 0.16
Transport facilitation \- \- \- \- 0.5 0.15
Cell environment \- \- \- \<0.05 0 0.06
Protein synthesis \<0.05 \- \< 0.0005 0.0001 0.2 0.06
Cell rescue \<0.05 \- \- \- 0 0.12
Signaling \- \- \- \- 0 0.12
Cell organization \<0.01 \<0.05 \- \- 0.08 0.12
Transcription \<0.05 \<0.01 \<0.01 \<0.001 0.68 0.3
Statistical significance (*P*-values), of the overall analysis of coexpression (*P*-exp), common 5\' regulatory motifs (*P*-mot), homogeneity in dispensability (*P*-hom) and lethality (*P*-let), for the phylogenetically distinct module cores (see text and Materials and methods for details). Not significant statistical results are denoted by -. *p*-core is the probability of finding lethal genes in the set of proteins without human homolog belonging to the significant cores. *p*-net is the probability of finding lethal genes in those proteins not found in humans which are part of each full network.
:::
|
PubMed Central
|
2024-06-05T03:55:51.848560
|
2004-11-1
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545784/",
"journal": "Genome Biol. 2004 Nov 1; 5(11):R93",
"authors": [
{
"first": "Juan F",
"last": "Poyatos"
},
{
"first": "Laurence D",
"last": "Hurst"
}
]
}
|
PMC545785
|
Background
==========
There are many practical applications that involve the grouping of a set of objects into a number of mutually exclusive subsets. Methods to achieve the partitioning of objects related by correlation or distance metrics are collectively known as clustering algorithms. Any algorithm that applies a global search for optimal clusters in a given dataset will run in exponential time to the size of problem space, and therefore heuristics are normally required to cope with most real-world clustering problems. This is especially true in microarray analysis, where gene-expression data can contain many thousands of variables. The ability to divide data into groups of genes sharing patterns of coexpression allows more detailed biological insights into global regulation of gene expression and cellular function.
Many different heuristic algorithms are available for clustering. Representative statistical methods include k-means, hierarchical clustering (HC) and partitioning around medoids (PAM) \[[@B1]-[@B3]\]. Most algorithms make use of a starting allocation of variables based, for example, on random points in the data space or on the most correlated variables, and which therefore contain an inherent bias in their search space. These methods are also prone to becoming stuck in local maxima during the search. Nevertheless, they have been used for partitioning gene-expression data with notable success \[[@B4],[@B5]\]. Artificial Intelligence (AI) techniques such as genetic algorithms, neural networks and simulated annealing (SA) \[[@B6]\] have also been used to solve the grouping problem, resulting in more general partitioning methods that can be applied to clustering \[[@B7],[@B8]\]. In addition, other clustering methods developed within the bioinformatics community, such as the cluster affinity search technique (CAST), have been applied to gene-expression data analysis \[[@B9]\]. Importantly, all of these methods aim to overcome the biases and local maxima involved during a search but to do this requires fine-tuning of parameters.
Recently, a number of studies have attempted to compare and validate cluster method consistency. Cluster validation can be split into two main procedures: internal validation, involving the use of information contained within the given dataset to assess the validity of the clusters; or external validation, based on assessing cluster results relative to another data source, for example, gene function annotation. Internal validation methods include comparing a number of clustering algorithms based upon a figure of merit (FOM) metric, which rates the predictive power of a clustering arrangement using a leave-one-out technique \[[@B10]\]. This and other metrics for assessing agreement between two data partitions \[[@B11],[@B12]\] readily show the different levels of cluster method disagreement. In addition, when the FOM metric was used with an external cluster validity measure, similar inconsistencies are observed \[[@B13]\].
These method-based differences in cluster partitions have led to a number of studies that produce statistical measures of cluster reliability either for the gene dimension \[[@B14],[@B15]\] or the sample dimension of a gene-expression matrix. For example, the confidence in hierarchical clusters can be calculated by perturbing the data with Gaussian noise and subsequent reclustering of the noisy data \[[@B16]\]. Resampling methods (bagging) have been used to improve the confidence of a single clustering method, namely PAM in \[[@B17]\]. A simple method for comparison between two data partitions, the *weighted-kappa*metric \[[@B18]\], can also be used to assess gene-expression cluster consistency. This metric rates agreement between the classification decisions made by two or more observers. In this case the two observers are the clustering methods. The *weighted-kappa*compares clusters to generate the score within the range -1 (no concordance) to +1 (complete concordance) (Table [1](#T1){ref-type="table"}). A high *weighted-kappa*indicates that the two arrangements are similar, while a low value indicates that they are dissimilar. In essence, the *weighted-kappa*metric is analogous to the adjusted Rand index used by others to compare cluster similarity \[[@B16],[@B19]\].
Despite the formal assessment of clustering methods, there remains a practical need to extract reliably clustered genes from a given gene-expression matrix. This could be achieved by capturing the relative merits of the different clustering algorithms and by providing a usable statistical framework for analyzing such clusters. Recently, methods for gene-function prediction using similarities in gene-expression profiles between annotated and uncharacterized genes have been described \[[@B20]\]. To circumvent the problems of clustering algorithm discordance, Wu *et al.*used five different clustering algorithms and a variety of parameter settings on a single gene-expression matrix to construct a database of different gene-expression clusters. From these clusters, statistically significant functions were assigned using existing biological knowledge.
In this paper, we confirm previous work showing gene-expression clustering algorithm discordance using a direct measurement of similarity: the *weighted-kappa*metric. Because of the observed variation between clustering methods, we have developed techniques for combining the results of different clustering algorithms to produce more reliable clusters. A method for clustering gene-expression data using resampling techniques on a single clustering method has been proposed for microarray analysis \[[@B19]\]. In addition, Wu *et al.*showed that clusters that are statistically significant with respect to gene function could be identified within a database of clusters produced from different algorithms \[[@B20]\]. Here we describe a fusion of these two approaches using a \'consensus\' strategy to produce both robust and consensus clustering (CC) of gene-expression data and assign statistical significance to these clusters from known gene functions. Our method is different from the approach of Monti *et al.*, in that different clustering algorithms are used rather than perturbing the gene-expression data for a single algorithm \[[@B19]\]. Our method is also distinct from the cluster database approach of Wu *et al*\[[@B20]\]. There, clusters from different algorithms were in effect fused if the consensus view of all algorithms indicated that the gene-expression profiles clustered independently of the method. In the absence of a defined rule base for selecting clustering algorithms, we have implemented clustering methods from the statistical, AI and data-mining communities to prevent \'cluster-method type\' biases. When consensus clustering was used with probabilistic measures of cluster membership derived from external validation with gene function annotations, it was possible to accurately and rapidly identify specific transcriptionally co-regulated genes from microarray data of distinct B-cell lymphoma types \[[@B21]\].
Results
=======
Cluster method comparison
-------------------------
Initially we assessed cluster method consistency for HC, PAM, SA and CAST using the *weighted-kappa*metric and a synthetic dataset of 2,217 gene-expression profiles over 100 time points that partitioned into 40 known clusters. The *weighted-kappa*values derived from the metric indicate the strength of agreement between two observers (Table [1](#T1){ref-type="table"}). To interpret two *weighted-kappa*scores, for example, from two cluster arrangements, the broad categories from Table [1](#T1){ref-type="table"} are used, together with an assessment of relative score differences. If the two scores in question were 0.2 and 0.4, one could say that the former is poor (worse) and the latter is fair (better), but not that one is twice as good as the other. To allow defined clusters to be extracted from the tree structure of HC we used the R statistical package \[[@B22]\] implementation of HC. This implementation uses the CUTTREE method to convert the tree structure into a specified number of clusters.
With the synthetic dataset, all clustering algorithms had a \'high\' *weighted-kappa*agreement (data not shown) \[[@B18]\]. It is possible that the highly stylized nature of synthetic data resulted in higher than expected cluster-method agreement compared to experimentally derived data. This effect has been observed previously \[[@B10],[@B12]\]. Therefore, we used a repeated microarray control element Amersham Score Card (ASC) dataset as a semi-synthetic validation standard. We also used an experimentally derived microarray dataset for cross-cluster-method comparison. To facilitate cross-method comparison, we used fixed parameters where appropriate (see Materials and methods). Consistent with other studies, we observed that clustering-method consistency varied between methods and datasets (Figure [1](#F1){ref-type="fig"}). As expected, the repeated gene/probe measurements present in the ASC dataset resulted in higher levels of cluster agreement between methods than the single gene probe B-cell data. With the ASC data there was in general a \'good\' level of agreement between all different clustering algorithms, with only CAST compared to HC scoring as \'moderate\'. This shows that most clustering methods are able to group highly correlated data accurately, and that repeated measurements of gene-expression values can aid cluster partitioning \[[@B12]\]. Nevertheless, even with such high gene-expression correlation not all cluster assignments were consistent. This effect is magnified with the single probe per gene B-cell lymphoma data, where the degree of agreement for cluster partitioning was less, with no comparison scoring above \'fair\'. This observation emphasizes the need for the current desired practice in microarray analysis of using many different clustering algorithms to explore gene-expression data, thereby not over-interpreting clusters on the basis of a single method \[[@B23]\].
Algorithms
----------
The partial agreement of the different clustering algorithms must reflect the clustering of highly similar gene-expression vectors regardless of the clustering methods used. Where algorithm-based inconsistency problems occur in other aspects of computational biology, such as protein secondary structure prediction, consensus algorithms are often used \[[@B24]\]. These can either report a full or a majority agreement. This consensus strategy has also been applied to explore the effect of perturbing the gene-expression data for a single clustering algorithm \[[@B19]\]. We have therefore designed a similar strategy to identify the consistently clustered gene-expression profiles in microarray datasets by producing a consensus over different clustering methods for a given parameter set (see Materials and methods). Extracting such consistently clustered robust data from a large gene-expression matrix is extremely useful, increasing overall analysis confidence.
### Robust clustering
We initially developed an algorithm called robust clustering (RC) for compiling the results of different clustering methods reporting only the co-clustered genes grouped together by all the different algorithms - that is, with maximum agreement across clustering methods. For two genes *i*and *j*, all clustering methods must have allocated them to the same cluster in order for them to be assigned to a robust cluster. This gives a higher level of confidence to the correct assignment of genes appearing within the same cluster. Robust clustering works by first producing an upper triangular *n*× *n*agreement matrix with each matrix cell containing the number of agreements among methods for clustering together the two variables, represented by the row and column indices (Figure [2](#F2){ref-type="fig"}). This matrix is then used to group variables on the basis of their cluster agreement (present in the matrix).
Robust clustering uses the agreement matrix to generate a list, *List*, which contains all the pairs where the appropriate cell in the agreement matrix contains a value equal to the number of clustering methods being combined (that is, full agreement). Starting with an empty set of robust clusters RC, where *RC*~*i*~is the *i*th robust cluster, the first cluster is created containing the elements of the first pair in *List*. Then the pairs in *List*are iterated through and checked to see if one of the members of the current pair is within any of the existing clusters, *RC*~*i*~.
If one element of the current pair is found and the other element of the pair is not in the same cluster, then the other element is added to that cluster. If neither element of the pair is found in an existing *RC*~*i*~in RC, then a new cluster is added to RC containing each element of the pair. When the end of the list is reached, the set of robust clusters, RC, is the output. The robust clustering algorithm is as follows:
Input:Agreement Matrix (*n*× *n*), *A*
\(1) Set *List*= all pairs (*x*, *y*) in the matrix, with agreement = the number of methods
\(2) Set *RC*to be an empty list of clusters
\(3) Create a cluster and insert the two elements (*x*, *y*) of the first pair in *List*into it
\(4) For *i*= 2 to size of *List*-1
(5) For *j*= 1 to number of Clusters in *RC*
(6) If *x*or *y*of *List*~*i*~is found within *RC*~*j*~
(7) If the other member of the pair *List*~*i*~is not found in *RC*~*j*~
(8) Add the other member to *RC*~*j*~
(9) End If
(10) Else If the other member of the pair *List*~*i*~is not found in *RC*~*j*~
(11) Add a new cluster to *RC*containing *x*and *y*
(12) End If
(13) End For
(14) End For
Output:Set of Robust Clusters *RC*
### Application of robust clustering
Robust clustering was applied to both the ASC and B-cell lymphoma datasets and the partitioning of the gene-expression profiles observed. As expected, the robust clusters do not contain all variables because of the underlying lack of consistent clustering by all methods. As a result, the *weighted-kappa*cannot be calculated. This metric requires both clustering arrangements being compared to be drawn from the same set of items. This is not the case with robust clustering because many items will not be assigned to a cluster. However, approximately 80% of the total ASC data variables and 25% of the B-cell lymphoma variables are assigned to a robust cluster. Robust clustering further subdivides the datasets into smaller clusters, with 24 rather than 13 clusters being defined for ASC, and 154 rather than 40 being defined for the B-cell lymphoma data (Table [2](#T2){ref-type="table"}). Robust clusters are therefore valuable for allowing a rapid \'drilling down\' in a gene-expression dataset to groups of genes whose coexpression pattern is identified in a manner independent of cluster method.
The robust clustering algorithm is, by definition, subject to discarding gene-expression vectors if only one clustering method performs badly in the co-clustering. This effect of single method under performance on a given dataset has been previously observed for single linkage hierarchical clustering \[[@B10],[@B13]\]. Therefore, to generate clusters with high agreement across methods but not so restrictive as to discard majority consistent variables, we adapted the algorithm to generate consensus clusters, making use of the same agreement matrix.
### Consensus clustering
Consensus clustering relaxes the full agreement requirement by taking a parameter, \'minimum agreement\', which allows different agreement thresholds to be explored. Rather than grouping variables on the basis of full agreement only, consensus clustering maximizes a metric, which rewards variables in the same cluster if they have high cluster method agreement and penalizes variables in the same cluster if they have low agreement. Consensus clustering maximizes agreement using the function *f*(*G*~*i*~) in Equation (1) to score each cluster of size *s*~*i*~
where *A*is the agreement matrix, *G*~*ij*~is the *j*th element of cluster *i*(*G*~*i*~) and *β*is a user-defined parameter (the agreement threshold), which determines whether the score for the cluster is increased or decreased. The score for a clustering arrangement is the sum of the scores of each cluster, which consensus clustering attempts to maximize.
If *β*is equal to *Min*, the minimum value in *A*, then the function is maximized when all variables are placed into the same cluster (that is, a single large cluster). Alternatively, when *β*is equal to *Max*, the maximum value in *A*, the function is maximized when each variable is placed into its own cluster. Essentially all clusters produced by Consensus Clustering are scored by *f*(*G*~*i*~), rewarding and preserving clusters with high agreement between members, while penalizing and discarding clusters containing low agreement between members. A value for *β*should lie between the minimum and the maximum agreement so as not to skew the scoring function. A suitable value for *β*is (*Max*+ *Min*)/2, where *Max*is the maximum value in *A*and *Min*is the minimum. For a uniformly distributed agreement matrix, (*Max*+ *Min*)/2 is the mean value; therefore we penalize values below the mean agreement and reward above it. For both the ASC and B-cell lymphoma data *β*was 2, as *Max*= 4 (four clustering algorithms giving complete agreement) and *Min*= 0 (no agreement). In order to maximize the scoring function for consensus clustering, a search over possible cluster membership is needed. There are many methods for performing a search and it was decided that SA was best because it is an efficient search/optimization procedure that does not suffer from becoming stuck in local maxima. The consensus algorithm is as follows:
Input: Agreement Matrix (*n*× *n*), *A;*Maximum Number of Clusters sought, *m;*Number of Iterations, *Iter;*Agreement Threshold, *β;*Initial Temperature, *θ*~0~*;*Cooling Rate, *c*
\(1) Generate a random number of empty clusters (\<m)
\(2) Randomly distribute the variables (genes) 1..*n*between the clusters
\(3) Score each cluster according to Equation (1)
\(4) For *i*= 1 to *Iter*do
(5) Either Split a cluster, Merge two clusters or Move a variable (gene) from one random cluster to another
(6) Set Δ*f*to difference in score according to Equation (1)
(7) If Δ*f*\< 0 Then
(8) Calculate probability, *p*, according to Equation (2)
(9) If *p*\> random(0,1) then undo operator
\(10) End If
(11) *θ*~*i*~= *cθ*~*i*-1~
\(12) End For
Output:Set of Consensus Clusters
Note that *random*(0,1) (line 9) returns a random uniformly distributed real number between 0 and 1.
The \'split\', \'merge\' and \'move\' operators (line 5) are as follows and used with equal probability:
Split a cluster:
Input: Cluster *g*of size *n*
\(1) Randomly shuffle the cluster
\(2) Set *i*to be a random whole number between 1 and *n*-1
\(3) Create two empty clusters *g*~1~and *g*~2~
\(4) Add elements 1..*i*of *g*to *g*~1~
\(5) Add elements i+1..*n*of *g*to *g*~2~
Output:Two new clusters *g*~1~and *g*~2~
Here the old cluster is deleted and the two new clusters are then added to the set of clusters.
Merge two clusters:
Input: Two Clusters *g*~1~and *g*~2~
\(1) A new cluster *g*is created by forming the union of *g*~1~and *g*~2~
Output: A new cluster *g*
Here the old clusters are deleted and new cluster is then added to the set of clusters.
Move a gene:
Input:A set of clusters *G*
\(1) Two random clusters *g*~1~and *g*~2~are chosen where the size of *g*~1~is greater than one
\(2) A random element of *g*~1~is moved into *g*~2~
Output:The updated set of clusters *G*
The probability (*p*) (line 8) is calculated by:
In the following experiments we found *θ*~0~= 100, *c*= 0.99994 and *iter*= 1,000,000 as the most efficient parameters for SA.
These parameter settings for SA are effectively determined by the *iter*setting. We denote the change in fitness during the SA algorithm as Δ*f*and the starting temperature as *θ*~0~which is always positive. From equation 2 it can be clearly seen that if Δ*f*= *θ*~0~then the (worse) solution will be accepted with probability 0.368 (*e*^-1^). As the temperature cools, this probability will reduce. Here we set *θ*~0~to be the average of Δ*f*over 1,000 trial evaluations, so that at the beginning of the algorithm, the average worse solution (Δ*f*= *θ*~0~) will be accepted with the probability stated above.
It can be seen from the consensus algorithm that during the *i*th stage of the SA algorithm *θ*~0~= *θ*~0~*c*^*i*^. The SA algorithm works by assuming that the temperature reduces to zero over an infinite number of iterations. As it is not practical to run the SA algorithm to infinity the method is usually terminated after a fixed number of iterations, (*iter*). At this time the temperature will not be zero, but very small and positive, say *ε*. Therefore,
Hence if some small positive value for *ε*is chosen, and the algorithm is to run for a defined number of iterations (*iter*), then the decay constant *c*is calculated as above.
### Application of consensus clustering
As consensus clustering relaxes the \'complete agreement\' criteria we would expect the majority but not necessarily all robust cluster members to be assigned to the same consensus clusters. This was indeed true for the B-cell data where consensus clustering of the datasets showed that 98.5% of the B-cell robust clusters were assigned correctly to their respective consensus clusters. With the more consistent ASC data 100% of the robust clusters were assigned to the correct consensus clusters.
The advantage of consensus clustering over all single-cluster methods was evident when comparing consensus clustering to the mean *weighted-kappa*score for each pairwise combination of individual clustering algorithms (derived from Figure [1](#F1){ref-type="fig"}). Comparisons for the ASC dataset (Figure [3a](#F3){ref-type="fig"}) and B-cell lymphoma data (Figure [3b](#F3){ref-type="fig"}) show that consensus clustering improves on all single methods regardless of dataset, except in the case of CAST compared to SA for the ASC dataset (Figure [3a](#F3){ref-type="fig"}). It is interesting to observe that consensus clustering has higher agreement with SA compared to SA agreement with all other methods in the B-cell data (Figure [3b](#F3){ref-type="fig"}). The reasons for this are unclear, but suggest that with datasets similar to the B-cell data, SA captures a reliably partitioned subset of the data. To determine if consensus clustering was consistently superior to the use of single clustering methods, particularly the stochastic methods CAST and SA, we performed 10 independent runs of CAST, SA and consensus clustering. From the resulting clusters we determined the mean *weighted-kappa*scores for 45 possible comparisons (that is, the number of unique pairs formed from 10 objects = 10 × 9/2) (Table [3](#T3){ref-type="table"}). Consensus clustering provided an extremely high degree of consistency over all 10 runs, with a mean *weighted-kappa*score of 0.96. Importantly, there was little variation between each of the 10 runs with a standard deviation of the mean *weighted-kappa*of 0.0015. SA had a similar low standard deviation, but produced lower inter-run consistency (mean *weighted-kappa*of 0.816). CAST was the least consistent of the methods (mean *weighted-kappa*of 0.646). The differences in the consensus clustering mean compared to SA and CAST are significant at greater than the 99.9% confidence level, thereby showing consensus clustering identifies a reliable data partition, which is significantly better than multiple runs of single clustering methods.
We wished to confirm that the benefit of consensus clustering was not simply due to the parameter settings chosen for the dataset used. This could be confirmed by extensively varying each algorithm\'s parameter settings and comparing cluster partitioning using the same dataset; however, the large number of combinations of possible parameter settings between all methods makes this unrealistic. An alternative approach is to compare all methods on additional datasets. We therefore tested consensus clustering on two different simulated datasets containing 60 defined clusters of genes. The first synthetic dataset was generated from a vector autoregressive process (VAR) and the second using a multivariate normal distribution (MVN). The number of genes in each cluster varied from 1 to 60, with the number of conditions (arrays) set to 20. The datasets therefore contained 1,830 genes over 20 conditions. As the structure of each dataset is known, the results of each clustering method can be evaluated for accuracy using the *weighted-kappa*metric. Cluster accuracy using the single methods ranged between a *weighted-kappa*of 0.505 to 0.7 (mean *weighted-kappa*of 0.614) (Table [4](#T4){ref-type="table"}). It is interesting to note that the single clustering methods performed differently on the two synthetic datasets, with HC, SA and CAST performing better on the MVN synthetic data and PAM better on the VAR synthetic data. Consensus clustering was superior to all single clustering algorithms with *weighted-kappa*scores of 0.725 and 0.729 for VAR and MVN respectively, demonstrating that consensus clustering is accurate regardless of subtleties in the data structure (Table [4](#T4){ref-type="table"}).
Interpretation of consensus clustering
--------------------------------------
Consensus clustering greatly improves the accuracy of identifying cluster group membership based solely on the gene-expression vector, but as with other clustering algorithms still produces essentially unannotated clusters which require further external validation by gene function analysis. To address this problem, we derived a probability score to test the significance of observing multiple genes with known function in a given cluster against the null hypothesis of this happening by chance. This identifies clusters of high functional group significance, aiding assignment of functions to unclassified genes in the cluster using the \'guilt by association\' hypothesis.
The probability score is based on the hypothesis that, if a given cluster, *i*, of size *s*~*i*~, contains *x*genes from a defined functional group of size *k*~*j*~, then the chance of this occurring randomly follows a binomial distribution and is defined by:
where *n*is the number of genes in the dataset. As *k*~*j*~and *x*may potentially be very large, *Pr*from the above equation would be difficult to evaluate. Therefore the normal approximation to the binomial distribution can be used as defined by:
Large positive values of *z*mean that the probability of observing *x*elements from functional group *j*in cluster *i*by chance is very small, (for example *z*\> 2.326 corresponds to a probability less than 1%). Note that we perform a one tailed test as we are only interested in the case where a significantly high number of co-clustered genes belong to the same functional group.
This cluster function probability score was used to identify statistically significant (at the 1% level) B-cell consensus clusters containing defined genes known to be associated with 10 functional groups \[[@B21]\]. To determine if consensus clustering was better able to identify important functional group clusters we determined the functional group probability scores produced by individual clustering algorithms analogous to the strategy of Wu *et al*. \[[@B20]\]. For each functional group, the mean lowest probability scores (using Equation (4)) were calculated for the signal clustering methods and compared to consensus clustering (Figure [4a](#F4){ref-type="fig"}). Consensus clustering always produced equivalent or lower probabilities for each functional group, indicating that it produced more informative clusters.
One potential confounding factor in this analysis is that consensus clustering achieves a lower probability score by finding smaller clusters. This would decrease the ability to associated new genes with a given functional group. In the worst case the number of genes defining a functional group (FG) would equal the cluster size (*s*~*i*~) (FG/*s*~*i*~= 1). Alternatively, single clustering methods may produce lower probability scores by increasing the cluster size, thereby pulling many genes into the cluster resulting in a FG/*s*~*i*~ratio tending towards zero. This would also reduce the usefulness of the clusters. We determined the cluster size and functional group size for two representative functional groups where the consensus clustering probability was similar to the single method probability score, namely the endoplasmic reticulum (ER) stress response (also known as the unfolded protein response) (ER/UPR) functional group, or the markedly better nuclear factor-κB (NFκB) functional group (Figure [4b](#F4){ref-type="fig"}). Apart from SA, all single clustering methods tended to produce larger clusters, thereby decreasing the FG/*s*~*i*~ratio. In the most extreme case of the ER/UPR functional group, the HC cluster size was 310 compared to the consensus clustering size of 40. SA tended to produce similar cluster sizes as consensus clustering but with higher overall probabilities. Therefore, consensus clustering identifies significant functional clusters while achieving a workable balance between large or small cluster sizes.
We further investigated the two groups NFκB and ER/UPR to assess what additional insights consensus clustering allowed. These two functional groups represent important B-cell functions at different stages of the B-cell development pathway. The consensus cluster associated with NFκB also contained genes either not previously associated with or only tentatively associated with NFκB activity in subsets of B-cell lymphomas. The gene-expression profiles from this consensus cluster were visualized by average linkage HC using the programmes Cluster and Treeview \[[@B5]\] (Figure [5](#F5){ref-type="fig"}) and clustered gene functions were investigated further using the annotation resources DAVID \[[@B25]\] and GeneCards \[[@B26]\]. From GeneCards each gene was identified in the complete human genome sequence using Ensembl \[[@B27]\] and 1,000 base pairs (bp) upstream of the predicted transcriptional start site extracted for promoter analysis using the program TESS from the Baylor College sequence analysis software BCM \[[@B28]\] (Figure [6](#F6){ref-type="fig"}).
This consensus cluster is predominantly overexpressed in the cell lines Raji, Pel-B, EHEB, Bonna-12 and L-428. These cell lines have in common the induction of the NFκB pathway, either through expression of Epstein-Barr virus LMP-1 protein (Raji, Pel-B, EHEB and Bonna-12) or the loss of function of the inhibitor of NFκB, namely IκB (L-428). This implies that many of these genes could be NFκB responsive. Twenty-four putative promoter regions were analyzed and NFκB-binding sites were identified in 12 of these. As expected, NFκB-binding sites were found in the CD40L receptor gene, *Bfl-1*, *BIRC3*, EBV-induced gene 3 (*EBI3*), and the genes for class I MHC-C and lymphotoxin *α*, as these have been previously characterized as NFκB responsive and were present in the initial NFκB-defined gene set. Interestingly, NFκB-binding sites were also found in six additional promoters for which accurate mapping of promoter transcription factor binding is not available (Figure [6a](#F6){ref-type="fig"}). All but four NFκB-binding sites conform precisely to the canonical consensus binding site (Figure [6b](#F6){ref-type="fig"}) \[[@B29],[@B30]\] and of the variants with T at position 1, two genes, *lymphotoxin α*and *BIRC3*are known to be NFκB responsive. Overall, this indicates that the six additional genes identified are likely to be NFκB responsive.
The consensus cluster associated with the ER/UPR functional group contained genes not previously associated with ER stress-induced upregulation. The gene-expression profiles were visualized and annotated as described for the NFκB functional group (Figure [7a](#F7){ref-type="fig"}). Annotation showed that of the 32 genes within the ER/UPR consensus cluster (23 defining the original functional group), 16% (5) were involved in calcium-ion binding within the ER and 13% (4) were involved with *N*-glycan biosynthesis. This functional group was overexpressed in cell lines of plasmablast or plasma-cell tumors, where physiological upregulation of the ER is required for cellular function. This process is controlled by two transcription factors, ATF6 and XBP1 \[[@B31]\]. The *ATF6*transcript was present as a defining signature gene in the ER/UPR functional group. This suggests that ATF6 and XBP1 may be responsible for upregulation of the calcium-ion binding and *N*-glycan biosynthetic genes. Two responsive elements have been defined for ATF6 and XBP1 respectively, the ER stress-response element (ESRE), comprising the binding site CCAATN~9~CCACG and the unfolded protein response element (UPRE), comprising the binding site TGACGTG(G) \[[@B32]\]. ATF6 and XBP1 can bind to the CCACG region of ESRE in conjunction with the general transcription factor NF-Y/CB. XBP1 can bind to the UPRE, but ATF6 can only bind to the UPRE when expressed to high (possibly non-physiological) levels \[[@B33]\]. ESRE sites were identified in two of the five calcium-ion binding proteins, namely, calnexin and the tumor rejection antigen (gp96) 1(TRA1) (Figure [7b](#F7){ref-type="fig"}). Interestingly, XBP1 (UPRE) binding sites were identified in two of the *N*-glycan biosynthetic genes but no ESRE sites were found. This suggests that these two groups of genes are regulated through two distinct mechanisms by transcription factors ATF6 and XBP1 as a result of ER stress.
Discussion
==========
Grouping data into sets based on a consistent property is a common occurrence in biological analysis. This has recently increased in importance with the production of large microarray datasets. Implicit in the experimental rationale is the fact that patterns of coexpressed genes should be identifiable in a gene expression matrix and these can be linked to shared biological processes. However, different clustering algorithms are known to partition data into different groups \[[@B10]-[@B13]\]. We also observe a similar lack of cluster-method concordance using a *weighted-kappa*metric. This metric effectively scores how well different cluster method pairs assign the same genes to the same clusters. The *weighted-kappa*metric readily shows that, even for highly correlated gene-expression profiles present in the ACS dataset, no two clustering algorithms have complete agreement, although the global search methods such as SA seem to produce the most consistent results. Overall this emphasizes that no single analysis method will identify all patterns in the gene-expression data; therefore multiple analyses should be performed and compared \[[@B23]\].
We and others recently described the use of consensus clustering to improve confidence \[[@B34]\] or as a re-sampling method for microarray analysis \[[@B19]\]. It was suggested that a natural extension of Monti *et al*. was to use a meta-consensus across different clustering algorithms rather than to re-sample over the same algorithm. Our results represent this extension and confirm the validity of consensus clustering. We have developed both robust and consensus clustering, with these methods offering specific advantages over the use of individual clustering algorithms for microarray analysis. The robust clustering algorithm is useful for creating clusters of genes with high confidence and is extremely effective for reducing the dimensionality of large gene-expression datasets. However, robust clustering can be restrictive in discarding genes that do not have full agreement. Consensus clustering overcomes this problem, requiring a minimum-agreement parameter to generate clusters based on the combined results of a number of existing clustering methods. This strategy enables the effective identification of cluster groups that are of high reliability and cluster method independent.
The choice of clustering algorithms and parameter settings is a major stumbling block for all gene array cluster analysis. The effect of varying the parameters depends on the cluster method used. The performance of cluster methods has been extensively investigated \[[@B12]\]. The authors show that model-based methods and certain partitional methods, when used with optimal distance matrices, perform well on synthetic and real-world data. From our study, SA, an optimization method, also performs well as a clustering method. Therefore, the individual algorithms used as input to consensus clustering should ideally consist of representative algorithms from optimization (for example, SA), graph theoretical (for example, CAST), model-based (for example, MCLUST \[[@B12]\]) and partitional (for example, HC). Some methods (for example, CAST, SA and MCLUST) can determine the number of clusters directly from the input data. However, some other methods require the number of clusters to be specified as a parameter (for example, PAM and HC). In principle, methods such as CAST, SA and MCLUST can be used to determine this parameter for methods such as PAM and HC.
Consensus clusters are likely to contain gene subsets that are co-regulated by common transcriptional control networks or are coexpressed to participate in cellular processes that together manifest a global phenotype of the cell or tissue. In either case, these clusters are of high biological value. To facilitate further analysis it is useful to know which clusters are involved in a given biological process. By supplying a list of genes from a given biological process or network, the use of the normal approximation to the binomial distribution of these genes over all consensus clusters, allows the identification of clusters of high functional significance. Similar statistical assignment of gene function based on cluster analysis was performed by Wu *et al.*using a database of clusters \[[@B20]\]. To assign significance Wu *et al.*used the hypergeometric distribution. This distribution can be formally shown to asymptotically become the binomial distribution when the population size increases. Therefore, when used on large gene-expression datasets our methods are directly analogous to Wu *et al*. However, consensus clustering has the advantage over a database of clusters by producing low-probability clusters containing a significant percentage of known elements from functional groups.
Two functionally significant clusters, the NFκB-responsive cluster and the ER/UPR cluster, were investigated further here. Within the NFκB-responsive cluster 50% of the putative promoters of genes investigated had canonical NFκB-binding sites within 1,000 base pairs of the transcriptional start site, suggesting that they are NFκB responsive \[[@B29],[@B30]\]. The majority of these genes had NFκB-binding sites within 500 bp of the transcriptional start sites, consistent with the location in other NFκB-responsive genes \[[@B35]\]. Of the remaining two genes with NFκB-binding sites greater than 800 bp from the transcription start site, one, *Bfl-2*, has been experimentally verified \[[@B36]\]. Analysis of the ER/UPR consensus cluster also provided information on gene regulatory elements, but more interestingly provides insights into the control and effect of the ERSR/UPR.
In cells, the presence of unfolded proteins in the ER is associated with induction of the ER/UPR. However, during the maturation of B-cells to antibody-secreting plasma cells, expansion of the ER to accommodate increased secretion of immunoglobulins is thought to be coupled to the final stages of plasma-cell maturation. The induction of the ER/UPR occurs via the coordinated activation of the transcription factors ATF6 and XBP1 \[[@B31],[@B33]\]. ATF6 is normally maintained as an inactive, ER-resident, transmembrane protein that is cleaved, after translocating to the Golgi upon ER stress, by the site proteases S1P and S2P \[[@B37],[@B38]\]. The cleaved transcriptionally active ATF6 is then free to translocate to the nucleus, where it can activate target genes such as XBP1 and the ER chaperon protein GRP78/BiP \[[@B39]\]. *XBP1*mRNA is cleaved by the ER stress activated protein IRE1 to yield the transcriptionally active form of XBP1, inducing further genes of the UPR \[[@B32]\]. The activation of both the ATF6 and IRE1/XBP1 pathways results in enhanced transcription of ESRE-responsive genes; however, only XBP1 appears able to transactivate the UPRE. The identification of ESRE binding sites in the promoter regions of genes for calcium-ion binding protein and UPRE binding sites in the promoter regions of *N*-glycan biosynthesis genes suggests that these genes are differentially regulated by ATF6/XBP1 and XBP1 respectively.
The only known UPRE target gene is ER degradation-enhancing *α*-mannoside-like protein (EDEM), whose induction depends solely on IRE1/XBP1 activity \[[@B33]\]. Induction of the two UPRE-containing genes, UDP-GlcNAc:dolichol phosphate *N*-acetylglucosamine-1 phosphate transferase (*DPAGT1*) and UDP-GlcNAc:*α*-6-D-mannoside *β*-1-2-*N*-acetylglucosaminyltransferase II (*MGAT2*), which catalyze essential steps in the biosynthetic pathway of complex *N*-linked glycans, supporting a clear link between the dolichol pathway and the UPR \[[@B40]\]. In addition, the ER/UPR functional group suggests that *DPAGT1*and *MGAT2*expression is regulated solely by the IRE1/XBP1 pathway. Altogether, these results show that consensus clustering and gene functional group analysis provide a highly accurate way of mining gene-expression data for novel insights into different genes within the cluster.
Robust and consensus clustering provide a platform for more efficient microarray analysis pipelines. There is effectively no limit to the number of different clustering algorithms that can be used to feed into the consensus clusters, and each clustering algorithm could be run under different parameter sets to fully explore a microarray dataset \[[@B19]\]. In addition, different distance matrices could be used as input into the range of clustering algorithms. In each case the consensus clustering algorithm effectively acts as the collation and interpretation point for the different individual analysis methods. This environment is ideal for use in parallel processing computer farms and the GRID \[[@B41]\]. In such an environment, each node of the farm could perform a range of analyses with a subset of clustering algorithms, with the master node compiling the consensus results. This would greatly increase computational speed and allow a thorough, single data entry point, access to an extensive range of clustering methods.
All areas of functional genomics that produce high-dimensional datasets with inherent patterns will require data partitioning to allow interpretation. Consensus clustering in the context of statistically defined functional groups could allow a consistent analysis platform for such diverse data types.
Materials and methods
=====================
Clustering methods
------------------
We implemented and compared a representative sample of methods from the statistical, AI and data-mining communities. The methods used were average linkage HC, PAM, SA and CAST. As all the clustering techniques use correlation between variables, we used the Pearson\'s correlation coefficient, *r*, to measure the linear relationship between two variables, *x*~1~and *x*~2~, where the variable can be either discrete or continuous \[[@B42]\]. HC and PAM methods were implemented using the statistical package R \[[@B22]\], while CAST and SA were implemented locally in C++.
HC is an agglomerative method that produces a hierarchical (binary) tree or dendrogram representing a nested set of data partitions. It has been applied successfully to many gene-expression datasets \[[@B43]\]. Sectioning a hierarchically clustered tree at a particular level leads to a partition with a number of disjointed groups, thereby yielding different clustering of the data. The tree was sectioned using the CUTTREE method, to yield 13 clusters for the ASC dataset and 40 for the B-cell dataset. The method PAM works by first selecting *m*out of *n*total objects that are the closest (according to a distance matrix) to the remaining (*n*- *m*) objects. The fitness of these medoids is calculated by placing the remaining (*n*- *m*) objects in a group according to the nearest medoid and summing the distances of the group members from this medoid. These *m*selected objects are the initial medoids. A *Swapping*procedure is then applied to reassort the objects until there is no improvement in the fitness of the medoids \[[@B3]\]. As with HC, PAM is set to search for 13 and 40 clusters. The choice of 13 clusters for the ASC data was determined by the number of repeated genes, whereas 40 clusters for the B-cell data was based on previous exploratory data analysis \[[@B21]\].
SA \[[@B6]\] is an iterative improvement search technique that starts with a random solution to a given problem, and then tries to increase its worth by a series of small changes in cluster membership. If such a small change is better than the previous solution, then further changes are made from this new point. However, if the new solution is worse than the old one, it is not discarded, but accepted with a certain probability. The measured worth of the SA clustering arrangement is based here on the EVM metric \[[@B44]\]. SA has recently been applied to the clustering of gene-expression data \[[@B45]\]. The performance related parameters for SA were set as follows: *θ*~0~= 100, *c*= 0.99994 and number of iterations = 1,000,000.
CAST \[[@B9]\] is a heuristic algorithm that uses an affinity measure to determine whether variables are assigned to clusters. It requires a threshold parameter, which determines whether variables are assigned or moved to new clusters. Once CAST is complete, a clean-up operation is applied to ensure that the affinity of every variable to its cluster is greater than a user defined threshold. The only parameter CAST needed was the affinity level, which was set to 0.5 as recommended \[[@B9]\].
Methods such as CAST and SA require the differences/relationships between a pair of observations, *x*~1~and *x*~2~, to be expressed as binary (*b*). As Pearson\'s correlation coefficient is bounded, it provides a good basis for defining a binary relationship between two variables as defined by:
where 0 \<*α*≤ 0 is a constant and  is a floor function that returns the largest integer less than or equal to the real number *y*.
Datasets for evaluation
-----------------------
Two datasets were used for evaluating the cluster methods. The first is a set of multiply repeated control element spots relating to the Amersham Score Card (ASC) probe set on the Human Genome Mapping Project Human Gen1 clone set array \[[@B46]\]. The ASC probes are present as a single row of 32 elements in each of the 24 array sub-grids. Of these elements, 13 gene probes consistently give signals above background in both the Cy5 and Cy3 channels. Therefore, each array has effectively 24 repeat measurements of 13 spots. After filtering for low signal-to-noise ratio (SNR) probes, a dataset of 30 arrays was examined by treating each positional repeat probe element across the 30 array set as an individual gene, which together with the remaining 23 same-gene probes per array, represents a highly correlated gene-expression profile. Therefore, we assume the repeated probes should cluster together; thus, this dataset becomes 308 genes/probe elements, which would cluster into 13 known groups, with each group having between 6 and 24 members after SNR filtering. In essence, the ASC data represent a semi-synthetic dataset for internal cluster method validation.
The second dataset consists of a series of 26 arrays (1,987 filtered genes) measuring gene-expression difference across a set of human B-cell lymphomas and leukemias \[[@B21]\]. The dataset is available via the URL indicated in Jenner et al. \[[@B21]\]. Each probe on the array detected a single gene transcript. This dataset contains a number of genes that correspond to known cellular functions, for example cell proliferation and NFκB response. The four clustering techniques described above were applied to both the datasets, with each method being set to find 13 clusters for the ASC and 40 clusters for the B-cell data.
Synthetic datasets
------------------
Two synthetic datasets were generated using a vector autoregressive process (VAR) or a multivariate normal distribution (MVN). The two datasets contained 1,830 genes and 20 conditions (arrays). The VAR process of order *p*is a linear multivariate time series defined by
where [*x*]{.underline}(*t*) is the *n*-dimensional vector of observations at time *t*, *A*~*i*~is the *n*× *n*autoregressive coefficient matrix at time lag *i*, and [*ε*]{.underline}(*t*) is the zero mean *n*-dimensional noise vector at time *t*(drawn from a normal distribution). Therefore [*x*]{.underline}(*t*) is a linear combination of the previous observations plus some random noise. For the synthetic dataset, each cluster was generated by a vector autoregressive model of order *p*= 1 and size *n*equal to the number of genes in the cluster.
For the MVN dataset, a vector of random variables [*x*]{.underline} has a MVN distribution if every linear combination of that vector is also normal. Under such conditions we use the notation [*x*]{.underline} \~ *N*([*μ*]{.underline}, Σ) to denote that [*x*]{.underline} follows the MVN distribution, where [*μ*]{.underline} is the mean vector and Σ is a positive definite matrix of covariance. The probability density function of [*x*]{.underline} is given by
where \|Σ\| = det(Σ). For the synthetic dataset, each cluster was drawn from an MVN distribution with varying mean [*μ*]{.underline} and covariance Σ.
*Weighted-kappa*metric
----------------------
To compare the resultant clusters for each method, a statistic known as *weighted-kappa*was used \[[@B18]\]. This metric rates agreement between the classification decisions made by two or more observers. In this case the two observers are the clustering methods. The classification from each observer for each unique pairing of variables (within the clusters) is divided into a 2 × 2 contingency table. Rows and columns within this table are indexed according to whether the two variables are in the same group or in different groups. The total number of comparisons, *N*, is defined in the following equation, where *Count*~*ij*~is the number of elements in the matrix cell indexed by (*i*,*j*),
and *n*is the number of variables (genes) in the clusters as this represents the number of unique variable pairings.
The *weighted-kappa*metric is calculated from the contigency table by:
where, *w*~*ij*~is the weights for each category comparison; *p*~*o*(*w*)~and *p*~*e*(*w*)~represent the observed weighted proportional agreement and the expected weighted proportional agreement; *Count*~*ij*~is the *i*th, *j*th element of the 2 × 2 contingency table; *N*is the sum of the elements within this table; *Row*(*i*) and *Col*(*i*) are the row and column totals for this table respectively and *K*~*w*~is the *weighted-kappa*value. The interpretation of *weighted-kappa*values indicates the strength of agreement between two observers (Table [1](#T1){ref-type="table"}) is used to compare cluster method agreement in both datasets.
Acknowledgements
================
This work is supported in part by the BBSRC, the EPSRC and the MRC in the UK. We would also like to thank Richard Jenner for the viral gene expression dataset and Antonia Kwan for preparing the ASC dataset. These methods are available on request as functions in the statistical package, R.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Pairwise comparison of consistency between different cluster algorithm data partitions using the *weighted-kappa*metric (Table 1) to score similarity. Each clustering algorithm was used to analyze the Amersham Score Card dataset (black bars) and the B-cell lymphoma dataset (gray bars), and the cluster-method agreement based on assigning the same genes to the same cluster was calculated and scored. HC, hierarchical clustering; CAST, cluster affinity search technique; PAM, partitioning around medoids; and SA, simulated annealing.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
A visual representation of the agreement matrix used as input to robust and consensus clustering. The *n*× *n*matrix is upper triangular. Each cell within the matrix, referenced by column *i*and row *j*, represents the number of clustering methods that have placed gene *i*and gene *j*into the same cluster. In other words, the number represents the agreement between clustering methods concerning gene *i*and gene *j*.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Comparison between consensus clustering and pairwise clustering. The *weighted-kappa*score for consensus clustering (solid line) calculated by comparing consensus clusters to the corresponding individual clustering algorithm is shown relative to mean pairwise *weighted-kappa*score for each single method compared to all other single methods (broken line) for **(a)**the ASC dataset, **(b)**the B-cell lymphoma dataset. The maximum and minimum *weighted-kappa*scores for the collection of single methods are indicated by the error bars. The definitions of *weighted-kappa*scores are derived from Table 1. The parameter settings for the clustering algorithms are: HC and PAM, 13 clusters for the ASC dataset and 40 for the B-cell dataset; CAST, affinity level 0.5; and SA, *θ*~0~= 100, *c*= 0.99994 and number of iterations = 1,000,000.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Probability scores and cluster size. **(a)**The lowest probability scores determined for clusters containing the following functional group signature genes were identified: AC, actin cytoskeleton; BST, B-cell signal transduction; EGT, ER/Golgi trafficking; ERUPR, ER stress/unfolded protein response; ICS, immunoglobulin class switching; IA, inflammation and adhesion; NFκB, NFκB signaling; OBS, other B-cell signaling; P, proliferation; RNA, RNA maturation and splicing. The mean (open diamond), standard error (green line) and standard deviation (thin black line and bars) of the minimum probability scores for SA, CAST, HC and PAM are shown together with the minimum probability score for the corresponding consensus cluster (red circle). **(b)**The cluster size (*s*~*i*~) (open circles) and number of defining functional group genes (FG) (open squares) for the NFκB signaling and ER/UPR functional groups are shown together with the FG/*s*~*i*~ratio (open diamonds).
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Visualization of average linkage HC using the programs Cluster and Treeview \[5\] of the NFκB responsive gene cluster identified from consensus clustering and functional annotation. The sample names correspond to different leukemia and lymphoma samples \[21\], with the NFκB-responsive gene cluster being predominantly expressed in the cell lines Raji, PEL-B, EHEB, BONNA-12 and L-428. Gene names with red circles represent those genes that contain one or more NFκB-binding sites in the region up to 1,000 bp upstream from the putative transcriptional start site.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
Location and consensus sequence of NFκB-binding sites. **(a)**Position of the NFκB-binding sites identified in the upstream 1,000-bp regions of each gene. The gray ovals represent the position of each binding site shown in **(b)**where the nucleotide sequence for each respective NFκB-binding site is shown relative to the consensus NFκB-binding site GGGRNNNYCY (R is G or A (purine), Y is T or C (pyrimidine) and N is any nucleotide).
:::
:::
::: {#F7 .fig}
Figure 7
::: {.caption}
######
Genes involved in the ER/UPR. **(a)**Visualization of the ER/UPR consensus cluster using Cluster and Treeview. The *ATF6*gene is indicated by a green circle. Gene names indicated by red circles represent those genes whose upstream 1,000-bp regions contain **(b)**the endoplasmic reticulum stress response element (ESRE), namely, *calnexin*and *TRA1*, or unfolded protein response element (UPRE) namely, *DPAGT1*and *MGAT2*.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
The *weighted-kappa*guideline
:::
*Weighted-kappa* Agreement strength
------------------ --------------------
0.0 ≤ *K*≤ 0.2 Poor
0.2 \<*K*≤ 0.4 Fair
0.4 \<*K*≤ 0.6 Moderate
0.6 \<*K*≤ 0.8 Good
0.8 \<*K*≤ 1.0 Very good
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Robust clusters
:::
Dataset ASC\* B-cell
----------------------------- ------- --------
Number of robust clusters 24 154
\% of variables assigned 79.2% 25%
Maximum robust cluster size 44 14
Minimum robust cluster size 2 2
Mean robust cluster size 10.2 3.2
\*Amersham Score Card dataset.
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Multiple runs of the stochastic clustering methods
:::
Method Mean\* Min^†^ Max^†^ SD^†^
--------------------------- -------- -------- -------- -------
CAST 0.646 0.448 0.769 0.092
Simulated annealing (SA) 0.816 0.794 0.838 0.015
Consensus clustering (CC) 0.960 0.922 0.982 0.010
\*Mean *weighted-kappa*scores; ^†^Min (minimum) and Max (maximum) and SD (standard deviation) of the *weighted-kappa*scores.
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Cluster partition *weighted-kappa*scores of two synthetic datasets
:::
Dataset HC PAM CAST SA CC
----------------------- ------- ------- ------- ------- -------
Vector autoregressive 0.505 0.700 0.537 0.614 0.725
Multivariate normal 0.697 0.605 0.591 0.667 0.729
HC, hierarchical clustering; PAM, partitioning around medoids; CAST, cluster affinity search technique; SA, simulated annealing; CC, consensus clustering.
:::
|
PubMed Central
|
2024-06-05T03:55:51.852311
|
2004-11-1
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545785/",
"journal": "Genome Biol. 2004 Nov 1; 5(11):R94",
"authors": [
{
"first": "Stephen",
"last": "Swift"
},
{
"first": "Allan",
"last": "Tucker"
},
{
"first": "Veronica",
"last": "Vinciotti"
},
{
"first": "Nigel",
"last": "Martin"
},
{
"first": "Christine",
"last": "Orengo"
},
{
"first": "Xiaohui",
"last": "Liu"
},
{
"first": "Paul",
"last": "Kellam"
}
]
}
|
PMC545795
|
Background
==========
Analyses of functional genomic data such as gene-expression microarray measurements are subject to what has been called the \'curse of dimensionality\'. That is, the number of variables of interest is very large (thousands to tens of thousands of genes), yet we have relatively few observations (typically tens to hundreds of samples) upon which to base our inferences and interpretations. Recognizing this, many investigators studying quantitative genomic data have focused on the use of either classical multivariate techniques for dimensionality reduction and ordination (for example, principal component analysis, singular value decomposition, metric scaling) or on various types of clustering techniques, such as hierarchical clustering \[[@B1]\], *k*-means clustering \[[@B2]\], self-organizing maps \[[@B3]\] and others. Clustering techniques in particular are based on the idea of assigning either variables (genes or proteins) or objects (such as sample units or treatments) to equivalence classes; the hope is that equivalence classes so generated will correspond to specific biological processes or functions. Clustering techniques have the advantage that they are readily computable and make few assumptions about the generative processes underlying the observed data. However, from a biological perspective, assigning genes or proteins to single clusters may have limitations in that a single gene can be expressed under the action of different transcriptional cascades and a single protein can participate in multiple pathways or processes. Commonly used clustering techniques tend to obscure such information, although approaches such as fuzzy clustering (for example, Höppner *et al*. \[[@B4]\]) can allow for multiple memberships.
An alternate mode of representation that has been applied to the study of whole-genome datasets is network models. These are typically specified in terms of a graph, *G*= {*V*,*E*}, composed of vertices (*V*; the genes or proteins of interest) and edges (*E*; either undirected or directed, representing some measure of \'interaction\' between the vertices). We use the terms \'graph\' and \'network\' interchangeably throughout this paper. The advantage of network models over common clustering techniques is that they can represent more complex types of relationships among the variables or objects of interest. For example, in distinction to standard hierarchical clustering, in a network model any given gene can have an arbitrary number of \'neighbors\' (that is *n*-ary relationships) allowing for a reasonable description of more complex inter-relationships.
While network models seem to be a natural representation tool for describing complex biological interactions, they have a number of disadvantages. Analytical frameworks for estimating networks tend to be complex, and the computation of such models can be quite hard (NP-hard in many cases \[[@B5]\]). Complex network models for very large datasets can be difficult to visualize; many graph layout problems are themselves NP-hard. Furthermore, because the topology of the networks can be quite complex, it is a challenge to extract or highlight the most \'interesting\' features of such networks.
Two major classes of network-estimation techniques have been applied to gene-expression data. The simpler approach is based on the notion of estimating a network of interactions by defining an association threshold for the variables of interest; pairwise interactions that rise above the threshold value are considered significant and are represented by edges in the graph, interactions below this threshold are ignored. Measures of association that have been used in this context include Pearson\'s product-moment correlation \[[@B6]\] and mutual information \[[@B7]\]. Whereas network estimation using this approach is computationally straightforward, an important weakness of simple pairwise threshold methods is that they fail to take into account additional information about patterns of interaction that are inherent in multivariate datasets. A more principled set of approaches for estimating co-regulatory networks from gene-expression data are graphical modeling methods, which include Bayesian networks and Gaussian graphical models \[[@B8]-[@B11]\]. The common representation that these techniques employ is a graph theoretical framework in which the vertices of the graph represent the set of variables of interest (either observed or latent), and the edges of the graph link pairs of variables that are not conditionally independent. The graphs in such models may be either undirected (Gaussian graphical models) or directed and acyclic (Bayesian networks). The appeal of graphical modeling techniques is that they represent a distribution of interest as the product of a set of simpler distributions taking into account conditional relationships. However, accurately estimating graphical models for genomic datasets is challenging, in terms of both computational complexity and the statistical problems associated with estimating high-order conditional interactions.
We have developed an analytical framework, called a first-order conditional independence (FOCI) model, that strikes a balance between these two categories of network estimation. Like graphical modeling techniques, we exploit information about conditional independence relationships - hence our method takes into account higher-order multivariate interactions. Our method differs from standard graphical models because rather than trying to account for conditional interactions of all orders, as in Gaussian graphical models, we focus solely on first-order conditional independence relationships. One advantage of limiting our analysis to first-order conditional interactions is that in doing so we avoid some of the problems of power that we encounter if we try to estimate very high-order conditional interactions. Thus this approach, with the appropriate caveats, can be applied to datasets with moderate sample sizes. A second reason for restricting our attention to first-order conditional relationships is computational complexity. The running time required to calculate conditional correlations increases at least exponentially as the order of interactions increases. The running time for calculating first-order interactions is worst case *O*(*n*^3^). Therefore, the FOCI model is readily computable even for very large datasets.
We demonstrate the biological utility of the FOCI network estimation framework by analyzing a genomic dataset representing microarray gene-expression measurements for approximately 5,000 yeast genes. The output of this analysis is a global network representation of coexpression patterns among genes. By comparing our network model with known metabolic pathways we show that many such pathways are well represented within our genomic network. We also describe an unsupervised algorithm for highlighting potentially interesting subgraphs of coexpression networks and we show that the majority of subgraphs extracted using this approach can be shown to correspond to known biological processes, molecular functions or gene families.
Results
=======
We used the FOCI network model to estimate a coexpression network for 5,007 yeast open reading frames (ORFs). The data for this analysis are drawn from publicly available microarray measurements of gene expression under a variety of physiological conditions. The FOCI method assumes a linear model of association between variables and computes dependence and independence relationships for pairs of variables up to a first-order (that is, single) conditioning variable. More detailed descriptions of the data and the network estimation algorithm are provided in the Materials and methods section.
On the basis of an edge-wise false-positive rate of 0.001 (see Materials and methods), the estimated network for the yeast expression data has 11,450 edges. It is possible for the FOCI network estimation procedure to yield disconnected subgraphs - that is, groups of genes that are related to each other but not connected to any other genes. However, the yeast coexpression network we estimated includes a single giant connected component (GCC, the largest subgraph such that there is a path between every pair of vertices) with 4,686 vertices and 11,416 edges. The next largest connected component includes only four vertices; thus the GCC represents the relationships among the majority of the genes in the genome. In Figure [1](#F1){ref-type="fig"} we show a simplification of the FOCI network constructed by retaining the 4,000 strongest edges. We used this edge-thresholding procedure to provide a comprehensible two-dimensional visualization of the graph; all the results discussed below were derived from analyses of the entire GCC of the FOCI network.
The mean, median and modal values for vertex degree in the GCC are 4.87, 4 and 2 respectively. That is, each gene shows significant expression relationships to approximately five other genes on average, and the most common form of relationship is to two other genes. Most genes have five or fewer neighbors, but there is a small number of genes (349) with more than 10 neighbors in the FOCI network; the maximum degree in the graph is 28 (Figure [2a](#F2){ref-type="fig"}). Thus, approximately 7% of genes show significant expression relationships to a fairly large number of other genes. The connectivity of the FOCI network is not consistent with a power-law distribution (see Additional data file 1 for a log-log plot of this distribution). We estimated the distribution of path distances between pairs of genes (defined as the smallest number of graph edges separating the pair) by randomly choosing 1,000 source vertices in the GCC, and calculating the path distance from each source vertex to every other gene in the network (Figure [2b](#F2){ref-type="fig"}). The mean path distance is 6.46 steps, and the median is 6.0 (mode = 7). The maximum path distance is 16 steps. Therefore, in the GCC of the FOCI network, random pairs of genes are typically separated by six or seven edges.
Coherence of the FOCI network with known metabolic pathways
-----------------------------------------------------------
To assess the biological relevance of our estimated coexpression network we compared the composition of 38 known metabolic pathways (Table [1](#T1){ref-type="table"}) to our yeast coexpression FOCI network. In a biologically informative network, genes that are involved in the same pathway(s) should be represented as coherent pieces of the larger graph. That is, under the assumption that pathway interactions require co-regulation and coexpression, the genes in a given pathway should be relatively close to each other in the estimated global network.
We used a pathway query approach to examine 38 metabolic pathways relative to our FOCI network. For each pathway, we computed a quantity called the \'coherence value\' that measures how well the pathway is recovered in a given network model (see Materials and methods). Of the 38 pathways tested, 19 have coherence values that are significant when compared to the distribution of random pathways of the same size (*p*\< 0.05; see Materials and methods). Most of the pathways of carbohydrate and amino-acid metabolism that we examined are coherently represented in the FOCI network. Of each of the major categories of metabolic pathways listed in Table [1](#T1){ref-type="table"}, only lipid metabolism and metabolism of cofactors and vitamins are not well represented in the FOCI network.
The five largest coherent pathways are glycolysis/gluconeogenesis, the TCA cycle, oxidative phosphorylation, purine metabolism and synthesis of N-glycans. Other pathways that are distinctive in our analysis include the glyoxylate cycle (6 of 12 genes in largest coherent subnetwork), valine, leucine, and isoleucine biosynthesis (10 of 15 genes), methionine metabolism (6 of 13 genes), phenylalanine, tyrosine, and tryptophan metabolism (two subnetworks each of 6 genes). Several coherent subsets of the FOCI network generated by these pathway queries are illustrated in the Additional data file 1.
Combined analysis of core carbohydrate metabolism
-------------------------------------------------
In addition to being consistent with individual pathways, a useful network model should capture interactions between pathways. To explore this issue we queried the FOCI network on combined pathways and again measured its coherence. We illustrate one such combined query based on four related pathways involved in carbohydrate metabolism: glycolysis/gluconeogenesis, pyruvate metabolism, the TCA cycle and the glyoxylate cycle.
Figure [3](#F3){ref-type="fig"} illustrates the largest subgraph extracted in this combined analysis. The combined query results in a subset of the FOCI network that is larger than the sum of the subgraphs estimated separately from individual pathways because it also admits non-query genes that are connected to multiple pathways. The nodes of the graph are colored according to their membership in each of the four pathways as defined by the Kyoto Encyclopedia of Genes and Genomes (KEGG). Many gene products are assigned to multiple pathways. This is particularly evident with respect to the glyoxylate cycle; the only genes uniquely assigned to this pathway are *ICL1*(encoding an isocitrate lyase) and *ICL2*(a 2-methylisocitrate lyase).
In this combined pathway query the TCA cycle, glycolysis/gluconeogenesis, and glyxoylate cycle are each represented primarily by a single two-step connected subgraph (see Materials and methods). Pyruvate metabolism on the other hand, is represented by at least two distinct subgraphs, one including {*PCK1*, *DAL7*, *MDH2*, *MLS1*, *ACS1*, *ACH1*, *LPD1*, *MDH1*} and the other including {*GLO1*, *GLO2*, *DLD1*, *CYB2*}. This second set of genes encodes enzymes that participate in a branch of the pyruvate metabolism pathway that leads to the degradation of methylglyoxal (methylglyoxal → L-lactaldehyde → L-lactate → pyruvate and methylglyoxal → (*R*)-*S*-lactoyl-glutathione → D-lactaldehyde → D-lactate → pyruvate) \[[@B12],[@B13]\]. In the branch of methylglyoxal metabolism that involves *S*-lactoyl-glutathione, methyglyoxal is condensed with glutathione \[[@B12]\]. Interestingly, two neighboring non-query genes, *GRX1*(a neighbor of *GLO2*) and *TTR1*(neighbor of *CYB2*), encode proteins with glutathione transferase activity.
The position of *FBP1*in the combined query is also interesting. The product of *FBP1*is fructose-1,6-bisphosphatase, an enzyme that catalyzes the conversion of beta-d-fructose 1,6-bisphosphate to beta-D-fructose 6-phosphate, a reaction associated with glycolysis. However, in our network it is most closely associated with genes assigned to pyruvate metabolism and the glyoxylate cycle. The neighbors of *FBP1*in this query include *ICL1*, *MLS1*, *SFC1*, *PCK1*and *IDP3*. With the exception of *IDP3*, the promoters of all of these genes (including *FBP1*) have at least one upstream activation sequence that can be classified as a carbon source-response element (CSRE), and that responds to the transcriptional activator Cat8p \[[@B14]\]. This set of genes is expressed under non-fermentative growth conditions in the absence of glucose, conditions characteristic of the diauxic shift \[[@B15]\]. Considering other genes in the vicinity of *FBP1*in the combined pathway query we find that *ACS1*, *IDP2*, *SIP4*, *MDH2*, *ACH1*and *YJL045w*have all been shown to have either CSRE-like activation sequences and/or to be at least partially Cat8p dependent \[[@B14]\]. The association among these Cat8p-activated genes persists when we estimate the FOCI network without including the data of DeRisi *et al*. \[[@B15]\], suggesting that this set of interactions is not merely a consequence of the inclusion of data collected from cultures undergoing diauxic shift.
The inclusion of a number of other genes in the carbohydrate metabolism subnetwork is consistent with independent evidence from the literature. For example, McCammon *et al*. \[[@B16]\] identified *YER053c*as among the set of genes whose expression levels changed in TCA cycle mutants.
Although many of the associations among groups of genes revealed in these subgraphs can be interpreted either in terms of the query pathways used to construct them or with respect to related pathways, a number of association have no obvious biological interpretation. For example, the tail on the left of the graph in Figure [3](#F3){ref-type="fig"}, composed of *LSC1*, *PTR2*, *PAD1*, *OPT2*, *ARO10*and *PSP1*has no clear known relationship.
Locally distinct subgraphs
--------------------------
The analysis of metabolic pathways described above provides a test of the extent to which known pathways are represented in the FOCI graph. That is, we assumed some prior knowledge about network structure of subsets of genes and asked whether our estimated network is coherent *vis-à-vis*this prior knowledge. Conversely, one might want to find interesting and distinct subgraphs within the FOCI network without the injection of any prior knowledge and ask whether such subgraphs correspond to particular biological processes or functions. To address this second issue we developed an algorithm to compute \'locally distinct subgraphs\' of the yeast FOCI coexpression network as detailed in the Materials and methods section. Briefly, this is an unsupervised graph-search algorithm that defines \'interestingness\' in terms of local edge topology and the distribution of local edge weights on the graph. The goal of this algorithm is to find connected subgraphs whose edge-weight distribution is distinct from that of the edges that surround the subgraph; thus, these locally distinct subgraphs can be thought of as those vertices and associated edges that \'stand out\' from the background of the larger graph as a whole.
We constrained the size of the subgraphs to be between seven and 150 genes, and used squared marginal correlation coefficients as the weighting function on the edges of the FOCI graph. We found 32 locally distinct subgraphs, containing a total of 830 genes (Table [2](#T2){ref-type="table"}). Twenty-four out of the 32 subgraphs have consistent Gene Ontology (GO) annotation terms \[[@B17]\] with *p*-values less than 10^-5^(see Materials and methods). This indicates that most locally distinct subgraphs are highly enriched with respect to genes involved in particular biological processes or functions. Members of the 21 largest locally distinct subgraphs are highlighted in Figure [1](#F1){ref-type="fig"}. The complete list of subgraphs and the genes assigned to them is given in Additional data file 2.
The five largest locally distinct subgraphs have the following primary GO annotations: protein biosynthesis (subgraphs A and B); ribosome biogenesis and assembly (subgraph C); response to stress and carbohydrate metabolism (subgraph K); and sporulation (subgraph N). Several of these subgraphs show very high specificity for genes with particular GO annotations. For example, in subgraphs A and B approximately 97% (32 out of 33) and 95.5% (64 out of 67) of the genes are assigned the GO term \'protein biosynthesis\'.
Subgraph P is also relatively large and contains many genes with roles in DNA replication and repair. Similarly, 21 of the 34 annotated genes in Subgraph F have a role in protein catabolism. Three medium-sized subgraphs (S, T, U) are strongly associated with the mitotic cell cycle and cytokinesis. Other examples of subgraphs with very clear biological roles are subgraph R (histones) and subgraph Z (genes involved in conjugation and sexual reproduction). Subgraph X contains genes with roles in methionine metabolism or transport.
Some locally distinct subgraphs can be further decomposed. For example, subgraph K contains at least two subgroups. One of these is composed primarily of genes encoding chaperone proteins: *STI1*, *SIS1*, *HSC82*, *HSP82*, *AHA1*, *SSA1*, *SSA2*, *SSA4*, *KAR2*, *YPR158w*, *YLR247c*. The other group contains genes primarily involved in carbohydrate metabolism. These two subgroups are connected to each other exclusively through *HSP42*and *HSP104*.
Three of the locally distinct subgraphs - Q, W and CC - are composed primarily of genes for which there are no GO biological process annotations. Interestingly, the majority of genes assigned to these three groups are found in subtelomeric regions. These three subgraphs are not themselves directly connected in the FOCI graph, so their regulation is not likely to be simply an instance of a regulation of subtelomeric silencing \[[@B18]\]. Subgraph Q includes 26 genes, five of which (*YRF1-2*, *YRF1-3*, *YRF1-4*, *YRF1-5*, *YRF1-6*) correspond to ORFs encoding copies of Y\'-helicase protein 1 \[[@B19]\]. Eight additional genes (*YBL113c*, *YEL077c*, *YHL050c*, *YIL177c*, *YJL225c*, *YLL066c*, *YLL067c*, *YPR204w*) assigned to this subgraph also encode helicases. This helicase subgraph is closely associated with subgraph P, which contains numerous genes involved in DNA replication and repair (see Figure [1](#F1){ref-type="fig"}). Subgraph W contains 10 genes, only one of which is assigned a GO process, function or component term. However, nine of the 10 genes in the subgraph (*PAU1*, *PAU2*, *PAU4*, *PAU5*, *PAU6*, *YGR294w*, *YLR046c*, *YIR041w*, *YLL064c*) are members of the seripauperin gene family \[[@B20]\], which are primarily found subtelomerically and which encode cell-wall mannoproteins and may play a role in maintaining cell-wall integrity \[[@B18]\]. Another example of a subgraph corresponding to a multigene family is subgraph CC, which includes nine subtelomeric ORFs, six of which encode proteins of the COS family. Cos proteins are associated with the nuclear membrane and/or the endoplasmic reticulum and have been implicated in the unfolded protein response \[[@B21]\].
As a final example, we consider subgraph FF, which is composed of seven ORFs (*YAR010c*, *YBL005w-A*, *YJR026w*, *YJR028w*, *YML040w*, *YMR046c*, *YMR051c*) all of which are parts of Ty elements, encoding structural components of the retrotransposon machinery \[[@B22],[@B23]\]. This set of genes nicely illustrates the fact that delineating locally distinct groups can lead to the discovery of many interesting interactions. There are only six edges among these seven genes in the estimated FOCI graph, and the marginal correlations among the correlation measures of these genes are relatively weak (mean *r*\~ 0.62). Despite this, the local distribution of edge weights in FOCI graph is such that this group is highlighted as a subgraph of interest. Locally strong subgraphs such as these can also be used as the starting point for further graph search procedures. For example, querying the FOCI network for immediate neighbors of the genes in subgraph FF yields three additional ORFs - *YBL101w-A*, *YBR012w-B*, and *RAD10*. Both *YBL101w-A*and *YBR012w-B*are Ty elements, whereas *RAD10*encodes an exonuclease with a role in recombination.
Discussion
==========
Comparisons with other methods
------------------------------
Comparing the performance of different methods for analyzing gene-expression data is a difficult task because there is currently no \'gold standard\' to which an investigator can turn to judge the correctness of a particular result. This is further complicated by the fact that different methods employ distinct representations such as trees, graphs or partitions that cannot be simply compared. With these difficulties in mind, we contrast and compare our FOCI method to three popular approaches for gene expression analysis - hierarchical clustering \[[@B1]\], Bayesian network analysis \[[@B10]\] and relevance networks \[[@B7],[@B24],[@B25]\]. Like the FOCI networks described in this report, both Bayesian networks and relevance networks represent interactions in the form of network models, and can, in principle, capture complex patterns of interaction among variables in the analysis. Relevance networks also share the advantage with FOCI networks that, depending on the scoring function used, they can be estimated efficiently for very large datasets.
Comparison with relevance networks
----------------------------------
Relevance networks are graphs defined by considering one or more scoring functions and a threshold level for every pair of variables of interest. Pairwise scores that rise above the threshold value are considered significant and are represented by edges in the graph; interactions below this threshold are discarded \[[@B25]\]. As applied to gene-expression microarray data, the scoring functions used most typically have been mutual information \[[@B7]\] or a measure based on a modified squared sample correlation coefficient \[[@B24]\]).
We estimated a relevance network for the same 5007-gene dataset used to construct the FOCI network. The scoring function employed was  with a threshold value of ± 0.5. The resulting relevance network has 13,049 edges and a GCC with 1,543 vertices and 12,907 edges. The next largest connected subgraph of the relevance network has seven vertices and seven edges. There are a very large number of connected subgraphs (3,341) that are composed of pairs or singletons of genes.
To compare the performance of the relevance network with the FOCI network we used the pathway query approach described above to test the coherence of the 38 metabolic pathways described previously. Of the 38 metabolic pathways tested, nine have significant coherence values in the relevance network. These coherent pathways include: glycolysis/gluconeogenesis, the TCA cycle, oxidative phosphorylation, ATP synthesis, purine metabolism, pyrimidine metabolism, methionine metabolism, amino sugar metabolism, starch and sucrose metabolism. Two of these pathways - amino sugar metabolism and starch and sucrose metabolism - are not significantly coherent in the FOCI network. However, there are 12 metabolic pathways that are coherent in the FOCI network but not coherent in the relevance network. On balance, the FOCI network model provides a better estimator of known metabolic pathways than does the relevance network approach.
Comparison with hierarchical clustering and Bayesian networks
-------------------------------------------------------------
To provide a common basis for comparison with hierarchical clustering and Bayesian networks, we explored the dataset of Spellman *et al*. \[[@B26]\] which includes 800 yeast genes measured under six distinct experimental conditions (a total of 77 microarrays; this data is a subset of the larger analysis described in this paper). Spellman *et al.*\[[@B26]\] analyzed this dataset using hierarchical clustering. Friedman *et al.*\[[@B10]\] used their \'sparse candidate\' algorithm to estimate a Bayesian network for the same data, treating the expression measurements as discrete values. For comparison with Bayesian network analysis we referenced the interactions highlighted in the paper by Friedman *et al*. and the website that accompanies their report \[[@B27]\]. For the purposes of the FOCI analysis we reduced the 800 gene dataset to 741 genes for which there were no more than 10 missing values. We conducted a FOCI analysis on these data using a partial correlation threshold of 0.33. The resulting FOCI network had 1599 edges and a GCC of 700 genes (the 41 other genes are represented by subgraphs of gene pairs or singletons).
On the basis of hierarchical clustering analysis of the 800 cell-cycle-regulated genes, Spellman *et al.*\[[@B26]\] highlighted eight distinct coexpressed clusters of genes. They showed that most genes in the clusters they identified share common promoter elements, bolstering the case that these clusters indeed correspond to co-regulated sets of genes (see \[[@B26]\] for description and discussion of these clusters).
Applying our algorithm for finding locally distinct subgraphs to the FOCI graph based on these same data (with size constraints min = 7, max = 75) we found 10 locally distinct subgraphs. Seven of these subgraphs correspond to major clusters in the hierarchical cluster analysis (the MCM cluster of Spellman *et al.*\[[@B26]\] is not a locally distinct subgraph). At this global level both FOCI analysis and hierarchical clustering give similar results. While the coarse global structure of the FOCI and hierarchical clustering are similar, at the intermediate and local levels the FOCI analysis reveals additional biologically meaningful interactions that are not represented in the clustering analysis. An example of interactions at an intermediate scale involves the clusters referred to as Y\' and CLN2 in Spellman *et al.*\[[@B26]\] Genes of the CLN2 cluster are involved primarily in DNA replication. The Y\' cluster contains genes known to have DNA helicase activity. The topology of the FOCI network indicates that these are relatively distinct subgraphs, but also highlights a number of weak-to-moderate statistical interactions between the Y\' and CLN2 genes (and almost no interactions between the Y\' genes and any other cluster). Thus the FOCI network estimate provides inference of more subtle functional relationships that cannot be obtained from the clustering family of methods.
An example at a more local scale involves the MAT cluster of Spellman *et al.*\[[@B26]\] This cluster includes a core set of genes whose products are known to be involved in conjugation and sexual reproduction. In the FOCI network one of the locally distinct subgraphs is almost identical to the MAT cluster, and includes *KAR4*, *STE3*, *LIF1*, *FUS1*, *SST2*, *AGA1*, *SAG1*, *MFα2*and *YKL177W*(*MFα1*is not included in the FOCI analysis because there were more than 10 missing values). The FOCI analysis additionally shows that this set of genes is linked to another subgraphs that includes *AGA2*, *STE2*, *MFA1*, *MFA2*and *GFA3*. This second set of genes are also involved in conjugation, sexual reproduction, and pheromone response. *AGA1*and *AGA2*form the bridge between these two subgraphs (the proteins encoded by these two genes, Aga1p and Aga2p, are subunits of the cell wall glycoprotein *α*-agglutinin \[[@B28]\]). These two sets of genes therefore form a continuous subnetwork in the FOCI analysis, whereas the same genes are dispersed among at least three subclusters in the hierarchical clustering. We interpret the difference as resulting from the fact that the FOCI network can include relatively weak interactions among variables, as long as the variables are not first order conditionally independent. For example, the marginal correlation between *AGA1*and *AGA2*is only 0.63, between *AGA1*and *GFA1*is 0.59, and between *AGA2*and *MFA1*only 0.61. Hierarchical clustering or other analyses based solely on marginal correlations will typically fail to highlight such relatively weak interactions among genes.
Because hierarchical clustering constrains relationships to take the form of strict partitions or nested partitions, this type of analysis seems best suited to highlight the overall coarse structure of co-regulatory relationships. The FOCI method, because it admits a more complex set of topological relationships, is well suited to capturing both global and local structure of transcriptional interactions.
Graphical models, like the FOCI method, exploit conditional independence relationships to derive a model that can be represented using a graph or network structure. Unlike the FOCI model, general graphical models represent a complete factorization of a multivariate distribution. In the case of Bayesian networks it is also possible to assign directionality to the edges of the network model. However, these advantages come at the cost of complexity - Bayesian networks are costly to compute - and generally this complexity scales exponentially with the number of vertices (genes). The estimation of a FOCI network is computationally much less complex than the estimation of a Bayesian network. Both methods allow for a richer set of potential interactions among genes than does hierarchical clustering. We therefore expect that both methods should be able to highlight biologically interesting interactions, at both local and global scales. Friedman *et al.*\[[@B10]\] analyzed the 800-gene dataset of Spellman *et al*. \[[@B26]\] and highlighted a number of relationships that are assigned high confidence in their analysis. Relationships that were recovered under both a multinomial and Gaussian model include *STE2*-*MFA2*, *CTS1*-*DSE2*(*YHR143w*), *OLE1*-*FAA4*, *KIP3*-*MSB1*, *SHM2*-*GCV2*, *DIP5*-*ARO9*and *SRO4*-*YOL007c*. All of these relationships, with the exception of *SRO5*-*YOL007c*, are present in the FOCI analysis of the same data.
Comparisons of the local topology of each network, based on examining the edge relationships for a number of query genes, suggests that the FOCI and Bayesian networks are broadly similar. There are of course, examples of biologically interpretable interactions that are present in the FOCI analysis but not in the Bayesian network and vice versa. For example, using a multinomial model, Friedman *et al.*demonstrated an interaction between *ASH1*and *FAR1*, both of which are known to participate in the mating type switch in yeast. This relationship is absent in the FOCI network. Similarly, the relationship between *AGA1*and *AGA2*that is highlighted in the FOCI analysis does not appear in the multinomial Bayesian network analysis.
Review of FOCI assumptions
--------------------------
As with all analytical tools, careful consideration of the assumptions underlying the FOCI network method is necessary to understand the limits of the inferences one can draw. For example, our current framework limits consideration to linear relationships as measured by correlations and partial correlations. These assumptions may be relaxed, allowing for other types of distributions and relationships among variables (for example, monotone and curvilinear relationships), but there is an inevitable trade-off to be made in terms of computational complexity and statistical power. However, as seen in our analysis, many biologically interesting relationships among gene expression measures appear to be approximately linear. Biologically speaking, it is important to keep in mind that the graphs resulting from a FOCI analysis of gene-expression measurements should properly be considered coexpression or co-regulation networks and not genetic regulatory networks *per se*. While the clusters and patterns of coexpression summarized by the FOCI network may result from particular regulatory dynamics, no causal hypothesis of regulatory interaction is implied by the network.
Conclusions
===========
Biology demands that the analytical tools we use for functional genomics should be able to capture and represent complex interactions; practical considerations stemming from the magnitude and scope of genomic data require the use of techniques that are computable and relatively efficient. The FOCI framework we have used for representing genomic coexpression patterns in terms of a weighted graph satisfies both these constraints. FOCI networks are readily computable, even for very large datasets. Comparisons with known metabolic pathways show that many key biological interactions are captured by FOCI networks, and the algorithm we provide for finding locally distinct subgraphs provides a mechanism for discovering novel associations based on local graph topology. The subgraphs and patterns of interactions that we are able to demonstrate based on such analyses are strongly consistent with known biological processes and functions, indicating that the FOCI network method is a powerful tool for summarizing biologically meaningful coexpression patterns. Furthermore, the kinds of interactions captured by network analysis are typically more natural than the clustering family of analyses where biased and unstable results can be forced by the algorithm. Secondary analysis based on the network properties also reveal additional subtle structure. For example, our procedure for finding locally distinct subgraphs reveals associated genes whose pairwise interactions may be globally weak but relatively strong compared to their local interactions. While the results reported here focus on the analysis of gene expression measurements, the FOCI approach can be applied to any type of quantitative data making it a generally suitable technique for exploratory analyses of functional genomic data.
Materials and methods
=====================
A statistical/geometrical model for estimating coexpression networks
--------------------------------------------------------------------
The approach we employ to estimate coexpression networks is based on a general statistical technique we have developed for representing the associations among a large number of variables in terms of a weighted, undirected graph. The technique is based on the consideration of so-called \'first-order\' conditional independence relationships among variables, hence we call the graphs that result from such analyses first-order conditional independence, FOCI, networks. The network representation that results from a FOCI analysis also has a dual geometrical interpretation in terms of proximity relationships defined with respect to the geometry of correlations and partial correlations. We outline the statistical and geometrical motivations underlying our approach below.
First-order conditional independence networks
---------------------------------------------
A FOCI network is a graph, *G*= {*V*,*E*}, where the vertex set, *V*, represents the variables of interest and the edge set, *E*, represents interactions among the variables. *e*~*ij*~is an edge in *G*, if and only if there is no other variable in the analysis, *k*(*k*≠ *k*≠ *k*) such that  or , where
 is a modified partial correlation between *i*and *j*conditioned on *k*.  takes values in the range -1 ≤  ≤ 1.  is approximately zero when *i*and *j*are independent conditional on *k*.  is positive when the marginal correlation, *ρ*~*ij*~, and the standard partial correlation, *ρ*~*ij*.*k*~, agree in sign, and is negative otherwise. Cases where the marginal and conditional correlations are of opposite sign are examples of \'Simpson\'s paradox\', which usually indicates that there is a lurking or confounding effect of the conditioning variable (see \[[@B29]\] for a general discussion of such relationships).
While true biological interactions may sometimes lead to inverted conditional associations, their interpretation can be complicated; therefore in the analysis presented above, we did not connect edges when the relationships became inverted. However, one can also keep such edges for subsequent analysis if there is reasonable functional justification. When such sign-reversed edges are ignored, we will call this the sign-restricted FOCI network. This definition means that variables *i*and *j*are connected in the FOCI network if there is no other variable in the analysis for which *i*and *j*are conditionally independent or which causes an association reversal. Because we restrict the conditioning set to single variables, these are so called \'first-order\' conditional interactions (marginal correlations correspond to zero-order conditional interactions; partial correlations given two conditioning variables are second-order conditional interactions, etc). If *i*and *j*are conditionally independent given *k*we write this as (*i*⊥ *j*\|*k*). Using an information theoretic interpretation suggested by Lauritzen \[[@B9]\], the statement (*i*⊥ *j*\|*k*) implies that if we observe the variable *k*, there is no additional information about *i*that we gain by also observing *j*(and vice versa). Because the edges of the FOCI network indicate pairs of variables that are not conditionally independent, one can interpret the FOCI graph as a summary of all the pairwise interactions that can not be \'explained away\' by any other single variable in the analysis.
Unlike standard graphical models, a FOCI network does not represent a factorization of a multivariate distribution into the product of simpler distributions. However, below we show that a sign-restricted FOCI graph has a unique geometric interpretation in terms of proximity relationships in the multidimensional space that represents the correlations among variables. This geometric interpretation suggests that the FOCI model should be a generally useful approach for exploratory analyses of very high-dimensional datasets.
Our FOCI approach is similar to a framework developed by de Campos and Huete \[[@B30]\] for estimating belief networks. These authors developed an algorithm based on the application of zero- and first-order conditional independence test to learn the \'prior skeleton\' of a Bayesian network, followed by a refinement procedure that uses higher-order interactions sparingly.
Geometrical model of first-order conditional independence
---------------------------------------------------------
Above we described the FOCI network model in statistical terms. Here we provide a geometrical interpretation of FOCI graphs. We show that a FOCI network is equivalent to a proximity graph of the variables of interest (genes in the current analysis). More specifically, we demonstrate that a sign-restricted FOCI network is a \'Gabriel graph\' in the geometric space that represents the relationships among the variables.
A Gabriel graph, introduced by Gabriel and Sokal \[[@B31]\], is a type of proximity graph. Let *B*(*x*,*r*) denote an open *n*-sphere centered at the point *x*with radius *r*, and let *d*(*p*,*q*) denote the Euclidean distance function. Given a set of points, *P*= {*p*~1~*p*~2~, \..., *p*~*n*~}, in an *n*-dimensional Euclidean space, (*p*~*i*~, *p*~*j*~) is an edge in the Gabriel graph if no other point, *p*~*k*~(*i*≠ *k*, *j*≠ *k*) in *P*falls within the diameter sphere defined by *B*((*p*~*i*~= *p*~*j*~)/2, *d*(*p*~*i*~, *p*~*j*~)/2). That is, *p*~*i*~and *p*~*j*~are connected in the Gabriel graph if no other point falls within the sphere that has the chord *p*~*i*~, *p*~*j*~as its diameter \[[@B32]\].
Geometry of marginal and partial correlations and conditional independence
--------------------------------------------------------------------------
One can represent random variables as vectors in the space of the observations (often called object space or subject space \[[@B33],[@B34]\]). In such a representation, a set of mean centered and standardized variables correspond to unit vectors whose heads lie on the surface of an *n*-dimensional hypersphere (where *n*is the number of observations). In this representation, the correlation between two random variables, *x*and *y*, is given by the cosine of the angle between their vectors. We will refer to this construction as the \'correlational hypersphere\'. The partial correlation between *x*and *y*given *z*is equivalent to the cosine of the angle between the residual vectors obtained by projecting *x*and *y*onto *z*. The vectors *x*, *y*and *z*form the vertices, A, B, and C, of a spherical triangle on that hypersphere with associated angles *γ*, *λ*, and *φ*. Then, *ρ*~*xy*.*z*~= cos(*φ*), *ρ*~*xz*.*y*~= cos(*λ*), and *ρ*~*yz*.*x*~= cos(*γ*) \[[@B35]\]. Given this geometric construction of partial correlations in terms of spherical triangles, conditional independence, defined as *ρ*~*xy*.*z*~= 0 for the multivariate normal, is obtained when cos(*φ*) = 0 (that is, when the *φ*= *π*/2). The set of *z*vectors that satisfy this condition defines a circle (actually a hypersphere of dimension *n*- 1) on the hypersphere whose diameter is the spherical chord between *x*and *y*. If the projection of *z*onto the hypersphere lies outside of this circle then *ρ*~*xy*.*z*~is positive, inside the circle *ρ*~*xy*.*z*~is negative (with *ρ*~*xy*.*z*~= -1 along the chord between *x*and *y*).
The sign-restricted FOCI network construction corresponds to the graph obtained by connecting variables *i*and *j*only if no third variable falls within the diameter sphere defined by *i*and *j*on the correlational hypersphere, or by the diameter sphere defined by *i*and -*j*when *r*~*ij*~\< 0 (allowing for deviations due to sampling). This is the same criteria of proximity that defines a Gabriel graph. A FOCI graph is therefore a summary of relative proximity relationships among the variables of interest, defined with respect to the geometry of correlations when restricted to the cases when the partial correlation signs are consistent with the marginal correlations.
FOCI network algorithm
----------------------
A simple algorithm for estimating a network based on first-order conditional independence relationships is described below. The results of this algorithm can be represented as a graph where the vertices represent the variables of interest (genes) and the edges represent interactions among variables that show at least first-order conditional dependence. A library of functions for estimating FOCI networks, implemented in the Python programming language, is available from the authors on request.
We use vanishing partial correlations \[[@B8],[@B36]\] to test whether pairs of genes are conditionally independent given any other single variable in the analysis. Strictly speaking, if the data are not multivariate normal, then zero partial correlations need not imply conditional independence, but rather conditional uncorrelatedness \[[@B37]\]. However, regardless of distributional assumptions, zero partial correlations among variates are of interest as long as the relationship between the variables has a strong linear component \[[@B38]\].
### FOCI algorithm
1\. **Estimate marginal associations.**For a set of *p*variables, indexed by *i*and *j*, calculate the *p*× *p*correlation matrix, *C*, where *C*~*i*,*j*~= corr(*i, j*) for all *i*, *j*; *i*= 1\...*p*, *j*= 1\...*p*.
2\. **Construct saturated graph.**Construct a *p*× *p*adjacency matrix, *G*. Let *G*~*i*,*j*~= 1 for all *i*, *j*.
3\. **Prune zero-order independent edges.**For each pair of variables, (*i*, *j*), if *C*~*i*,*j*~\<*T*~*crit*~(or some appropriately chosen function, *f*(*C*~*i*,*j*~) \<*T*~*crit*~), where *T*~*crit*~is a threshold value for determining marginal/conditional independence (see below), then set *G*~*i*,*j*~= 0. *G*defines a marginal independence graph.
4\. **Estimate first-order relationships.**For each pair of variables (*i*, *j*) in *G*calculate , the minimum partial correlation between *i*and *j*, conditioned on each of the other variables in the analysis taken one at a time.  for all *k*such that *i*≠ *k*and *j*≠ *k*and (*i*, *k*) and (*j*, *k*) are both edges in *G*.  is the sample modified partial correlation coefficient as defined in equation (1).
5\. **Prune first-order independent edges.**If  \<*T*~*crit*~(or *f*() \<*T*~*crit*~then set *G*~*i*,*j*~= 0.
The resulting adjacency matrix *G*, can be represented as an undirected graph, with *p*vertices, whose edge set is defined by the non-zero elements in *G*. The edges of this graph can be represented as either unweighted (all edges having equal weight) or with weights defined by some function of corr(*i, j*) or . If we assume multivariate normality we can use Fisher\'s z-transformation \[[@B39]\] to normalize the expected distribution of correlation/partial correlations and use standard tables of the normal distribution to define *T*~*crit*~for a given edge-wise false-positive rate. Alternatively, one can define *T*~*crit*~by other methods such as via permutation analysis to define a null distribution for . While the FOCI approach requires that one define a critical threshold for determining conditional independence, this threshold is in theory a function of the sample size and the null distribution of  rather than the somewhat fuzzier distinction between \'strong\' and \'weak\' correlation that most pairwise network estimation approaches require.
Estimating the yeast FOCI coexpression network
----------------------------------------------
We used the FOCI network algorithm to estimate a coexpression network for the budding yeast, *Saccharomyces cerevisae*. The data used in our analysis are drawn from publicly available microarray measurements of gene expression described in DeRisi *et al.*\[[@B15]\], Chu *et al.*\[[@B40]\] and Spellman *et al.*\[[@B26]\]. These data represent relative measurements of gene expression taken at different points in the cell cycle in yeast cultures synchronized using a variety of different mechanisms \[[@B26]\] or in the context of specific physiological process such as diauxic shift \[[@B15]\] or sporulation \[[@B40]\]. The data were log~2~-transformed, duplicate and missing data were removed and any ORFs listed as \'dubious\' in the *Saccharomyces*Genome Database as of December 2003 were filtered out. The final dataset consisted of expression measurements for 5,007 ORFs represented by 87 microarrays (see Rifkin *et al.*\[[@B41]\] for a full description of the pretreatment of these data). The mean centered data were treated as continuous variables for the purposes of our analysis.
Microarray measurements, especially spotted microarrays, are subject to a variety of systematic effects such as those due to dye biases and print-tip effects, and a number of methods have been devised to normalize and correct for such biases \[[@B42],[@B43]\]. However, the data analyzed here include both spotted DNA microarray measurements and expression measurements based on Affymetrix arrays (experiments of Cho *et al.*\[[@B44]\] as reported by Spellman *et al.*\[[@B26]\]), making it difficult to apply a consistent correction. Another consideration is that the assemblage of experiments considered by Spellman *et al.*\[[@B26]\], have been frequently used to illustrate the utility of new analytical methods \[[@B7],[@B10],[@B45]\]. To facilitate comparison with previous reports we have chosen to analyze these data without any transformations other than the log-transformation and mean-centering described above.
As noted above, zero partial correlations are exactly equivalent to conditional independence only for multivariate normal distributions. However, from the perspective of exploratory analyses, the more important assumption is that the relationships among the gene expression measures are predominantly linear. We tested each of these assumptions as follows. We used a Cramer-von Mises statistic \[[@B46]\] to test for the normality of each vector of gene expression measurements. Approximately 59% of the univariate distributions of the variables are consistent with normality (*p*\< 0.05). While a majority of the univariate distributions are approximately normal, a significant proportion of the trivariate distributions are clearly not multivariate normal. As a crude test of linearity for bivariate relationships we calculated linear regressions for 10,000 random pairs of gene expression measures (randomly choosing one of the pair as the dependent variables), and performed runs tests \[[@B47]\] for randomness of the signs of the residuals from each regression. Significant deviations from non-linearity in the bivariate relationships should manifest themselves as non-random runs of positive or negative residuals. For approximately 95% of the runs tests we can not reject the null hypothesis of randomness in the signs of the residuals (*p*\< 0.05). We therefore conclude that the assumption of quasi-linearity is valid for a large number of the pairwise relationships.
Given these observations, in order to define an appropriate partial correlation threshold, *T*~*crit*~, for these data we considered both permutation tests and false-positive rates based on asymptotic expectations for the distribution of first-order partial correlations (see above). Permutation tests were carried out by independently randomizing the values for each gene expression variable such that each gene had the same mean and variance as its original observation vector, but both the marginal and partial correlations had an expected value of zero. We then sampled 1,000 such randomized variables and examined the distribution of  for every pair of variables in this sample. For *p*≤ 0.001 the permutation test indicates a value of *T*~*crit*~\~ 0.3. The asymptotic threshold for *p*≤ 0.001 based on Fisher\'s z-transform is *T*~*crit*~\~ 0.3. We used the slightly more conservative value of *T*~*crit*~\~ 0.34.
Metabolic pathways
------------------
We used 38 metabolic pathways as documented in KEGG release 29.0, January 2004 \[[@B48],[@B49]\] to test the biological relevance of the estimated yeast coexpression network. These pathways are listed in Table [1](#T1){ref-type="table"}. In our analysis we only considered metabolic pathways for which more than 10 pathway genes were represented in the gene expression dataset described above. The metabolic pathways we studied are not independent, as there are a number of genes whose products participate in two or more metabolic processes. However, for the purposes of the present analysis we have treated each pathway as independent.
Testing the coherence of pathways using pathway queries
-------------------------------------------------------
We used the following method to compare our FOCI network to the metabolic pathways from KEGG. We say that a subset of vertices, *H*, is two-step connected in the graph *G*if no vertex in *H*is more than two edges away from at least one other element of *H*. Given a set of genes assigned to a pathway (the query genes), we computed the set of two-step connected subgraphs for the query genes in the GCC of our yeast coexpression network. This procedure yields one or more subgraphs that are composed of query (pathway) genes plus non-query genes that are connected to at least two pathway genes. We used two steps as a criterion for our pathway queries because our estimate of the distribution of path distances (Figure [2b](#F2){ref-type="fig"}) indicated that more than 99% of gene pairs in our network are separated by a distance greater than two steps. Therefore, two-step connected subgraphs in our coexpression network represent sets of genes which are relatively close to each other with respect to the topology of the graph as a whole.
Suppose we have a set of query genes from a known pathway denoted as *P*= {g~1~,*g*~2~,\...*g*~*k*~}. We construct the two-step connected graph of the elements of *P*from our FOCI estimated network denoted as *F*~*P*~⊃ *P*. That is, *F*~*P*~is a subgraph from the FOCI network that contains elements of *P*and its neighbors according to the two-step connected criteria described above. *F*~*P*~may itself be composed of one or more connected components. We define *F*~*Pmax*~as the connected component of *F*~*P*~that has the greatest overlap with *P*. If the FOCI network was completely coherent with respect to *P*, than *F*~*P*~should constitute a single connected component (that is, *F*~*Pmax*~= *F*~*P*~) whose vertex set completely overlaps *P*(that is, \|*F*~*p*~∩ *P*\| = \|*P*\|). For cases in which the query pathway is less than perfectly represented in the estimated network we measure the degree of coherence as \|*F*~*Pmax*~∩ *P*\| / \|*P*\|). We refer this ratio the \'coherence value\' of the pathway *P*in the network of interest. However, we note that in a completely connected graph (that is, every vertex is connected to every other vertex), every possible pathway query would be maximally coherent but so would any random set of genes. It is therefore necessary to compare the coherence of a given pathway to the distribution of coherence values for random pathways composed of the same number of genes drawn from the same network. We estimated this distribution by using a randomization procedure in which we used 1,000 replicate random pathways to estimate the distribution of coherence values for pathways of different sizes. In Table [1](#T1){ref-type="table"}, pathways that are significantly more coherent than at least 95% of random pathways are marked with an asterisk.
Locally distinct subgraphs of coexpression networks
---------------------------------------------------
We describe an algorithm for extracting a set of \'locally distinct\' subgraphs from an edge-weighted graph. We assume that the edge-weights of the graph are measures of the strength of association between the variables of the interest. We define a locally distinct subgraph as a subgraph in which all edges within the subgraph are stronger than edges that connect subgraph vertices to vertices not within the subgraph. Such subgraphs are \'locally distinct\' because they are defined not by an absolute threshold on edge strengths, but rather by a consideration of the local topology of the graph and the distribution of edge weights. We describe an algorithm for finding locally distinct subgraphs below.
### An algorithm for finding locally distinct subgraphs
Let *G*= {*V*, *E*} and *w*:*E*→ **R**be an edge-weighted graph where *w*(*e*) is the edge weight function, and \|*V*\| = *p*and \|*E*\| = *q*. Define an ordering on *E*, *O*(*E*) = (*e*~1~,*e*~2~,\...,*e*~*q*~), such that *w*(*e*~*i*~) ≥ *w*(*e*~*j*~) for all *i*≤ *j*(that is, order the edges from strongest to weakest). Let *G*(*τ*)*=*{*V*, *E*(*τ*)} be a subgraph of *G*obtained by deleting all edges, *e*, such that *w*(*e*) \<*e*~*τ*~. *G*(*τ*) an edge-level graph. Also let  denote the *k*connected components of *G*(*τ*). Let Ω = *C*~1~∪ *C*~2~∪ ... *C*~*n*~. Define *L*~*α*,*ζ*~= {*l*~1~,*l*~2~,\...,*l*~*m*~} where *l*~*i*~⊆ Ω, *l*~*i*~∩ *l*~*j*~=  (*i*≠ *j*) and *α*≤\|*l*~*i*~\|≤ *ζ*. That is, *L*~*α*,*ζ*~is a collection of disjoint subgraphs of *G*, where every *l*~*i*~is a connected component of some *G*(*τ*) and the size of *l*~*i*~is between *α*and *ζ*. We call the elements of *L*~*α*,*ζ*~the *α*,*ζ*-constrained locally distinct subgraphs of *G*. We say *L*~*α*,*ζ*~is optimal if \|*l*~*i*~∪ *l*~*j*~... *l*~*m*~\| is maximal and \|*L*~*α*,*ζ*~\| is minimal. Our goal is to find the optimal *L*~*α*,*ζ*~for the graph *G*given the constraints *α*and *ζ*. A simple algorithm for calculating the *L*~*α*,*ζ*~is as follows:
1. let *L*← , *i*= 0
2. while *i*≤ *q*:
3. calculate *G*(*i*) and *C*~*i*~
4. for  in *C*~*i*~:
5. if  :
6. for *l*in *L*:
7. if  :
8. *L*← *L*- {*l*}
9. 
10. *i*= *i*+ 1
11. *L*~*α*,*ζ*~← *L*
The algorithm is straightforward. At each iteration, *i*, we calculate the connected components of the edge-level graph, *G*(*i*), and add those components which satisfy the size constraints to the candidate list *L*. Lines 6-8 of the algorithm serve to eliminate from *L*any non-maximal components.
Biological significance of locally distinct subgraphs
-----------------------------------------------------
We applied the locally distinct subgraph algorithm to our yeast FOCI coexpression network. We used pairwise marginal correlations as the edge-weighting function, and set the size constraints as *α*= 7, *ζ*= 150. The subgraph search given these constraints yielded 32 locally distinct subgraphs (see Table [2](#T2){ref-type="table"} and Additional data file 2). For each locally distinct subgraph found we used the SGD Gene Ontology (GO) term finder of the *Saccharomyces*Genome Database \[[@B50],[@B51]\] to search the set of genes in each subgraph for significant shared GO terms. We excluded from the term finder search any genes for which no biological process or molecular function term was assigned. Table [2](#T2){ref-type="table"} summarizes the primary GO terms assigned to each subgraph and the number of genes labeled with that GO term is shown in parentheses. The *p*-values in Table [2](#T2){ref-type="table"} indicate the frequency at which one would expect to find the same number of genes assigned to the given GO term in a random assemblage of the same size.
Additional data files
=====================
Additional data are available with the online version of this article. Additional data file [1](#s1){ref-type="supplementary-material"} provides supplementary figures illustrating the connectivity distribution (on a log-log scale) of the estimated yeast FOCI network and additional examples of coherent subgraphs of the FOCI network generated by querying with known metabolic pathways. Additional data file [2](#s2){ref-type="supplementary-material"} contains a table detailing each of the 32 locally distinct subgraphs generated from the yeast FOCI network via the unsupervised graph search algorithm described in the text. A listing is provided for each locally distinct subgraphs describing yeast ORFs assigned to that subgraph and the Yeast GO Slim annotations associated with each ORF.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
Supplementary figures illustrating the connectivity distribution (on a log-log scale) of the estimated yeast FOCI network and additional examples of coherent subgraphs of the FOCI network generated by querying with known metabolic pathways
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
A table detailing each of the 32 locally distinct subgraphs generated from the yeast FOCI network via the unsupervised graph search algorithm described in the text
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
This research was facilitated by an NSF Minority Postdoctoral Research Fellowship (P. Magwene) and by NIH Grant 1P20GM069012-01 and a Penn Genomic Institute grant (J. Kim). We thank members of the Kim lab for constructive comments and critiques of the methods described in this paper.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Simplification of the yeast FOCI coexpression network constructed by retaining the 4,000 strongest edges (= 1,729 vertices). The colored vertices represent a subset of the locally distinct subgraphs of the FOCI network; letters are as in Table 2, and further details can be found there. Some of the locally distinct subgraphs of Table 2 are not represented in this figure because they involve subgraphs whose edge weights are not in the top 4,000 edges.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Topological properties of the yeast FOCI coexpression network. Distribution of **(a)**vertex degrees and **(b)**path lengths for the network.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Largest connected subgraph resulting from combined query on four pathways involved in carbohydrate metabolism: glycolysis/gluconeogenesis (red); pyruvate metabolism (yellow); TCA cycle (green); and the glyoxylate cycle (pink). Genes encoding proteins involved in more than one pathway are highlighted with multiple colors. Uncolored vertices represent non-pathway genes that were recovered in the combined pathway query. See text for further details.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Summary of queries for 38 metabolic pathways against the yeast FOCI coexpression network
:::
Pathway Number of genes(in KEGG) Size of largest coherent subnetwork(s)
----------------------------------------------------- -------------------------- ----------------------------------------
**Carbohydrate metabolism**
Glycolysis/gluconeogenesis 41 (47) 18\*
Citrate cycle (TCA cycle) 27 (30) 18\*
Pentose phosphate pathway 20 (27) 6\*
Fructose and mannose metabolism 39 (46) 4
Galactose metabolism 25 (30) 8\*
Ascorbate and aldarate metabolism 11 (13) 3
Pyruvate metabolism 32 (34) 8\*
Glyoxylate and dicarboxylate metabolism 12 (14) 6\*
Butanoate metabolism 27 (30) 7\*
**Energy metabolism**
Oxidative phosphorylation 53 (76) 31\*
ATP synthesis 21 (30) 7\*
Nitrogen metabolism 24 (27) 3
**Lipid metabolism**
Fatty acid metabolism 13 (17) 3
**Nucleotide metabolism**
Purine metabolism 87 (99) 34\*
Pyrimidine metabolism 72 (80) 15\*
Nucleotide sugars metabolism 11 (14) 2
**Amino acid metabolism**
Glutamate metabolism 25 (27) 3
Alanine and aspartate metabolism 26 (27) 7\*
Glycine, serine and threonine metabolism 36 (42) 7\*
Methionine metabolism 13 (14) 6\*
Valine, leucine and isoleucine biosynthesis 15 (16) 10\*
Lysine biosynthesis 16 (20) 3
Lysine degradation 26 (30) 4
Arginine and proline metabolism 20 (24) 5\*
Histidine metabolism 20 (25) 3
Tyrosine metabolism 27 (34) 2
Tryptophan metabolism 20 (25) 2
Phenylalanine, tyrosine and tryptophan biosynthesis 21 (23) 6\*
**Metabolism of complex carbohydrates**
Starch and sucrose metabolism 118 (139) 29
N-Glycans biosynthesis 43 (49) 13\*
O-Glycans biosynthesis 18 (20) 2
Aminosugars metabolism 16 (20) 2
Keratan sulfate biosynthesis 18 (20) 2
**Metabolism of complex lipids**
Glycerolipid metabolism 56 (68) 12\*
Inositol phosphate metabolism 87 (103) 10
Sphingophospholipid biosynthesis 101 (118) 11
**Metabolism of cofactors and vitamins**
Vitamin B6 metabolism 11 (14) 2
Folate biosynthesis 14 (17) 1
The values in the second column represent the number of pathway genes represented in the GCC of the yeast FOCI graph, with the total number of genes assigned to the given pathway in parentheses. The third column indicates the number of pathway genes in the largest coherent subgraph resulting from each pathway query. Pathways represented by coherent subgraphs that are significantly larger than are expected at random (*p*\< 0.05) are marked with asterisks.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Summary of locally distinct subgraphs of the yeast FOCI coexpression network
:::
Subgraph Number of genes Number unkown Major GO terms *p*-value
---------- ----------------- --------------- ----------------------------------------------------------------------------------------------------------- --------------
A 33 0 Protein biosynthesis (32) **1.82e-30**
B 67 2 Protein biosynthesis (64) **2.20e-61**
C 124 26 Ribosome biogenesis and assembly (74) **2.10e-89**
D 10 0 Glycolysis/gluconeogenesis (8) **6.29e-20**
E 7 1 Carboxylic/organic acid metabolism (4) **5.07e-05**
F 41 7 Ubiquitin dependent protein catabolism (21) **1.37e-31**
G 14 4 Cell organization and biogenesis (7) 1.60e-04
H 7 0 Main pathways of carbohydrate metabolism (4) **2.46e-07**
I 13 0 Electron transport (7) **2.00e-15**
J 13 0 Glutamate biosynthesis/TCA cycle (4) **7.09e-10**
K 71 25 Response to stress (17); carbohydrate metabolism (13) **3.94e-11**
L 10 4 Response to stress (2) 3.35e-02
N 149 51 Sporulation (27) **2.23e-29**
M 5 2 Mitochondrial matrix (5); mitochondrial ribosome (4) **2.83e-09**
O 7 2 Meiosis (4) **3.77e-07**
P 52 13 Cell proliferation (32); DNA replication and chromosome cycle (28) **1.12e-28**
Q 26 21 Telomerase-independent telomere maintenance (5) **1.82e-14**
R 7 0 Chromatin assembly/disassembly (7) **4.25e-18**
S 14 5 Cell wall (4); bud (4) **4.47e-05**
T 24 8 Cell proliferation (15); mitotic cell cycle (9) **6.54e-16**
U 21 4 Cell separation during cytokinesis (4); cell proliferation (9); cell wall organization and biogenesis (5) **5.27e-10**
V 12 4 Metabolism (7) 2.48e-02
W 10 9 Nine of ten are members of the seripauperin gene family NA
X 9 0 Sulfur amino acid metabolism (6); amino acid metabolism (3) **3.33e-13**
Y 7 1 Cell growth and maintenance (6) 7.50e-04
Z 19 2 Conjugation with cellular fusion (13) **1.82e-21**
AA 8 4 Biotin biosynthesis (2) **1.81e-06**
BB 7 0 Response to abiotic stimulus (2) 1.48e-02
CC 9 5 Six of nine members belong to COS family of subtelomerically encoded proteins NA
DD 18 7 Cell growth and/or maintenance (8) 4.43e-03
EE 11 3 Vitamin B6 metabolism (2) **2.58e-05**
FF 7 0 Ty element transposition (7) **6.01e-14**
The columns of the table summarize the total size of the locally distinct subgraph, the number of genes in the subgraph that are unannotated (according to the GO Slim annotation from the Saccharomyces Genome Database of December 2003), the primary GO term(s) associated with the subgraph, and a p-value indicating the frequency at which one would expect to find the same number of genes assigned to the given GO term in a random assemblage of the same size.
:::
|
PubMed Central
|
2024-06-05T03:55:51.857852
|
2004-11-30
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545795/",
"journal": "Genome Biol. 2004 Nov 30; 5(12):R100",
"authors": [
{
"first": "Paul M",
"last": "Magwene"
},
{
"first": "Junhyong",
"last": "Kim"
}
]
}
|
PMC545796
|
Rationale
=========
Since complete genome sequences have become available, the amount of annotated genes has increased dramatically. These advances have allowed the systematic comparison of the gene content of different organisms, leading to the conclusion that organisms share the majority of their genes with only relatively few species-specific genes. On this basis, one can develop strategies to infer gene annotations from model species to less experimentally tractable organisms. However, such functional inferences require the definition of species-independent annotation policies.
In this regard, the Gene Ontology consortium \[[@B1]\] has been created to develop a unified view of gene functional annotations for different model organisms. Three structured vocabularies (or ontologies) have been proposed, which allow the description of molecular functions, biological processes and cellular locations of any gene product, respectively. Whereas the majority of GO terms are common to several organisms, some of them are specific to a few organisms only, enabling the description of some aspects of gene function which are specific to few lineages only. Within each of these ontologies, the terms are organized in a hierarchical way, according to parent-child relationships in a directed acyclic graph (DAG). This allows a progressive functional description, matching the current level of experimental characterization of the corresponding gene product. The hierarchical organization of the gene ontology is particularly well adapted to computational processing and is used for the functional annotations of gene products of several model organisms such as budding yeast \[[@B2]\], *Drosophila*\[[@B3]\], mouse \[[@B4]\], nematode \[[@B5]\] and *Arabidopsis*\[[@B6]\]. More recently, GO annotations for human genes have been proposed in the context of the GOA project \[[@B7]\].
In parallel, the recent development of new high-throughput methods has generated an enormous amount of functional data and has motivated the development of dedicated analysis tools. For instance, one might wonder whether genes detected as being coexpressed in a DNA chip experiment are related in terms of molecular or cellular function. In practical terms, we address here the following generic questions. First, are there statistically over- or under-represented GO terms associated with a given gene set, compared to the distribution of these terms among the annotations of the complete genome? Second, among a particular gene set, are there closely functionally related gene subsets? And third, are there genes having GO similarities with a given probe gene?
To formulate such questions properly in a well defined mathematical framework, we have developed a set of methods and tools, collectively called GOToolBox, to process the GO annotations for any model organism for which they are available (Figure [1](#F1){ref-type="fig"}).
All the programs are written in PERL and use the CGI and DBI modules. All the ontology data and the gene-GO terms associations are taken from the GO consortium website. These data are structured in a PostGreSQL relational database, which is updated monthly. Statistics are calculated using the R statistical programming environment. The web implementation of the GOToolBox is accessible at \[[@B8]\].
Features
========
In this section, we describe the five main functionalities of the GOToolBox suite. Two of them (GO-Proxy and GO-Family) are not encompassed by any other GO-processing tool currently available (see also \'Comparison of the GOToolBox with other GO-based analysis programs\').
Dataset creation
----------------
The first step in analyzing gene datasets consists in retrieving, for each individual gene of the dataset, all the corresponding GO terms and their parent terms using the Dataset creation program. The genomic frequency of each GO term associated with genes present in the dataset is then calculated. The resulting information is structured and stored in a data file, available for download on the GOToolBox server for one week. This file contains also the counts of terms within a reference gene dataset (genome or user-defined), and can then be used as an input for the GO-Stats and GO-Proxy programs described below.
Ontology options
----------------
An optional tool, GO-Diet, allows either the reduction of the term dataset to a slim GO hierarchy (either one proposed by the GO consortium or a user-defined one) or the restriction of the considered terms to a chosen depth of the ontology. It is also possible to filter terms based on the way these have been assigned to the gene products (evidence code). This tool is useful to decrease the number of GO terms associated with a gene dataset, thereby facilitating the analysis of the results of programs described below, particularly when the input gene list and/or the number of associated GO terms is large. Note that the GO-Diet program can generate a gene-term association file in the TLF format, allowing the use of GO terms as gene labels with the TreeDyn tree drawing program \[[@B9]\]. The GO-Diet options are proposed in the Dataset-Creation form.
GO term statistics
------------------
Frequencies of terms within the dataset are calculated and compared with reference frequencies (for example with genomic frequencies or with the frequencies of these terms in the complete list of genes spotted on an array). This procedure allows the delineation of enrichments or depletions of specific terms in the dataset. The probability of obtaining by chance a number *k*of annotated genes for a given term among a dataset of size *n*, knowing that the reference dataset contains *m*such annotated genes out of *N*genes, is then calculated. This test follows the hypergeometric distribution described in Equation 1:
where the random variable *X*represents the number of genes within a given gene subset, annotated with a given GO term. Implemented in the GO-Stats tool, this formula permits the automatic ranking of all annotation terms, as well as the evaluation of the significance of their occurrences within the dataset. An illustration of such an approach is given in \'Mining biological data\'. A typical GO-Stats output is presented in Figure [2](#F2){ref-type="fig"}.
GO-based gene clustering
------------------------
The goal of the GO-Proxy tool is to group together functionally related genes on the basis of their GO terms. The rationale sustaining our method is that the more genes have common GO terms, and the less they have specific associated terms, the more likely they are to be functionally related. For any two genes of the gene set, the program calculates an annotation-based distance between genes, taking into account all GO terms that are common to the pair and terms which are specific to each gene. Indeed, any two genes can have 0, 1 or several shared GO terms (common terms) and a variable number of terms specific for each gene (specific terms). This distance is based on the Czekanowski-Dice formula (Equation 2):
In this formula, *x*and *y*denote two genes, *Terms(x)*and *Terms(y)*are the lists of their associated GO terms, *\#*stands for \'number of \' and Δ for the symmetrical difference between the two sets. This distance formula emphasizes the importance of the shared GO terms by giving more weight to similarities than to differences. Consequently, for two genes that do not share any GO terms, the distance value is 1, the highest possible value, whereas for two genes sharing exactly the same set of GO terms, the distance value is 0, the lowest possible value. All possible binary pairs of genes from the dataset are considered, resulting in a distance matrix.
Next this matrix is processed with a clustering algorithm, such as the WPGMA algorithm, and a functional classification tree is drawn, in which the leaves correspond to input genes. On the basis of this tree, classes can be defined, for instance by using partition rules, and the statistical relevance of the terms associated with each class is calculated using the method described for GO-Stats. The Czekanowski-Dice distance and the corresponding clustering have already proved their effectiveness in delineating protein functional classes derived from the analysis of protein-protein interaction graphs \[[@B10]\].
Finding GO-related genes
------------------------
A last tool, GO-Family, aims at finding genes having shared GO terms with a user-entered gene, on the basis of a functional similarity calculation. It searches the genomes either of one or several supported species (five at the moment). Given an input gene name, the program retrieves the associated GO terms and compares them with those of all other genes by calculating a functional similarity percentage. The program then returns the list of similar genes, sorted by score. By similar genes, we mean either genes having more than one common associated term, or genes which have different associated terms but one or more common parent terms.
When measuring the similarity percentage *S*between the input gene A and another gene G, one can identify terms that are common to the two genes (Tc), and terms that are specific to A (Ta) and G (Tg). Three different similarity measures have been implemented and proposed to the user:
Si = (Tc/(Ta+Tc)) × 100 (3)
Sp = (Tc/(Ta+Tg+Tc)) × 100 (4)
Scd = (1 - ((Ta+Tg)/(Ta+Tg+2Tc))) × 100 (5)
respectively called similarity percentage relative to the input gene (*Si*), similarity percentage relative to the pair of genes (*Sp*) and Czekanowski-Dice proximity percentage (*Scd*). The results are then ranked by decreasing similarity values. A typical GO-Family output is presented in Figure [3](#F3){ref-type="fig"}.
Mining biological data with the GOToolBox
=========================================
In this section, we provide two examples showing how combinations of several GO analysis tools can be used to validate or further delineate gene functional classifications.
Application of GOToolBox to the study of protein-protein interaction networks
-----------------------------------------------------------------------------
PRODISTIN \[[@B10]\] is a functional classification method for proteins, based on the analysis of a protein-protein interaction network, that aims to compare and predict a cellular role for proteins of unknown function. Given a set of proteins and a list of interactions between them, a distance is calculated between all possible pairs of proteins. A distance matrix is then generated, to which the NJ clustering algorithm is applied. A classification tree is then built, within which functional classes are defined, based on the annotation terms associated with the proteins involved in known biological processes. GO-Diet and GO-Stats are useful at two steps of the analysis (Figure [4a](#F4){ref-type="fig"}).
The first is to generate the GO annotation set necessary to define the functional classes of proteins. In this particular study devoted to the yeast interactome, the term dataset was fitted to the fourth ontology level using GO-Diet. We chose to work at this particular level because it was previously shown to provide a good representation of the complexity of the cellular functions of the proteins described by the biological process annotations \[[@B10]\]. The second step is to estimate the relevance of the annotations associated with the resulting classes using associated GO terms. The GO-Stats program can be used in this framework, using as reference dataset the list of proteins given as an input to PRODISTIN (Figure [4b](#F4){ref-type="fig"}).
As shown in Table [1](#T1){ref-type="table"}, the classes issued from PRODISTIN can be associated to one or to several GO terms. In the latter case, the calculated annotation biases emphasize the most relevant terms for the functional assignment of the class (first row in Table [1](#T1){ref-type="table"}), allowing the ranking of the annotation terms. When the class is associated with a single GO term (second and third rows in table [1](#T1){ref-type="table"}), GO-Stats estimates the probability of obtaining a class with the same size and functional coherence associated by chance with this GO term. For instance, in Table [1](#T1){ref-type="table"}, the term \'RNA metabolism\' is clearly over-represented in the second class, whereas this is certainly not true in the case of the \'cell cycle\' class.
Functional clustering of sets of transcriptional factor targets
---------------------------------------------------------------
GO can also be used to split gene sets into coherent functional subclasses on the basis of shared annotation terms. As an illustration, we have analyzed a gene set encompassing putative targets of the Engrailed transcription factor in *Drosophila melanogaster*. These genes were identified on the basis of *in vivo*UV cross-linking and chromatin immunoprecipitation experiments (X-ChIP) \[[@B11]\]. These experiments led to the cloning and sequencing of several hundreds of DNA fragments, allowing the computational identification of a well conserved DNA pattern, which was closely related to the known engrailed consensus. In order to delineate potential functional biases among engrailed targets, we have used Go-Diet and Go-Proxy to cluster the corresponding genes on the basis of \'Biological Processes\' GO annotations.
In the first step, the set of putative target genes has been fed to the dataset-creation program and slimmed down by cutting the annotations to the fourth level of the Gene Ontology, using GO-Diet. This eliminates the poorly informative terms. In a second step, the resulting dataset has been processed with GO-Proxy, leading to 11 classes as shown in Table [2](#T2){ref-type="table"}. Finally, for each of these classes, the probability of obtaining it by chance has been calculated, enabling the evaluation of the significance of the corresponding class relative to the initial gene dataset. In this analysis, the GOToolBox suite has proved to be very useful to define different functionally related sub-groups within a set of genes harbouring different functions (D.M., F. Maschat and B.J., unpublished work).
Comparison of GOToolBox with other GO-based analysis programs
=============================================================
In this study, we have described the GOToolBox suite, which performs five main tasks: gene dataset creation, selection and fitting of ontology level (GO-Diet), statistical analysis of terms associated with gene sets (GO-Stats), GO-based gene clustering (GO-Proxy), and gene retrieval based on GO annotation similarity (GO-Family). Recently, several web-based GO-processing tools have been developed to display, query or process GO annotations. In this section, we are interested in comparing GOToolBox to several GO-processing programs. As shown in Table [3](#T3){ref-type="table"}, comparisons were performed with 12 web-based programs listed on the official GO site \[[@B12]\].
Functionalities unique to the GOToolBox suite
---------------------------------------------
First, it should be highlighted that, to the best of our knowledge, no other program performing all five functions proposed in GOToolBox exists at present. Furthermore, the GO-Proxy and GO-Family tasks are unique to GOToolBox. These two functionalities are potentially very useful to the biologist. Indeed, on the one hand, the GO-Proxy implementation of a gene-to-gene distance calculation based on several GO terms allows the determination of classes consisting of functionally related genes. This feature should prove useful in all cases where the user wishes to identify functional subgroups within a list of genes of interest. On the other hand, the ability to search for genes similar to a user-defined gene on the basis of related GO terms (GO-Family) is also unique among all GO processing tools. When used to find functionally similar genes within a given species, the GO-Family program is often able to find paralogs as well as other genes with related functions, independently of sequence similarities. Similarly, when used to find functionally similar genes in other species, the program can successfully identify genes with related functions, including orthologs. In addition, the GO-Family program could be very valuable in the context of genome annotation: it could be used by database annotators to verify the coherence of the annotations of genes with known related functions, which if correctly annotated, would indeed be expected to be detected by the program.
Because of the presence of these two programs in our suite, we are inclined to think that GOToolBox represents a major improvement over other GO-based Web tools.
Comparison of statistical analyses performed by all GO-based Web tools
----------------------------------------------------------------------
Numerous programs have been developed to provide statistical evaluation of the occurrence of GO terms (Table [3](#T3){ref-type="table"}). We compared these programs to GO-Stats at two levels: the statistics used to calculate the enrichment/depletion of GO terms, and the availability of different features, such as the output types and the GO terms filtering utilities to create the gene dataset.
As shown in Table [3](#T3){ref-type="table"} (column 3), four different approaches to calculating the probability of having *x*genes annotated for a given GO category have been implemented in various dedicated programs: hypergeometric distribution, binomial distribution, Fisher exact test and Chi-square test.
The two latter are non-parametric tests and are therefore less powerful than *P*-value calculations obtained with both the hypergeometric and the binomial distributions. In particular, the Chi-square test seems to be the less efficient, because it only gives valid results for large gene datasets, and it does not distinguish between over and under-represented terms \[[@B13]\].
The binomial distribution permits us to calculate the probability of obtaining *x*genes annotated for a given GO category when randomly picking *k*times one gene among *N*genes, leaving the possibility that one gene can be picked many times, which is not the correct situation in our case. It is important to note that when *N*is large, the hypergeometric distribution tends to give the same results as the binomial distribution. On average, the hypergeometric distribution seems to be both the most adapted model and the most powerful statistical test.
To compare the results obtained with the different methods for *P*-value calculation, we have implemented these methods in the GO-Stats module of GOToolBox, excepted the Chi-square test for reasons explained above. The implementation of these tests in GO-Stats permits us to compare the methods without having to deal with problems due to program-specific input formats, data update, and supported/unsupported organism species, as is often the case when using different programs. In addition, this gives great flexibility to the user, allowing he or she to use different statistical methods. We verified that (as might be expected) different programs using the same statistical methods give the same results. This was essentially true, with slight variations probably due to the use of different versions of GO by some programs (data not shown). Therefore, the comparison between programs mainly relies on the number of possible statistical tests that are available. As shown in Table [3](#T3){ref-type="table"}, three programs (GOToolBox, GFINDer \[[@B13]\] and CLENCH \[[@B14]\]) propose the same three possible statistical tests, whereas all other programs have implemented only one method.
However, among these three programs, *GOToolBox*is the only one in which a multiple testing correction is implemented to adjust *P*-values and provide a correction for the occurrence of false positives. We choose the Bonferroni correction since it appears to be the most stringent in assessing the significance of enrichment/depletion
Comparison of other features proposed by GO-based web tools
-----------------------------------------------------------
In addition to the statistical tests used by the different programs, the presence of functional features offering flexibility to the end-user can also be considered as a criterion for program comparison. Features such as the GO terms filtering utilities and output types proposed by different programs are worth comparing (Table [3](#T3){ref-type="table"}, last two columns).
The GO terms filtering functions allows one to restrict the number of GO terms associated with each gene in the dataset, to facilitate interpretation of the results. Many ways to perform this restriction are possible: either mapping the terms on a slim ontology or fitting the terms to a given level (depth) of the ontology hierarchy. As shown in Table [3](#T3){ref-type="table"}, only GOToolBox allows the use of both these filtering methods. They have been implemented and are accessible under the \'Create Dataset\' form. In addition, in GOToolBox it is possible to restrict the number of terms associated with each gene, by taking into account only terms inferred in a particular way (for instance, terms inferred from direct assay) and to combine the filtering methods with the slim mapping or the level fitting described above.
As far as the output types are concerned, several programs propose a tabulated output file with terms ranked according to their *P*-values, (with the exception of GoMiner \[[@B15]\] and GOTM \[[@B16]\], therefore precluding the interpretation of the results in these cases). However, a positive attribute of GO Term Finder \[[@B17]\], GOTM and GoMiner over GOToolBox is that they propose directed acyclic graph (DAG) graphics for visualization of results. At the moment, GO-Stats allows the visualization of relationships between terms in tabulated output only, but a future version of GOToolBox will also incorporate a DAG graphical output option.
In conclusion, the GOToolBox is a multipurpose, flexible and evolvable software suite that compares favorably to all existing GO-based web-analysis programs. Its two unique features, GO-Proxy and GO-Family, enable new kinds of analyses to be carried out, based on the functional annotations of gene datasets These new functionalities are likely to be very useful to many biologists wanting to extract novel and meaningful biological information from gene datasets.
Acknowledgements
================
The authors would like to thank Badih Ghattas for helpful discussions. This project is supported by two grants from the Action Bioinformatique inter-EPST, awarded to D.T. and B.J., respectively. D.M. and C.B. are respectively indebted to the French Ministère de l\'Education, de la Recherche et de la Technologie, and to the Fondation pour la Recherche Médicale for financial support.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Flowchart of the GOToolBox programs.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Typical output from the GO-Stats program. From the input of a group of *Drosophila*genes, GO-stat returns a series of GO terms associated with them (columns 1 and 3). The terms are ranked according to a *P*-value representing their statistical relevance (column 8). The output also lists additional useful information: column 2 describes the depth at which a given GO term is found in the GO hierarchy (note that some terms can be found at several levels simultaneously; for example, GO:0009586). Columns 4 and 6 list the numbers of genes annotated for a given term in the reference and the user sets, respectively. Columns 5 and 7 list the corresponding occurrence frequencies. Finally, the last column indicates whether a given GO term is enriched (E) or depleted (D), based on the term frequency ratio (column 7/column 5). Note that hyperlinks to GO terms definitions by the GO consortium are provided (underlined in column 3). In such an output, all GO terms associated with the input genes are listed in the table. To visualize the hierarchy between these terms, an interactive functional feature is provided with GO-Stats: by clicking on a term (radio button on the left of GO terms list), all its parent terms in the list are highlighted. Finally, when working in the program, moving the mouse pointer on the GO ID column will make all the genes associated with a given GO term appear in a box.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Typical output from the GO-Family program. In this figure, we have asked for all the genes from human, mouse and nematode that share more than 45% functional similarity with an input gene: the *Drosophila*gene *engrailed*. The output is composed of four columns: rank, name of similar gene, percentage of similarity and species from which the similar gene is issued.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Use of the GOToolBox programs in the PRODISTIN framework. **(a)**Flowchart of the programs used in the PRODISTIN pipeline. The \'Dataset creation\' program and GO-Diet are used to generate a slimmed protein annotation file in a suitable format (tlf). This tlf file can be used as input both for PRODISTIN and for the tree-visualization program TreeDyn (not shown in the figure). In a second step, when functional classes have been generated by PRODISTIN, the GO-Stats tool allows the evaluation of the relevance of the class annotation term. **(b)**Histograms showing the distribution of the relevance values for the 79 classes issued from PRODISTIN (probability is described in the Features section).
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Examples of class relevance evaluation
:::
Original class annotations Most relevant term among class annotations Associated probability Number of proteins in the class Number of class proteins annotated for the term Number of proteins annotated for the term in the reference set
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------- ------------------------ --------------------------------- ------------------------------------------------- ----------------------------------------------------------------
Conjugation with cellular fusion; perception of abiotic stimulus; cell surface receptor linked signal transduction; sensory perception; response to pheromone during conjugation with cellular fusion Conjugation with cellular fusion 3.85E-09 11 8 32
RNA metabolism RNA metabolism 2.76E-09 32 17 70
Cell cycle Cell cycle 0.07383 6 3 115
The second and third columns are the results of the GO-Stats program, whereas all other columns are the results of a PRODISTIN analysis (see text for details).
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Classes found by *GO-Proxy*in a set of transcriptional putative targets and their statistical evaluation
:::
Class-associated term Number of genes in the class Probability
---------------------------------------------------------------- ------------------------------ -------------
Neurophysiological process 8 5.642e-9
Nucleobase, nucleoside, nucleotide and nucleic acid metabolism 7 2.609e-8
Cell growth and/or maintenance 14 1.011e-7
Protein metabolism 7 0.000001
Organismal movement 5 0.000006
Organogenesis 6 0.000030
Phosphorus metabolism 4 0.000110
Cell adhesion 3 0.000302
Response to external stimulus 3 0.001510
Signal transduction 5 0.002765
Signal transduction 3 0.034355
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Summary of the functionalities offered by GOToolBox and other GO processing tools
:::
Program References Statistics Multiple testing correction Output Ontology options
----------------------- ------------ ------------------------------------------- ------------------------------ --------------- ----------------------
eGOn \[18\] Fisher exact test \- TAB/RANK/TREE ALL/EVID
CLENCH \[14\] Hypergeometric Binomial Chi Square \- TAB/RANK ALL/SLIM
FatiGO \[19\] Fisher exact test Westfall/Benjamini/Yekutieli TAB/RANK/TREE LEVEL
FuncAssociate \[20\] Fisher exact test *P*-value adjustment TAB/RANK ALL
FuncSpec \[21\] Hypergeometric Bonferroni TAB/RANK ALL
GeneMerge \[22\] Hypergeometric Bonferroni TAB/RANK ALL
GFINDer \[13\] Hypergeometric Fisher exact test Binomial \- TAB/RANK ALL
GoMiner \[15\] Fisher exact test \- TAB/DAG/TREE ALL
Gostat \[23\] Fisher exact test Holm/Benjamini/Yekutieli TAB/RANK LEVEL
GO Term-Finder (CPAN) \[17\] Hypergeometric Bonferroni/Benjamini TAB/RANK/DAG ALL
GO Term-Finder (SGD) \[17\] Binomial \- TAB/RANK/DAG ALL
GOTM \[16\] Hypergeometric \- TAB/TREE/DAG ALL
GOToolBox This paper Hypergeometric Fisher exact test Binomial Bonferroni TAB/RANK ALL/SLIM/ LEVEL/EVID
In the output column, TREE, DAG, RANK and TAB refer respectively to tree-based output, directed acyclic graph visualization, *P*-value based ranking of terms, and results organized in a table. In the Ontology options column, terms listed refer to the way a gene set-GO term association can be built: ALL stands for \'all terms are taken into account (including parent terms)\'; SLIM for \'mapping of the terms on a slim ontology\'; LEVEL for \'fit the terms to a given depth of the ontology\'; and EVID for \'filter terms according to the type of evidence which indicates how annotation has been associated to the gene\'. See text for more details.
:::
|
PubMed Central
|
2024-06-05T03:55:51.865095
|
2004-11-26
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545796/",
"journal": "Genome Biol. 2004 Nov 26; 5(12):R101",
"authors": [
{
"first": "David",
"last": "Martin"
},
{
"first": "Christine",
"last": "Brun"
},
{
"first": "Elisabeth",
"last": "Remy"
},
{
"first": "Pierre",
"last": "Mouren"
},
{
"first": "Denis",
"last": "Thieffry"
},
{
"first": "Bernard",
"last": "Jacq"
}
]
}
|
PMC545797
|
Rationale
=========
Most eukaryotic genes contain introns that are spliced from the precursor mRNA (pre-mRNA). The correct interpretation of splicing signals is essential to generate authentic mature mRNAs that yield correct translation products. As an important post-transcriptional mechanism, gene function can be controlled at the level of splicing through the production of different mRNAs from a single pre-mRNA (reviewed in \[[@B1]\]). The general mechanism of splicing has been well studied in human and yeast systems and is largely conserved between these organisms. Plant RNA splicing mechanisms remain comparatively poorly understood, due in part to the lack of an *in vitro*plant splicing system. Although the splicing mechanisms in plants and animals appear to be similar overall, incorrect splicing of plant pre-mRNAs in mammalian systems (and vice versa) suggests that there are plant-specific characteristics, resulting from coevolution of splicing factors with the signals they recognize or from the requirement for additional splicing factors (reviewed in \[[@B2],[@B3]\]).
Genome projects are accelerating research on splicing. For example, with the majority of splicing-related genes already known in human and budding yeast, these gene sequences were used to query the *Drosophila*and fission yeast genomes in an effort to identify potential homologs \[[@B4],[@B5]\]. Most of the known genes were found to have homologs in both *Drosophila*and fission yeast. The availability of the near-complete genome of *Arabidopsis thaliana*\[[@B6]\] provides the foundation for the simultaneous study of all the genes involved in particular plant structures or physiological processes. For example, Barakat *et al.*\[[@B7]\] identified and mapped 249 genes encoding ribosomal proteins and analyzed gene number, chromosomal location, evolutionary history (including large-scale chromosomal duplications) and expression of those genes. Beisson *et al.*\[[@B8]\] catalogued all genes involved in acyl lipid metabolism. Wang *et al.*\[[@B9]\] surveyed more than 1,000 *Arabidopsis*protein kinases and computationally compared derived protein clusters with established gene families in budding yeast. Previous surveys of *Arabidopsis*gene families that contain some splicing-related genes include the DEAD box RNA helicase family \[[@B10]\] and RNA-recognition motif (RRM)-containing proteins \[[@B11]\]. At present, the *Arabidopsis*Information Resource (TAIR) links to more than 850 such expert-maintained collections of gene families \[[@B12]\].
Here we present the results of computational identification of potentially all or nearly all *Arabidopsis*genes involved in pre-mRNA splicing. Recent mass spectrometry analyses revealed more than 200 proteins associated with human spliceosomes (\[[@B13]-[@B17]\], reviewed in \[[@B18]\]). By extensive sequence comparisons using known plant and animal splicing-related proteins as queries, we have identified 74 small nuclear RNA (snRNA) genes and 395 protein-coding genes in the *Arabidopsis*genome that are likely to be homologs of animal splicing-related genes. About half of the genes occur in multiple copies in the genome and appear to have been derived both from chromosomal duplication events and from duplication of individual genes. All genes were classified into gene families, named and annotated with respect to their inferred gene structure, predicted protein domain structure and presumed function. The classification and analysis results are available as an integrated web resource, the database of *Arabidopsis*Splicing Related Genes (ASRG), which should facilitate genome-wide studies of pre-mRNA splicing in plants.
ASRG: a database of *Arabidopsis*splicing-related genes
=======================================================
Our up-to-date web-accessible database comprising the *Arabidopsis*splicing-related genes and associated information is available at \[[@B19]\]. The web pages display gene structure, alternative splicing patterns, protein domain structure and potential gene duplication origins in tabular format. Chromosomal locations and spliced alignment of cognate cDNAs and expressed sequence tags (ESTs) are viewable via links to the *Arabidopsis*genome database AtGDB \[[@B20]\], which also provides other associated information for these genes and links to other databases. Text-search functions are accessible from all the web pages. Sequence-analysis tools including BLAST \[[@B21]\] and CLUSTAL W \[[@B22]\] are integrated and facilitate comparison of splicing-related genes and proteins across various species.
*Arabidopsis*snRNA genes
========================
A total of 15 major snRNA and two minor snRNA genes were previously identified experimentally in *Arabidopsis*\[[@B23]-[@B28]\]. These genes were used as queries to search the *Arabidopsis*genome for other snRNA genes. A total of 70 major snRNAs and three minor snRNAs were identified by this method. In addition, a single *U4atac*snRNA gene was identified by sequence motif search. We assigned tentative gene names and gene models as shown in Table [1](#T1){ref-type="table"}, together with chromosome locations and similarity scores relative to a representative query sequence. The original names for known snRNAs were preserved, following the convention atUx.y, where x indicates the U snRNA type and y the gene number. Computationally identified snRNAs were named similarly, but with a hyphen instead of a period separating type from gene number (atUx-y). Putative pseudogenes were indicated with a \'p\' following the gene name. Pseudogene status was assigned to gene models for which sequence similarity to known genes was low, otherwise conserved transcription signals are missing and the gene cannot fold into typical secondary structure. A recent experimental study of small non-messenger RNAs identified 14 tentative snRNAs in *Arabidopsis*by cDNA cloning (\[[@B29]\], GenBank accessions 22293580 to 22293592 and 22293600, Table [1](#T1){ref-type="table"}). All these newly identified snRNAs were found in the set of our computationally predicted genes.
Conservation of major snRNA genes
---------------------------------
As shown in Table [1](#T1){ref-type="table"}, each of five major snRNA genes (U1, U2, U4, U5 and U6) exists in more than 10 copies in the *Arabidopsis*genome. U2 snRNA has the largest copy number, with a total of 18 putative homologs identified. Both U1 and U5 snRNAs have 14 copies, U6 snRNA has 13 copies, and U4 snRNA has only 11 copies. Sequence comparisons within *Arabidopsis*snRNA gene families showed that the U6 snRNA genes are the most similar, and the U1 snRNA genes are the most divergent. Eight active U6 snRNA copies are more than 93% identical to each other in the genic region, whereas active U1 snRNAs are on average only 87% identical. The U2 and U4 snRNAs are also highly conserved within each type, with more than 92% identity among the active genes. Details about the individual snRNAs and the respective sequence alignments are displayed at \[[@B30]\].
Previous studies identified two conserved transcription signals in most major snRNA gene promoters: USE (upstream sequence element, RTCCACATCG (where R is either A or G) and TATA box \[[@B24]-[@B27]\]. All 14 U5 snRNAs have the USE and TATA box. Furthermore, their predicted secondary structures are similar to the known structure of their counterparts in human, indicating that all these genes are active and functional (structure data not shown; for a review of the structures of human snRNAs, see \[[@B31]\]). Similarly, we identified 17 U2, 10 U1, nine U4, and nine U6 snRNA genes as likely active genes, with a few additional genes more likely to be pseudogenes because of various deletions. U4-10 and U6-7 do not have the conserved USE in the promoter region, but their U4-U6 interaction regions (stem I and stem II) are fairly well conserved. U2-16 is also missing the USE but has a secondary structure similar to other U2 snRNAs. These genes may be active, but differences in promoter motifs suggest that their expression may be under different control compared with other snRNAs homologs. The U2-17 snRNA has all conserved transcription signals, but 20 nucleotides are missing from its 3\' end. The predicted secondary structure of U2-17 is similar to that of other U2 snRNAs, with a significantly shorter stem-loop in the 3\' end as a result of the deletion. We are not sure if the U2-17 snRNA is functional, but the conserved transcription signals imply that it may be active.
Other conserved transcription signals were also identified in most active snRNAs, including the sequence element CAANTC (where N is either A, C, G or T) in U2 snRNAs (located at -6 to -1) \[[@B23]\], and the termination signal CAN~3-10~AGTNNAA in U snRNAs (U1, U2, U4 and U5) transcribed by RNA polymerase II (Pol II) \[[@B23],[@B24],[@B32]\]. The previously identified monocot-specific promoter element (MSP, RGCCCR, located upstream of USE) in U6.1 and U6.26 \[[@B33]\] is also found in five other U6 snRNA genes (U6.29, U6-2, U6-3, U6-4, U6-5). In all seven U6 snRNAs the consensus MSP sequence extends by two thymine nucleotides to RGCCCRTT. Although the MSP does not contribute significantly to U6 snRNA transcription initiation in *Nicotiana plumbaginifolia*protoplasts \[[@B33]\], the extended consensus may imply a role in gene expression regulation in *Arabidopsis*.
Low copy number of minor snRNA genes
------------------------------------
The minor snRNAs are functional in the splicing of U12-type (AT-AC) introns. Four types of minor snRNAs, which correspond to four types of major snRNAs, exist in mammals. U11 is the analog of U1, U12 is the analog of U2, U4atac is the analog of U4, and U6atac is the analog of U6. The U5 snRNA seems to function in both the major and minor spliceosome \[[@B34]\]. Two minor snRNAs (atU12 and atU6atac) were experimentally identified in *Arabidopsis*\[[@B28]\]. Both have the conserved USE and TATA box in the promoter region. We identified another *U6atac*gene (*atU6atac-2*) by sequence mapping. This gene has a USE and a TATA box in the promoter region. The *atU6atac-2*gene is more than 90% similar to *atU6atac*in both its 5\' and 3\' ends, with a 10-nucletotide deletion in the central region. The putative U4atac-U6atac interaction region in atU6atac-2 is 100% conserved with the interaction region previously identified in atU6atac \[[@B28],[@B35]\].
U11 and U4atac have not been experimentally identified in *Arabidopsis*. BLAST searches using the human U11 and U4atac homologs as queries against the *Arabidopsis*genome failed to find any significant hits, indicating divergence of the minor snRNAs in plants and mammals. Using the strategy described below, we successfully identified a putative *Arabidopsis U4atac*gene. It is a single-copy gene containing all conserved functional domains. We also found a single candidate U11 snRNA gene (chromosome 5, from 17,492,101 to 17,492,600) that has the USE and TATA box in the promoter region. This gene also contains a putative binding site fr Sm protein and a region that could pair with the 5\' splice site of the U12-type intron.
Identification of an *Arabidopsis U4atac*snRNA gene
---------------------------------------------------
Like U4 snRNA and U6 snRNA, human U4atac and U6atac snRNAs interact with each other through base pairing \[[@B36]\]. The same interaction is expected to exist between the *Arabidopsis*homologs. Therefore, we deduced the tentative AtU4atac stem II sequence (CCCGTCTCTGTCAGAGGAG) from AtU6atac snRNA and searched for matching sequences in the *Arabidopsis*genome. Hit regions together with flanking regions 500 base-pairs (bp) upstream and 500 bp downstream were retrieved and screened for transcription signals (USE and TATA box). One sequence was identified that contains both the USE and TATA box in appropriate positions, as shown in Figure [1](#F1){ref-type="fig"}.
The tentative *U4atac*snRNA gene contains not only the stem II sequence, but also the stem I sequence that presumably base-pairs with U6atac snRNA stem I. Furthermore, a highly conserved Sm-protein-binding region exists at the 3\' end. The predicted secondary structure is nearly identical to hsU4atac, with a relative longer single-stranded region (data not shown). With the highly conserved transcriptional signals, functional domains and secondary structure, this candidate gene is likely to be a real U4atac snRNA homolog. We named it AtU4atac and assigned At4g16065 as its tentative gene model because it is located between gene models At4g16060 and At4g16070 on chromosome 4.
Tandem arrays of snRNAs genes
-----------------------------
Some snRNAs genes exist as small groups on the *Arabidopsis*chromosomes \[[@B6]\]. We identified 10 snRNA gene clusters: seven U1-U4 snRNA clusters, one U2-U5 snRNA cluster, and a tandem duplication for both U2 snRNA (U2-10) and U5 snRNA (U5.1) (Figure [2](#F2){ref-type="fig"}). All seven *Arabidopsis*U1-U4 clusters have the U1 snRNA gene located upstream of the U4 snRNA gene, with a 180-300-nucleotide intergenic region. Five of the U1-U4 arrays are located on chromosome 5 (U1a/U4.1, U1-4/U4-5, U1-8/U4-7, U1-9/U4-8, and U1-13p/U4.3p), and the remaining two on chromosome 1 (U1-10/U4-6 and U1-14p/U4-10). The U2-17 and U5-10 occur in tandem array on chromosome 5, separated by fewer than 200 nucleotides.
*Arabidopsis*splicing-related protein-coding genes
==================================================
Most of the proteins involved in splicing in mammals and *Drosophila*are known \[[@B4],[@B37],[@B38]\]. In addition, recent proteomics studies revealed many novel proteins associated with human spliceosomes (reviewed in \[[@B18]\]). Using all these animal proteins as query sequences, we identified a total of 395 tentative homologs in *Arabidopsis*. Sequence-similarity scores and comparison of gene structure and protein domain structure were used to assign the genes to families. Each gene was assigned a tentative name based on the name of its respective animal homolog. Different homologs within a gene family were labeled by adding an Arabic number (1, 2, and so on) to the name. Close family members with similar gene structure were indicated by adding -a, -b, and -c to the name. The 395 genes were classified into five different categories according to the presumed function of their products. Ninety-one encode small nuclear ribonucleoprotein particle (snRNP) proteins, 109 encode splicing factors, and 60 encode potential splicing regulators. Details of EST evidence, alternative splicing patterns, duplication sources and domain structure of these genes are listed in Table [2](#T2){ref-type="table"}. We also identified 84 *Arabidopsis*proteins corresponding to 54 human spliceosome-associated proteins. The remaining 51 genes encode proteins with domains or sequences similar to known splicing factors, but without enough similarity to allow unambiguous classification. These two categories are not discussed in detail here, but information about these genes is available at our ASRG site \[[@B39]\].
The majority of snRNP proteins are conserved in *Arabidopsis*
-------------------------------------------------------------
There are five snRNPs (U1, U2, U4, U5 and U6) involved in the formation of the major spliceosome, corresponding to five snRNAs. Five snRNPs (U1 snRNP, U2 snRNP, U5 snRNP, U4/U6 snRNP and U4.U6/U5 tri-snRNP) have been isolated experimentally in yeast or human \[[@B40]-[@B45]\]. Each snRNP contains the snRNA, a group of core proteins, and some snRNP-specific proteins. Most of these proteins are conserved in *Arabidopsis*. All U snRNPs except U6 snRNP contain seven common core proteins bound to snRNAs. These core proteins all have an Sm domain and have been called Sm proteins. The U6 snRNP contains seven LSM proteins (\'like Sm\' proteins). Another LSM protein (LSM1) is not involved in binding snRNA (reviewed in \[[@B46]\]).
As shown in Table [2](#T2){ref-type="table"}, all Sm and LSM proteins have homologs in *Arabidopsis*, and eight of them are duplicated. It is likely that these genes existed as single copies in the ancestor of animals and plants, but duplicated within the plant lineage. Only one of the 24 genes (*LSM5*, At5g48870) has been characterized experimentally in *Arabidopsis*. The *LSM5*gene was cloned from a mutant supersensitive to ABA (abscisic acid) and drought (*SAD1*\[[@B47]\]). *LSM5*is expressed at low levels in all tissues and its transcription is not altered by drought stress \[[@B47]\]. cDNA and EST evidence exist for all other core protein genes, indicating that all 24 genes are active.
There are 63 *Arabidopsis*proteins corresponding to the 35 snRNP-specific proteins used as queries in our genome mapping. Very few of them have been characterized experimentally, including U1-70K, U1A and a tandem duplication pair of SAP130 \[[@B48]-[@B50]\]. *U1-70K*was reported as a single-copy essential gene. Expression of *U1-70K*antisense transcript under the *APETALA3*promoter suppressed the development of sepals and petals \[[@B51]\]. We identified an additional homolog of *U1-70K*(At2g43370) and named it *U1-70K2*. The U1-70K2 proteins showed 48% similarity to the U1-70K protein according to Blast2 results. Both genes retain the sixth intron in some transcripts, a situation which would produce truncated proteins \[[@B48]\]. Interestingly, we found that five of the 10 *Arabidopsis*U1 snRNP proteins, including the U1-70K-coding genes, may undergo alternative splicing.
Several genes in U2, U5, U4/U6 and U4.U6/U5 snRNPs, but none in U1 snRNP, occur in more than three copies in the *Arabidopsis*genome. The atSAP114 family has five members, including two that occur in tandem (*atSAP114-1a*and *atSAP114-1b*). Three members have EST/cDNA evidence (Table [2](#T2){ref-type="table"}). Interestingly, the predicted atSAP114p (At4g15580) protein contains a RNase H domain at the amino-terminal end, and thus *atSAP114p*shares similarity to At5g06805, a gene annotated as encoding a non-LTR retroelement reverse transcriptase-like protein. It is likely that the *atSAP114p*gene is a pseudogene that originated by retroelement insertion. There are three copies of the gene for the tri-snRNP 65 kilodalton (kDa) subunit, which are clustered on chromosome 4. Both the U4/U6 90 kDa protein and the U4/U6 15.5 kDa protein also have three gene copies, and the 116 kDa and 200 kDa subunits in U5 snRNP have four copies apiece.
The yeast U1 snRNP contains several specific proteins that are not present in mammalian U1 snRNPs \[[@B52]\]. As in mammals, *Arabidopsis*also lacks homologs of Prp42, a component of U1 snRNP in yeast \[[@B53]\]. However, *Arabidopsis*has two copies of the gene for Prp39, which are similar to Prp42. Furthermore, *atPrp39a*can produce a shorter protein isoform with a novel amino-terminal sequence by exon skipping. It is possible that the duplicates and alternative isoforms of plant U1 snRNP proteins are functional homologs of the yeast-specific proteins.
Several proteins specific to the minor spliceosome are also conserved in *Arabidopsis*. The human 18S U11/U12 snRNP contains several proteins found in U2 snRNP as well as seven novel proteins \[[@B14]\]. Four of the seven U11/U12-specific proteins (U11/U12-35K, 25K, 65K and 31K) are conserved in *Arabidopsis*, while the remaining three (59K, 48K and 20K) have no clear homologs. Interestingly, all four *Arabidopsis*genes are single copy in the genome, and three of them are apparently alternatively spliced (Table [2](#T2){ref-type="table"}).
Splicing factors are slightly different in *Arabidopsis*than in other organisms
-------------------------------------------------------------------------------
We divided the splicing factors into eight subgroups according to recent human spliceosome studies \[[@B13],[@B14],[@B16],[@B18]\]: splice-site selection proteins; SR proteins; 17S U2 associated proteins; 35S U5 associated proteins; proteins specific to the BΔU1 complex; exon junction complex (EJC) proteins; second-step splicing factors and other known splicing factors. We focused our analysis on the first two subgroups because their functions in splicing are well established. A total of 109 proteins in *Arabidopsis*were identified, corresponding to 67 human queries from all eight subgroups. Most of the proteins are conserved among eukaryotes, but some human proteins have no obvious homologs in the *Arabidopsis*genome, and some novel splicing factors appear to exist in *Arabidopsis*. About 43% of genes encoding splicing factors are duplicated in the genome, whereas some proteins, such as SF1/BBP (branchpoint-binding protein, which facilitates U2 snRNP binding in fission yeast \[[@B54]\]) and cap-binding proteins (CBP20 and CBP80, possibly involved in cap proximal intron splicing \[[@B55]\]), derive from single-copy genes \[[@B56]\]. These single-copy gene products may work with all pre-mRNAs, including the ones with U12-type introns. Surprisingly, mutation of CBP80 (*ABH1*) is not lethal and is non-pleiotropic. The *abh1*plants show ABA-hypersensitive closure of stomata and reduced wilting during drought \[[@B57]\].
Many splicing factors have been identified previously in *Arabidopsis*, including two U2AF65, two U2AF35, and 18 SR proteins \[[@B58]-[@B67]\]. The U2AF35-related protein atUrp, which could interact with U2AF65 and position RS-domain-containing splicing factors \[[@B68]\], is also present in the *Arabidopsis*genome. Although the *Urp*gene is expressed ubiquitously in human tissues, no ESTs from this gene were found in *Arabidopsis*. Three copies of *PTB/hnRNP-I*genes were identified in *Arabidopsis*. The PTB protein competes for the poly-pyrimidine tract with the U2AF large subunit, thus negatively regulating splicing \[[@B69]\].
We also identified a homolog related to atU2AF^65^(At2g33440) and an additional SR protein (At2g46610). The U2AF^65^-related protein (atULrp, At2g33440) has 247 amino acids and shares over 40% similarity with the carboxy terminal region of the two atU2AF^65^homologs. Only one RRM can be identified in atULrp, in contrast to three RRMs and one amino-terminal RS domain in atU2AF^65^proteins, and there is no apparent RS domain in atULrp. No animal homolog of atULrp could be identified. The function of this one-RRM U2AF^65^-related protein is not clear. As it lacks other functional motifs, it might act as a competitor of U2AF65. A two-RRM U2AF^65^protein can be produced through alternative splicing. The 11th intron of atU2AF^65^a can be retained (see RAFL full-length cDNA, gi:19310596) to produce a truncated protein with only the first two RRMs. Interestingly, the last RRM in atU2AF^65^a contains several amino-acid variations from the consensus pattern such that it could not be detected by InterPro and NCBI-CDD searches using default values, also suggesting that perhaps only the first two RRMs are essential.
The additional SR protein belongs to the atRSp31 family and was named atRSp32 (At2g46610). It shares 70% identity and 78% similarity with atRSp31. The protein is 250 amino acids in length and contains two RRMs and some RS dipeptides in the carboxy-terminal region. The gene structure of *atRSp32*is similar to that of *atRSp31*. Two other genes (*atRSp40*and *atRSp41*) are in the same family and also have similar exon and intron sizes (see gene structure information at \[[@B70]\]). Similarly to the previous classification of 18 SR proteins \[[@B61]\], the 19 SR proteins (including SR45) can be grouped into four large families of four to five members according to sequence similarity, gene structure and protein domain structure.
The atRSp31 family (atRSp31, atRSp32, atRSp40 and atRSp41) belongs to a novel plant SR family and has no clear animal ortholog. Other families include the SC35 (or SRrp/TASR2) family, SF2/ASF family, and the 9G8 family. *Arabidopsis*has a single copy of the SC35 ortholog and four SC35-like proteins (atSR33, atSCL30a, atSCL30 and atSCL28), which appear to have diverged significantly from SC35. It seems that this divergence predates the split of plants and animals because a similar SC35-like gene family exists in the human genome (SRrp35 and SRrp40). The SRrp35 and SRrp40 were found to antagonize other SR proteins *in vitro*and function in 5\' splice-site selection \[[@B71]\]. SF2/ASF has four copies (atSR1/SRp34, atSRp30, atSRp34a and atSRp34b) with similar gene structures and domains. Human 9G8 protein has five homologs in *Arabidopsis*, with three (atRSZp21, atRSZp22 and atRSZp22a) containing one CCHC-type zinc finger and two (atRSZ33, atRSZ32) containing two CCHC-type zinc fingers in addition to an RRM and an RS domain. Interestingly, several SR proteins (atRSZp21, atRSZp22, SR45 and SCL33) were found to interact with atU1-70K, and some SR proteins can interact with each other, thus forming a complicated interaction network to facilitate splice-site selection and spliceosome assembly \[[@B3],[@B61]-[@B63]\]. atSR45 was initially regarded as a novel plant SR protein \[[@B63]\], but by virtue of sequence-similarity scores it actually may be the ortholog of the human *RNPS1*gene, which encodes an EJC protein. Other human SR proteins (SRp20, SRp30c, SRp40, SRp54, SRp55 and SRp75) lack clear orthologs in *Arabidopsis*. We conclude that SR protein families evolved differently in animals and plants from three to four common ancestors, including SC35, SF2/ASF and 9G8/RSZ. The SRrp (SC35-like in plants) family may not be classical SR proteins but they play important roles in splice-site selection.
Proteins in other subgroups, such as 17S U2 snRNP-associated proteins, 35S U5 snRNP-associated proteins, and protein specific to the BΔU1 complex, are also conserved in *Arabidopsis*. The BΔU1 complex is the spliceosome complex captured immediately before catalytic activation. Most proteins in the 35S U5 snRNP are absent in the BΔU1 complex but present in the active B complex, indicating the important roles of 35S U5 snRNP-associated proteins in spliceosome activation \[[@B13]\]. Conservation of these proteins in *Arabidopsis*revealed the same pathway of spliceosome activation in plants. A subcomplex named Prp19 complex in 35S U5 snRNP has a critical role in spliceosome activation \[[@B13],[@B72]\]. All proteins in the human Prp19 complex have homologs in *Arabidopsis*, including a chromosomal duplication pair of *Prp19*genes and a single copy of the *CDC5*gene. For the BΔU1 complex, six human genes have homologs, and five of them are single copy in *Arabidopsis*. Two genes (*NPW38BP/SNP70*and *p220*(*NPAT*)) in the human BΔU1 complex have no apparent *Arabidopsis*homologs.
*Arabidopsis*also lacks an SMN protein complex. In human, the SMN protein (survival of motor neurons) can interact with a series of proteins including Gemin2, Gemin3 (a helicase), Gemin4, Gemin5 and Gemin6 to form an SMN complex, which has important roles in the biogenesis of snRNPs and the assembly of the spliceosome through direct interactions with Sm proteins and snRNA \[[@B73]\]. Although the SMN protein exists in the fission yeast genome (GenBank accession CAA91173), no SMN complex members can be identified in the *Arabidopsis*genome.
Splicing regulators are expanded in *Arabidopsis*
-------------------------------------------------
Splicing regulators are proteins that can either modify splicing factors or compete with splicing factors for their binding site. Important splicing regulators are hnRNP proteins and SR protein kinases. The exact role of phosphorylation of SR proteins in splicing is not yet clear, but SR protein kinases are well conserved and exist as multiple copies in *Arabidopsis*. A total of eight SR protein kinases were identified in *Arabidopsis*, including three Lammer/CLK kinases (AFC1, AFC2 and AFC3), two SRPK1 homologs, and three SPRK2 homologs. The three Lammer/CLK kinases were identified previously, and AFC2 was shown to phosphorylate SR protein *in vitro*\[[@B63],[@B74]\]. Overexpression of tobacco AFC2 homolog PK12 in *Arabidopsis*changed the alternative splice patterns of several genes, including *atSRp30*, *atSR1*/*atSRp34*and *U1-70K*\[[@B75]\], indicating that these SR proteins may function to modulate splicing in plants.
The heterogeneous nuclear ribonucleoproteins (hnRNPs) bind to splice sites and to binding sites for splicing factors on nascent pre-mRNAs, thus competing with splicing factors to negatively control splicing (reviewed in \[[@B76]\]). Humans have about 20 hnRNP proteins, many of which function in splicing. A total of 35 potential hnRNP proteins possibly related to splicing was found in *Arabidopsis*by sequence-similarity searches, including a superfamily of glycine-rich RNA-binding proteins. This family contains 21 members similar to human hnRNP A1 and hnRNP A2/B1. It can be further divided into two subfamilies. One includes eight proteins containing one RRM, and another has 13 members with two RRMs. 12 of these proteins were identified previously, including AtGRP7, AtGRP8, UBA2a, UBA2b, UBA2c and AtRNPA/B1-6 \[[@B11],[@B77],[@B78]\]. AtGRP7 was found to be able to influence alternative splicing of its own transcripts as well as *AtGRP8*transcripts \[[@B79]\]. UBA2 proteins can interact with UBP1 and UBA1 proteins, which have three RRMs and one RRM respectively, to recognize U-rich sequences in the 3\' untranslated region (UTR) and stabilize mRNA \[[@B78]\]. Although the overexpression of UBA2 did not stimulate splicing of a reporter gene in tobacco protoplasts \[[@B78]\], we cannot rule out the possibility that it could be involved in splicing of other genes.
Other human hnRNPs related to splicing also have homologs in *Arabidopsis*. BLAST searches of the human (CUG)n triplet repeat RNA-binding protein (CUG-BP) against all *Arabidopsis*proteins revealed three putative homologs, including atFCA. atFCA and CUG-BP share similarity within the RRMs and a region approximately 40 amino acids in length. An additional protein (At2g47310) related to FCA was identified and named FCA2, as it shares about 50% similarity with FCA. The FCA proteins have two RRMs and a WW domain, which interact with the FY protein, a homolog of yeast polyadenylation factor Psf2p \[[@B80],[@B81]\]. The FCA-FY complex negatively regulates the FCA protein by favoring a polyadenylation site from the third intron of FCA pre-mRNA \[[@B80],[@B82]\]. FCA may be a multifunctional protein involved in mRNA processing, as human CUG-BP can function in both alternative splicing and deadenylation \[[@B83]\]. We also list 15 previously identified hnRNP-like proteins and two additional homologs as possible splicing regulators. The UBP1 proteins can strongly enhance splicing of some introns in protoplasts \[[@B84]\], whereas UBA1, RBP45 and RBP47 proteins have no similar function \[[@B78],[@B85]\].
Unclassified splicing protein candidates
----------------------------------------
In addition to the 260 proteins in the above three categories, there are also 84 *Arabidopsis*proteins corresponding to human spliceosome-associated proteins identified in recent proteomic studies \[[@B15]-[@B18]\]. Some of these proteins function in other processes, such as transcription, polyadenylation and even translation. Their association with spliceosomes provides evidence for the coupling of splicing and other processes. Other proteins have no known functions. Only 35.8% of the proteins in this category are duplicated in *Arabidopsis*. We also identified a total of 51 *Arabidopsis*protein-coding genes similar to known splicing proteins. They have conserved domains and some level of sequence similarity to known splicing factors. We did not include these two categories in Table [2](#T2){ref-type="table"}, but detailed information about them is available at ASRG \[[@B39]\].
Distribution and duplication of *Arabidopsis*splicing-related genes
===================================================================
The distribution of *Arabidopsis*snRNA and splicing-related proteins across the genome is shown in Figure [2](#F2){ref-type="fig"} and at the ASRG website. Overall, the genes appear evenly distributed on the chromosomes, with several small gene clusters. Only four snRNA genes are located on chromosome 2, three of which are U2 snRNA genes. No U4 snRNA gene is located on chromosome 4. For the protein-coding genes, most functional categories have members located on each chromosome. The only exception is the SR protein kinase family, which has no member on chromosome 1. Interestingly, chromosome 1 contains the most snRNP proteins and splicing factors, but has the fewest splicing regulators. Several gene clusters encoding splicing-related proteins were also identified. Some clusters, such as tandemly duplicated gene pairs, include genes from the same category. One cluster located on chromosome 4 includes four genes encoding tri-snRNP proteins (atTri65a, atTri65b, atTri65c and atTri15.5c, homologs of tri-snRNP 65-KD protein and 15.5 KD protein). Two other clusters, *atU2A-atCdc5*and *atCUG-BP1-atU1C*, include genes from different functional categories. No clear clusters of genes for snRNA-splicing-related proteins were identified. Although about one third of snRNA genes are located near other protein-coding genes, none of their neighboring genes is related to splicing. As a caveat, we should point out that our snRNA gene determination strongly suggests annotation errors in overlapping protein-coding gene models. Thus, atU2-1, atU2.3, atU4.2, atU4-11p, atU5-13 and atU6.26 overlap gene models At1g16820, At3g57770, At3g06895, At1g68390, At5g53740 and At3g13857, respectively, but none of these models is well supported by cDNA or EST evidence (see displays linked at ASRG \[[@B30]\]).
The 260 proteins in the first three categories could be grouped into 130 families, 66 of which consist of multiple members. The duplication rate is over 50%, which is higher than the 44% duplication rate of *Arabidopsis*transcription factors \[[@B86]\]. As shown in Table [3](#T3){ref-type="table"}, about 50% of genes encoding snRNP proteins, 43% of splicing factors, and 78% of splicing regulators have duplications. The much higher duplication rate of splicing regulators may reflect diversification in splicing control.
At least 130 duplication events are required to yield the 260 proteins from 130 families given one single-copy ancestor per family. Thirty-three duplication events (about a quarter of the total) are likely to be the result of chromosome duplications. The chromosomal duplication ratio is 18.9-27.5% among the three groups (see Table [3](#T3){ref-type="table"}). Some snRNA genes pairs, such as *U2-14*/*U2-10*, *U5-3*/*U5-5*and (*U6.1 U6.26*)/(*U6-8p U6-9p*), may also have been produced by chromosome duplication. The C.D.2-3 region (chromosome duplication region between chromosomes 2 and 3, see \[[@B87]\]) has the most splicing-related gene pairs. Six genes in this region on chromosome 2 were duplicated in the same order on chromosome 3. EST evidence shows that all these genes are expressed. Three U5 snRNA genes (*U5.1*, *U5.1b*and *U5-4*) and four U2 snRNA genes (*U2.2*, *U2.3*, *U2.4*and *U2.6*) also are located in the same region on chromosome 3. No U5 and U2 homologs exist in the corresponding region on chromosome 2, suggesting that the snRNA duplication events in that region may have happened after the chromosome duplication event, or that the snRNA duplicates were lost subsequent to chromosome duplication.
Chromosomal duplication rather than individual gene duplication appears to be the predominant mode of amplification for some types of genes. As shown in Table [2](#T2){ref-type="table"}, the 24 genes encoding core proteins have nine duplication pairs, five of which can be attributed to chromosomal duplications. The 19 SR protein genes include eight duplication pairs, six of which are probably the results of chromosomal duplications. At least five chromosomal duplication events contributed to the superfamily of 21 hnRNP glycine-rich RBD and A/B genes. It is not clear why these functional categories have high chromosomal duplication ratios. It is possible that chromosomal duplication could create positive selection to maintain similar copy numbers of other genes encoding proteins that interact with the products of already duplicated genes.
Alternative splicing of *Arabidopsis*splicing-related genes
===========================================================
According to EST/cDNA alignments, 80 of the 260 protein coding genes show 66 alternative splicing events. This rate (30.8%) is much higher than the overall frequency of alternative splicing in *Arabidopsis*, which is about 13% using the same criteria (2,747 genes out of 20,446 genes with EST/cDNA evidence; B.-B.W. and V.B., unpublished work). As shown in Table [4](#T4){ref-type="table"}, the snRNP protein-coding genes have the lowest alternative splicing ratio (24.2%), whereas the ratios for splicing factor and splicing regulator genes are both over 33%. More than half of the genes encoding EJC proteins, proteins specific for the BΔU1 complex, SR proteins, U11/U12 snRNP-specific proteins and U1 snRNP proteins undergo alternative splicing.
Among different types of alternative splicing, intron retention is the most abundant of the alternative transcripts identified for the 260 classified splicing-related genes. As shown in Table [4](#T4){ref-type="table"}, 44 of the total 80 alternative splicing genes (about 55%) involve intron retention, 28 (35%) involve alternative acceptor-site selection and 15 (18.7%) are due to exon skipping. Compared with the corresponding ratio in all *Arabidopsis*alternative splicing events (55.3% intron retention, 23.4% alternative acceptor-site selection and 6.3% exon skipping; B.-B.W. and V.B., unpublished work), the ratio of intron retention in splicing-related genes is similar and the ratio of exon skipping is higher. Interestingly, only one of the 20 splicing regulator genes processed by alternative splicing (about 5%) shows exon skipping, indicating that exon skipping is an important post-transcriptional method for controlling the expression of splicing factor coding genes but not the splicing regulator genes.
Discussion
==========
Previous studies had determined 30 snRNA genes and 46 protein-coding genes related to splicing in *Arabidopsis*(see Tables [1](#T1){ref-type="table"} and [2](#T2){ref-type="table"}). In this study, we have computationally identified an additional 44 snRNA genes (Table [1](#T1){ref-type="table"}) and 349 protein-coding genes (Table [2](#T2){ref-type="table"}) that also may be involved in splicing. Among the five types of U snRNAs, U6 is the most conserved and U1 is the least conserved. We identified seven U1-U4 snRNA gene clusters. We were surprised to see so many U1-U4 clusters in *Arabidopsis*. In *Drosophila*, four snRNA clusters were reported \[[@B4]\], but none of them includes U1-U4 gene pairs. It is likely that a U1-U4 snRNA cluster existed in a progenitor of the current *Arabidopsis*genome, which was duplicated several times to form the extant seven clusters. The non-clustered U1 and U4 snRNA genes may have arisen by individual gene duplication or gene loss in duplicated clusters.
Among the proteins involved in splicing, most animal homologs are conserved in plants, indicating an ancient, monophylytic origin for the splicing mechanism. A striking feature of plant splicing-related genes is their duplication ratio. Fifty percent of the splicing genes are duplicated in *Arabidopsis*. The duplication ratio of the splicing-related genes increases from genes encoding snRNP proteins to genes encoding splicing regulators. These data strongly suggest that the general splicing mechanism is conserved, but that the control of splicing may be more diverse in plants.
The high duplication ratio of *Arabidopsis*splicing-related genes could be the result of evolutionary selection. Unlike animals, which can move around to maintain more homogeneous physiological conditions, plants are exposed to a larger range of stress conditions such as heat and cold. The duplicates will more probably be maintained in the genome as their functions become diversified, and potentially plant-specific, to ensure the fidelity of splicing under such varied conditions. Chromosome duplication has produced several Sm proteins, SR proteins and hnRNP proteins in *Arabidopsis*, which in turn could create positive selective pressures influencing the rate of duplication for functionally related genes. Because chromosome duplication occurred differentially within each plant lineage, we would expect different duplication patterns of these genes in, for example, monocots and dicots.
To confirm the above hypothesis, we searched the recently sequenced rice genome using the five *Arabidopsis*SC35 and SC35-like proteins as probes. Eight distinct genome loci were found to encode SC35 and SC35-like proteins, including three homologs of atSC35, two homologs of atSR33/SCL33 and atSCL30a, two homologs of atSCL30, and one homolog of atSCL28. Five of the eight rice genes are currently annotated in GenBank with accession numbers BAC79909 (osSC35a), BAD09319 (osSC35b), AAP46199 (osSR33-1), BAC799901 (osSCL30a/osSR33-2), and BAD19168 (osSCL30-1). As shown in the phylogenetic tree displayed in Figure [3](#F3){ref-type="fig"}, the two rice SC35 genes and atSC35 are likely to be orthologs of the animal SC35 gene. The other sequences cluster in SC35-like (SRrp/TASR) clades, indicating that the SC35 and SRrp/TASR genes diverged before the divergence of monocot and dicot plants (the divergence presumably happened even before the divergence of animals and plants, as described earlier). In addition, there are species-specific duplications. Thus, the *Arabidopsis*chromosomal duplication pair atSR33 - atSCL30a forms a clade, while their rice copies (osSR33-1 and osSCL30a) form another clade. Also there are additional duplications for the rice SC35 and SCL30 genes. We are currently working to identify all rice splicing related genes. The complete sets of these genes in two plant species should provide a good foundation for assessing similarities and differences in splicing mechanisms used by monocot and dicot plants.
As introns evolve rapidly, the mechanism to recognize and splice them should either evolve correspondingly or be flexible enough to accommodate the changes. It seems that plants deploy the most economic and practical way by keeping a largely conserved splicing mechanism and a very flexible recognition and control mechanism. Direct evidence comes from the presence of plant-specific splicing proteins, such as the novel SR protein family and the superfamily of hnRNP A/B. The absence of SMN complex and some yeast U1 snRNP proteins in *Arabidopsis*indicates that other organisms also have integrated new proteins or pathways into the splicing mechanism over the course of evolution relative to other eukaryotes. Other evidence supporting the conserved splicing but flexible regulating mechanism include differential conservation among U snRNAs (U1 snRNAs are less conserved than U6 snRNAs) and high alternative splicing frequency in U1 snRNP proteins, SR proteins and hnRNP proteins. The SR proteins and U1 snRNP proteins are involved in early steps of splicing and 5\' and 3\' splice-site selection; multiple isoforms of these proteins may be functionally significant in the control of splicing.
It is interesting to note that the overall alternative splicing frequency in splicing related genes is much higher than the frequency averaged over all *Arabidopsis*genes. More than half of SR proteins and U1 snRNP proteins show alternative splicing. Alternative splicing might increase protein diversity derived from splicing-related genes, which would further add flexibility to the splicing mechanism. The high frequency of alternative transcripts from splicing related genes raises another interesting question - how is splicing regulated in these splicing-related genes? One possible answer is that some splicing-related genes may be autoregulated. Accumulation of one transcript would feed back to inhibit/promote other isoforms. Several splicing-related genes have been reported to be regulated in this way. For example, AtGRP7 (hnRNP A/B superfamily) is a circadian clock-regulated protein which negatively autoregulates its expression \[[@B79]\]. When the AtGRP7 protein accumulates over the circadian cycle, it promotes production of alternative transcripts which use a cryptic 5\' splice site. As a result of message instability, the alternative transcripts contain pre-mature stop codons and do not accumulate to high levels, thus decreasing the level of AtGRP7 protein \[[@B79]\]. atSRp30 has similar effects on its own transcripts \[[@B65]\]. Another possible answer is that some splicing-related genes might regulate the splicing of other splicing-related genes. For example, overexpression of *AtGRP7*and *atSRp30*is known to affect the splicing of *AtGRP8*and *atSR1*, respectively \[[@B65],[@B79]\]. A third possibility is that the environment could affect the alternative splicing pattern. A good example is the *SR1*gene. The ratio of two transcripts from the *SR1*gene (SR1B/SR1) increases in a temperature-dependent manner \[[@B67]\]. Generally, heat or cold stress could cause intron retention in some splicing regulators, which could further alter the splicing pattern of other genes. The fourth possible regulators are intronless genes. Combining all these possibilities, a pathway to regulate splicing could be inferred as follows: environmental changes → splicing pattern changes in some specific splicing-related genes and/or intronless genes → expression pattern changes (including splicing pattern changes) in general splicing related genes → changes in splicing patterns for specific genes.
Conclusions
===========
A large number of *Arabidopsis*splicing-related genes were computationally identified in this study by means of sequence comparisons and motif searches, including a tentative *U4atac*snRNA gene containing all conserved motifs, a new SR protein-coding gene (*atRSp32*) belonging to the atRSp31 family, and several genes related to genes encoding known splicing-related proteins (atULrp and atFCA2). A web-accessible database containing all the *Arabidopsis*splicing related genes has been constructed and will be expanded to other organisms in the near future. This compilation should provide a good foundation to study the splicing process in more detail and to determine to what extent these genes are conserved across the entire plant kingdom. Our data show that about 50% of the splicing-related genes are duplicated in *Arabidopsis*. The duplication ratios for splicing regulators are even higher, indicating that the splicing mechanism is generally conserved among plants, but that the regulation of splicing may be more variable and flexible, thus enabling plants to respond to their specific environments.
Materials and methods
=====================
Search for *Arabidopsis*snRNAs
------------------------------
Sequences of the 15 experimentally identified major snRNAs were downloaded from GenBank. The two minor snRNAs sequences were compiled from the literature \[[@B28]\]. These genes were used to search against the *Arabidopsis*genome at the AtGDB BLAST server \[[@B88]\] and at the SALK T-DNA Express web server \[[@B89]\]. Our initial analysis was based on Release 3.0 of the *Arabidopsis*genome (GenBank accession numbers NC\_003070.4, NC\_003071.3, NC\_003074.4, NC\_003075.3, and NC\_003076.4). Local BLAST \[[@B21]\] was used to derive the locations of the snRNA homologs from more recently sequenced regions of the genome. Criteria used for local BLAST were \'e 1 -F F -W 7\' (cutoff eval is 1, dust filter on, with a minimum word size of 7). Human and maize snRNAs were also included as query sequences, and all hits with e-values less than 10^-5^were regarded as possible homologs. A total of 70 major snRNAs and three minor snRNAs were identified by this method. Each major snRNA type has 10-18 copies in the genome. A tentative gene name and gene model were assigned to each snRNA gene after comparison with the snRNAs identified in MATDB \[[@B90]\]. Sequence-similarity values were based on BLAST alignments.
Search for *Arabidopsis*splicing-related proteins
-------------------------------------------------
A three-round BLAST search strategy was used to identify *Arabidopsis*splicing related protein-coding genes. First, sequences of splicing-related proteins from human and *Drosophila*were downloaded from GenBank according to several recent proteomic studies \[[@B15]-[@B18]\] and the website compilation of Stephen Mount\'s group available at \[[@B91]\]. Human hnRNP proteins identified in a recent review \[[@B76]\] were downloaded from GenBank. All these sequences were used as queries in a local BLAST search against *Arabidopsis*annotated proteins (obtained from TIGR at \[[@B92]\]). All hits with an e-value less than 10^-10^were collected as candidates. Many of these candidates had highly significant e-values (usually 10^-30^or below and much lower than other hits). These candidates were regarded as true homologs.
In the second step, all identified true homologs were used to query the *Arabidopsis*protein set again. An e-value of 10^-20^was used as a cutoff value to find possible paralogs of the true homologs. Sequences identified in both rounds of BLAST hits were regarded as main candidates for splicing related proteins.
Finally, the main candidates were queried against GenPept and all annotated human proteins (obtained from Ensembl \[[@B93]\]). All candidates with significant similarity to proteins unrelated to splicing were removed from the main candidate list, and all candidates with significant similarity to proteins related to splicing were regarded as true splicing-related genes and were promoted to the status of true homologs. The remaining candidates were regarded as unclassified splicing-related proteins. BLAST results were initially analyzed by MuSeqBox \[[@B94]\]). Two custom scripts were written to read MuSeqBox output files, largely automating the search procedure.
Gene structure and chromosomal locations
----------------------------------------
The gene structure and chromosomal locations for the genes encoding splicing-related proteins were retrieved from AtGDB \[[@B95]\]. The chromosomal locations of the snRNA genes were inferred from the BLAST results. The location maps (Figure [1](#F1){ref-type="fig"}) were generated using the AtGDB advanced search function \[[@B96]\]. Spliced alignments of ESTs and cDNAs generated by GeneSeqer \[[@B97]\] were used to verify gene models. Gene structure information was used as an important criteria to group homologs into gene families.
Protein domains
---------------
InterProScan 3.3 was downloaded from \[[@B98]\] and was subsequently used to search protein domain databases using default parameters \[[@B99]\]. A Perl script was written to process the text results from InterProScan. Protein domain information was used in comparisons of homologs from different species. The search of the National Center for Biotechnology Information Conserved Domain Database (NCBI-CDD) \[[@B100]\] was conducted manually for certain genes to confirm the InterPro results.
Duplication source
------------------
The gene families with multiple copies were inspected to determine whether they were likely to have derived from chromosome-duplication events. Gene models of the duplicated gene were searched against the gene list of each chromosome redundancy region at MATDB \[[@B101]\]. If the gene and its duplicate were both in the list, they were regarded as a chromosome duplication pair. Otherwise, they were assumed to be produced by random gene duplication.
Identification of alternative splicing
--------------------------------------
All *Arabidopsis*ESTs and cDNAs were aligned against the genome using the spliced alignment program GeneSeqer as made available through AtGDB \[[@B102]\]. We retrieved the intron and exon coordinates of the reliable cognate alignments from the database. Scripts were written to identify introns that overlap with other introns or exons. We defined the alternative splicing cases as follows: alternative donor (AltD): an intron has the same 3\'-end coordinate but different 5\'-end coordinate as another overlapping intron; alternative acceptor (AltA): an intron has the same 5\'-end coordinate but different 3\'-end coordinate as another intron; alternative position (AltP): an intron has different 5\'-end and 3\'-end coordinates as another overlapping intron; exon skipping (ExonS): an annotated intron completely contains an alternatively identified exon in the same transcription direction; intron retention (IntronR): an annotated intron is completely contained by an alternatively identified exon.
Database and interface construction
-----------------------------------
Details about each splicing-related gene were saved in a MySQL database. PHP scripts were written to interact with the database and generate the interface web pages. Text and BLAST searches were implemented by Perl-cgi scripts.
Acknowledgements
================
We thank Shannon Schlueter for help with the web page and database design and implementation. We are also grateful to Shailesh Lal, Carolyn Lawrence and Michael Sparks for discussions and critical reading of the manuscript and to the anonymous reviewers for excellent suggestions. This work was supported in part by a grant from the ISU Plant Sciences Institute and NSF grants DBI-0110189 and DBI-0110254 to V.B.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Sequence alignments of U4atac and U6atac snRNAs. The tentative *Arabidopsis*U4atac snRNA was aligned against the human U4atac snRNA (U62822) using CLUSTAL W \[22\]. Possible sequence domains are indicated by different background colors, with cyan indicating transcription signals (USE, upstream sequence element; TATA, TATA box), green indicating the region involved in the stem-loop-stem structure, and pink indicating the domain that binds Sm proteins. The corresponding interaction region in U6atac snRNA is also marked in green. Red background indicates G-T base-pairs in the stem-loop structure. Grey letters indicate the genome sequence upstream and downstream of the putative U4atac gene. Asterisks (upper panel) and black shading (lower panel) show conserved positions in the alignment.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Chromosomal locations of *Arabidopsis*snRNAs. Chromosomes 1 to 5 are represented to scale by the long thick lines in dark green. The small bars above the chromosomes indicate the presence of an snRNA gene in that region. Different colors represent different snRNA types: red, U1 snRNA; magenta, U2 snRNA; blue, U4 snRNA; green, U5 snRNA; yellow, U6 snRNA; black, minor snRNA. The seven U1-U4 snRNA gene clusters (red-blue) and the single U2-U5 snRNA gene cluster (magenta-green) are indicated by red circles.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Phylogenetic tree of the SC35 protein family. The phylogenetic tree was constructed on the basis of protein sequence alignments of the SC35 homologs in human, *Drosophila*, *Arabidopsis*and rice. The GenBank accession numbers for the sequences are as follows: hsSC35, Q01130; hsSRrp40, AAL57514; hsSRrp35: AAL57515; dmSC35, AAF53192; atSC35, NP\_851261; atSR33/SCL33, NP\_564685; atSCL30a, NP\_187966; atSCL30, NP\_567021; atSCL28, NP\_197382; osSC35a, BAC79909; osSC35b, BAD09319; osSR33-1, AAP46199; osSCL30a/SR33-2, BAC799901; osSCL30-2, BAD19168. The sequences were aligned using CLUSTALW \[22\] with default parameters, and the phylogenetic tree was produced according to the neighbor-joining method using PAM substitution model distances as implemented in the PHYLIP package \[103\].
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
*Arabidopsis*snRNA genes
:::
Gene GeneID Chromosome Strand From To Length (nucleotides) e-value Similarity GenBank ID
--------------- ----------------- ------------ -------- ---------- ---------- ---------------------- --------- -------------------------- ---------------
*atU1a\** At5g49054 5 \- 19903323 19903158 166 1E-89 1-166, 100% gi17660
*atU1-2* At4g23415 4 \+ 12225621 12225786 166 1E-58 1-166, 92% gi22293582
*atU1-3* At5g51675 5 \+ 21013986 21014149 164 4E-55 3-166, 91%
*atU1-4* At5g25774 5 \- 8972971 8972807 165 2E-51 1-166, 90% gi22293583
*atU1-5* At1g08115 1 \- 2538238 2538073 166 1E-46 1-166, 89% gi22293581
*atU1-6* At3g05695 3 \+ 1681815 1681977 163 4E-40 4-166, 87%
*atU1-7* At3g05672 3 \+ 1657766 1657928 163 4E-40 4-166, 87% gi22293580
*atU1-8* At5g27764 5 \+ 9832576 9832740 165 1E-39 1-166, 87%
*atU1-9* At5g26694 5 \- 9494594 9494430 165 1E-27 1-166, 84%
*atU1-10* At1g11884 1 \- 4007396 4007236 161 1E-18 4-61, 93%; 80-166, 88%
*atU1-11p* At4g16645 4 \+ 9370786 9370841 56 7E-17 4-59, 94%
*atU1-12p* At4g23565 4 \- 12298871 12298802 70 1E-15 94-163, 90%
*atU1-13p* At5g49524 5 \- 20112431 20112275 157 2E-14 4-50, 91%; 91-166, 88%
*atU1-14p* At1g35354 1 \+ 12986822 12986908 87 1E-06 10-60, 88%; 84-118, 88%
*atU2-1* At1g16825 1 \+ 5758381 5758575 195 2E-88 1-196, 96%
*atU2.2\** At3g57645 3 \+ 21357718 21357913 196 1E-107 1-196, 100% gi17661
*atU2.3* At3g57765 3 \- 21408595 21408400 196 1E-95 1-196, 97% gi17662
*atU2.4* At3g56825 3 \- 21052994 21052800 195 5E-86 1-196, 95% gi17663
*atU2.5* At5g09585 5 \+ 2975013 2975208 196 7E-79 1-196, 93% gi17664
*atU2.6* At3g56705 3 \+ 21015472 21015667 196 1E-83 1-196, 94% gi17665
*atU2.7* At5g61455 5 \- 24730829 24730634 196 5E-86 1-196, 95% gi17666
*atU2-8* At5g67555 5 \+ 26966884 26967079 196 5E-86 1-196, 95%
*atU2.9* At4g01885 4 \+ 815273 815466 194 2E-82 1-194, 94% gi17667
*atU2-10* At2g02938 2 \+ 849777 849972 196 3E-93 1-196, 96% gi22293586
*atU2-10b/12* At2g02940 2 \+ 852859 853054 196 3E-93 1-196, 96%
*atU2-11* At1g09805/09895 1 \- 3180736 3180547 190 8E-85 1-190, 95%
*atU2-13* At2g20405 2 \+ 8809169 8809364 196 3E-81 1-196, 94% gi22293584
*atU2-14* At1g14165 1 \+ 4842274 4842469 196 3E-81 1-196, 94% gi22293585
*atU2-15* At5g62415 5 \+ 25083790 25083985 196 4E-74 1-196, 92%
*atU2-16* At5g57835 5 \- 23448717 23448522 196 2E-67 1-196, 92%
*atU2-17* At5g14545 5 \- 4690105 4690008 98 3E-44 1-98, 97%
*atU2-18p* At3g26815 3 \+ 9881236 9881303 68 2E-14 1-68, 89%
*atU4.1\** At5g49056 5 \- 19902970 19902817 154 4E-80 1-154, 99% gi17673
*atU4.2* At3g06900 3 \- 2178343 2178190 154 2E-75 1-154, 98% gi17674
*atU4.3p* At5g49526 5 \- 20112072 20112030 43 2E-11 15-57, 95% gi17675
*atU4-4* At1g49242/49235 1 \- 18222354 18222201 154 2E-75 1-154, 98% gi22293588
*atU4-5* At5g25776 5 \- 8972618 8972465 154 1E-70 1-154, 96%
*atU4-6* At1g11886 1 \- 4007020 4006867 154 1E-70 1-154, 96% gi22293587
*atU4-7* At5g27766 5 \+ 9832934 9833083 150 7E-66 1-150, 96%
*atU4-8* At5g26996 5 \- 9494230 9494081 150 7E-66 1-150, 96%
*atU4-9* At1g79965 1 \+ 30086031 30086168 138 9E-47 18-154, 92%
*atU4-10* At1g35356 1 \+ 12987189 12987313 125 3E-34 1-124, 90%
*atU4-11p* At1g68395 1 \+ 25647322 25647396 75 9E-07 18-37, 100%; 60-102, 90%
*atU5.1\** At3g55865 3 \- 20740607 20740503 105 6E-35 1-105, 94% gi17676
*atU5.1b* At3g55855 3 \- 20736881 20736780 102 7E-38 1-102, 96% gi22293592
*atU5-2* At1g65115 1 \+ 24194482 24194586 105 1E-39 1-105, 96%
*atU5-3* At1g70185 1 \+ 26433396 26433497 102 7E-38 1-102, 96% gi22293590
*atU5-4* At3g55645 3 \+ 20653843 20653947 105 3E-37 1-105, 95%
*atU5-5* At1g24105/24095 1 \- 8525204 8525103 102 2E-35 1-102, 95% gi22293591
*atU5-6* At1g04475 1 \- 1215831 1215730 102 2E-35 1-102, 95% gi22293589
*atU5-7* At4g02535 4 \- 1114629 1114528 102 1E-30 2-103, 93%
*atU5-8* At3g25445 3 \- 9227212 9227116 97 1E-20 5-101, 89%
*atU5-9* At1g79545 1 \- 29928543 29928447 97 1E-20 5-101, 89%
*atU5-10* At5g14547 5 \- 4690412 4690370 43 3E-12 24-67, 97%
*atU5-11* At5g54065 5 \- 21957066 21957023 44 2E-10 20-64, 95%
*atU5-12* At1g71355 1 \+ 26895255 26895298 44 2E-10 20-64, 95%
*atU5-13* At5g53745 5 \- 21829988 21829943 46 3E-09 24-70, 93%
*atU6.1\** At3g14735 3 \+ 4951596 4951697 102 1E-51 1-102, 100% gi16516
*atU6.26* At3g13855 3 \+ 4561111 4561212 102 2E-49 1-102, 99% gi16517
*atU6.29* At5g46315 5 \+ 18804616 18804717 102 2E-49 1-102, 99% gi16518
*atU6-2* At5g62995 5 \+ 25296825 25296926 102 1E-51 1-102, 100%
*atU6-3* At4g27595 4 \+ 13782215 13782316 102 1E-51 1-102, 100%
*atU6-4* At4g03375 4 \- 1483121 1483020 102 1E-51 1-102, 100%
*atU6-5* At4g33085 4 \- 15965258 15965158 101 8E-37 1-101, 94%
*atU6-6* At4g35225 4 \+ 16754836 16754931 96 1E-32 1-102, 93%
*atU6-7* At2g15532 2 \+ 6784793 6784869 77 7E-25 4-80, 93%
*atU6-8p* At1g52605 1 \+ 19596398 19596476 96 2E-19 4-99, 87%
*atU6-9p* At1g53465 1 \- 19960538 19960485 54 9E-09 21-74, 88%
*atU6-10p* At3g45705 3 \+ 16792802 16792888 87 2E-06 1-46, 89%; 62-100, 89%
*atU6-11p* At5g11085 5 \- 3522167 3522143 25 9E-06 1-25, 100%
*atU12\** At1g61275 1 \+ 22606785 22606960 176 1E-95 1-176, 100% ^†^gi22293600
*atU6atac\** At5g40395 5 \- 16183534 16183413 122 1E-63 1-122, 100% ^†^
*atU6atac-2* At1g21395 1 \- 7491489 7491378 112 5E-20 1-65, 95%; 81-110, 93%
*atU4atac* At4g16065 4 \+ 9096374 9096532 159 N/A N/A
Chromosomal locations were determined by conducting BLAST searches against the *Arabidopsis*genome (Release 5.0). \*The gene used for query in the BLAST search; ^†^atU12 and atU6atac sequences, which were experimentally identified \[28\]. Their sequences were compiled manually from the cited paper. The GenBank gi numbers for the chromosome sequences used are as follows: chromosome 1, 42592260; chromosome 2, 30698031; chromosome 3, 30698537; chromosome 4, 30698542; chromosome 5, 30698605.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
*Arabidopsis*splicing-related proteins
:::
Human homologs *Saccharomyces cerevisiae* Gene name GeneID Chromosome Tnb AltS Chromosomal duplication Protein domain Reference
---------------------------------------------- ---------------------------- --------------------------- ----------- ------------ ----- ----------------------------------- ------------------------- --------------------------------------------------------------------------- -----------
**1.1 Sm core proteins**
SmB SmB1 *atSmB-a* At5g44500 5 7 \>4-5a Sm, 1
*atSmB-b* At4g20440 4 21 IntronR (1); \>4-5a Sm, 1
SmD1 SmD1 *atSmD1-a* At3g07590 3 7 IntronR (1); Sm, 1
*atSmD1-b* At4g02840 4 13 Sm, 1
SmD2 SmD2 *atSmD2-a* At2g47640 2 7 AltA (1); AltD (1); Sm, 1
*atSmD2-b* At3g62840 3 25 AltA (1); Sm, 1
SmD3 SmD3 *atSmD3-a* At1g76300 1 9 \>1-1c Sm, 1
*atSmD3-b* At1g20580 1 7 \>1-1c Sm, 1
SmE SmE *atSmE-a* At4g30330 4 2 \>2-4b Sm, 1
*atSmE-b* At2g18740 2 10 AltA (1); \>2-4b Sm, 1
SmF SmF *atSmF* At4g30220 4 6 Sm, 1
SmG SmG *atSmG-a* At2g23930 2 13 Sm, 1
*atSmG-b* At3g11500 3 9 Sm, 1
LSM2 LSm2 *atLSM2* At1g03330 1 7 Sm, 1
LSM3 LSm3 *atLSM3a* At1g21190 1 6 \>1-1c Sm, 1
*atLSM3b* At1g76860 1 16 \>1-1c Sm, 1
LSM4 LSm4 *atLSM4* At5g27720 5 13 Sm, 1
LSM5 LSm5 *atLSM5 /SAD1* At5g48870 5 7 AltA (1); Sm, 1 \[47\]
LSM6 LSm6 *atLSM6a* At3g59810 3 7 \>2-3 Sm, 1
*atLSM6b* At2g43810 2 5 \>2-3 Sm, 1
LSM7 LSm7 *atLSM7* At2g03870 2 6 Sm, 1
LSM8 LSm8 *atLSM8* At1g65700 1 9 Sm, 1
LSM1 LSm1 *atLSM1a* At1g19120 1 8 Sm, 1
*atLSM1b* At3g14080 3 9 IntronR (1); Sm, 1
**1.2 U1 snRNP specific proteins**
U1A Subunit Mud1 *atU1A* At2g47580 2 14 ExonS (1); RRM, 2 \[49\]
U1C Subunit Yhc1 *atU1C* At4g03120 4 5 C2H2, 1; mrCtermi, 3
U1-70K Snp1 *atU1-70K* At3g50670 3 32 IntronR (1); RRM, 1 \[48\]
\- Prp39 *atPrp39a* At1g04080 1 12 ExonS (6); HAT, 7; TPR-like, 1
*atPrp39b* At5g46400 5 1 HAT, 4;
FBP11 Prp40 *atPrp40a* At1g44910 1 10 IntronR (1); WW, 2; FF, 5
FBP11 Prp40 *atPrp40b* At3g19670 3 5 WW, 2; FF, 5
Luc7-like protein Luc7 *atLuc7a* At3g03340 3 6 DUF259, 1
*atLuc7b* At5g17440 5 8 DUF259, 1
Related to Luc7-like protein Luc7 *atLuc7-rl* At5g51410 5 7 IntronR (1); DUF259, 1
**1.3 17S U2 snRNP specific proteins**
U2A\' Subunit Lea1p *atU2A* At1g09760 1 21 LRR 4;
U2B\" Subunit Msl1p *atU2B\"a* At1g06960 1 6 AltD (1); \>1-2a RRM, 2
*atU2B\"b* At2g30260 2 13 AltA (1); IntronR (1); \>1-2a RRM, 2;
SF3a120/SAP114 Subunit Prp21p *atSAP114-1a* At1g14650 1 17 AltB (1); SWAP/Surp, 2; Ubiquitin, 1
*atSAP114-1b* At1g14640 1 SWAP/Surp, 2
*atSAP114-2* At5g06520 5 SWAP/Surp, 4
*atSAP114-3* At4g16200 4 1 SWAP/Surp, 3
*atSAP114p* At4g15580 4 SWAP/Surp, 3; Ubiquitin, 1
SF3a60/SAP61 Subunit Prp9p *atSAP61* At5g06160 5 10 AltD (1); C2H2, 1
SF3a66/SAP62 Subunit Prp11p *atSAP62* At2g32600 2 13 C2H2, 1;
SF3b120/SAP130 Subunit Rse1p *atSAP130a* At3g55200 3 6 CPSF\_A, 1; WD40-like, 1 \[50\]
*atSAP130b* At3g55220 3 7 CPSF\_A, 1; WD40-like, 1 \[50\]
SF3b150/SAP145 Subunit Cus1p *atSF3b150* At4g21660 4 16 PSP, 1; DUF382, 1
*atSF3b150p* At1g11520 1
SF3b160/SAP155 Subunit Hsh155 *atSAP155* At5g64270 5 11 HEAT, 1; ARM, 2; SAP\_155, 1
SF3b53/SAP49 Subunit Hsh49p *atSAP49a* At2g18510 2 20 RRM, 2
*atSAP49b* At2g14550 2 RRM, 2
p14 Snu17p *atP14-1* At5g12190 5 7 RRM, 1;
*atP14-2* At2g14870 2 RRM, 1;
SF3b 14b /PHP5A Rds3p *atSF3b\_14b-a* At1g07170 1 10 \>1-2a UPF0123, 1;
*atSF3b\_14b -b* At2g30000 2 8 \>1-2a UPF0123, 1;
SF3b 10 *SF3b10a* At4g14342 4 11 SF3b10, 1;
*SF3b10b* At3g23325 3 6 SF3b10, 1;
**1.4 U5 snRNP specific proteins**
15 kD Subunit Dib1p *atU5-15* At5g08290 5 28 DIM1, 1; Thioredoxin\_2; 1
40 kD Subunit *atU5-40* At2g43770 2 21 WD-40, 7;
100 kD Subunit Prp28p *atU5-100KD* At2g33730 2 13 DEAD, 1; Helicase\_C, 1
102 KD/Prp6-like Prp6p *atU5-102KD* At4g03430 4 18 Ubiquitin, 1; TPR, 3; HAT, 15; TPR-like, 2; Prp1\_N, 1
116 kD Subunit /elongation Snu114p *atU5-116-1a* At1g06220 1 19 ExonS (1); EFG\_C, 1; GTP\_EFTU, 1; GTP\_EFTU\_D2; 1; Small\_GTP, 1; EFG\_IV, 1;
*atU5-116-1b* At5g25230 5 EFG\_C, 1; GTP\_EFTU, 1; GTP\_EFTU\_D2; 1; EFG\_IV, 1;
*atU5-116-2* At1g56070 1 214 EFG\_C, 1; GTP\_EFTU, 1; GTP\_EFTU\_D2; 1; EFG\_IV, 1;
*atU5-116-3* At3g22980 3 3 EFG\_C, 1; GTP\_EFTU, 1; Small\_GTP, 1;
200 kD Subunit/Helicase Brr2p *atU5-200-1* At5g61140 5 11 IntronR (1); DEAD, 2; Helicase\_C, 2; Sec63, 2; ARM, 1
*atU5-200-2a* At1g20960 1 23 DEAD, 2; Helicase\_C, 2; Sec63, 2
*atU5-200-2b* At2g42270 2 5 DEAD, 2; Helicase\_C, 2; Sec63, 2
*atU5-200-3* At3g27730 3 DEAD, 1; Sec63, 1; RuvA domain 2-like, 1
220 kD Subunit Prp8p *atU5-220/Prp8a* At1g80070 1 33 Mov34, 1
*atU5-220/Prp8b* At4g38780 4 2 Mov34, 1
**1.5 U4/U6 snRNP specific proteins**
U4/U6-90K / SAP90 Prp3p *atSAP90-1* At1g28060 1 10
*atSAP90-2* At3g55930 3
*atSAP90-3* At3g56790 3
U4/U6-60K / SAP60 Prp4p *atSAP60* At2g41500 2 8 WD-40, 7; SFM, 1; WD40-like, 1
U4/U6-20K / CYP20 *atTri-20* At2g38730 2 11 Pro\_isomerase, 1
U4/U6-61KD Prp31 *atU5-61/Prp31a* At1g60170 1 26 Nop, 1
*atU5-61/Prp31b* At3g60610 3 Nop, 1
U4/U6-15.5K Snu13p *atU4/U6-15.5a* At5g20160 5 18 IntronR (2); Ribosomal\_L7Ae, 1
*atU4/U6-15.5b* At4g12600 4 14 Ribosomal\_L7Ae, 1
*atU4/U6-15.5c* At4g22380 4 9 Ribosomal\_L7Ae, 1
**1.6 Tri-snRNP specific proteins**
Tri-65 KD Snu66p *atTri65a* At4g22350 4 7 UCH; 1; ZnF\_UBP, 1
*atTri65b* At4g22290 4 20 UCH; 1; ZnF\_UBP, 1; Pentaxin, 1
*atTri65c* At4g22410 4 UCH; 1; ZnF\_UBP, 1
Tri-110 KD SAD1 *atTri110* At5g16780 5 7 SART-1, 1
Tri-27 kD/RY1 *atTri-27 kD/RY1* At5g57370 5 14
hSnu23/FLJ31121 Snu23p *atSnu23* At3g05760 3 7 ZnF\_U1, 1;
**1.7 18S U11/U2 snRNP specific proteins**
U11/U12-35K *atU11/U12-35kD* At2g43370 2 7 IntronR (1); RRM, 1
U11/U12-25K (-99 protein) *atU11/U12-25K* At3g07860 3 6 IntronR (2); C2H2, 1;
U11/U12-65K *atU11/U12-65K* At1g09230 1 15 AltA (1); RRM, 2;PHOSPHOPANTETHEINE, 2;
U11/U12-31K (MADP1) *atU11/U12-31K* At3g10400 3 5 RRM, 1;CCHC, 1;
**2.1 Splice site selection**
U2AF35 *atU2AF35a/AUSa* At1g27650 1 26 RRM, 1; CCCH, 2;
*atU2AF35/AUSb* At5g42820 5 8 RRM, 1; CCCH, 2; \[58\]
U2AF65 Mud2 *atU2AF65b/AULa* At1g60900 1 10 RRM, 3; \[58\]
*atU2AF65a/AULb* At4g36690 4 29 AltA (1); IntronR (2); RRM, 2; \[58\]
*atULrp* At2g33440 2 2 RRM, 1
*AUL3p* At1g60830 1
U2AF35 related protein *atUrp* At1g10320 1 RRM, 1; CCCH, 2;
SF1/BBP *atSF1/BBP* At5g51300 5 23 IntronR (1); RRM, 1; CCHC, 2; KH, 1;
CBP20 Cbc1 *atCBP20* At5g44200 5 8 RRM, 1 \[56\]
CBP80 Cbc2p *atCBP80* At2g13540 2 21 MIF4G, 1; ARM, 3 \[56\]
PTB/hnRNP I *atPTB1* At1g43190 1 26 RRM, 4;
*atPTB2a* At3g01150 3 21 AltD (1); ExonS (1); RRM, 2
*atPTB2b* At5g53180 5 17 ExonS (1); RRM, 2
**2.2 SR proteins**
SC35 *atSC35* At5g64200 5 32 AltD (1); RRM, 1; \[61\]
SRrp40/TASR-2 *atSR33/atSCL33* At1g55310 1 12 IntronR (1); \>1-3b RRM, 1 \[63\]
*atSCL30a* At3g13570 3 32 ExonS (2); IntronR (4); \>1-3b RRM, 1 \[61\]
*atSCL30* At3g55460 3 14 ExonS (1); RRM, 1 \[61\]
*atSCL28* At5g18810 5 5 RRM, 1 \[61\]
SF2/ASF *atSR1/atSRp34* At1g02840 1 37 AltA (1); IntronR (1); \>1-4 RRM, 2 \[64,67\]
*atSRp34a* At4g02430 4 13 AltA (1); ExonS (1); IntronR (4); \>1-4 RRM, 2
*atSRp34b* At3g49430 3 3 ExonS (1); IntronR (1); RRM, 2
*atSRp30* At1g09140 1 15 AltA (1); RRM, 2 \[65\]
9G8 *atRSZp22/atSRZ22* At4g31580 4 26 \>2-4e RRM, 1; CCHC, 1 \[63,66\]
*atRSZp22a* At2g24590 2 7 \>2-4e RRM, 1; CCHC, 1 \[63,66\]
*atRSzp21/atSRZ21* At1g23860 1 18 RRM, 1; CCHC, 1 \[63,66\]
*atRSZ33* At2g37340 2 30 IntronR (1); \>2-3 RRM, 1; CCHC, 2 \[61\]
*atRSZ34* At3g53500 3 36 AltA (1); IntronR (3); \>2-3 RRM, 1; CCHC, 2 \[61\]
\- *atRSp32* At2g46610 2 23 AltD (1); IntronR (1); \>2-3 RRM, 2
*atRSp31* At3g61860 3 17 AltA (1); \>2-3 RRM, 2 \[59\]
*atRSp41* At5g52040 5 34 AltA (1); \>4-5b RRM, 2 \[59\]
*atRSp40/atRSP35* At4g25500 4 15 ExonS (1); IntronR (1); \>4-5b RRM, 2 \[59\]
**2.3 17S U2 associated proteins**
hPrp43 Prp43p *atPrp43-1* At5g14900 5 HA2, 1
*atPrp43-2a* At3g62310 3 17 AltA (1); \>2-3 DEAD, 1; Helicase\_C, 1; HA2, 1
*atPrp43-2b* At2g47250 2 14 \>2-3 DEAD, 1; Helicase\_C, 1; HA2, 1
SR140 *atSR140-1* At5g25060 5 11 Surp, 1;RRM, 1;, 1;RPR, 1;
*atSR140-2* At5g10800 5 2 Surp, 1;RRM, 1;RPR, 1;
SPF45 *atSPF45* At1g30480 1 9 D111/G-patch domain, 1; RRM, 1;
SPF30 *atSPF30* At2g02570 2 9 AltA (1); Tudor, 1;
**2.4 35S U5 associated proteins**
hPrp19\* Prp19p *atPrp19a* At1g04510 1 18 \>1-2a WD-40, 7; Ubox, 1;
*atPrp19b* At2g33340 2 27 IntronR (1); \>1-2a WD-40, 7; Ubox, 1;
CDC5\* Cef1 *atCDC5* At1g09770 1 12 SANT, 2; \[104\]
PRL1\* Prp46p *atPRL1* At4g15900 4 14 WD-40, 2;WD40like, 1;
*atPRL2* At3g16650 3 6 WD-40, 2;WD40like, 1;
AD-002\* Cwc15p *atAD-002* At3g13200 3 22 Cwf\_Cwc\_15, 1;
HSP73/HSPA8\* *HSP73-1* At3g12580 3 35 Hsp70, 1;
*HSP73-2* At5g42020 5 51 IntronR (1); Hsp70, 1;
*HSP73-3* At5g02500 5 553 IntronR (1); Hsp70, 1;
SPF27/BCAS2\* *atSPF27* At3g18165 3 15 BCAS2, 1;
beta catenin-like 1\* *atCTNNNBL1* At3g02710 3 12 Armadillo, 1;ARM, 1;
hSyf1 Syf1p *atSyf1* At5g28740 5 7 TPR, 1;HAT, 10;TPRlike, 3;
hSyf3/CRN Syf3 *atCRN1a* At5g45990 5 TPR, 1; HAT, 14; TPR-like, 2
*atCRN1b* At3g13210 3 TPR, 1; HAT, 12; TPR-like, 2
*atCRN1c* At5g41770 5 13 TPR, 1; HAT, 14; TPR-like, 2
*atCRN2* At3g51110 3 8 TPR, 1; HAT, 9; TPR-like, 1
hIsy1 Isy1p *atlsy1* At3g18790 3 10 Isy1, 1;
GCIP p29 Syf2 *atGCIPp29* At2g16860 2 12
SKIP Prp45p *atSKIP* At1g77180 1 28 SKIP/SNW, 1;
hECM2 Ecm2p *atECM2-1a* At1g07360 1 21 \>1-2a RRM, 1;CCCH, 1;
*atECM2-1b* At2g29580 2 10 \>1-2a RRM, 1;CCCH, 1;
*atECM2-2* At5g07060 5 CCCH, 1;
KIAA0560 *atAquarius* At2g38770 2 11
MGC23918 *atMGC23918* At3g05070 3 7
G10 Cwc14p *atG10* At4g21110 4 12 G10, 1;
Cyp E *atCypE1a/CYP2* At2g21130 2 4 \>2-4c Pro\_isomerase, 1
*atCypE1b* At4g38740 4 59 \>2-4c Pro\_isomerase, 1;
*atCypE2a/ROC3* At2g16600 2 39 \>2-4a Pro\_isomerase, 1
*atCypE2b* At4g34870 4 80 \>2-4a Pro\_isomerase, 1;
PPIase-like 1 *atPPIase-like1* At2g36130 2 10 Pro\_isomerase, 1;
**2.5 Proteins specific for BΔU1**
NPW38 *atNPW38* At2g41020 2 16 AltD (1); IntronR (1); WW, 2;
N-CoR1 *atN-CoR1* At3g52250 3 3 SANT, 2;Homeodomain\_like, 2;
hPrp4 kinase *atPRP4K-1* At3g25840 3 13 ExonS (1); Pkinase, 1;TyrKc, 1;S\_Tkc, 1;, 1;Kinase\_like, 1;
*atPRP4K-2* At1g13350 1 5 IntronR (1); Pkinase, 1;TyrKc, 1;S\_Tkc, 1;, 1;Kinase\_like, 1;
*atPRP4K-3* At3g53640 3 Pkinase, 1;TyrKc, 1;S\_Tkc, 1;, 1;Kinase\_like, 1;
FBP-21 *atFBP21* At1g49590 1 12 ExonS (1); IntronR (3); C2H2, 1;
TBL1-rp 1 *atTBL1-rp1* At5g67320 5 14 WD-40, 5;Peptidase\_S9A\_N, 1;LisH, 1;WD40like, 1;
Smc-1 *atSmc1* At3g54670 3 12 ATP\_GTP\_A\_BS, 1;SMC\_N, 1;SMC\_C, 1;ABC\_transporter, 1;SMC\_hinge, 1;
**2.6 Exon junction complex (EJC) proteins**
ALY Yra1p *atALY-1a* At5g02530 5 19 IntronR (1); RRM, 1;
*atALY-1b* At5g59950 5 16 IntronR (1); RRM, 1;
*atALY-2a* At5g37720 5 17 \>1-5b RRM, 1;
*atALY-2b* At1g66260 1 38 ExonS (1); \>1-5b RRM, 1;
Y14 *atY14* At1g51510 1 10 IntronR (1); RRM, 1;RBM8, 4;
Srm160-like *atSRM102* At2g29210 2 18 AltA (1); PWI, 1
Magoh *atMagoh* At1g02140 1 19 Mago\_nashi, 1;
Nuk-34/eIF4A3/DDX48 *atDDX48/eIF4A3-1* At3g19760 3 50 \>1-3a DEAD, 1;Helicase\_C, 1;
*atDDX48/eIF4A3-2* At1g51380 1 5 \>1-3a DEAD, 1;Helicase\_C, 1;
RNPS1 *atSR45/atRNPS1* At1g16610 1 27 AltA (1); RRM, 1 \[63\]
UAP56 *atUAP56a* At5g11200 5 21 AltA (1); DEAD, 1; Helicase\_C, 1
*atUAP56b* At5g11170 5 25 DEAD, 1; Helicase\_C, 1
pinin *atPinin* At1g15200 1 9 AltA (1); Pinin/SDK/memA, 1;
**2.7 Second step splicing factors**
Prp22 Prp22 *atPrp22-1* At3g26560 3 11 DEAD, 1; Helicase\_C, 1; S1, 1; HA2, 1;
*atPrp22-2* At1g26370 1 5 DEAD, 1; Helicase\_C, 1; HA2, 1
*atPrp22-3* At1g27900 1 15 DEAD, 1; Helicase\_C, 1; HA2, 1
Prp17 Prp17p *atPrp17-1* At1g10580 1 10 WD-40, 7;
*atPrp17-2* At5g54520 5 5 AltA (1); WD-40, 6;
Prp18 Prp18 *atPrp18-1* At1g03140 1 16 Prp18, 1; SFM 1;
*atPrp18-2* At1g54590 1 Prp18, 1
Slu7 Slu7p *atSLU7-1a* At1g65660 1 6
*atSLU7-1b* At4g37120 4 11
*atSLU7-2* At3g45950 3
Prp16 Prp16p *atPrp16* At5g13010 5 22 DEAD, 1; Helicase\_C, 1; HA2, 1
**2.8 Other known splicing factors**
SRm300 *atSRM300like* At3g23900 3 5 AltD (1); RRM, 1; Filamin/ABP280 repeat, 1
hTra-2/SFRS10 *atTra/SFRS1* At1g07350 1 25 ExonS (1); IntronR (3); RRM, 1
Prp2 *atPrp2-1a* At1g32490 1 9 \>1-2c DEAD, 1; Helicase\_C, 1; HA2, 1
*atPrp2-1b* At2g35340 2 \>1-2c DEAD, 1; Helicase\_C, 1; HA2, 1
*atPrp2-2* At4g16680 4 DEAD, 1; Helicase\_C, 1; HA2, 1
Prp5 *atPrp5-1a* At3g09620 3 DEAD, 1; Helicase\_C, 1
*atPrp5-1b* At1g20920 1 11 DEAD, 1; Helicase\_C, 1
*atPrp5-2* At2g47330 2 9 DEAD, 1; Helicase\_C, 1
hDbr1 dbr1 *atDbr1* At4g31770 4 12 Metallophos, 1; DBR1, 1
**3.1 SR protein kinase**
Lammer/CLK kinase *AFC1* At3g53570 3 11 AltA (1); IntronR (3); PKinase, 1; TyrKc, 1; S\_Tkc, 1; PKinase-like, 1 \[74\]
*AFC2* At4g24740 4 9 ExonS (1); PKinase, 1; TyrKc, 1; S\_Tkc, 1; PKinase-like, 1 \[74\]
*AFC3* At4g32660 4 9 AltD (1); IntronR (1); PKinase, 1; TyrKc, 1; S\_Tkc, 1; PKinase-like, 1 \[74\]
SRPK1 *atSRPK1a* At2g17530 2 7 \>2-4a PKinase, 1; TyrKc, 1; S\_Tkc, 1; PKinase-like, 1
*atSRPK1b* At4g35500 4 10 \>2-4a PKinase, 1; TyrKc, 1; S\_Tkc, 1; PKinase-like, 1
SRPK2 *atSRPK2a* At5g22840 5 2 PKinase, 1; TyrKc, 1; S\_Tkc, 1; PKinase-like, 1
*atSRPK2b* At3g53030 3 7 PKinase, 1; TyrKc, 1; S\_Tkc, 1; PKinase-like, 1
*atSRPK2c* At3g44850 3 1 PKinase, 1; TyrKc, 1; PKinase-like, 1
**3.2 Glycine-rich RNA binding protein**
HnRNP A/B *atGRBP1a* At1g18630 1 5 \>1-1c RRM, 1
*atGRBP1b* At1g74230 1 14 \>1-1c RRM, 1; Eggshell, 4
*atGRBP1c* At4g13850 4 17 \>3-4 RRM, 1
*atGRBP1d* At3g23830 3 12 AltA (1); \>3-4 RRM, 1
*atGRBP1e* At5g61030 5 8 RRM, 1; PfkB\_Kinase, 1
*atGRBP2* At2g16260 2 RRM, 1
*AtGRP7/atGRBP3a* At2g21660 2 182 AltD (1); IntronR (3); \>2-4c RRM, 1 \[77\]
*AtGRP8/atGRBP3b* At4g39260 4 67 AltB (1); AltD (1); IntronR (5); \>2-4c RRM, 1 \[77\]
**3.3 hnRNP A/B family**
hnRNP A/B *AtRNPA/B\_1* At4g14300 4 4 RRM, 2 \[11\]
*AtRNPA/B\_2* At2g33410 2 13 RRM, 2 \[11\]
*AtRNPA/B\_3* At5g55550 5 13 IntronR (3); \>4-5c RRM, 2 \[11\]
*AtRNPA/B\_4* At4g26650 4 21 \>4-5c RRM, 2 \[11\]
*AtRNPA/B\_5* At5g47620 5 12 AltD (2); RRM, 2 \[11\]
*AtRNPA/B\_6* At3g07810 3 18 AltA (1); RRM, 2; FKBP\_PPIASE\_2, 2 \[11\]
*AtRNPA/B\_7* At1g58470 1 6 RRM, 2
*AtRNPA/B\_8a* At5g40490 5 3 RRM, 2; Eggshell, 4
*AtRNPA/B\_8b* At1g17640 1 RRM, 2
*AtRNP\_N1* At3g13224 3 16 IntronR (1); RRM, 2;HUDSXLRNA, 2;
*UBA2a* At3g56860 3 23 IntronR (1); \>2-3 RRM, 2 \[78\]
*UBA2b* At2g41060 2 9 \>2-3 RRM, 2 \[78\]
*UBA2c* At3g15010 3 10 IntronR (1); RRM, 2 \[78\]
**3.4 Other hnRNPs (with animal homologs)**
hnRNP E1/E2 *at-hnRNP-E* At3g04610 3 10 KH, 3;
hnRNP F/ hnRNP H *at-hnRNP-F/ AtRNPH/F\_1* At5g66010 5 9 AltA (1); RRM, 2 \[11\]
*at-hnRNP-H/ AtRNPH/F\_2* At3g20890 3 RRM, 2 \[11\]
hnRNP G *at-hnRNP-G1* At5g04280 5 6 RRM, 1; CCHC, 1
*at-hnRNP-G2* At3g26420 3 35 AltA (1); RRM, 1; CCHC, 1
*at-hnRNP-G3* At1g60650 1 7 RRM, 1; CCHC, 1
hnRNP P2 *at-hnRNP-P* At1g50300 1 9 RRM, 1; ZnF\_RBZ, 2
hnRNP R/Q *hnRNP-R1* At4g00830 4 19 RRM, 3;
*hnRNP-R2* At3g52660 3 1 \>2-3 RRM, 3;
*hnRNP-R3 / AtRNPA/B\_9* At2g44710 2 13 \>2-3 RRM, 3
CUG-BP *AtCUG-BP1* At4g03110 4 4 AltA (1); IntronR (1); RRM, 3; HUDSXLRNA, 4 \[11\]
*AtCUG-BP2* At1g03457 1 9 AltA (1); RRM, 3; HUDSXLRNA, 4 \[11\]
(CUG-BP) *atFCA1* At4g16280 4 13 AltB (1); IntronR (1); RRM, 2; WW, 1 \[81\]
*atFCA2* At2g47310 2 6 RRM, 2; WW, 1
**3.5 Other plant hnRNPs**
*AtUBP1a* At1g54080 1 48 AltA (1); \>1-3b RRM, 3 \[84\]
*AtUBP1c* At3g14100 3 13 \>1-3b RRM, 3 \[84\]
*AtUBP1b* At1g17370 1 17 RRM, 3 \[84\]
*UBA1a* At2g22090 2 15 \>2-4c RRM, 1 \[78\]
*UBA1b* At2g22100 2 2 \>2-4c RRM, 1 \[78\]
*UBA1c* At2g19380 2 1 RRM, 1; C2H2, 3 \[78\]
*atRBP45a* At5g54900 5 42 \>4-5c RRM, 3 \[85\]
*atRBP45c* At4g27000 4 52 \>4-5c RRM, 3 \[85\]
*AtRBP45b* At1g11650 1 53 RRM, 3 \[85\]
*atRBP45d* At5g19350 5 10 RRM, 3
*AtRBP47a* At1g49600 1 10 \>1-3a RRM, 3 \[85\]
*AtRBP47b* At3g19130 3 21 \>1-3a RRM, 3 \[85\]
*AtRBP47c* At1g47490 1 23 IntronR (1); RRM, 3 \[85\]
*AtRBP47c\'* At1g47500 1 12 RRM, 3 \[85\]
*Ath1* At4g16830 4 34 HANP4\_PAI-RBP1, 1 \[105\]
*Ath2* At4g17520 4 29 \>4-5a HANP4\_PAI-RBP1, 1 \[105\]
*Ath3* At5g47210 5 67 IntronR (1); \>4-5a HANP4\_PAI-RBP1, 1
Gene names were kept consistent with names used in previous publications or derived from the names of the respective homologs (yeast names are given in the S.c. column, where available). The Tnb column gives the numbers of cognate cDNAs and ESTs supporting the gene structure. The AltS column indicates evidence for alternative splicing, including alternative donor site (AltD), alternative acceptor site (AltA), alternative position (AltP, both acceptor and donor sites are different), exon skipping (ExonS), and intron retention (IntronR). Chromosomal duplication indicates a known chromosome duplication region. Functional groups of proteins are separated by long lines spanning all columns. Different members in the group are separated by short lines starting at the *Arabidopsis*gene name. Genes duplicated in *Arabidopsis*are clustered together with no line between them. Dash line separate the Prp19 complex from other 35S U5 associated proteins and \* indicates proteins in that complex. Abbreviations for domains are as follows: ABC\_transporter: ABC transporter; Armadillo: Armadillo; ARM: ARM repeat fold; ATP\_GTP\_A\_BS: ATP/GTP-binding site motif A (P-loop); BCAS2: Breast carcinoma amplified sequence 2; C2H2: Zn-finger, C2H2 matrin type; C2H2: Zn-finger, C2H2 type; CCCH: Zn-finger, C-x8-C-x5-C-x3-H type; CCHC: Zn-finger, CCHC type; CPSF\_A: CPSF A subunit, C-terminal; Cwf\_Cwc\_15: Cwf15/Cwc15 cell cycle control protein; DBR1: Lariat debranching enzyme, C-terminal; DEAD: ATP-dependent helicase, DEAD-box; DEAD: DEAD/DEAH box helicase; DIM1: Pre-mRNA splicing protein; DUF259: Protein of unknown function DUF259; DUF382: Protein of unknown function DUF382; EFG\_C: Elongation factor G, C-terminal; EFG\_IV: Elongation factor G, domain IV; Eggshell: Eggshell protein; FF: FF domain; FKBP\_PPIASE\_2: Peptidylprolyl isomerase, FKBP-type; G10: G10 protein; GTP\_EFTU\_D2: Elongation factor Tu, domain 2; GTP\_EFTU: Protein synthesis factor, GTP-binding; HA2: Helicase-associated region; HANP4\_PAI-RBP1, 1: Hyaluronan/mRNA binding protein; HAT: RNA-processing protein, HAT helix; Helicase\_C: Helicase, C-terminal; Homeodomain\_like: Homeodomain-like; Hsp70: Heat shock protein Hsp70; HUDSXLRNA: Paraneoplastic encephalomyelitis antigen; Isy1: Isy1-like splicing; Kinase\_like: Protein kinase-like; LisH: Lissencephaly type-1-like homology motif; LRR: Leucine-rich repeat,; Mago\_nashi: Mago nashi protein; Metalloph: Metallophosphoesterase; MIF4G: Initiation factor eIF-4 gamma, middle; Mov34: Mov34/MPN/PAD-1; mrCtermi: Molluscan rhodopsin C-terminal tail; Nop: Pre-mRNA processing ribonucleoprotein, binding region; Peptidase\_S9A\_N: Peptidase S9A, prolyl oligopeptidase, N-terminal beta-propeller domain; PfkB\_Kinase: Carbohydrate kinase, PfkB; PHOSPHOPANTETHEINE: Phosphopantetheine attachment site; Pinin/SDK/memA: Pinin/SDK/memA protein; Pkinase: Protein kinase; Pro\_isomerase: Peptidyl-prolyl cis-trans isomerase, cyclophilin type; Prp18: Prp18 domain; Prp1\_N: PRP1 splicing factor, N-terminal; PSP: PSP, proline-rich; PWI: Splicing factor PWI; RBM8: RNA binding motif protein 8; Ribosomal\_L7Ae: Ribosomal protein L7Ae/L30e/S12e/Gadd45; RPR: Regulation of nuclear pre-mRNA protein; RRM: RNA-binding region RNP-1 (RNA recognition motif); S1: RNA binding S1; SANT: Myb DNA-binding domain; SAP\_155: Splicing factor 3B subunit\_1; SART-1: SART-1 protein; Sec63: Sec63 domain; SF3b10: Splicing factor 3B subunit 10; SFM: Splicing factor motif; SKIP/SNW: SKIP/SNW domain; Small\_GTP: Small GTP-binding protein domain; SMC\_C: Structural maintenance of chromosome protein SMC, C-terminal; SMC\_hinge: SMCs flexible hinge; SMC\_N: SMC protein, N-terminal; Sm\_like\_riboprot: Small nuclear-like ribonucleoprotein; Sm: Small nuclear ribonucleoprotein (Sm protein); S\_Tkc: Serine/threonine protein kinase; Surp: SWAP/Surp; Thioredoxin\_2: Thioredoxin domain 2; TPRlike: TPR-like; TPR: TPR repeat; Tudor: Tudor domain; TyrKc: Tyrosine protein kinase; Ubox: Zn-finger, modified RING; UCH: Peptidase C19, ubiquitin carboxyl-terminal hydrolase family 2; UPF0123: Protein of unknown function UPF0123; UPF0123: Protein of unknown function UPF0123; WD-40: G-protein beta WD-40 repeat; WD40like: WD40-like; WW: WW/Rsp5/WWP domain; ZnF\_RBZ: Zn-finger, Ran-binding; ZnF\_U1: Zn-finger, U1-like; ZnF\_UBP: Zn-finger in ubiquitin thiolesterase.
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Duplication source involving *Arabidopsis*splicing-related proteins
:::
Genes Family\* Single/multiple Duplication ratio Duplication events Chromosomal duplications Chromosomal duplication ratio
-------------------- ------- ---------- ----------------- ------------------- -------------------- -------------------------- -------------------------------
snRNP proteins 91 54 27/27 50.0% 37 7 18.9%
Splicing factors 109 58 33/25 43.1% 51 14 27.5%
Splicing regulator 60 18 4/14 77.8% 42 11 26.2%
Total 260 130 64/66 50.8% 130 32 24.6%
\*Family indicates both single copy gene and multiple-copy gene families. The Chromosomal duplication ratio column gives the fraction of all duplication events caused by chromosomal duplications.
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Alternative splicing in splicing-related genes
:::
Genes AltA AtlD AltP ExonS IntronR Overall Ratio
-------------------- ------- ------ ------ ------ ------- --------- --------- -------
snRNP proteins 91 6 3 1 3 11 22 23.2%
Splicing factors 109 14 5 0 11 21 38 34.9%
Splicing regulator 60 8 4 2 1 12 20 33.3%
Total 260 28 12 3 15 44 80 30.8%
The column entries are the numbers of genes in which the respective alternative splicing events can occur. AltA, alternative acceptor site; AltD, alternative donor site; AltP, alternative intron position (both acceptor and donor sites are different); ExonS, exon skipping; IntronR, intron retention. The Overall and Ratio columns give the number and fraction of genes with any type of alternative splicing, respectively.
:::
|
PubMed Central
|
2024-06-05T03:55:51.868771
|
2004-11-29
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545797/",
"journal": "Genome Biol. 2004 Nov 29; 5(12):R102",
"authors": [
{
"first": "Bing-Bing",
"last": "Wang"
},
{
"first": "Volker",
"last": "Brendel"
}
]
}
|
PMC545798
|
Background
==========
Global technologies in the budding yeast *Saccharomyces cerevisiae*have changed the face of biological study from the investigation of individual genes and proteins to a systems-biology approach involving integration of global gene expression with protein-protein and protein-DNA information \[[@B1]\]. These data, when combined with phenotypic profiling of the deletion mutant library of nonessential genes, allow an unparalleled assessment of the responses of yeast to environmental stressors \[[@B2]-[@B4]\]. In this study, we used these two genomic approaches to study the response of yeast to arsenic, a toxicant present worldwide, affecting millions of people \[[@B5]\].
Arsenic, a ubiquitous environmental pollutant found in drinking water, is a metalloid and human carcinogen affecting the skin and other internal organs \[[@B6]\]. It is also implicated in vascular disorders, neuropathy, diabetes and as a teratogen \[[@B7]\]. Furthermore, arsenic compounds are also used in the treatment of acute promyelocytic leukemia \[[@B8]-[@B10]\]. Consequently, the potential for future secondary tumors resulting from such therapy necessitates an understanding of the mechanisms of arsenic-mediated toxicity and carcinogenicity. However, even though a number of arsenic-related genes and processes related to defective DNA repair, increased cell proliferation and oxidative stress have been described, the exact mechanisms of arsenic-related disease remain elusive \[[@B11]-[@B19]\]. This is, in part, due to the lack of an acceptable animal model that faithfully recapitulates human disease \[[@B15]\].
A number of proteins involved in metalloid detoxification have been described in different organisms, including *Saccharomyces cerevisiae*. Bobrowicz *et al*. \[[@B20]\] found that Arr1 (also known as Yap8 and which is a member of the YAP family that shares a conserved bZIP DNA-binding domain) confers resistance to arsenic by directly or indirectly regulating the expression of the plasma membrane pump Arr3 (also known as Acr3), another mechanism for arsenite detoxification of yeast in addition to the transporter gene, *YCF1*\[[@B21]\]. Arr3 is 37% identical to a *Bacillus subtilis*putative arsenic-resistance protein and encodes a small (46 kilodalton (kDa)) efflux transporter that extrudes arsenite from the cytosol \[[@B22],[@B23]\]. Ycf1, on the other hand, is an ATP-binding cassette protein that mediates uptake of glutathione-conjugates of AsIII into the vacuole \[[@B21],[@B22]\]. Until recently, very little was known about arsenic-specific transcriptional regulation of detoxification genes. Wysocki *et al*. \[[@B24]\] found that Yap1 and Arr1 (called Yap8 in their paper) are not only required for arsenic resistance, but that Arr1 enhances the expression of Arr2 and Arr3 while Yap1 stimulates an antioxidant response to the metalloid. Menezes *et al*. \[[@B25]\], on the other hand, found that arsenite-induced expression of Arr2 and Arr3, as well as Ycf1, is likely to be regulated by both Arr1 (called Yap 8 in their paper) and Yap1.
Although Arr1 and Yap1 seem specifically suited for arsenic tolerance, the other seven YAP-family proteins are still worthy of investigation in light of the fact that each one regulates a specific set of genes involved in multidrug resistance with overlaps in downstream targets. One such interesting protein is Cad1 (Yap2). Although Yap1 and Cad1 are nearly identical in their DNA-binding domains, Yap1 controls a set of genes (including Ycf1) involved in detoxifying the effects of reactive oxygen species, whereas Cad1 controls genes that are over-represented for the function of stabilizing proteins in an oxidant environment \[[@B26]\]. However, Cad1 also has a role in cadmium resistance. As arsenic has metal properties, it is conceivable that Cad1 might play a greater part in arsenic tolerance and perhaps more so than the oxidative-stress response gene, *YAP1*.
Understanding the role of AP-1-like proteins (such as YAP family members) in metalloid tolerance was one of the goals in this study within the realm of the larger objective - using an integrative experimental and computational approach to combine gene expression and phenotypic profiles (multiplexed competitive growth assay) with existing high-throughput molecular interaction networks for yeast. As a consequence we uncovered the pathways that influence the recovery and detoxification of eukaryotic cells after exposure to arsenic. Networks were analyzed to identify particular network regions that showed significant changes in gene expression or systematic phenotype. For each data type, independent searches were performed against two networks: the network of yeast protein-protein and protein-DNA interactions, corresponding to signaling and regulatory effects (the regulatory network); and the network of all known biochemical reactions in yeast (the metabolic network). For the gene-expression analysis, we found several significant regions in the regulatory network, suggesting that Yap1 and Cad1 have an important role. However, no significant regions in the metabolic network were found. In order to test the functional significance of Yap1 and Cad1, we used targeted gene deletions of these and other genes, to test a specific model of transcriptional control of arsenic responses.
In contrast to the gene-expression data, the phenotypic profile analysis revealed no significant regions in the regulatory network, but two significant metabolic networks. Furthermore, we found that phenotypically sensitive pathways are upstream of differentially expressed ones, indicating that metabolic pathway associations can be discerned between phenotypic and transcriptional profiling. This is the first study to show a relationship between transcriptional and phenotypic profiles in the response to an environmental stress.
Results and discussion
======================
Transcript profiling reveals that arsenic affects glutathione, methionine, sulfur, selenoamino-acid metabolism, cell communication and heat-shock response
----------------------------------------------------------------------------------------------------------------------------------------------------------
Before gene-expression analysis of arsenic responses in *S. cerevisiae*, we performed a series of dose-response studies. We found that treatment of wild type cells with 100 μM and 1 mM AsIII had a negligible effect on growth, but that these cells still exhibited a pronounced transcriptional response (see Additional data files 1 and 2). Microarray analysis of biological replicates (four chips per replicate experiment) of the high-dose treated cells (1 mM AsIII) clustered extremely well together when using Treeview (see Materials and methods, and Additional data file 2). The lower dose time-course (100 μM AsIII) showed the beginning of gene-expression changes at 30 minutes, with the robust changes occurring at 2 hours, or one cell division (see Additional data file 2). The 2 hour, 100 μM dose clustered together with the 30 minute, 1 mM biological replicates and was in fact so similar to them that an experiment of one set of four chips for the 2 hour lower dose was deemed sufficient. Furthermore, when combining the three datasets (2 hour, 100 μM AsIII and each 30 minute, 1 mM AsIII replicate data) and using a 95% confidence interval (see Materials and methods) we found 271 genes that were not only statistically significant in at least 75% of the total data (9 out of 12 chips), but also that the direction and level of expression of these genes were similar between the datasets. The lower dose time-course also included a 4 hour treatment, or two cell divisions. This experiment demonstrated the greatest degree of variability, indicating either a cycling effect or the cell\'s return to homeostasis, which was further exemplified by a decrease in the transcriptional response (see Additional data file 2).
Genes were categorized by Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway and Simplified Gene Ontology (biological process, cellular component and molecular function) (Table [1](#T1){ref-type="table"}). In total, 829 genes out of 6,240 had significantly altered expression (see Materials and methods) in at least one experimental condition. The categories significantly enriched for differentially expressed genes in the KEGG pathways were glutathione, methionine, sulfur and selenoamino-acid metabolism, and in the Simplified Gene Ontology (biological process), cell communication and heat-shock response (Table [1](#T1){ref-type="table"}).
Network mapping of transcript profiling data finds a stress-response network involving transcriptional activation and protein degradation
-----------------------------------------------------------------------------------------------------------------------------------------
We used the Cytoscape network visualization and modeling environment together with the ActiveModules network search plug-in to carry out a comprehensive search of the regulatory and metabolic networks \[[@B27],[@B28]\]. The former consists of the complete yeast-interaction network of 20,985 interactions, in which 5,453 proteins are connected into circuits of protein-protein or protein-DNA interactions \[[@B29],[@B30]\]. For each protein in this network, we defined a network neighborhood containing the protein and all its directly interacting partners. In the metabolic network, based on a reconstruction by Forster *et al*. \[[@B31]\] with 2,210 metabolic reactions and 584 metabolites, nodes represent individual reactions and edges represent metabolites. A shared metabolite links two reactions. We searched for sequences of related reactions governed by sensitive proteins (enzymes) in the phenotypic profiling data. To aid visualization, these sequences of reactions were combined to create metabolic pathways. We then identified the neighborhoods associated with significant changes in expression using the ActiveModules plug-in. This process resulted in the identification of seven significant neighborhoods in the regulatory network, centered on nodes Fhl1, Pre1, Yap1, Cad1, Hsf1, Msn2 and Msn4 (Figure [1](#F1){ref-type="fig"}). Together these neighborhoods narrow the significant data to 20% of the genes with the most significant changes in expression across one or more arsenic conditions (see Materials and methods and Additional data file 2). We did not find the emergence of any significant neighborhoods in the metabolic network.
The highest-scoring regulatory network neighborhood was defined by the transcription factor Fhl1 (Figure [1a](#F1){ref-type="fig"}). Its expression did not change significantly, but it was the highest-scoring node as judged by the significant expression changes observed for its surrounding neighborhood. Fhl1 controls a group of proteins important for nucleotide and RNA synthesis, as well as the synthesis and assembly of ribosomal proteins \[[@B32]\] which, from our data, are downregulated by arsenic exposure. Downregulation of ribosomal proteins in response to environmental stress has been reported previously \[[@B33],[@B34]\], but to our knowledge this is the first association of Fhl1 as a key control element in this process. It seems likely that the repression of *de novo*protein synthesis in response to arsenic allows energy to be diverted to the increased expression of genes involved in stress responses and protection of the cell. One such pathway may involve sulfur metabolism, which leads to glutathione synthesis. In fact, included in Figure [1](#F1){ref-type="fig"} is Met31 (Figure [1e](#F1){ref-type="fig"}), a transcriptional regulator of methionine metabolism, which interacts with Met4, an important activator of the sulfur-assimilation pathway that is probably involved in the glutathione-requiring detoxification process. While the differential expression of this neighborhood was not strictly significant according to ActiveModules (see Materials and methods), it has high biological relevance in light of the statistically significant alteration in expression categorized using KEGG pathways (Table [1](#T1){ref-type="table"}).
Another high-scoring neighborhood comprises part of the proteasome protein complex (Figure [1b](#F1){ref-type="fig"}). The components of the proteasome are likely to be upregulated to meet the increased demand for protein degradation brought about by the binding of AsIII to the sulfhydryl groups on proteins and/or glutathione that subsequently interfere with numerous enzyme systems such as cellular respiration \[[@B7],[@B15]\]. In this paper, we will propose that this occurs through indirect oxidative stress as a result of the depletion of glutathione.
The role of transcription factors Yap1 and Cad1 and the metalloid stress response
---------------------------------------------------------------------------------
Many of the central proteins in the significant neighborhoods uncovered by ActiveModules were transcription factors (Figure [1a,c-f](#F1){ref-type="fig"}). Although some of these proteins were not differentially expressed themselves, they were still high-scoring nodes because of the highly significant expression of their targets. This is also important to keep in mind as we discuss later which genes might be sensitive to arsenic, but not necessarily differentially expressed, and why many genes that are differentially expressed do not display sensitive phenotypes when deleted.
Transcription factors Msn2, Yap1, Msn4, Cad1 and Hsf1 were the central proteins for many of the significant neighborhoods found (Figure [1c,d,f](#F1){ref-type="fig"}). Together with several genes previously implicated in oxidative-stress responses, these neighborhoods compose a stress-response network \[[@B24],[@B26],[@B35]-[@B39]\]. Of particular interest are Yap1 and Cad1, because of the high number of shared downstream targets (Figure [1c,f](#F1){ref-type="fig"}).
When overexpressed, Yap1 confers resistance to several toxic agents, and Yap1 mutants are hypersensitive to oxidants \[[@B33],[@B40]-[@B44]\]. Conversely, Cad1 responds strongly to cadmium, but not to hydrogen peroxide (H~2~O~2~) \[[@B26],[@B35]\]. Following arsenic exposure, Yap1 is induced at least fourfold, with many of its downstream targets showing high levels of induction (see Additional data file 3). Several of its targets are among the most highly upregulated genes (as high as 178-fold for *OYE3*(encoding a NADPH dehydrogenase)). Moreover, Yap1 regulates *GSH1*, which encodes γ-glutamylcysteine synthetase (an enzyme involved in the biosynthesis of antioxidant glutathione), *TRX2*(the antioxidant thioredoxin), *GLR1*(glutathione reductase) and drug-efflux pumps *ATR1*and *FLR1*\[[@B35],[@B45]-[@B50]\]. It should be noted that *GSH1*and *ATR1*are examples of several genes also targeted by Cad1. All of these specified Yap1 targets are induced after arsenic exposure, recapitulating the toxicant\'s role as a likely oxidant. During the course of this work, Wysocki *et al*. \[[@B24]\] also implicated Yap1 in arsenic tolerance.
As Cad1 and Yap1 share many downstream targets, the genes defined by these transcription factors are very similar. To determine which transcription factor is playing the most active role in the high level of differential expression for this group (see Figure [1c,f](#F1){ref-type="fig"}), we tested the roles of both activators by treatment of *yap1Δ*and *cad1Δ*deletion strains with 100 μM AsIII for 2 hours (Additional data file 4). Surprisingly, we did not find that Cad1 was involved in regulation in response to arsenic-mediated stress. The *yap1Δ*strain was not only sensitive to AsIII by phenotypic profiling (Additional data file 5) but also defective in the induction of several downstream enzymes with antioxidant properties (Figure [2a,b](#F2){ref-type="fig"}). Conversely, the *cad1Δ*strain displayed an almost identical profile to wild type, eliminating it as a strong factor in the arsenic response (Figure [2a,b](#F2){ref-type="fig"}). A list of arsenic-mediated genes with at least a twofold difference in expression compared to wild type for *yap1Δ*and *cad1Δ*is provided (Additional data files 6 and 7). These were generated using Rosetta Resolver with a *p*-value less than 0.001 (see Materials and methods for more detail). Also, Additional data files 8 and 9 contain tables of genes failing to be induced or repressed (or showing such a decrease in expression that they no longer make significantly expressed gene lists) in the *yap1Δ*and *cad1Δ*experiments, compared to the parent experiment, after treatment with 100 μM AsIII for 2 hours. These are lists of genes that would be potentially regulated by Yap1 and Cad1 in the presence of arsenic.
The proteasome responds to arsenic, and Rpn4 mediates a transcriptional role
----------------------------------------------------------------------------
Treatment of yeast with as little as 100 μM AsIII for 2 hours resulted in the induction of at least 14 ubiquitin-related and proteasome gene products (Figure [1b](#F1){ref-type="fig"} and Figure [3](#F3){ref-type="fig"}). The eukaryotic proteasome consists of a 20S protease core and a 19S regulator complex, which includes six AAA-ATPases known as regulatory particle triple-A proteins (RPT1-6p) \[[@B51],[@B52]\]. Proteins are targeted for degradation by the proteasome via the covalent attachment of ubiquitin to a lysine side chain on the target protein (Figure [3](#F3){ref-type="fig"}). Conjugating enzymes then function together with ubiquitin-ligase enzymes to adhere to the target protein, and are tailored to carry out specific protein degradation in DNA repair, growth control, cell-cycle regulation, receptor function and stress response, to name a few \[[@B53],[@B54]\]. The apparent importance of Yap1 in response to possible oxidative damage by arsenic indicated a potential role for Rpn4 (induced eightfold, Figure [3](#F3){ref-type="fig"}). This is a 19S proteasome cap subunit, which also acts as a transcriptional activator of the ubiquitin-proteasome pathway and a variety of base-excision and nucleotide-excision DNA repair genes \[[@B34],[@B55],[@B56]\].
Rpn4 is required for tolerance to cytotoxic compounds and may regulate multidrug resistance via the proteasome \[[@B57]\]. Moreover, Owsianik *et al*. \[[@B57]\] identified an YRE (Yap-response element) site present in the *RPN4*promoter. This YRE was found to be functional and important for the transactivation of *RPN4*by Yap1 in response to oxidative compounds, such as H~2~O~2~. However, we also located the Rpn4-binding sequence, TTTTGCCACC, 47 bases distant from the open reading frame (ORF) of *YAP1*, indicating that Yap1 not only activates Rpn4, but that Rpn4 may in fact activate Yap1 \[[@B58]\]. In support of this hypothesis we found that relative to wild type, the level of Yap1 induction was lower in the *rpn4Δ*strain under arsenic stress conditions, whereas Rpn4 was equally induced in the *yap1Δ*strain (Additional data file 10).
With respect to wild type, the profile of *rpn4Δ*after treatment with arsenic was the most dramatically altered, save for *arr1Δ*(Figure [2](#F2){ref-type="fig"} and Additional data files 11 and 12). These data suggest that arsenic modification of sulfhydryl groups on proteins leads to protein inactivation and therefore degradation via the 26S proteasome. Another scenario is that the proteasome, and/or its proteases, is sensitive to arsenic-related events, leading to dysfunctional protein turnover and an increased requirement for 26S proteasome subunits. A similar idea was proposed for the direct methylating agent, methylmethane sulfonate \[[@B34]\].
*ARR1* transcriptional responses
--------------------------------
Arr1 is structurally related to Yap1 and Cad1 \[[@B20],[@B24]\]. However, little is known about how Arr1 may be involved in oxidative stress and/or multidrug resistance. Furthermore, Arr1 is not well represented by the interactions present in the yeast regulatory network. However, studies by Bobrowicz *et al*. \[[@B20],[@B59]\] show that the transcriptional activation of Arr3 requires the presence of the Arr1 gene product. Moreover, a report by Bouganim *et al*. \[[@B60]\] supports our finding that Yap1 also is important for arsenic resistance. They show that overproduction of Yap1 blocks the ability of Arr1 to fully activate Arr3 expression at high doses of arsenite, suggesting that Yap1 can compete for binding to the promoter of the Arr1 target gene, *ARR3*. While this paper was being written, Tamas and co-workers \[[@B24]\] showed that Arr1 transcriptionally controls Arr2 and Arr3 expression from a plasmid containing their promoters fused to the *lacZ*gene and measuring β-galactosidase activities. This was done by growing the cells for 20 hours with a low dose of metalloid and spiking the concentration to 1 mM AsIII for the last 2 hours of incubation. These experiments showed that *ARR1*deletion resulted in complete loss of Arr3-*lacZ*induction, whereas *YAP1*deletion did not significantly affect induction. Similar results were obtained for the Arr2-*lacZ*induction assay and the authors concluded that Yap1 has a role in metalloid-dependent activation of oxidative stress response genes, whereas the main function of Arr1 seems linked to the control of Arr2 and Arr3. Interestingly, this study was shortly followed by another from Menezes *et al*. \[[@B25]\] which found contrasting results when looking at mRNA and Northern-blot analysis. In this study, the induction of Arr2 and Arr3, after treatment with 2 mM AsIII for up to 90 minutes, did not occur in either the *ARR1*-deleted strain or the *YAP1*-deleted strain. These authors conclude that the requirement for both *YAP1*and *ARR1*is vital to yeast in the function of regulating and inducing genes important for arsenic detoxification. Finally, transcription profiling experiments presented here show that the arsenic transport proteins Arr2 and Arr3 are still expressed (2.9-fold induction for Arr2 and 1.8-fold for Arr3, respectively) in the *ARR1*mutant, but show defective induction in the *yap1Δ*strain treated in parallel (Additional data files 4 and 10). These results indicate that Yap1 may control Arr2 and Arr3 when yeast is subjected to 100 μM AsIII for 2 hours.
Our results and those of Menezes *et al*. \[[@B25]\], in contrast to the results of Tamas and colleagues \[[@B24]\], might be explained by the following. Our and Menezes *et al.*\'s studies looked at genes in the normal chromosome context rather than genes ectopically expressed from a plasmid; in addition, in our study, we treated the yeast with 100 μM AsIII while Wysocki *et al*. \[[@B24]\] started with a low dose, but spiked the concentration to 1 mM AsIII in the last 2 hours of incubation. However, Menezes *et al*. \[[@B25]\] used an even higher dose (2 mM AsIII for a time-course ending at 90 minutes) and obtained more similar results to ours, with the exception that their Northern-blot analysis, which can sometimes miss relatively small changes, indicated an apparent lack of induction of *ARR2*or *ARR3*in either the *ARR1*- or *YAP1*-deleted strains. Taken together, these data indicate that both *ARR1*and *YAP1*are important genes involved in the process of arsenite detoxification in the yeast cell, but because of the different strains and treatment protocols used between these three studies, further experiments are warranted to resolve the differences.
Other interesting results from our transcription profiling of the *arr1Δ*and parent strains after arsenic treatment (Figure [2a,d](#F2){ref-type="fig"} and Additional data files 13 and 14), included large differences in expression as a whole and in particular the inability of *arr1Δ*to induce serine biosynthesis-related genes such as *SER3*, and sulfur and methionine amino-acid metabolism genes including *SAM4*. Conversely, *arr1Δ*failed to repress *SAM3*, as well as *CIT2*, a glutamate biosynthesis gene, when compared to the parent profile.
These observations indicate that Arr1 may regulate sulfur-assimilation enzymes that are necessary for arsenic detoxification. This is particularly interesting considering that the ActiveModules algorithm identified the node Met31 (Figure [1e](#F1){ref-type="fig"}), the transcriptional regulator of methionine metabolism which interacts with Met4, an important activator of the sulfur-assimilation pathway that is likely to be involved in the glutathione-requiring detoxification process. Sulfur metabolism was also a functional category in the Simplified Gene Ontology found to be significantly enriched by the hypergeometric statistical test (see Materials and methods) (Table [1](#T1){ref-type="table"}). Furthermore, phenotypic profiling results discussed later show the importance of serine and glutamate metabolism in the sensitivity response to arsenic. Lastly, it is important to note that *arr1Δ*also displays loss of expression of a number of ubiquitin-proteasome-related gene products, sharing similar expression patterns with *rpn4Δ*(Additional data files 13 and 14) and suggesting that it may have a role in protein degradation as well.
Arsenic treatment stimulates cysteine and glutathione biosynthesis and leads to indirect oxidative stress
---------------------------------------------------------------------------------------------------------
Our arsenic-treatment experiments revealed the strong induction of over 20 enzymes in the KEGG sulfur amino acid and glutathione biosynthesis pathways (Table [1](#T1){ref-type="table"}). This is consistent with the hypothesis that glutathione acts as a first line of defense against arsenic by sequestering and forming complexes with the toxic metalloid \[[@B21]\].
Dormer *et al*. \[[@B61]\] showed that *GSH1*induction by cadmium is dependent on the presence of Met4, Met31, Met32 and Cbf1 in the transcriptional complex of MET genes. Met4 and Met32 are also differentially expressed in response to arsenic and interact with Met31, which defines a network neighborhood as shown in Figure [1e](#F1){ref-type="fig"}. The biological impact of the sulfur-related stress response was further exemplified by comparisons of our arsenic profiles to H~2~O~2~profiles (400 μM H~2~O~2~) from Causton *et al*. \[[@B62]\] (Table [2](#T2){ref-type="table"}). Although we found many expected similarities between arsenic and H~2~O~2~gene-expression profiles in regard to oxidative-stress response genes, sulfur and methionine metabolism genes, in response to H~2~O~2~, were either repressed or did not change (Table [2](#T2){ref-type="table"}). Furthermore, a study by Fauchon *et al*. \[[@B63]\] showed that yeast cells treated for 1 hour with 1 mM of the metal Cd^2+^, responded by converting most of the sulfur assimilated by the cells into glutathione, thus reducing the availability of sulfur for protein synthesis. Our arsenic profile showed a similar response to the sulfur-assimilation profile seen with Cd^2+^(Table [2](#T2){ref-type="table"}). As a consequence, arsenic may be conferring indirect rather than direct oxidative stress mediated by the depletion of glutathione, thus inhibiting the breakdown of increasing amounts of H~2~O~2~by glutathione peroxidase (*GPX2*, up 13-fold) (Figure [4](#F4){ref-type="fig"}) \[[@B21],[@B64]\].
Phenotypic profiling defines arsenic-sensitive strains and maps to the metabolic network
----------------------------------------------------------------------------------------
To identify genes and pathways that confer sensitivity to arsenic, we identified deletion mutants with increased sensitivity to growth inhibition using a deletion mutant library of nonessential genes (4,650 homozygous diploid strains) \[[@B65],[@B66]\]. Each strain contains two unique 20-bp sequences (UPTAG and DOWNTAG) enabling their growth to be analyzed *en masse*and the fitness contribution of each gene to be quantitatively assayed by hybridization to high-density oligonucleotide arrays. The top 50 sensitive deletion strains included: *THR4*, *SER1*, *SER2*, *CPA2*, *CPA1*, *HOM2*, *HOM3*, *HOM6*, *ARG1*, *YAP1*, *CDC26*, *ARR3*, *CIN2*, *ARO1*, *ARO2*and *ARO7*. A listing of the rank order for all sensitivities is available (Additional data file 5).
Only 10% of the top 50 sensitive mutant strains were significantly differentially expressed in the transcript profile. This lack of direct correlation between gene expression and fitness data is consistent with data from our own and other laboratories \[[@B2],[@B4],[@B65]\]. At least three factors may contribute to this discrepancy. First, some highly expressed genes when deleted are nonviable (around 1,000 genes) and are therefore unable to be scored for fitness. Some examples of highly expressed, yet nonviable, genes under arsenic stress are *ERO1*(7- to 10-fold induced), *HCA4*(5- to 9-fold induced), and *DCP1*(9- to 22-fold induced). Second, there are redundant pathways mediated by multiple genes, such that deletion of one does not lead to sensitivity. *OYE2*, *OYE3*, and a large number of reductases fall into this category. Finally, gene products that do not change significantly, mediate important biological responses and thus when deleted could sensitize the cell to a specific stressor. *ARO1*, *ARO2*, *THR4*and *HOM2*are examples of genes that are not differentially expressed but are very sensitive to arsenic.
Like the gene-expression data, the phenotypic data was subjected to searches performed against the regulatory network of yeast protein-protein and protein-DNA interactions as well as the metabolic network of all known biochemical reactions in yeast. Unlike the transcription profile, the phenotypic data analysis revealed no significant regions in the regulatory network, but did map to two statistically significant metabolic networks. The first significant pathway was amino acid synthesis/degradation with the terminal products being L-threonine and L-homoserine, beginning with precursors such as L-arginine, fumarate and oxaloacetate (Figure [5a](#F5){ref-type="fig"}). These products function in serine, threonine and glutamate metabolism. The second network indicated the importance of the shikimate pathway, which is essential for the production of aromatic compounds in plants, bacteria and fungi (Figure [5b](#F5){ref-type="fig"}). The shikimate pathway operates in the cytosol of yeast and utilizes phosphoenol pyruvate and erythrose 4-phosphate to produce chorismate through seven catalytic steps. It is a pathway with multiple branches, with chorismate representing the main branch point, and various branches giving rise to many end products. Interestingly, chorismate is also used for the production of ubiquinone, *p*-aminobenzoic acid (PABA) and folates, which are donors to homocysteine \[[@B67]-[@B69]\].
Relationship between gene-expression and phenotypic profiles
------------------------------------------------------------
Combining transcript profiling and phenotypic profiling provides deeper insights into the biology of arsenic responses. Until now there has been a lack of correlation between the differential expression of genes and sensitivity of deletion mutants \[[@B2],[@B4],[@B65]\] and this was the case in the present study. However, by mapping each dataset to the regulatory and metabolic networks, we have uncovered the likely reason for this lack of congruence. Our data show that many of the most sensitive genes (Additional data file 5; top 50 ranks) are involved in serine and threonine metabolism, glutamate, aspartate and arginine metabolism, or shikimate metabolism, which are pathways upstream of the differentially expressed sulfur, methionine and homocysteine metabolic pathways, respectively. These downstream pathways are important for the conversion to glutathione, necessary for the cell\'s defense from arsenic (Figures [4](#F4){ref-type="fig"}, [5a](#F5){ref-type="fig"}, [6](#F6){ref-type="fig"} and Table [1](#T1){ref-type="table"}). This overlap of sensitive upstream pathways and differentially expressed downstream pathways provides the link between transcriptional and phenotypic profiling data (Figures [4](#F4){ref-type="fig"} and [6](#F6){ref-type="fig"}).
Thus, we believe our work shows that the deletion of an individual gene can lead to a change in sensitivity to an agent only if the protein product of that gene is important for some process (for example, amino-acid synthesis or a transcription factor required for the increased expression of genes needed to protect against the agent). On the other hand, expression profiling shows the end product of the cell\'s response to arsenic. Therefore, an agent such as arsenic might cause a transcription factor (Yap1, for example) to increase the expression of as many as 50 genes, 20 of which might help to protect against the agent. However, deletion of any of the 50 would not be expected to have an effect on the response to arsenic. The effect of gene deletion would be on the transcription factor itself (whose expression might not be affected by the agent). Thus, in the case of arsenic exposure, we conclude that phenotypic profiling interrogates genes upstream of the genes that ultimately protect against arsenic toxicity and that the downstream targets that demonstrate differential expression probably share redundant functions and are not vulnerable in the phenotypic profiling (Figure [6](#F6){ref-type="fig"}).
Conclusions
===========
Systems biology represents an important set of methods for understanding stress responses to environmental toxicants, such as arsenic. In this study we have catalogued the centers of activity associated with arsenic exposure in yeast, identifying the key neighborhoods of activity in the regulatory and metabolic networks using the visualization tools and algorithms in Cytoscape. The transcriptional profile mapped to the regulatory network, revealing several important nodes (Fhl1, Msn2, Msn4, Yap1, Cad1, Pre1, Hsf1 and Met31) as centers of arsenic-induced activity. From these results we can conclude that arsenic detoxification in yeast focuses around: nucleotide and RNA synthesis; methionine metabolism and sulfur assimilation; protein degradation; and transcriptional regulation by proteins that form a stress-response network. In summary, protein synthesis in response to arsenic allows energy to be diverted toward the genes channeling sulfur into glutathione, which then leads to indirect oxidative stress by depleting glutathione pools and alters protein turnover. These processes require regulation by transcription factors, the understanding of which we refined by analysis of specific knockout strains. Our experiments, in fact, confirmed that the transcription factors Yap1, Arr1 and Rpn4 strongly mediate the cell\'s adaptation to arsenic-induced stress but that Cad1 has negligible impact. Finally, contrary to the gene-expression analyses, the phenotypic profiling data mapped to the metabolic network. The two significant metabolic networks unveiled were shikimate and serine, threonine and glutamate biosynthesis. Our goal was to integrate the computational identification of these important pathways found via transcript and phenotypic profiling by regulatory and metabolic network mapping. In doing so, we have shown that genes that confer sensitivity to arsenic are in pathways that are upstream of the genes that are transcriptionally controlled by arsenic and share redundant functions.
Materials and methods
=====================
Strains, media and growth conditions
------------------------------------
*S. cerevisiae*strain BY4741 (*MAT***a**, *his3Δ*, *leu2Δ0*, *met15Δ0*, *uraΔ0*) was used and grown in synthetic complete medium at 30°C. Cells were grown to a density of 1 × 10^7^cells per ml. Cultures were split into two; NaAsO~2~ (100 μM and 1 mM in two biological repeats) was added to one culture, and both were incubated at 30°C for 0.5, 2 or 4 h. Cells were pelleted and washed in distilled water before RNA extraction. Deletion strains (*yap1Δ*, *cad1Δ*, *arr1Δ*and *rpn4Δ*) of the same background were obtained from Research Genetics, confirmed and treated the same way, for 2 h and 100 μM NaAsO~2~.
RNA extraction
--------------
For the cDNA hybridization experiments, total RNA was isolated using an acid-phenol method. Pellets were resuspended in 4 ml lysis buffer (10 mM Tris-HCL pH 7.5, 10 mM EDTA, 0.5% SDS). Four milliliters of acid (water-saturated, low pH) phenol was added followed by vortexing. The lysing cell solutions were incubated at 65°C for 1 h with occasional vigorous vortexing and then placed on ice for 10 min before centrifuging at 4°C for 10 min. The aqueous layers were re-extracted with phenol (room temperature, no incubation) and extracted once with chloroform. Sodium acetate was then added to 0.3 M with 2 volumes of absolute ethanol, placed at -20°C for 30 min, and then spun. Pellets were washed two or three times with 70% ethanol followed by Qiagen Poly(A)^+^RNA purification with the Oligotex oligo (dT) selection step. Total RNA for the specific knockout strains and parent experiment was isolated by enzymatic reaction, following the RNeasy yeast protocol (Qiagen).
Microarray hybridizations and analyses
--------------------------------------
A cDNA yeast chip, developed in-house at National Institute of Environmental Health Sciences (NIEHS), was used for gene-expression profiling experiments. A complete listing of the ORFs on this chip is available at \[[@B70]\]. cDNA microarray chips were prepared as previously described \[[@B71],[@B72]\]. The cDNA was spotted as described \[[@B73]\]. Each poly(A) RNA sample (2 μg) was labeled with Cy3- or Cy5-conjugated dUTP (Amersham) by a reverse transcription reaction using the reverse transcriptase SuperScript (Invitrogen), and the primer oligo(dT) (Amersham). The hybridizations and analysis were performed as described Hewitt *et al*. \[[@B74]\] except that genes having normalized ratio intensity values outside of a 95% confidence interval were considered significantly differentially expressed. Lists of differentially expressed genes were deposited into the NIEHS MAPS database \[[@B75]\]. Genes that were differentially expressed in at least three of the four replicate experiments were compiled and subsequently clustered using the Cluster/Treeview software \[[@B76]\]. GeneSpring (Silicon Genetics) and Cytoscape \[[@B28]\] were used to further analyze and visualize the data.
The knockout experiments were conducted on an Agilent yeast oligo array platform. Samples of 10 μg total RNA were labeled using the Agilent fluorescent direct label kit protocol and hybridizations were performed for 16 h in a rotating hybridization oven using the Agilent 60-mer oligo microarray-processing protocol. Slides were washed as indicated and scanned with an Agilent scanner. Data was gathered using the Agilent feature extraction software, using defaults for all parameters, save the ratio terms. To account for the use of the direct label protocol, error terms were changed to: Cy5 multiplicative error = 0.15; Cy3 multiplicative error = 0.25; Cy5 additive error = 20; Cy3 additive error = 20.
GEML files and images were exported from the Agilent feature extraction software and deposited into Rosetta Resolver (version 3.2, build 3.2.2.0.33) (Rosetta Biosoftware). Two arrays for each sample pair, including a fluor reversal, were combined into ratio experiments in Rosetta Resolver. Intensity plots were generated for each ratio experiment and genes were considered \'signature genes\' if the *p*-value was less than 0.001. *p*-values were calculated using the Rosetta Resolver error model with Agilent error terms. The signature genes were analyzed with GeneSpring. The entire in-house and Agilent-based dataset is available in the Additional data files.
Ontology enrichment
-------------------
Genes have previously been categorized into various ontologies and pathways. If a particular pathway is enriched for genes that are significantly expressed in response to a process, we conclude that the pathway is likely to be involved in this process. In total, 829 genes out of 6,240 had a significant alteration in expression in at least one experimental condition. Along with the size of each functional category, a statistical measure for the significance of the enrichment was calculated by using a hypergeometric test. The level of significance for this test was determined using the Bonferroni correction, where the α value was set at 0.05 and the number of tests conducted for KEGG pathway and Simplified Gene Ontology (biological process) were 27 and 11, respectively.
Network searches
----------------
The ActiveModules algorithm was used to identify neighborhoods in the regulatory network corresponding to significant levels of differential expression. In this search, if a protein has many neighbors, it is likely that at random a few will show significant changes in expression and these could be selected as a significant sub-network. Neighborhood scoring is a method we used to correct for this bias. In this scheme, a significant sub-network must contain either all or none of the neighbors of each protein. The significance then represents an aggregate over all neighbors of a protein. This prevents the biased selection of a few top-scoring proteins out of a large neighborhood in the search for significant sub-networks. For an in-depth description of this algorithm see Ideker *et al*. \[[@B1]\].
In defining the network used in the metabolic analysis, edges corresponding to metabolites linking more than 175 reactions were eliminated. This excludes metabolic cofactors such as ATP, NADH and H~2~O from the search. Scores for each ORF were generated by mapping the fitness significance value to a Z-score. To assign scores to the individual reactions, Förster\'s mapping from ORF to reaction was used to generate a list of ORFs for each reaction. The Z-scores of these ORFs were then aggregated into a single score for that reaction using the following equation:
We used a dynamic programming algorithm adapted from Kelley *et al*. \[[@B77]\] to identify high-scoring paths in this network. Briefly, the highest-scoring path of length (*n*) ending at each node is determined by combining the scores of the individual node and the highest-scoring path of length (*n*- 1) ending at a neighbor node using the following formula:
Since a node with many neighbors is more likely to belong to a high-scoring path by random chance, the score of the neighboring path is corrected against the extreme-value statistic with the number of observations equal to the number of neighbors.
The significances of the top-scoring networks were determined by comparison to a distribution of the top-scoring networks from random data (reaction scores randomized with respect to the nodes of the network). After running the path finding/scoring algorithm, the score of the single highest-scoring path was added to the null distribution. This process was repeated for 10,000 interactions. This null distribution was then used to determine an empirical *p*-value, which represents the null hypothesis that there is no significant correlation between the topology of the metabolic network and the assignment of significance values to nodes in that network.
Specific deletion experiment filter on fold-change comparisons
--------------------------------------------------------------
The intensity plots were generated from each experiment in Rosetta Resolver. A gene was considered a signature gene if the *p*-value was less than 0.001 and if the fold-change value was greater than or equal to twofold. Signature genes were then broadcasted on the intensity plot and exported as text files. Lists were imported into GeneSpring. The \'Filter on Fold Change\' function was used to compare the parent control vs. parent AsIII experiment with each deletion (AsIII) experiment. The gene list selected for each filter on fold change analysis was a combination of the parent signature gene list and the signature gene list of the AsIII-treated deletion being analyzed at the time. For example, if the comparison was being done between parent (AsIII-treated) and Yap1 (AsIII-treated), the list used in the analysis was the combination of the parent signature genes and the Yap1 signature genes. The filter on fold change function reports genes that were selected from the one condition (parent) that had normalized data values that were greater or less than those in the other condition (deletion under investigation) by a factor of twofold. Each resulting gene list was saved. All the resulting gene lists were combined and an annotated gene list was exported for use in Eisen\'s Cluster/Treeview package (described earlier). The format of the exported data was the natural log. The gene tree generated for the paper was generated in GeneSpring. Each filter on fold change was saved as an annotated gene list.
Generation of specific deletion experiment \'minus\' lists
----------------------------------------------------------
Signature gene lists were generated in Rosetta Resolver from intensity plots as described above. Each signature gene list was saved as a \'Bioset\' in Resolver. The parent Bioset was compared to each deletion Bioset using the \'Minus\' function. This function finds those members in Bioset group 1 (parent) that do not exist in Bioset group 2 (deletion). Each of the resulting lists was saved as a new Bioset. The new \'minus\' Bioset was broadcasted on its corresponding intensity plot and exported as a text file. This was repeated for each experiment with fine-tuning of the data using GeneSpring.
Phenotypic profiling
--------------------
Homozygous diploid deletion strains and pooling of the strains were done as described \[[@B66]\]. Aliquots were grown until logarithmic phase, diluted to OD~600~0.05-0.1, split into tubes and treated with arsenic for 1-2 h at 1 mM, 2 mM and 5 mM concentrations. Similar responses were observed at each concentration, so the results were pooled. These cultures and a mock-treated sample were maintained in logarithmic phase growth by periodic dilution for 16-18 h. UPTAG and DOWNTAG sequences were separately amplified from genomic DNA of the drug and mock-treated samples by PCR using biotin-labeled primers as described previously \[[@B66]\]. The amplification products were combined and hybridized to Tags3 arrays (Affymetrix). Procedures for PCR amplification, hybridization and scanning were done as described \[[@B66]\], and according to the manufacturer\'s recommendation when applicable. The images were quantified by using the Affymetrix Microarray Suite software. UPTAG and DOWNTAG values were separately normalized, ratioed (treated sample signal/control) and filtered for intensities above background \[[@B78]\].
Additional data files
=====================
The following additional data files are available with the online version of this article and at \[[@B79]\]. Additional data file [1](#s1){ref-type="supplementary-material"} shows the dose-response curve of *S. cerevisiae*strain BY4741 (*MAT***a**, *his3Δ*, *leu2Δ0*, *met15Δ0*, *uraΔ0*) grown in synthetic complete medium at 30°C after treatment with arsenic. Treatment with 1 mM, 2 mM and 5 mM AsIII resulted in a negligible effect on growth (after 18 h) and survival (1 h treatment followed by plating and colony formation counting), but still exhibited a pronounced transcriptional response (see Additional data file [2](#s2){ref-type="supplementary-material"}). Additional data file [2](#s2){ref-type="supplementary-material"} contains a figure showing all genes found to be significant by MAPS analysis (see Materials and methods) which were compiled across the four arrays, averaged and subsequently clustered with Cluster/Treeview software (Eisen *et al.*\[[@B76]\]). The dendogram highlighted in pink depicts the zoomed in region shown to the right of the entire tree. Genes in red are induced and genes in green are repressed. A table depicts the numbers of genes changing in each experiment at both the 95% and 99% confidence intervals (see Materials and methods). Additional data file [3](#s3){ref-type="supplementary-material"} contains the primary raw cDNA data from all the experiments. Additional data file [4](#s4){ref-type="supplementary-material"} contains the primary raw data for all the deletion strain experiments. Additional data file [5](#s5){ref-type="supplementary-material"} contains the sensitivity (phenotypic profiling) data ranked on the basis of four experiments, 1 mM (2x), 2 mM and 5 mM AsIII, and assigned a new uniform distribution of *p*-values. Every gene in this table has a percentile rank. In the case that there was slow growth in the wild type, then a default value of 0.5 was assigned. The rankings on this table were used for the metabolic networking. Additional data files [6](#s6){ref-type="supplementary-material"} and [7](#s7){ref-type="supplementary-material"} contain data produced by applying the \'Filter on Fold Change\' function in GeneSpring after importing the significant gene lists generated using Rosetta Resolver with a *p*-value less than 0.001 (see Materials and methods for more detail). The control parent vs. parent experiment (100 μM AsIII for 2 h) was compared with the *yap1Δ*(Additional data file [6](#s6){ref-type="supplementary-material"}) and *cad1Δ*(Additional data file [7](#s7){ref-type="supplementary-material"}) profiling experiments treated in parallel (for details see Materials and methods). Additional data files [8](#s8){ref-type="supplementary-material"} and [9](#s9){ref-type="supplementary-material"} contain tables of genes (\'Minus\' lists) that failed to be induced or repressed (or showed such a decrease in expression that they no longer make significantly expressed gene lists), compared to the parent experiment, in the *yap1Δ*(Additional data file [8](#s8){ref-type="supplementary-material"}) and *cad1Δ*(Additional data file [9](#s9){ref-type="supplementary-material"}) experiments after treatment with 100 μM AsIII for 2 h. Additional data file [10](#s10){ref-type="supplementary-material"} contains a figure showing that Yap1 is likely to regulate Arr2 and Arr3 after 2 h 100 μM AsIII but that it does not regulate Rpn4 under arsenic-induced stress. The self-organized heat map labeling and conditions in this figure are the same as for Figure [2](#F2){ref-type="fig"}. (a) The Yap1 knockout strain fails completely to induce Arr2 (0.834 average fold-change) whereas the Arr1 knock-out induces Arr2 (2.90 average fold-change). (b) The Arr1 knockout induction is more elevated compared to the Yap1 knock-out (1.8 and 1.1 average fold-change, respectively). (c) Yap1 is induced 2.7 fold in the Rpn4 knock-out. (d) The wild type parent strain shows an averaged induction of 4.7 fold. (e) Rpn4 is induced 3.7 fold in the Yap1 knock-out compared to 4.1 fold induction in the wild type parent strain. In the presence of arsenic, Yap1 does not appear to regulate Rpn4. Additional data file [11](#s11){ref-type="supplementary-material"}, as explained for Additional data files [6](#s6){ref-type="supplementary-material"} and [7](#s7){ref-type="supplementary-material"}, compares the control parent vs. parent experiment (100 μM AsIII for 2 h) to the *rpn4Δ*profiling experiment treated in parallel. Additional data [12](#s12){ref-type="supplementary-material"} contains a table of genes (\'Minus\' list) that fail to be induced or repressed, compared to the parent experiment, in the *rpn4Δ*experiment after treatment with 100 μM AsIII for 2 h. Additional data file [13](#s1){ref-type="supplementary-material"}, as explained for Additional data files [6](#s6){ref-type="supplementary-material"} and [7](#s7){ref-type="supplementary-material"}, is from comparing the control parent vs. parent experiment (100 μM AsIII for 2 h) to the *arr1Δ*profiling experiment treated in parallel. Additional data file [14](#s14){ref-type="supplementary-material"} contains a table of genes (\'Minus\' list) that fail to be induced or repressed, compared to the parent experiment, in the *arr1Δ*experiment after treatment with 100 μM AsIII for 2 h. Additional data file [15](#s15){ref-type="supplementary-material"} contains the self-organized clustering of specific deletion and parent strain experiments (*yap1Δ*vs. *yap1Δ*2 h 100 μM AsIII, *cad1Δ*vs. *cad1Δ*2 h 100 μM AsIII, *rpn4Δ*vs. *rpn4Δ*2 h 100 μM AsIII, *arr1Δ*vs. *arr1Δ*2 h 100 μM AsIII, parent vs. parent with 2 h 100 μM AsIII, as well as the parent strain vs. each deletion strain without arsenic). Additional data files [16](#s16){ref-type="supplementary-material"}, [17](#s17){ref-type="supplementary-material"}, [18](#s18){ref-type="supplementary-material"} and [19](#s19){ref-type="supplementary-material"} contain the gene lists of differential expression in knockout strains *yap1Δ*, *cad1Δ*, *rpn4Δ*and *arr1Δ*, respectively, compared to the parent without arsenic treatment. Additional data file [20](#s20){ref-type="supplementary-material"} contains every gene mentioned in this paper and the corresponding gene product descriptions. The primary microarray data will be submitted to the Gene Expression Omnibus (GEO) database at \[[@B80]\].
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
The dose-response curve of *S. cerevisiae*strain, BY4741
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
A self-organized tree of arsenite treated yeast experiments and a table depicting the numbers of significant genes
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 3
The primary raw cDNA data from all the experiments
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 4
The primary raw data for all the deletion experiments
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 5
The ranked arsenite sensitivity (phenotypic profiling) data
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 6
Genes two-fold or more differentially expressed after arsenite in the Yap1 deletion strain compared to the parent
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 7
Genes two-fold or more differentially expressed after arsenite in the Cad1 deletion strain compared to the parent
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 8
Genes failing to be induced or repressed by arsenite in the Yap1 deleted strain
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 9
Genes failing to be induced or repressed by arsenite in the Cad1 deleted strain
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 10
Under arsenite-treated conditions, Yap1 might regulate Arr2 and Arr3, and does not regulate Rpn4
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 11
Genes two-fold or more differentially expressed after arsenite in the Rpn4 deletion strain compared to the parent
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 12
Genes failing to be induced or repressed by arsenite in the Rpn4 deleted strain
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 13
Genes two-fold or more differentially expressed after arsenite in the Arr1 deletion strain compared to the parent
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 14
Genes failing to be induced or repressed by arsenite in the Arr1 deleted strain
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 15
Self-organized clustering of deletion strains with AsIII treatment and parent strain vs. deletion strains without arsenic
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 16
Gene list of two-fold differential expression in *yap1Δ* vs. parent without arsenic treatment
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 17
Gene list of two-fold differential expression in *cad1Δ* vs. parent without arsenic treatment
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 18
Gene list of two-fold differential expression in *rpn4Δ* vs. parent without arsenic treatment
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 19
Gene list of two-fold differential expression in *arr1Δ* vs. parent without arsenic treatment
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 20
A file of all the genes mentioned in the paper
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank Sherry Grissom, Eric Steele, Dmitry Gordenin and Gopalakrishnan Karthikeyan for technical assistance, James Brown for help with the analysis of the phenotypic profiling, and Rick Paules and Jennifer Fostel for critical review of the manuscript. This work was in part supported by grant CA 67166 (J.M.B.) from the US National Cancer Institute.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Arsenic-induced signaling and regulatory mechanisms involve transcriptional activators and the proteasome. **(a-d)**Significant network neighborhoods (*p*\< 0.005) uncovered by the ActiveModules algorithm, with the search performed at depth 1 (all nodes in the network are the nearest neighbors of one central node): (a) *FHL1*center; (b) *PRE1*center and proteasome complex; (c) *YAP1*and *CAD1*centers; (d) *HSF1*center. **(e)**An additional network centered on *MET31*with functional relevance to the arsenic response, which, however, did not reach significance in this analysis, *p*\< 0.11. **(f)**An overview of the network relationships between major arsenic-responsive transcription factors. Shades of red, induced; shades of green, repressed; blue boxed outline, significant expression; orange arrows, protein-DNA interaction; blue dashed lines, protein-protein interactions. The 2 h, 100 μM AsIII condition was used for the visual mappings. Many of the genes mapped to the network neighborhoods and displayed in this figure are boxed for the sake of clarity and space, but are mostly significantly differentially expressed.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Yap1 but not Cad1 is important for mediating the cell\'s adaptation to arsenic. **(a)**Self-organized heat map (dendograms were removed and boxes 1-3 indicate specific clusters) of 6,172 genes selected from the various indicated conditions. AsIII-treated parent wild type strain with normalized data values that are greater or less than those in condition(s) knocked-out Yap1, Cad1, Rpn4, or Arr1 treated with AsIII, by a factor of twofold. All knockouts tested revealed altered profiles compared to the wild type, except for *cad1Δ*. **(b)***yap1Δ*(condition 2) loses induced expression of stress response genes found in box 1, such as *SIR4*, *ISU2*, *MSN1*, *ATR1*, *CYT2*, *MDH1*, *AAD6*, *AAD4*, *TRR1*, *FLR1*, *GLR1*and *GRE2*. **(c)***rpn4Δ*(condition 4) loses induced expression of ubiquitinating and proteasomal genes found in box 3 - *UBP6*, *PRE8*, *PRE4*, *PRE7*and *PRE1*. **(d)***arr1Δ*(condition 5) loses repressed expression of sulfur amino-acid metabolism gene *SAM3*and glutamate biosynthesis gene *CIT2*, among others (box 2). *arr1Δ*also loses induced expression of serine biosynthesis gene *SER3*, sulfur amino-acid metabolism gene *SAM4*, cell-cycle regulator *ZPR1*, spindle-checkpoint subunit *MAD2*, ribonucleotide reductase *RNR1*and RNA polymerase I transcription factor *RRN9*, to name a few (box 3). Red, induced; green, repressed. For a comprehensive list of genes affected in all knockout experiments, see the Additional data files with the online version of this paper.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
The ubiquitin (Ub) and proteasome system responds to arsenic-mediated toxicity. *S. cerevisiae*ubiquitin and proteasome pathways show differential expression in a number of key genes, including that for the proteasomal activator *RPN4*. Induction is denoted by red boxes with fold-change ranges representing the 2 h, 100 μM AsIII and 0.5 h, 1 mM AsIII experiments, respectively.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Gene-expression profiling links sulfur assimilation, methionine and glutathione pathways. Selected genes in these pathways are represented as red for induced (2 h, 100 μM AsIII and 0.5 h, 1 mM AsIII, respectively) and green for repressed. Genes in white boxes are not differentially expressed. The pathways in the blue ovals are upstream of methionine, cysteine and glutathione, and are sensitive to arsenic. The downstream pathways employ numerous redundant enzymes that are differentially expressed, but are not sensitive. LT, late time-point, 4 h, 100 μM AsIII experiment; h, human; y, yeast.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
LinearActivePaths analysis finds that virtually all genes in active metabolic networks confer sensitivity to arsenic when deleted. **(a)**Serine, threonine, glutamate amino-acid synthetic pathways; **(b)**the shikimate pathway. The paths that compose these networks all have individual *p*-values of \< 0.05. The coloration for these figures is based on red for any gene ranked in the top 50 significant genes, yellow for 51-100, and green for \>101.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
Global model of the arsenic response: combining phenotypic data with gene-expression profiles reveals synergistic pathways leading to yeast detoxification mechanisms. Serine, threonine, aspartate and arginine, as well as shikimate metabolisms, in light blue, represent pathways that are judged as sensitive by phenotypic profiling. Yap1, colored light blue and red, is an example of a transcription factor that is both sensitive and confers induced gene expression. Deletion analysis confirms its role in arsenic-mediated control of the stress response. Red and green represent pathways or genes that are differentially expressed but not sensitive by phenotypic profiling. This schematic diagram demonstrates how the deletion of an individual gene leads to a change in sensitivity if the protein product of that gene is important in a biological process for adaptation to arsenic. On the other hand, expression profiling shows the end product of the cell\'s response to arsenic. Many of these downstream targets share redundant functions and are not vulnerable in the phenotypic profiling. The expression changes lead to the cell\'s response to indirect oxidative stress and mechanisms for detoxification. The arrows A, B, C and D represent the multiple branchpoints between redundant pathways. Note that the transport protein, Arr3, which extrudes AsIII out of the cell, is both sensitive and highly differentially expressed.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Pathways enriched for genes significantly expressed in response to arsenic
:::
Category Differentially expressed genes Pathway size *p*-value Significant
------------------------------------------ -------------------------------- -------------- ----------------- -------------
**KEGG pathway**
Cell cycle reference pathway 8 87 0.9072 False
Galcatose metabolism 5 15 0.0391 False
**Glutathione metabolism** **6** **11** **0.0014** **True**
MAPK signaling pathway 7 55 0.609 False
**Methionine metabolism** **8** **11** **1.07E-05** **True**
Proteasome 9 30 0.0127 False
Purine metabolism 14 139 0.8991 False
Pyrmidine metabolism 8 80 0.8515 False
**Sulfur metabolism** **7** **7** **7.15E-07** **True**
Serine, threonine and glycine metabolism 8 25 0.0125 False
Citrate cycle 4 22 0.3345 False
Starch and sucrose 9 31 0.0159 False
Pyruvate 4 25 0.4292 False
Reductive carboxylate 5 16 0.0508 False
Second messenger signaling 3 19 0.472 False
Valine, leucine, isoleucine 2 13 0.5313 False
Circadian rhythm 2 19 0.7398 False
Porphyrin and chlorophyll metabolism 7 74 0.8782 False
**Selenoamino-acid metabolism** **10** **12** **8.36E-08** **True**
Ubiquitin-mediated proteolysis 2 29 0.9133 False
Cysteine metabolism 2 4 0.088 False
Fructose and mannose 6 15 0.0093 False
Carbon fixation 3 15 0.3207 False
Alanine and aspartate 2 24 0.8477 False
Glutamate 3 19 0.472 False
Methane 2 4 0.088 False
**Gene Ontology (biological process)**
Biological process 72 436 0.0244 False
**Cell communication** **72** **270** **\<1.00E-008** **True**
Cell growth and maintenance 47 268 0.0231 False
Cell surface linked signal transduction 14 91 0.3197 False
Developmental processes 5 32 0.4233 False
**Heat-shock response** **14** **22** **5.40E-08** **True**
Intracellular signaling 9 47 0.1635 False
Serine threonine kinase signaling 5 38 0.5815 False
Signal transduction 26 172 0.2656 False
ATPase 3 78 0.9988 False
Cyclin 4 29 0.5499 False
Transcript profiling reveals that arsenic affects glutathione, methionine, sulfur, selenoamino-acid metabolism, cell communication and heat-shock response. Genes were categorized by KEGG pathway and Simplified Gene Ontology. In total, 829 genes out of 6,240 had a significant alteration in expression in at least one experimental condition. Along with the size of each functional category, a statistical measure for the significance of the enrichment was calculated by using a hypergeometric test. The level of significance for this test (True-shown in bold, False) was determined using the Bonferroni correction, where the α value is set at 0.05 and 27 and 11 tests were done for KEGG pathway and Simplified Gene Ontology, respectively.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Genes affiliated to sulfur metabolism
:::
---------------------------------------------------- -------------------- ------------- -------- ------------ ------------------------------------------------------
Gene Open reading frame AsIII Cd^2+^ H~2~O~2~ Enzyme
Fold change
Sulfate assimilation
*MET3* YJR010W 5.0-20.0 10.8 *-1.0-3.0* ATP sulfurylase
*MET14* YKL001C 2.0-14.0 9.6 *-2.0-3.0* APS kinase
*MET16* YPR167C 2.0-12.0 9.1 NC PAPS reductase
*MET22 (*0.5 h, 1 mM AsIII) YOL064C 3.3 2.5 NC Diphosphonucleoside phosphohydrolase
*MET10* YFR030W 3.0-5.0 5.2 NC Sulfite reductase alpha
*MET5/ECM17* YJR137C 2.0-5.0 4.5 *-2.0-4.0* Sulfite reductase beta
*MET1/20* YKR069W NC 6 NC Uroporphyrinogen III methylase
*MET8* YBR213W NC 7 *-2.0-4.0* Siroheme synthase
Sulfide incorporation and transulfuration pathways
*MET2* YNL277W 4.0-3.0 4.6 NC Homoserine transacetylase
*MET25/17* YLR303W NC 4.8 NC O-Acetylhomoserine sulfhydrylase
*STR4/CYS4* YGR155W 2.0-2.7 2.5 *-2.0-3.0* Cystathionine B-synthase
*STR1/CYS3*(4 h, 100 μM AsIII) YAL012W *-3.5* 13.4 NC Cystathionine-lyase
*STR3* YGL184C 2.2-3.9 13.5 2.0-4.0 Cystathionine G-synthase
*YFR055W* YFR055W *-9.0-5.0* *-1.1* NC Cystathionine-lyase
Methionine and AdoMet biosynthesis
*MET6* YER091C 1.0-3.5 NC *-2.0-5.0* N5-Methyltetrahydrofolate homocysteine transferase
*MET7* YOR241W NC *-1.6* NC Tetrahydrofolyl polyglutamate synthase
*MET13* YGL125W NC NC *-1.0-3.0* Methylene tetrahydrofolate reductase
*SAM1*(4 h, 100 μM AsIII) YLR180W 2.5 NC *-2.0-9.0* AdoMet synthetase
*SAM2*(4 h, 100 μM AsIII) YDR502C 3.8 NC *-2.0-4.0* AdoMet synthetase
*MHT1* YLL062C 5.0-2.8 10.6 NC S-methylmethionine: homocysteine S-methyltransferase
Sulfur compound uptake
*SUL1* YBR294W 5.4-2.4 20 NC Sulfate transporter
*SUL2* YLR092W 2.5-2.8 3 NC Sulfate transporter
*MUP1*(2 h and 4 h, 100 μM AsIII) YGR055W 5.0-14.0 2 *-1.0-4.0* Methionine permease, high affinity
*MUP3* YHL036W 8.0-7.0 7 Methionine permease, low affinity
Regulatory genes
*MET4*(2 h, 100 μM AsIII) YNL103W 2 1.5 NC bZIP
*MET28* YIR017C NC 5 NC bZIP
*CBF1* YJR060W NC *-2* NC bHLH
*MET30* YIL046W 5.0-1.6 7 NC WD40 repeats F box
*MET31* YPL039W NC 1 *-1.0-7.0* Zinc finger
*MET32* YDR253C 6.0-3.6 14 *-1.0-4.0* Zinc finger
---------------------------------------------------- -------------------- ------------- -------- ------------ ------------------------------------------------------
Arsenic treatment stimulates a sulfur response in yeast. Gene expression data comparisons between arsenic, cadmium, and H~2~O~2~-treated *Saccharomyces cerevisiae*reveal arsenic and cadmium mediated sulfur responses, but none with hydrogen peroxide. AsIII column, 2 h, 100 μM and 0.5 h, 1 mM (combined biological replicates), unless noted; cadmium column, 1 h, 1 mM \[63\]; H~2~O~2~column, 10, 20, 40, 60, 120 min, 400 μM \[62\]. Numbers in ordinary typeface denote induction; (-) and italicized numbers denote repression; NC, no change.
:::
|
PubMed Central
|
2024-06-05T03:55:51.888656
|
2004-11-29
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545798/",
"journal": "Genome Biol. 2004 Nov 29; 5(12):R95",
"authors": [
{
"first": "Astrid C",
"last": "Haugen"
},
{
"first": "Ryan",
"last": "Kelley"
},
{
"first": "Jennifer B",
"last": "Collins"
},
{
"first": "Charles J",
"last": "Tucker"
},
{
"first": "Changchun",
"last": "Deng"
},
{
"first": "Cynthia A",
"last": "Afshari"
},
{
"first": "J Martin",
"last": "Brown"
},
{
"first": "Trey",
"last": "Ideker"
},
{
"first": "Bennett",
"last": "Van Houten"
}
]
}
|
PMC545799
|
Background
==========
Protein-protein interactions have an essential role in a wide variety of biological processes. A wealth of data has emerged to show that most proteins function within networks of interacting proteins, and that many of these networks have been conserved throughout evolution. Although some of these networks constitute stable multi-protein complexes while others are more dynamic, they are all built from specific binary interactions between individual proteins. Maps depicting the possible binary interactions among proteins can therefore provide clues not only about the functions of individual proteins but also about the structure and function of entire protein networks and biological systems.
One of the most powerful technologies used in recent years for mapping binary protein interactions is the yeast two-hybrid system \[[@B1]\]. In a yeast two-hybrid assay, the two proteins to be tested for interaction are expressed with amino-terminal fusion moieties in the yeast *Saccharomyces cerevisiae*. One protein is fused to a DNA-binding domain (BD) and the other is fused to a transcription activation domain (AD). An interaction between the two proteins results in activation of reporter genes that have upstream binding sites for the BD. To map interactions among large sets of proteins, the BD and AD expression vectors are placed initially into different haploid yeast strains of opposite mating types. Pairs of BD and AD fused proteins can then be tested for interaction by mating the appropriate pair of yeast strains and assaying reporter activity in the resulting diploid cells \[[@B2]\]. Large arrays of AD and BD strains representing, for example, most of the proteins encoded by a genome, have been constructed and used to systematically detect binary interactions \[[@B3]-[@B6]\]. Most large-scale screens have used such arrays in a library-screening approach in which the BD strains are individually mated with a library containing all of the AD strains pooled together. After plating the diploids from each mating onto medium that selects for expression of the reporters, the specific interacting AD-fused proteins are determined by obtaining a sequence tag from the AD vector in each colony.
High-throughput two-hybrid screens have been used to map interactions among proteins from bacteria, viruses, yeast, and most recently, *Caenorhabditis elegans*and *Drosophila melanogaster*\[[@B4]-[@B10]\]. Analyses of the interaction maps generated from these screens have shown that they are useful for predicting protein function and for elaborating biological pathways, but the analyses have also revealed several shortcomings in the data \[[@B11]-[@B13]\]. One problem is that the interaction maps include many false positives - interactions that do not occur *in vivo*. Unfortunately, this is a common feature of all high-throughput methods for generating interaction data, including affinity purification of protein complexes and computational methods to predict protein interactions \[[@B11]-[@B14]\]. A solution to this problem has been suggested by several studies that have shown that the interactions detected by two or more different high-throughput methods are significantly enriched for true positives relative to those detected by only one approach \[[@B11]-[@B13]\]. Thus it has become clear that the most useful protein-interaction maps will be those derived from combinations of cross-validating datasets.
A second shortcoming of the large-scale screens has been the high rate of false negatives, or missed interactions. This is evident from comparing the high-throughput data with reference data collected from published low-throughout studies. Such comparisons with two-hybrid maps from yeast \[[@B13]\] and *C. elegans*\[[@B5]\], for example, have shown that the high-throughput data rarely covers more than 13% of the reference data, implying that only about 13% of all interactions are being detected. The finding that different large datasets show very little overlap, despite having similar rates of true positives, supports the conclusion that high-throughput screens are far from saturating \[[@B10],[@B12]\]. For example, three separate screening strategies were used to detect hundreds of interactions among the approximately 6,000 yeast proteins, and yet only six interactions were found in all three screens \[[@B10]\]. These results suggest that many more interactions might be detected simply by performing additional screening, or by applying different screening strategies to the same proteins. In addition, anecdotal evidence has suggested that the use of two-hybrid systems based on different fusion moieties may broaden the types of protein interactions that can be detected. In one study, for example, screens performed using the same proteins fused to either the LexA BD or the Gal4 BD produced only partially overlapping results, and each system detected biologically significant interactions missed by the other \[[@B15]\]. Thus, the application of different two-hybrid systems and different screening strategies to a proteome would be expected to provide more comprehensive datasets than would any single screen.
We set out to map interactions among the approximately 14,000 predicted *Drosophila*proteins by using two different yeast two-hybrid systems (LexA- and Gal4-based) and different screening strategies. Results from the screens using the Gal4 system have already been published \[[@B6]\]. In that study, Giot *et al*. successfully amplified 12,278 *Drosophila*open reading frames (ORFs) and subcloned a majority of them into the Gal4 BD and Gal4 AD expression vectors by recombination in yeast. They screened the arrays using a library-screening approach and detected 20,405 interactions involving 7,048 proteins. To extend these results we subcloned the same amplified *Drosophila*ORFs into vectors for use in the LexA-based two-hybrid system, and constructed arrays of BD and AD yeast strains for high-throughput screening. Our expectation was that maps generated with these arrays would include interactions missed in previous screens, and would also partially overlap the Gal4 map, providing opportunities for cross-validation.
Initially, we screened for interactions involving proteins that are primarily known or suspected to be cell-cycle regulators. We chose cell-cycle proteins as a starting point for our interaction map because cell-cycle regulatory systems are known to be highly conserved in eukaryotes, and because previous results have suggested that the cell-cycle regulatory network is centrally located within larger cellular networks \[[@B16]\]. This is most evident from examination of the large interaction maps that have been generated for yeast proteins using yeast two-hybrid and other methods. Within these maps there are more interactions between proteins that are annotated with the same function (for example, \'Pol II transcription\', \'cell polarity\', \'cell-cycle control\') than between proteins with different functions, as expected for a map depicting actual functional connections between proteins. Interestingly, however, certain functional groups have more inter-function interactions than others. Proteins annotated as \'cell-cycle control\', in particular, were frequently connected to proteins from a wide range of other functional groups, suggesting that the process of cell-cycle control is integrated with many other cellular processes \[[@B16]\]. Thus, we set out to further elaborate the cell-cycle regulatory network by identifying new proteins that may belong to it, and new connections to other cellular networks.
Results
=======
Construction of an extensive protein interaction map centered on cell-cycle regulators by high-throughput two-hybrid screening
------------------------------------------------------------------------------------------------------------------------------
We used the same set of 12,278 amplified *Drosophila*full-length ORFs from the Gal4 project \[[@B6]\] to generate yeast arrays for use in a modified LexA-based two-hybrid system (see Materials and methods). In the LexA system the BD is LexA and the AD is B42, an 89-amino-acid domain from *Escherichia coli*that fortuitously activates transcription in yeast \[[@B17]\]. In the version that we used, both fusion moieties are expressed from promoters that are repressed in glucose so that their expression can be repressed during construction and amplification of the arrays \[[@B18]\]. Previous results have shown that this prevents the loss of genes encoding proteins that are toxic to yeast, and that interactions involving such proteins can be detected by inducing their expression only on the final indicator media \[[@B18],[@B19]\]. The ORFs were subcloned into the two vectors by recombination in yeast as previously described \[[@B3],[@B6]\], and the yeast transformants were arrayed in a 96-well format. The resulting BD and AD arrays each have approximately 12,000 yeast strains, over 85% of which have a full-length *Drosophila*ORF insert (see Materials and methods). For all strains involved in an interaction reported here, the plasmid was isolated and the insert was sequenced to verify the identity of the ORF.
As a first step toward generating a LexA-based protein-interaction map, we chose 152 BD-fused proteins that were either known or homologous to regulators of the cell cycle or DNA damage repair (see Additional data file 2). We used all 152 proteins as \'baits\' to screen the 12,000-member AD array. We used a pooled mating approach \[[@B19]\] in which individual BD bait strains are first mated with pools of 96 AD strains. For pools that are positive with a particular BD, the corresponding 96 AD strains are then mated with that BD in an array format to identify the particular interacting AD protein(s). We had previously shown that this approach is very sensitive and allows detection of interactions involving proteins that are toxic to yeast or BD fused proteins that activate transcription on their own \[[@B19]\]. Moreover, the final assay in this approach is a highly reproducible one-on-one assay between an AD and a BD strain, in which the reporter gene activities are recorded to provide a semi-quantitative measure of the interaction. Using this approach we detected 1,641 reproducible interactions involving 93 of the bait proteins. We also performed library screening \[[@B6]\] with a subset of the 152 baits that did not activate the reporter genes on their own. This resulted in the detection of 173 additional interactions with 57 bait proteins. Thirty-nine interactions were found by both approaches, and these involved 21 of the 44 BD genes active in both approaches. There were 95 BD genes for which interaction data was obtained by the pooled mating approach, and 59 active BD genes in the library screening approach. The average number of interactions was 18 per BD gene in the pooled mating data, while the library screening data had an average of only four interactions per active BD gene. The average level of reporter activation for the 39 interactions that were detected in both screens was significantly higher than the average of all interactions (see Additional data file 3), suggesting that the weaker interactions are more likely to be missed by one screen or another, even though they are reproducible once detected.
Altogether we detected interactions with 106 of the 152 baits, which resulted in a protein-interaction map with 1,814 unique interactions among the products of 488 genes (see Additional data file 3). The map includes interactions that were already known or that could be predicted from known orthologous or paralogous interactions (see below). The map also includes a large number of novel interactions, including many involving functionally unclassified proteins.
Evaluation of the LexA-based protein interaction map
----------------------------------------------------
As is common with data derived from high-throughput screens, the number of novel interactions detected was large, making direct *in vivo*experimental verification impracticable. Thus, we set out to assess the quality of the data by examining the topology of the interaction map, by looking for enrichment of genes with certain functions, and by comparing the LexA map with other datasets. First we examined the topology of the interaction map, because recent studies have shown that cellular protein networks have certain topological features that correlate with biological function \[[@B20]\]. In our interaction map, the number of interactions per protein (*k*) varies over a broad range (from 1 to 84) and the distribution of proteins with *k*interactions follows a power law, similar to previously described protein networks \[[@B6],[@B21]\]. Most (98%) of the proteins in the map are linked together into a single network component by direct or indirect interactions (Figure [1a](#F1){ref-type="fig"}). The network has a small-world topology \[[@B22]\], characterized by a relatively short average distance between any two proteins (Table [1](#T1){ref-type="table"}) and highly interconnected clusters of proteins. Removal of the most highly connected proteins from the map does not significantly fragment the network, indicating that the interconnectivity is not simply due to the most promiscuously interacting proteins (Figure [1b](#F1){ref-type="fig"}). In other interaction maps generated with randomly selected baits, proteins with related functions tend to be clustered into regions that are more highly interconnected than is typical for the map as a whole \[[@B5],[@B6],[@B16]\]. Moreover, interactions within more highly interconnected regions of a protein-interaction map tend to be enriched for true positives \[[@B6],[@B23]-[@B25]\]. Thus, the overall topology of the interaction map that we generated is consistent with that of other protein networks, and in particular, with the expectation for a network enriched for functionally related proteins.
Next we assessed the list of proteins in the interaction map to look for enrichment of proteins or pairs of proteins with particular functions. An interaction map with a high rate of biologically relevant interactions should have a high frequency of interactions between pairs of proteins previously thought to be involved in the same biological process. Among the 488 proteins in the map, 153 have been annotated with a putative biological function using the Gene Ontology (GO) classification system \[[@B26],[@B27]\]. Because we used a set of BD fusions enriched for cell-cycle and DNA metabolic functions, we expected to see similar enrichments in the list of interacting AD fusions, as well as more interactions between genes with these functions. Both of these expectations are borne out. In the list of BD genes, both cell-cycle and DNA metabolism functions are enriched approximately 17-fold compared to similarly sized lists of randomly selected proteins (*P*\< 0.00002). In the AD list, these two functions are enriched four- and threefold, respectively (Table [2](#T2){ref-type="table"}). The frequency with which interactions occur between pairs of proteins annotated for DNA metabolism is five times more than expected by chance; similarly, cell-cycle genes interact with each other six times more frequently than expected (*P*\< 0.001). Thus, the enrichment for proteins and pairs of interacting proteins annotated with the same function suggests that many of the novel interactions will be biologically significant. It also suggests that the map will be useful for predicting the functions of novel proteins on the basis of their connections with proteins having known functions, as described for other interaction maps \[[@B16],[@B28]\].
Comparison of the *Drosophila*protein-interaction maps
------------------------------------------------------
Direct comparison of the LexA cell-cycle map with the Gal4 data revealed that only 28 interactions were found in common between the two screens (Table [1](#T1){ref-type="table"}). Moreover, more than a quarter of the proteins in the LexA map were absent from the Gal4 proteome-wide map. Among the 106 baits that had interactions in the LexA map, for example, 60 failed to yield interactions in the Gal4 proteome-wide map, even though all but six of these were successfully cloned in the Gal4 arrays \[[@B6]\] (see Additional data file 6). Similarly, 46 of the 152 LexA baits that we used failed to yield interactions from our work, yet 14 of these had interactions in the Gal4 map. Thus, the lack of overlap between the two datasets is partly due to their unique abilities to detect interactions with specific proteins. Nevertheless, for the 347 proteins common to both maps, the two screens combined to detect 1428 interactions, and yet only 28 of these were in both datasets. This indicates that the two screens detected mostly unique interactions even among the same set of proteins. Comparison with a set of approximately 2,000 interactions recently generated in an independent two-hybrid screen \[[@B29]\] showed only three interactions in common with our data, in part because only eight of the same bait proteins were used successfully in both screens.
Although only 28 interactions were found in both the Gal4 map and our map, this rate of overlap is significantly greater than expected by chance (*p*\< 10^-6^; Table [1](#T1){ref-type="table"}). To show this, we generated 10^6^random networks having the same BD proteins, total interactions and topology as the LexA map, and found that none of these random maps shared more than two interactions in common with the Gal4 map. To assess the relative quality of the 28 common interactions we used the confidence scores assigned to them by Giot *et al*. \[[@B6]\]. They used a statistical model to assign confidence scores (from 0 to 1), such that interactions with higher scores are more likely to be biologically relevant than those with lower scores. The average confidence scores of the 28 interactions in common with our LexA data (0.63), was higher than the average for all 20,439 Gal4 interactions (0.34), or for random samplings of 28 Gal4 interactions (0.32; *P*\< 0.0001), indicating that the overlap of the two datasets is significantly enriched for biologically relevant interactions. Thus, the detection of interactions by both systems could be used as an additional measure of reliability. The surprisingly small number of common interactions, however, severely limits the opportunities for cross-validation, and suggests that both datasets are far from comprehensive.
An alternative explanation for the small proportion of common interactions is the possible presence of a large number of false positives in one or both datasets. The estimation of false-positive rates is challenging, in part because it is difficult to prove that an interaction does not occur under all *in vivo*conditions, and also because the number of potential false positives is enormous. Nevertheless, the relative rates of false positives between two datasets can be inferred by comparing their estimated rates of true positives \[[@B11]-[@B13]\]. To compare true-positive rates between the LexA and Gal4 datasets, we looked for their overlap with several datasets that are thought to be enriched for biologically relevant interactions (Table [3](#T3){ref-type="table"}). These include a reference set of published interactions involving the proteins that were used as baits in both the LexA and Gal4 screens; interactions between the *Drosophila*orthologs of interacting yeast or worm proteins (orthologous interactions or \'interlogs\' \[[@B30],[@B31]\]); and between proteins encoded by genes known to interact genetically, which are more likely to physically interact than random pairs of proteins \[[@B32],[@B33]\]. As expected, the overlap with these datasets is enriched for higher confidence interactions. The average confidence scores for the Gal4 interactions in common with the yeast interlogs, worm interlogs and *Drosophila*genetic interactions are 0.63, 0.68 and 0.80, respectively, substantially higher than the average confidence scores for all Gal4 interactions (0.34). This supports the notion that these datasets are enriched for true-positive interactions relative to randomly selected pairs of proteins. We found that the fractions of LexA- and Gal4-derived interactions that overlap with these datasets are similar (Table [3](#T3){ref-type="table"}). For example, 25 (1.4%) of the 1814 LexA interactions and 294 (1.4%) of the 20,439 Gal4 interactions have yeast interlogs. This suggests that the LexA and Gal4 two-hybrid datasets have similar percentages of true positives, and thus similar rates of false positives. They also appear to have similar rates of false negatives, which may be over 80% if calculation is based on the lack of overlap with published interactions (Table [3](#T3){ref-type="table"}). This supports the explanation that the main reason for the lack of overlap between the datasets is that neither is a comprehensive representation of the interactome, and suggests that a large number of interactions remain to be detected.
Biologically informative interactions
-------------------------------------
Further inspection of the LexA cell-cycle interaction map revealed biologically informative interactions and additional insights for interpreting high-throughput two-hybrid data. For example, we expected to observe interactions between cyclins and cyclin-dependent kinases (Cdks), which have been shown to interact by a number of assays. Our interaction map includes six proteins having greater than 40% sequence identity to Cdk1 (also known as Cdc2). A map of all the interactions involving these proteins reveals that they are multiply connected with several cyclins (Figure [2](#F2){ref-type="fig"}). For example, all of the known cyclins in the map interacted with at least two of the Cdk family members. The map includes 20 interactions between five Cdks and six known cyclins plus one uncharacterized protein, CG14939, which has sequence similarity to cyclins. Only one of these interactions (Cdc2c-CycJ) is known to occur *in vivo*\[[@B34]\], and several others are thought not to occur *in vivo*(for example Cdc2-CycE \[[@B35]\]). Similarly, the Gal4 interaction map has three Cdk-cyclin interactions \[[@B6]\], including one known to occur *in vivo*(Cdk4-CycD) and two that do not occur *in vivo*\[[@B35]\].
Thus, while some of these interactions are false positives in the strictest sense, the data is informative nevertheless, as it clearly demonstrates a high incidence of paralogous interactions - where pairs of interacting proteins each have paralogs, some combinations of which also interact *in vivo*. Such patterns are consistent with potential interactions between members of different protein families, even though they do not reveal the precise pair of proteins that interact *in vivo*. This class of informative false positives may be common in two-hybrid data where the interaction is assayed out of biological context. Experimentally reproducible interactions, whether or not they occur *in vivo*, can be used to discover interacting protein motifs or domains \[[@B6],[@B36]\]. They can also suggest functional relationships between protein families and guide experiments to establish the actual *in vivo*interactions and functions of specific pairs of interacting proteins.
The Cdk subgraph also illustrates that proteins with similar interaction profiles may have related functions or structural features. To look for other groups of proteins having similar interaction profiles we used a hierarchical clustering algorithm to cluster BD and AD fusion proteins according to their interactions (see Materials and methods). The resulting clustergram reveals several groups of proteins with similar interaction profiles (Figure [3](#F3){ref-type="fig"}). One of the most prominent clusters (Figure [3](#F3){ref-type="fig"}, circled in blue) includes three related proteins involved in ubiquitin-mediated proteolysis, SkpA, SkpB and SkpC. Skp proteins are known to interact with F-box proteins, which act as adaptors between ubiquitin ligases, known as SCF (Skp-Cullin-F-box) complexes, and proteins to be targeted for destruction by ubiquitin-mediated proteolysis \[[@B37]\]. A map of the interactions involving the Skp proteins shows a group of 21 AD proteins that each interact with two or three of the Skp proteins (Figure [4](#F4){ref-type="fig"}). This group is highly enriched for F-box proteins, including 13 of the 15 F-box proteins in the AD list; the other two F-box proteins interacted with only one Skp (Figure [4](#F4){ref-type="fig"}). Several of the interactions in common with the Gal4 data are also in the Skp cluster, and 12 out of 16 of these involve proteins that interact with two or more Skp proteins.
Thus, the Skp cluster provides another example of how proteins with similar interaction profiles may be structurally or functionally related, and how such clusters may be enriched for biologically relevant interactions. This is consistent with previous results showing that protein pairs often have related functions if they have a significantly larger number of common interacting partners than expected by chance \[[@B24],[@B38]\]. These groups of proteins are likely to be part of more extensive functional clusters that could be identified by more sophisticated topological analyses (for example \[[@B39]-[@B44]\]. Maps showing several other major clusters derived from the cluster-gram are shown in Additional data file 7.
The interaction profile data is statistically confirmed by domain-pairing data, which shows that certain pairs of domains are found within interacting pairs of proteins more frequently than expected by chance (Table [4](#T4){ref-type="table"}). These include the Skp domain and F-box pair, the protein kinase and cyclin domains, and several less obvious pairings. For example, the cyclin and kinase domains are observed to be associated with various zinc-finger and homeodomain proteins, and the kinase domain with a number of nucleic-acid metabolism domains (Table [4](#T4){ref-type="table"}). A similar analysis of the Gal4 data, performed by Giot *et al*. \[[@B6]\], revealed a number of significant domain pairings, including the Skp/F-box and the kinase/cyclin pairs and several others found in the LexA dataset. Therefore, although the number of proteins in the LexA dataset is relatively small, domain associations are observed in the data, demonstrating that a high-density interaction map, with a high average number of interactions per protein, provides insight into patterns of domain interactions that is equally valuable as that obtained from a proteome-wide map.
Discussion
==========
Proteome-wide maps depicting the binary interactions among proteins provide starting points for understanding protein function, the structure and function of protein complexes, and for mapping biological pathways and regulatory networks. High-throughput approaches have begun to generate large protein-interaction maps that have proved useful for functional studies, but are also often plagued by high rates of false positives and false negatives. Several analyses have shown that the set of interactions detected by more than one high-throughout approach is enriched for biologically relevant interactions, suggesting that the application of multiple screens to the same set of proteins results in higher-confidence, cross-validated interactions \[[@B11]-[@B13]\]. Such cross-validation has been limited, however, by the lack of overlap among high-throughput datasets. Here we describe initial efforts to complement a recently published *Drosophila*protein interaction map that was generated using the Gal4 yeast two-hybrid system \[[@B6]\]. We constructed yeast arrays for use in the LexA-based two-hybrid system by subcloning approximately 12,000 *Drosophila*ORFs, using the same PCR amplification products used in the Gal4 project, into the LexA two-hybrid vectors. Initially, we used a novel pooled mating approach \[[@B19]\] to screen one of the 12,000-member arrays with 152 bait proteins related to cell cycle regulators. By using both a different screening approach and a different two-hybrid system, we expected to increase coverage and to validate some of the interactions detected by the Gal4 screens.
The level of coverage for a high-throughput screen can be estimated by determining the percentage of a reference dataset that was detected; reference sets have been derived from published low-throughput experiments, for example, which are considered to have relatively low false-positive rates. High-throughput two-hybrid data for yeast and *C. elegans*proteins were shown to cover only about 10-13% of the corresponding reference datasets \[[@B5],[@B10],[@B13]\]. Two factors may contribute to this lack of coverage. First, some interactions cannot be detected using the yeast two-hybrid system, even though they could be detected in low-throughput studies using other methods. Examples include interactions that depend on certain post-translational modifications, that require a free amino terminus or that involve membrane proteins. Second, high-throughput yeast two-hybrid screens often fail to test all possible combinations of interactions; in other words, the screens are not saturating or complete.
Although the relative contribution of these two factors is difficult to estimate, results from screens to map interactions among yeast proteins suggest that the major reason for the lack of coverage is that the screens are incomplete. Complete screens would identify all interactions that could possibly be detected by a given method; ideally therefore, two complete screens using the same method would identify all the same interactions. However, the rate of overlap among the different yeast proteome screens is low, even though they used very similar two-hybrid systems. Moreover, the overlap between screens is not significantly greater than the rate at which they overlap any reference set \[[@B4],[@B10]\]. This is true even when only higher-confidence interactions are considered; for example, two large interaction screens of yeast proteins detected 39% and 65% of a higher-confidence dataset, respectively, but only 11% of the reference set was detected by both screens \[[@B12]\]. These results indicate that the lack of coverage in high-throughput two-hybrid data is largely due to incomplete screening, and that significantly larger datasets than those currently available will be needed before different datasets can be used to cross-validate interactions.
The rates of coverage and completeness from our high-throughput two-hybrid screening with *Drosophila*proteins are consistent with those for the yeast proteins. We used the LexA system to detect 1,814 reproducible interactions to complement the 20,439 interactions previously detected in a proteome-wide screen using the Gal4 system \[[@B6]\]. The overlap between the LexA and Gal4 screens is less than 2% of each dataset, whereas their overlap with a reference set was 17% and 14%, respectively, and only 2% of the reference set was detected by both screens (Table [2](#T2){ref-type="table"}). Taken together, these results suggest that, like the yeast interaction data, both *Drosophila*datasets are far from complete and that many more interactions could be detected by additional two-hybrid screening.
The actual number of interactions that might be detected by complete two-hybrid screening might be roughly estimated from the partially overlapping datasets, as was performed for accurate estimation of the number of genes in the human genome \[[@B45],[@B46]\]. In this approach, the overlap of two subsets, given that one subset is a homogeneous random sample of the whole, is sufficient to estimate the size of the whole. To make such an estimate with high-throughput two-hybrid data, however, it is necessary to first filter out false positives, as they are mostly different for the two datasets, as suggested by the fact that the nonoverlapping data has a lower rate of true positives than the overlapping data. Giot *et al*. estimated that at least 11% of the Gal4 interactions are likely to be biologically relevant, based on the prediction accuracy of their statistical model \[[@B6]\]. We found by comparison with other datasets that the rates of true positives are not substantially different between the LexA and Gal4 data (Table [3](#T3){ref-type="table"}). Thus, if we use 11% as the minimal rate of true positives in each dataset, we obtain 200 true interactions from the LexA screen and 2,248 from the Gal4 screens. If we further assume that all of the 28 common interactions are true positives, we can estimate that complete screens should be able to detect around 16,000 true positive interactions (200 × 2,248/28). If each screening approach has a false-positive rate of 89%, then around 150,000 interactions from each approach would be required in order to create complete, cross-validating datasets, where the overlap would be comprised of true positives. This estimate is highly sensitive to both the frequency of true positives in the two datasets, and the number of positives in the overlap between the datasets; for example, if true-positive frequency is underestimated by only twofold, there will be four times as many interactions.
False-positive interactions have been classified as technical or biological \[[@B5]\]. A technical false positive is an artifact of the particular interaction assay, and the two proteins involved do not actually interact under any setting. A biological false positive is one in which the two proteins genuinely and reproducibly interact in a particular assay, but the interaction does not take place in a biological setting; for example, the interacting proteins may never be temporally or spatially co-localized *in vivo*. Using the approach described here, the interactions are shown to be reproducible during the one-on-one two-hybrid assays that are used to record reporter activity scores, suggesting that we have minimized the frequency of technical false positives.
We suggest that the biological false positives might be further classified as informative and non-informative. Informative false positives are interactions that do not occur *in vivo*, but that nevertheless have some biological basis for being detected and are potentially useful for guiding future experiments. In our data, for example, the Cdk and Skp proteins each interact with a different group of targets, which in turn interact with multiple Cdk or Skp proteins. From this data alone, we would accurately predict that Cdk proteins interact with cyclins, and that Skp proteins interact with F-box proteins, even though only some of the specific combinations are true *in vivo*partners. Similarly, from analysis of domain pairs in the LexA dataset, other patterns are evident, such as homeobox domains being associated with both protein kinase and cyclin domains (Table [4](#T4){ref-type="table"}). Additional information or experimentation would be needed to determine which of the specific paralogous interactions function *in vivo*. Co-affinity purification, for example, might be used to directly test all possible pairs of paralogous interactions implied by the two-hybrid map. Alternatively, the genes encoding each possible pair of proteins could be examined for correlated expression patterns, for example, to suggest more likely pairs or to exclude pairs that are not coexpressed.
Conclusions
===========
We used high-throughput screening to detect 1,814 protein interactions involving many proteins with cell-cycle and related functions. The resulting interaction map is similar in quality to other large interaction maps and is predominated by previously unidentified interactions. The majority of the proteins in the map have not been assigned a biological function, and the map provides a first clue about the potential functions of these proteins by connecting them with characterized proteins or pathways. High-throughput interaction data such as this should allow researchers to quickly identify possible patterns of protein interactions for use in selecting additional functional assays to perform on their gene(s) of interest. This narrows down the number of potential assays necessary to establish function for a given gene from hundreds to just a handful; conversely, when studying a specific function, such as the cell cycle, interaction data can identify which few genes, selected from thousands, may have a role in the process. Just as the sequencing of various genomes has not allowed unambiguous ascription of biological function to the majority of the identified genes, mapping of an interactome by high-throughput methods does not allow final assignment of interaction capacity or of higher functionality to a protein. This requires additional experiments, guided by these and other high-throughput data. The results presented here show that extending and combining different two-hybrid datasets will allow further refinement of the selection of functional analyses to be performed for each protein of the proteome.
Materials and methods
=====================
Plasmids and strains
--------------------
Yeast two-hybrid vectors used are related to those originally described for the LexA system \[[@B17]\]. The vector for expressing amino-terminal LexA DNA-binding domain (BD) fusions was pHZ5-NRT, which expresses fusions from the regulated *MAL62*promoter \[[@B18]\]. The vector for expressing amino-terminal activation domain (AD) fusions from the *GAL1*promoter was pJZ4-NRT, which was constructed from pJG4-5 \[[@B17]\] by replacing the *ADH1*terminator with the *CYC1*terminator and inserting the 5\' and 3\' recombination tags (5RT1 and 3RT1 \[[@B18]\]) into the cloning site downstream from the AD coding region. Construction details can be found in Additional data file 1. Maps and sequences are available at \[[@B47]\]. Yeast (*S. cerevisiae*) strain RFY231 (MAT *trp1*::*hisG his3 ura3-1 leu2*::3Lexop-*LEU2*) and RFY206 (Mat**a***his3Δ200 leu2-3 lys2Δ201 ura3-52 trp1Δ*::*hisG*) were previously described \[[@B2],[@B48]\]. RFY206 containing the *lacZ*reporter plasmid pSH18-34 \[[@B49]\] is referred to here as strain Y309.
Yeast two-hybrid arrays
-----------------------
Two yeast arrays were constructed by homologous recombination (gap repair) in yeast \[[@B3]\]. We began with the 13,393 unique PCR products, which were generated using gene-specific primer pairs corresponding to the predicted *Drosophila*ORFs, from ATG to stop codon, described in Giot *et al*. \[[@B6]\]. For the AD array, we co-transformed RFY231 with each PCR product along with pJZ4-NRT that had been linearized with *Eco*RI and *Bam*HI, and selected recombinants on glucose minimal media lacking tryptophan. Five colonies from each transformation were picked and combined into a well of a 96-well plate. For the BD array, we co-transformed Y309 with each PCR product along with pHZ5-NRT that had been linearized with *Eco*RI and *Bam*HI, and selected recombinants on glucose minimal medium lacking histidine and uracil. BD clones used in the screens and AD clones showing positive interactions were sequenced to verify the ORF identities. See Additional data files for details.
Two-hybrid screening
--------------------
The BD fused proteins used as baits in our screens are listed in Additional data file 2. The AD array was screened using a two-phase pooled mating approach \[[@B19]\]. First, pools containing the 96 AD strains from each plate in the AD array were constructed by scraping strains grown on agar plates, dispersing in 15% glycerol, and aliquoting into a 96-well format; the 142 pools, representing approximately 13,000 AD strains, were arrayed on two 96-well plates. In the first phase, individual BD strains were mated with the 142 AD pools by dispensing 5-μl volumes of each culture onto YPD plates using a Biomek FX robot (Beckman Coulter). After 2 days growth at 30°C, yeast were replicated to medium selective for diploids, which have the AD, BD and *lacZ*reporter plasmids, and containing both galactose and maltose to induce expression of the AD and BD fusions, respectively. The plates also lacked leucine to assay for expression of the *LEU2*reporter, and contained X-Gal (40 μg/ml) to assay for expression of *lacZ*. These plates were photographed after 5 days at 30°C and interactions were scored as described \[[@B19]\]. In the second phase of screening, single BD strains were mated with the appropriate panel(s) of 93 AD strains corresponding to the pools that were positive in the first phase. The *LEU2*and *lacZ*reporters were assayed on separate plates: growth on plates lacking leucine was scored from 0 (no growth) to 3 (heavy growth); the extent of blue on the X-Gal plates was scored from 0 (white) to 5 (dark blue). After re-testing interactions (see Additional data files) the AD plasmids from interacting AD strains were rescued in bacteria and clones were sequenced to verify insert identity. Cloned plasmids were then reintroduced into RFY231 and used in all possible combinations of one-on-one mating operations with the appropriate BD strains to repeat the interaction assay a third time. The same set of BDs was also used to screen a pool of all approximately 13,000 AD strains using a library screening approach as described in the Additional data files. All interaction data from both screens are listed in Additional data file 3 and are also available at \[[@B47],[@B50]\] and at IntAct \[[@B51]\] in the Proteomics Standards Initiative - Molecular Interactions (PSI-MI) standard exchange format \[[@B52]\].
Data analysis
-------------
The interaction profiles for the BD fused proteins and AD fused proteins were independently clustered and are plotted in Figure [3](#F3){ref-type="fig"} using Genespring software (Silicon Genetics). Protein-interaction map graphs in Figures [1](#F1){ref-type="fig"}, [2](#F2){ref-type="fig"} and [4](#F4){ref-type="fig"} and Additional data file 7 were drawn with a program developed by Lana Pacifico (L. Pacifico, F. Fotouhi and R.L.F., unpublished work) available at \[[@B47]\]. To determine *Drosophila*interlogs of yeast or worm interactions, a list of *Drosophila*proteins belonging to eukaryotic clusters of orthologous groups (KOGs) \[[@B53]\] was obtained from the National Center for Biotechnology Information (NCBI) \[[@B54]\]. Each fly protein was assigned one or more KOG IDs, based on the cluster(s) to which it belongs. A list of interactions among yeast (*S. cerevisiae*) proteins, derived mostly from high-throughput yeast two-hybrid screens \[[@B4],[@B55]\] and from the determination of proteins in precipitated complexes \[[@B56],[@B57]\], was obtained from the Comprehensive Yeast Genome Database \[[@B58],[@B59]\].
For the interactions determined by precipitation of complexes, two lists were generated. One list includes the binary interactions between the bait protein and every protein that was co-precipitated, but not between the precipitated proteins (hub and spoke model). The second list included all possible binary interactions among the members of a complex (matrix model). The lists were each used to generate a list of interactions between KOG pairs, which in turn was used to generate a list of potential interactions between pairs of *Drosophila*proteins belonging to those KOGs. Similarly, *Drosophila*-worm (*C. elegans*) interlogs were determined using the list of interactions between worm proteins determined by high-throughput yeast two-hybrid screening \[[@B5]\]. *Drosophila*genetic interactions were obtained from Flybase \[[@B27],[@B60]\]. To compare the two-hybrid data with other datasets we generated random interaction maps having the same BD proteins, total interactions and topological properties as the LexA or Gal4 data. The AD clones in each interaction list were indexed, an array of the same number of genes as the AD clones was randomly fetched from the *Drosophila*Release 3.1 genome \[[@B61]\] and these genes were used to replace the original AD clones at the same indexed positions.
Fifty thousand such random networks were generated for each two-hybrid dataset, and then compared with the yeast interlogs, worm interlogs, and genetic interactions to determine the amount of overlap expected by chance. *P*values represented the number of times that the observed number of overlapping interactions was detected in 50,000 iterations of a random network, divided by 50,000. In most cases *P*\< 0.0002 (see Additional data file 6). Additional methods are in Additional data file 1. To compare the number of common interactions between the LexA and Gal4 maps with the number expected by chance, we generated 10^6^random LexA maps and found that they never contained more than two interactions in common with the Gal4 map; thus, the *P*-value for the 28 common interactions is significantly less than 10^-6^.
Additional data files
=====================
The following additional data are available with the online version of this paper. Additional data file [1](#s1){ref-type="supplementary-material"} contains Supplementary materials and methods; Additional data file [2](#s2){ref-type="supplementary-material"} contains Supplementary Table 1, BD \'baits\' used in the LexA screens; Additional data file [3](#s3){ref-type="supplementary-material"} contains Supplementary Table 2, Interactions detected in the LexA screens; Additional data file [4](#s4){ref-type="supplementary-material"} contains Supplementary Table 3, Enrichment of Gene Ontology classes, complete list; Additional data file [5](#s5){ref-type="supplementary-material"} contains Supplementary Table 4, Enrichment of Domain pairs, complete list; Additional data file [6](#s6){ref-type="supplementary-material"} contains Supplementary Table 5, *P*-values for overlap among datasets, and Supplementary Table 6, Interactions from the LexA and Gal4 screens that successfully used the same BD bait proteins; Additional data file [7](#s7){ref-type="supplementary-material"} is a PDF containing Supplementary Figure 1, Interaction maps of other clusters; Additional data file [8](#s8){ref-type="supplementary-material"} is a PDF containing Supplementary Figure 2, Proteins clustered by interaction profile; Additional data file [9](#s9){ref-type="supplementary-material"} contains the legends to Supplementary Figures 1 and 2.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
Supplementary Materials and methods
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
Supplementary Table 1: BD \'baits\' used in the LexA screens
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 3
Supplementary Table 2: Interactions detected in the LexA screens
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 4
Supplementary Table 3: Enrichment of Gene Ontology classes (the complete list)
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 5
Supplementary Table 4: Enrichment of Domain pairs (the complete list)
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 6
Supplementary Table 5: *P*-values for overlap among datasetsa nd Supplementary Table 6: Interactions from the LexA and Gal4 screens that successfully used the same BD bait proteins
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 7
Supplementary Figure 1: Interaction maps of other clusters
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 8
Supplementary Figure 2: Proteins clustered by interaction profile
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 9
The legends to Supplementary Figures 1 and 2
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank Ari Firestone for developing the array of AD pools and members of the Finley laboratory for helpful discussions and technical assistance. We also thank Kyle Gardenour and Jodi Parrish for critical comments on the manuscript. We are particularly grateful to Mike McKenna for his help in the initial stages of this project. This work was supported by NIH grants HG01536 and GM62403.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
A protein interaction map centered on cell cycle regulators. **(a)**The entire map includes 1,814 unique interactions (lines) among the proteins encoded by 488 genes (circles). The map has five distinct networks; one network contains 479 (98%) of the proteins, one has three proteins, and three have two proteins (upper right, green circles). **(b)**The interconnectedness of the map does not depend strongly on the proteins with the most interactions. The map shown comprises data filtered to remove proteins with more than 30 interactions (*k*\> 30), leaving 792 interactions among 343 proteins. This produced only one additional network, which has two proteins (green circles on the left of (b)); 97% of the proteins still belong to a single large network. Further deletion of proteins with *k*\> 20 removes an additional 469 interactions, which creates only four additional small networks and leaves 85% of the proteins in a single network (data not shown). A high-resolution version of this figure with live links to gene information can be drawn using a program available at \[47\].
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
A map of the interactions involving cyclin-dependent kinases (Cdks). All the interactions involving at least one of the six Cdks (Cdc2, Cdc2c, Cdk4, Cdk5, Cdk7) and Eip63E (red nodes) are shown. All the Cdks except Cdk7 interacted with at least two cyclins (red text). All the cyclins interacted with at least two Cdks, with the exception of the novel cyclin-like protein CG14939, which only interacted with Eip63E. Other known or paralogous interactions include, Cdc2c-dap, Cdc2-twe, and the interactions of Cdc2 and Cdc2c with CG9790, a Cks1-like protein. Proteins are depicted according to whether they appear in the map only as BD fusions (squares), only as AD fusions (circles), or as both BD and AD fusions (triangles). Proteins connected to more than one Cdk are green. Interactions are colored if they involve proteins contacting two Cdks (red), three Cdks (blue), or five Cdks (green).
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Proteins clustered by their interaction profiles. BD fused proteins (*y*-axis) and AD fused proteins (*x*-axis) were independently clustered according to the similarities of their interaction profiles using a hierarchical clustering algorithm (see Materials and methods). An interaction between a BD and AD protein is indicated by a small colored square. The squares are colored according to the level of two-hybrid reporter activity, which is the sum of LEU2 (0-3) and lacZ (0-5) scores, where higher scores indicate more reporter activity (1, yellow; 5+, red). The cluster circled in blue (center) corresponds to interactions involving SkpA, SkpB and SkpC BD fusions, which are mapped in Figure 4. Maps of other clusters (circled in green) are shown in Additional data file 7. The large cluster at upper left is due primarily to AD proteins that interact with many different BD proteins. A larger version of the figure with the gene names indicated in the axes is in Additional data file 8.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
A map of the interactions in the Skp cluster. All the interactions with the BD fusions SkpA, SkpB and SkpC, are shown. Proteins (green) interacting with more that one Skp paralog are enriched for proteins possessing an F-box domain (red text). Other colors and shapes are as in Figure 2.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Comparison of *Drosophila*protein-interaction maps generated by high-throughput yeast two-hybrid methods
:::
LexA cell-cycle map\* Gal4 proteome-wide map^†^ Common
------------------------ ----------------------- --------------------------- --------
Interactions 1,814 20,439 28
Proteins 488 6,951 347
Proteins as BD fusions 106 3,616 46
Proteins as AD fusions 403 5,425 250
Proteins as AD and BD 21 2,090 8
Degree exponent^‡^ 1.72 1.91 NA
Mean path length^§^ 3.3 4.1 NA
\*The LexA interactions are from this study, listed in Additional data file 3. ^†^The Gal4 interactions are from Giot *et al*. \[6\]. The chance of observing more than two common interactions between the Gal4 map and a random network with the same topological properties as the LexA map is \< 10^-6^(see Materials and methods). ^‡^The degree exponent and mean path length are topological properties of the networks. The degree exponent is γ in the equation P(*k*) = *k*^-γ^, where *k*is the degree or number of interactions per protein, and P(*k*) is the distribution of proteins with *k*interactions. ^§^The mean path length is the shortest number of links between a pair of proteins, averaged over all pairs in the network.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Enrichment of the most frequently classified gene functions
:::
Description BD genes AD Genes Same-pair interactions
-------------------------------------------------- ---------- ---------- ------------------------ ---------- ------- ---------- --------------- --------- -------- ---------- ------------- ----------
Protein modification 30 2.92 \<0.00002 10.3 21 11.12 0.00210 1.9 25 14.86 0.09916 1.7
**Cell cycle** **22** **1.27** **\<0.00002** **17.3** 19 **4.83** **\<0.00002** **3.9** **26** **4.40** **0.00006** **5.9**
**DNA metabolism** **14** **0.79** **\<0.00002** **17.7** **6** **2.99** **0.03006** **2.0** **6** **1.15** **0.00860** **5.2**
Transcription 9 2.04 0.00002 4.4 14 7.77 0.01134 1.8 7 1.85 0.00242 3.8
Gametogenesis 9 1.49 \<0.00002 6.0 13 5.69 0.00172 2.3 7 1.53 0.00072 4.6
Neurogenesis 8 1.91 0.00018 4.2 12 7.29 0.03142 1.6 14 3.75 0.00168 3.7
Cell-surface receptor-linked signal transduction 8 2.48 0.00088 3.2 11 9.39 0.23272 1.2 5 3.05 0.12498 1.6
**DNA repair** **6** **0.45** **\<0.00002** **13.4** **7** **1.71** **0.00030** **4.1** **3** **0.28** **0.00064** **10.8**
Intracellular signaling cascade 6 0.65 0.00002 9.3 6 2.44 0.01036 2.5 3 0.98 0.03602 3.1
Imaginal disk development 5 0.80 0.00022 6.3 9 3.04 0.00092 3.0 3 0.45 0.00266 6.7
Average 11.7 1.48 0.00022 9.2 11.8 5.63 0.03209 2.4 9.9 3.23 0.02769 4.71
The top 10 most frequently classified BD gene functions, derived from GO biological process level 4 (see Materials and methods), are shown. The number of proteins or pairs of proteins in our experimental data (Exp) with each GO function is shown, alongside the average number of times the function would appear in a random interaction map (Rand) having the same topology and number of proteins (see Materials and methods), and the ratio of Exp/Rand. The functions listed are significantly enriched in the BD list, to *P*\< 0.001, and most to *P*\< 0.0003. Cell cycle, DNA metabolism and DNA repair (highlighted) are the three most proportionally enriched classifications in the BD list, These classes are also enriched for self-associations in the interaction list, with cell cycle and DNA metabolism around six- and fivefold enriched, while DNA repair is approximately 11-fold more self-associated than expected by chance. Of these three, DNA metabolism is not significantly enriched in the AD gene list (*P*\> 0.03), while the other two classifications are approximately fourfold enriched. A complete list of all functions and function pairs found in the interaction data is in Additional data file 4.
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Overlap of two-hybrid data with datasets enriched for true positives
:::
Interactions Overlap with LexA map (*N*= 1,814) Overlap with Gal4 map (*N*= 20,439) Overlap in common
------------------------------- -------------- ------------------------------------ ------------------------------------- -------------------
Yeast interlogs (hub/spoke)\* 67,238 23 (1.26%) 251 (1.23%) 4
Yeast interlogs (matrix)\* 244,202 25 (1.38%) 294 (1.44%) 4
Worm interlogs\* 37,863 3 (0.17%) 61 (0.30%) 0
Drosophila genetic^†^ 2,751 4 (0.22%) 22 (0.11%) 1
Reference set^‡^ 47 8 (0.44%) 6 (.03%) 2
Ref set (common BD)^§^ 20 3 (0.17%) 2 (.01%) 0
\*Yeast (*S. cerevisiae*) and worm (*C. elegans*) interlogs are predicted interactions between the *Drosophila*orthologs of interacting yeast and worm proteins; \'hub/spoke\' and \'matrix\' refer to the methods used to derive predicted binary interactions from the protein complex data (see Materials and methods). ^†^Genetic interactions were obtained from Flybase \[27\]. ^‡^The Reference set includes published interactions involving any of the 106 BD proteins in the LexA data. ^§^The subset of reference interactions involving proteins successfully used as BDs in both the Gal4 and LexA screens is also shown; no interactions from the reference set were found in both the LexA and Gal4 screens using the same BD baits. The chance of finding the indicated number of overlapping interactions with a random set of interactions was \<10^-4^for all but the LexA overlaps with worm interlogs (*P*\< 0.1436) or genetic interactions (*P*\< 0.0024) (Additional data file 6).
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Domain pair enrichment
:::
AD domain BD domain Domain pairings
------------- ----------- ----------------- ---- ----------- ---------------- ---- ----- ---- ----------- ---- ------ ----- -----------
Cyclin 8 0.5 16 \<0.00002 Protein kinase 30 1.7 18 \<0.00002 38 0.6 60 \<0.00002
F-box 17 1.2 15 \<0.00002 Skp1 4 0.1 75 \<0.00002 34 0.3 123 \<0.00002
F-box 17 1.2 15 \<0.00002 Skp1\_POZ 4 0.1 65 \<0.00002 34 0.3 123 \<0.00002
Homeobox 9 2.9 3 0.00080 Protein kinase 30 1.7 18 \<0.00002 33 3.7 9 0.00002
Extensin\_2 20 11.0 2 0.00316 Protein kinase 30 1.7 18 \<0.00002 33 14.0 2 0.01536
Cyclin\_C 4 0.3 15 \<0.00002 Protein kinase 30 1.7 18 \<0.00002 26 0.3 76 \<0.00002
Drf\_FH1 11 4.3 3 0.00128 Protein kinase 30 1.7 18 \<0.00002 19 5.5 3 0.01278
Cyclin 8 0.5 16 \<0.00002 RIO1 11 0.3 39 \<0.00002 19 0.3 59 \<0.00002
Rrm 12 4.3 3 0.00032 Protein kinase 30 1.7 18 \<0.00002 18 5.5 3 0.01692
The top 10 domain pairs observed in the interaction list are shown. As expected from interaction profiles (see text), cyclin and protein kinase domains are significantly associated, as are F-box and Skp domains. RIO1 is a recently described kinase domain \[62\] while the Extensin\_2 domain is a proline-rich sequence. Drf\_FH1 is the Diaphanous-related formin domain, a low-complexity 12-residue repeat found in proteins involved with cytoskeletal dynamics and the Rho-family GTPases \[63\], and the Rrm is an RNA-recognition motif. There are also additional associations between protein kinase domains and nucleic acid metabolism domains (see Additional data file 5). These data demonstrate the capacity of relatively small sets of proteins to generate high-confidence domain associations. A complete list of all domains and domain pairs found in the interaction data is in Additional data file 5.
:::
|
PubMed Central
|
2024-06-05T03:55:51.894823
|
2004-11-26
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545799/",
"journal": "Genome Biol. 2004 Nov 26; 5(12):R96",
"authors": [
{
"first": "Clement A",
"last": "Stanyon"
},
{
"first": "Guozhen",
"last": "Liu"
},
{
"first": "Bernardo A",
"last": "Mangiola"
},
{
"first": "Nishi",
"last": "Patel"
},
{
"first": "Loic",
"last": "Giot"
},
{
"first": "Bing",
"last": "Kuang"
},
{
"first": "Huamei",
"last": "Zhang"
},
{
"first": "Jinhui",
"last": "Zhong"
},
{
"first": "Russell L",
"last": "Finley"
}
]
}
|
PMC545800
|
Background
==========
For an increasing number of otherwise uncharacterized protein sequences from genome-sequencing projects, function assignment is attempted solely with *in silico*prediction methods, as reliable and cost-effective large-scale experimental methods are not available. In addition to sequence homology and annotation transfer considerations \[[@B1]\], these function assignments increasingly rely on algorithms that recognize protein-sequence features responsible for posttranslational modifications, subcellular localization and interactions with specific domains of other proteins.
Although considerable effort has been invested in achieving low false-positive prediction rates, our experience with tools for recognizing glycosyl phosphatidylinositol (GPI) lipid \[[@B2],[@B3]\] and myristoyl \[[@B4]-[@B6]\] anchor attachment sites and for predicting potential targets for PTS1-dependent translocation to peroxisomes \[[@B7]\] shows that a small but noticeable number of proteins without appropriate biological context (for example with contradictory subcellular localization or in taxa without the modifying enzyme or receptor) are systematically hit by these tools. For example, we found more than a dozen metazoan lysozymes \[[@B7],[@B8]\], known extracellular proteins, that are predicted to have carboxyl termini with a functional peroxisomal targeting signal 1 (PTS1) region.
Are these false-positive predictions? All three of the sequence-analysis tools mentioned above check query sequences for a recognition pattern that is explicitly described in terms of its physical properties and it is possible to check the concordance between pattern descriptions and query sequence individually. Nevertheless, this visual inspection is frequently unable to rationalize the findings as false-positive predictions, as all known components of the pattern appear to be present. Even in the case of high accuracy of the prediction tool, an erroneous prediction cannot be excluded. Alternatively, these predicted sequence motifs may occur by chance and be functional in an appropriate test system, but still have no biological meaning because the necessary cellular context is absent *in vivo*. Only experimental tests can resolve this contradiction. As a case study, we report the results of an experimental analysis that demonstrates the existence of naturally occurring peroxisomal targeting signals in several known non-peroxisomal proteins. We also discuss the evolutionary perspective of functional localization signals in unrelated proteins as well as the consequences for experimental localization determination and function prediction from sequence.
The major mechanism for targeting proteins to the matrix of peroxisomes, which are membrane-bounded organelles \[[@B9]\] of eukaryotic cells, is initiated in the cytoplasm by interaction of the receptor protein peroxin 5 (PEX5) with the carboxy-terminal signal PTS1 on the target protein \[[@B10],[@B11]\]. This signal consists of three regions of sequence comprising approximately 12 residues \[[@B12],[@B13]\]. It is composed of the most carboxy-terminal tripeptide (classically, the -SKL terminus), preceded by a region of around four residues (which interact with the surface at the mouth of the PEX5 binding cavity), and a solvent-accessible (or easily unfoldable) stretch of around five residues further upstream. The PTS1-prediction program \'PTS1\' \[[@B14]\] identifies PTS1 signals in query protein sequences by evaluating their carboxy-terminal ends with respect to features necessary for interaction with the tetratricopeptide repeats of PEX5. The predictor\'s scoring function searching for this motif within the 12 carboxy-terminal residues achieves an estimated sensitivity of 90% and a selectivity above 99% \[[@B7]\].
Results
=======
The carboxyl termini of several non-peroxisomal proteins interact with PEX5
---------------------------------------------------------------------------
Screening of SWISS-PROT \[[@B15]\] entries with the PTS1 predictor identified proteins from several families that are clearly not peroxisomal but score highly and are predicted as PEX5 targets \[[@B7],[@B8]\]. We were not able to rationalize these results as false predictions as the proteins\' carboxyl termini did not deviate from the generalized PTS1 sequence pattern \[[@B13]\]. To verify whether these proteins could indeed interact with PEX5, we tested the carboxyl termini of seven representative proteins in a yeast two-hybrid system: hen egg-white lysozyme (P00698, secreted); dog lysozyme C from milk (P81708); tyrosinase from human (P14679, a melanosomal type I membrane protein); frog tyrosinase (Q04604); *Drosophila*sevenless (P13368, a large transmembrane protein required for photoreceptor development); precursor of lysosomal bovine cathepsin D (P80209); and a mitochondrial ribosomal protein from yeast (P12687). We also examined the carboxyl terminus of a mouse dihydrofolate reductase construct with an added SKL peptide, which has been shown not to be imported into yeast peroxisomes \[[@B16],[@B17]\].
Depending on their taxonomic origin, the carboxyl termini of the eukaryotic sequences were assayed for interaction with the tetratricopeptide repeat domains of either human or yeast PEX5 using published methodologies \[[@B12]\]. The query sequences, along with prediction scores and measured β-galactosidase activities, are summarized in Table [1](#T1){ref-type="table"}. The results show that all peptide sequences interact with the PTS1-receptor PEX5 in the two-hybrid system. Hence, the carboxy-terminal sequences of these assayed non-peroxisomal proteins fulfill the requirements to function as PTS1 signals.
The accessibility of the PTS1-like carboxyl terminus is critical
----------------------------------------------------------------
The fact that the peroxisomal translocation machinery fails to import naturally occurring mature proteins carrying PTS1 signals into peroxisomes *in vivo*could be explained by the non-accessibility of their carboxyl termini. These could either be hidden in the native structure of the mature protein or of its functional complexes, or competing translocation machineries could lead to a removal of the respective proteins from the cytosol before their recognition by PEX5.
The first possibility is exemplified by DHFR-SKL. The carboxy-terminal 16 residues of the DHFR-SKL construct (EKGIKYKFEVYEK**SKL**, sequences appended to DHFR are in bold type, see results in Table [1](#T1){ref-type="table"}) interact with yeast PEX5 in the two-hybrid test but *in vivo*the complete construct is not imported into peroxisomes, thus confirming the prediction \[[@B16],[@B17]\]. For comparison, it should be noted that two other DHFR-derived constructs with slightly longer carboxyl termini (IKYKFEVYEK**GGKSKL**and IKYKFEVYEK**KNIESKL**) are predicted to be peroxisomally targeted. Their scores calculated with the PTS1 predictor \[[@B7]\] are 13.2 and 9.9, respectively (compare with data in Table [1](#T1){ref-type="table"}). They were experimentally shown \[[@B17]\] to be translocated to peroxisomes. In the native three-dimensional structure of DHFR \[[@B18]\], the carboxyl terminus is part of a β-sheet that is buried in the fold, deprived of flexibility and accessibility. Seemingly, this structure prevents the carboxy-terminal appended residues SKL in the construct from entering the PEX5 binding cavity, whereas slightly longer carboxyl termini may do. In our two-hybrid test system, the carboxy-terminal 16-mers are always considered exposed as, in the non-native sequence environment of the carboxyl terminus of the GAL4 activation domain, they are free from interfering or blocking structural features. Thus, DHFR-SKL fails to be imported into peroxisomes because its carboxyl terminus is sequestered in the structure of the mature protein.
Competing targeting signals prevent translocation into peroxisomes despite the presence of PTS1-like carboxyl termini
---------------------------------------------------------------------------------------------------------------------
Alternatively, functional PTS1 signals can be overruled by other localization signals \[[@B7]\]. For instance, distribution of the mammalian alanine-glyoxylate amino transferase (AGT) between peroxisomes and mitochondria is regulated by the variable occurrence of an amino-terminal mitochondrial targeting signal in the mature protein (depending on the usage of two alternative transcription initiation sites) \[[@B19],[@B20]\].
Does a naturally occurring PTS1-like carboxyl terminus of a clearly non-peroxisomal protein that is capable of interacting with PEX5 indeed lead to *in vivo*import of the respective protein, provided that a potentially overruling sequence signal is eliminated? A set of three target proteins with amino-terminal leader sequences was chosen from Table [1](#T1){ref-type="table"}. Chicken lysozyme (SWISS-PROT id P00698), a secreted enzyme, is one of the best characterized proteins and has an apparently accessible carboxyl terminus as deduced from its three-dimensional structure (Protein Data Bank (PDB) number 1H6M \[[@B21]\]). The corresponding carboxy-terminal 16-mer produces moderate β-galactosidase activity in the yeast two-hybrid assay (most of the other proteins in Table [1](#T1){ref-type="table"} appear to interact even more strongly with PEX5). Human tyrosinase (P14679) is a melanosomal marker protein that functions in the formation of pigments such as melanins. Yeast 60S ribosomal protein L2 (P12687), or MRP7, is a component of the large subunit of the mitochondrial ribosome.
Green fluorescent protein (GFP) was appended to the amino terminus of each of the selected proteins. It can be assumed that translocation into the endoplasmic reticulum (ER) or mitochondria is disrupted by the resulting shift of the signal peptide from the amino terminus to the center of the protein. The resulting molecules are expected to be redirected into peroxisomes if their carboxyl termini can act as PTS1 signals. Targeting of the GFP-constructs *in vivo*was indeed confirmed by co-localization with a peroxisomal DsRed2-SKL construct in COS7 cells for the metazoan enzymes (Figure [1](#F1){ref-type="fig"}) and with DsRed-SKL in yeast cells for the *Saccharomyces cerevisiae*protein (Figure [2](#F2){ref-type="fig"}). Thus, the PTS1 signals at the carboxyl termini of the assayed proteins are normally suppressed by alternative amino-terminal targeting sequences. A similar mechanism can be inferred for other eukaryotic SWISS-PROT proteins listed in Table [1](#T1){ref-type="table"}, although steric carboxy-terminal accessibility or other factors might also play a role.
Functional PTS1 sequences can occur in organisms without peroxisomes
--------------------------------------------------------------------
The occurrence of silent PTS1s without a targeting role raises the question of whether such signals can also evolve in organisms that do not carry peroxisomes. To test this hypothesis, we extended Table [1](#T1){ref-type="table"} with a set of four predicted carboxyl termini from prokaryotic enzymes: *Escherichia coli*glutamate-1-semialdehyde 2,1-aminomutase (P23893), *E. coli*transaldolase A (P78258), *Methanopyrus kandleri*riboflavin synthase (NCBI-Refseq accession NP\_613646) and *Archaeoglobus fulgidus*2-nitropropane dioxygenase (NCBI-Refseq accession NP\_070998). Indeed, these proteins harbor carboxyl termini that qualify as PTS1 signals (lower part of table [1](#T1){ref-type="table"}). As confirmation, for the bacterial protein glutamate-1-semialdehyde 2,1-aminomutase (GSA) we used the same methodology for subcellular localization determination as for yeast MRP7. The resulting GFP-GSA construct is also imported into peroxisomes (Figure [2](#F2){ref-type="fig"}), demonstrating that its PTS1-like carboxyl terminus is functional in the mature protein.
Discussion
==========
In families of orthologous proteins, peroxisomal location and its targeting signal in the amino-acid sequence are not necessarily conserved. For example, in plants the five enzymes of the glyoxylate cycle are localized to peroxisomes, but in *S. cerevisiae*three of the five (aconitase, isocitrate lyase, and the respective malate dehydrogenase isoform) could not be found in peroxisomes \[[@B22]\]. Thus, it is not surprising to find sporadically occurring PTS1 signals in protein families (see some examples in Table [1](#T1){ref-type="table"}).
In dually localized proteins such as AGT \[[@B23]\], the PTS1 signal has a biological role as a targeting signal. However, the carboxyl termini of the proteins from Table [1](#T1){ref-type="table"} do not seem to fulfill any specific targeting function. We suggest that these PTS1 signals occur as a result of neutral mutation. The presence of a functional PTS1 signal would not lead to evolutionary pressure in this context because mislocalization is prevented by overriding the function of these sequences either by alternative exposure of amino-terminal signals or by steric carboxy-terminal inaccessibility.
The case of lysozyme is particularly noteworthy because a large number of homologous proteins were systematically hit when performing a SWISS-PROT screen using the prediction tool (30 cases with putative PTS1s and 46 other lysozyme carboxyl termini are shown in Figure [3](#F3){ref-type="fig"}). Because of the close relationship of the originating species and the occurrence of several isozymes, the lysozyme sequences in the multiple alignment share a high degree of similarity. The PTS1 carboxyl termini seem to be a mimicry of the sequence needed to support structural features of the protein. The cysteine at the antepenultimate position, which is present as part of a disulfide bridge \[[@B21]\] in the final secreted form of lysozyme, happens to fulfill the need for a small residue at the respective PTS1 location. The PTS1 is mostly functional, with a positively charged or amidic penultimate amino acid and the correct hydrophobic carboxy-terminal residue, which is the case for a large proportion of the lysozymes. Note that the disulfide bridge will not be formed in our GFP-lysozyme test case because translocation of the fusion protein into the endoplasmic reticulum is prevented.
We conclude that a PEX5-interacting sequence can evolve simply by mutational alterations in the carboxy-terminal region of a protein. Although shuffling of a carboxy-terminal exon cannot be excluded for other examples, the fact that the open reading frames (ORFs) of the carboxy-terminal exons for human tyrosinase (GenBank accession AP000720.4), fly sevenless (GenBank accession AE003484.2) and chicken lysozyme (GenBank accession AF410481.1) reach far into the functional domains of their proteins, rather supports an evolutionary mechanism of several point substitutions. The occurrence of functional PTS1 sequences in non-eukaryotic species further supports a stochastic model for the evolution of PEX5-interacting protein carboxyl termini.
In non-globular regions of proteins, sequences that code for targeting to other subcellular compartments, or for posttranslational modifications, might appear in similar ways during evolution. For example, the sequence motif coding for amino-terminal *N*-myristoylation of glycines behaves as an exchangeable functional module, as protein families do exist where it has been substituted by alternative sequence determinants that facilitate membrane association \[[@B6]\]. This is exemplified by the *Arabidopsis thaliana*Rab5 ortholog Ara7 and its paralog Ara6. Ara7 is geranylgeranylated on carboxy-terminal cysteines just as Rab5 is in other species. However, the closely related paralog Ara6 lacks the carboxy-terminal cysteines and has an experimentally verified amino-terminal myristoylation motif \[[@B24]\].
Many of these signals seem to remain silent under normal physiological conditions (as is the case for the PTS1 signal in some metazoan lysozymes) but have the potential to become important in some future evolutionary scenarios or in pathological situations. Alternatively, the PTS1 signal might have become obsolete and the corresponding sequence segment is now subject to evolutionary alterations. Apparently, the cell exploits only a fraction of the potential molecular capabilities of its proteins.
Futhermore, subcellular targeting is organized in a hierarchy of cellular recognition mechanisms. The co-translational sorting into the ER serves as a first decision node. Posttranslational processes such as interaction with chaperones, folding, and covalent modifications are concomitant with the appropriate exposure of targeting signals. The amino-terminal signals are made first and are therefore favored when it comes to recognition by receptors. PEX5 needs only to categorize the remaining unsorted proteins with accessible carboxyl termini into \'stay here\' or \'let\'s go into peroxisomes\'. This might also explain why the PTS1 signal is comparatively short and permissive for a wide range of residues.
Clearly, the fact that functional sequences for subcellular targeting occur in unrelated proteins needs to be considered for prediction-tool development. The construction of a negative learning set (sequences without the specific localization signal) on the basis of proteins with differing cellular localization is problematic. For example, a set of non-peroxisomal but organellar localized \[[@B25]\], viral \[[@B26]\] or bacterial sequences might contain a considerable number of proteins that potentially interact with PEX5. Thus, such a set does not directly qualify for automated learning procedures or the assessment of false-positive prediction \[[@B27],[@B28]\].
Surprisingly, when Maurer-Stroh and Eisenhaber applied their myristoylation site predictor for eukaryotic proteins to bacterial proteomes \[[@B5]\], systematic hits were found despite the absence of known amino-terminal *N*-myristoyltransferases (NMT) in bacteria. Are these false-positive predictions? A literature search revealed that myristoylation by host NMTs has physiological relevance for several secreted proteins of intracellular bacterial parasites \[[@B5]\]. Thus, the sequence motif coding for amino-terminal *N*-myristoylation is typical for eukaryotes but occurs also in bacteria. In many cases, it remains without phenotypic effect for bacteria but may become evolutionarily important in the case of host-parasite interactions.
In the case of the endothelin-converting enzyme 1 and the neprilysin-like zinc metallopeptidase family, the carboxy-terminal CXAW motif is a valid prenylation motif. This carboxy-terminus is functionally hidden because the protein is exported to the extracellular side of the cytomembrane and the carboxy-terminal residues are apparently involved in folding and enzyme function \[[@B29]\].
Clearly, the accessibility of the recognition motif in the substrate protein to the respective receptor or protein-modifying enzyme is a major issue. For PTS1 signal prediction from the amino-acid sequence, carboxy-terminal exposure needs to be assessed both from the steric point of view as well as in the context of competing translocation mechanisms. Analyzing only the carboxy-terminal dodecamer peptide \[[@B7],[@B13]\] might not suffice for reliable prediction of accessibility to the receptor, but a full solution would require sufficiently accurate three-dimensional structure prediction.
In databases, it should also be routine to flag proteins that contain several competing targeting signals with differing priority. Finally, silent localization signals might become active in mutant protein constructs and lead to non-native localizations, an issue that needs to be assessed especially in localization screens of proteins with uniformly incorporated fluorescent dyes such as GFP. It cannot be excluded that the subcellular location of a considerable number of proteins has not been correctly determined in published large-scale studies that rely on this methodology \[[@B30],[@B31]\].
To conclude, sequence segments coding for subcellular targeting or for posttranslational modifications can occur in proteins that are not substrates in either of these processes. Accurate prediction techniques reveal candidate proteins carrying hidden sequence signals. Many of these can be experimentally confirmed. In the case of the PTS1 predictor program, there is no reasonable argument to assume a difference in prediction accuracies for real and hidden PTS1s as, in both cases, productive interaction of the carboxyl terminus with PEX5 is the criterion for a functional PTS1.
Materials and methods
=====================
Cloning procedures
------------------
Oligonucleotides were purchased from MWG Biotech (Munich, Germany). The *E. coli*strain DH5α, Bethesda Research Laboratories) was used for all transformations and plasmid isolations. For the yeast two-hybrid-assay, the hybridized oligonucleotide pairs coded for the carboxy-terminal 16-mers of the selected proteins flanked by *Bam*HI (5\') and *Eco*RI (3\') restriction sites. Each oligonucleotide pair was introduced into a *Bam*HI-*Eco*RI-digested pGAD.GH fragment, generating plasmids containing the Gal4p activation domain in addition to the desired carboxy-terminal 16-mer extension (Gal4pAD-16mer). All pGAD.GH constructs were sequenced (VBC Genomics, Vienna, Austria). The plasmids pAH987 and hP87 contain the binding domain of Gal4p fused to the TPR domain of *S. cerevisiae*or *Homo sapiens*PEX5, respectively (Gal4pBD-TPR) \[[@B12]\].
Chicken cDNA for the amplification of lysozyme was generated from chicken oviduct using Tripure (Invitrogen) according to the manufacturer\'s instructions. Reverse transcription was performed using RNA-PCR Core Kit (Applied Biosystems) following the manufacturer\'s instructions. For the amplification of tyrosinase, we used cDNA from the melanoma cell line 29 WUBI (generous gift of Walter Berger, Vienna). The coding regions of lysozyme and tyrosinase were gained by PCR (for oligonucleotide primers see Table [2](#T2){ref-type="table"}) using the Advantage cDNA Polymerase Mix kit from Clontech and the GeneAmp PCR-system from Perkin Elmer. The PCR-fragments were cloned into the pCR2.1 vector (Invitrogen) by T/A cloning and sequenced as control (VBC Genomics). The fragments containing the lysozyme or tyrosinase coding regions were excised with *Eco*RI/*Bam*HI and ligated into pEGFP-C1 (Clontech). The DsRed2-SKL construct was obtained by PCR using Pfu-polymerase (Promega) and the plasmid pDsRed2-C1 (Clontech) as template (for oligonucleotides, see Table [2](#T2){ref-type="table"}). The PCR fragment and the plasmid were both cut with *Eco*47-3/*Xho*I and the PCR fragment encoding the carboxy-terminal SKL was introduced to replace the original DsRed2 end sequence. The final plasmid encodes the DsRed2-SKL protein under the control of the cytomegalovirus promoter.
Standard procedures were used for cloning of the GFP-MRP7 and GFP-GSA constructs including control sequencing (VBC Genomics). The plasmids expressing GFP and GFP-SKL under control of the *MLS1*promoter were described previously \[[@B32]\]. The DNA fragment coding for DsRed-SKL was obtained by PCR (for oligonucleotides, see Table [2](#T2){ref-type="table"}; template pDsRed, Clontech) and cloned (*Bam*HI-and partially with *Pst*I) after the *MLS1*promoter in the vector YEplac181. DNA fragments coding for MRP7 and GSA were obtained by PCR (see Table [2](#T2){ref-type="table"} for oligonucleotide sequences) and cloned (*Bam*HI-*Sph*I) in-frame with GFP to give rise to the expression of GFP-MRP7 and GFP-GSA, respectively, all of them under the control of the *MLS1*promoter.
Yeast two-hybrid assay
----------------------
According to the Matchmaker two-hybrid protocol, yeast strain PCY3 (*MAT*α, *his3*Δ200, *ade2*-101, *trp1*Δ63, *leu2*, *gal4*Δ, *gal80*Δ, *lys2*::*GAL1-HIS3*, *ura3*::*GAL1-lacZ*) \[[@B12]\] was transformed with the Gal4pAD-16mer constructs (plasmid pGAD.GH) together with either pAH987 or hP87. Yeast transformants were selected and grown on minimal medium containing 2% glucose and supplemented with bases and amino acids as required (SC-leu-trp). For quantitative measurement of β-galactosidase activity in accordance with published techniques \[[@B12]\], yeast cells were grown in selective medium (SC-leu-trp) overnight at 30°C, diluted to *A*~600~= 0.3 into the same medium and finally harvested at absorptions of *A*~600~between 0.9 and 1.1.
*In vivo*localization study in COS7 cells
-----------------------------------------
COS7 cells were transfected with the pEGFP-C1-constructs and DsRed2-SKL by electroporation using 920 μF and 220 mV (Gene pulser II, Bio-Rad), grown on coverslips for 36 h, washed, fixed with 0.5% formaldehyde in PBS for 15 min and covered with geltol. Cells were analyzed using the Olympus BX51 fluorescence microscope (60 × enlargement).
*In vivo*localization study in yeast cells
------------------------------------------
The yeast strain used in this study is *S. cerevisiae CB80*(*MATa*, *ura3-52*, *leu2-1*, *trp1-63*, *his3-200*). Yeast transformants were selected and grown on minimum medium containing 0.67% yeast nitrogen bases without amino acids (Difco Laboratories), 2% glucose and amino acids (20-150 μg/ml) as required (SC-leu-ura). For fluorescence microscopy, yeast cells were grown at 30°C with shaking in selective media with 0.5% glucose as sole carbon source until the glucose concentration was very low (0.05%, usually 16 h), harvested by centrifugation and resuspended in the original volume of induction medium containing 0.67% yeast nitrogen bases without amino acids, 0.1% yeast extract, 30 mM potassium phosphate pH 6.0, 0.125% oleate, 0.2% Tween-80 and amino acids as required. Cells were grown for 16 h in induction medium and observed live for fluorescence. Briefly, cells were collected by centrifugation and washed twice in water. Cell pellets were resuspended in induction medium without oleate and aliquots were spotted onto multitest slides (ICN Biochemicals) previously coated with concanavalin A (6 mg/ml, Sigma). Cells were allowed to attach for 5 min at room temperature and the slides were washed twice with induction medium and a coverslip applied for observation. Fluorescence was viewed with a Zeiss Axioplan 2 fluorescence microscope using a 63 × (1.4 NA) lens. Digital images were captured with a Quantix CCD camera using Lightview software without further modification. The pictures were mounted and false-color overlays were made in Adobe Photoshop.
Acknowledgements
================
We wish to acknowledge the skilled technical assistance of Michael Schuster (Medical University, Vienna) and Peter Steinlein (Institute of Molecular Pathology, Vienna) as well as Sebastian Maurer-Stroh (Institute of Molecular Pathology, Vienna) for helpful literature suggestions. G.N. and F.E. are grateful for generous support from Boehringer Ingelheim. This research has been partially funded by the Austrian National Bank (P15037 to F.E.) and by the Fonds zur Förderung der Wissenschaftlichen Forschung Österreichs (P15037 to F.E., P15510 to J.B., P14956 to A.H.), by the Austrian Gen-AU BIN (to F.E.) and by the Austrian Ministry for Economics BMWA (to F.E.).
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Targeting of GFP-tyrosinase and GFP-lysozyme to peroxisomes in human cells. Fluorescence of human COS7 cells expressing **(a)**GFP-lysozyme or DsRed2-SKL; **(b)**GFP-tyrosinase and DsRed2-SKL; or **(c)**GFP-lysozyme and DsRed2-SKL. Cells were observed 36 h after transfection (magnification 60 ×). Separate small images of the GFP fluorescence (green) and DsRed2 fluorescence (red) are shown to the left of each main picture, in which the two fluorescent images are overlaid. Areas in which red and green fluorescence coincide show as yellow. (a) Control experiments reveal that expression of GFP-lysozyme is an adjunct to the cellular punctuate fluorescence pattern independently of the presence of DsRed2-SKL. The figures show a punctate fluorescence pattern for GFP fusions with (b) human tyrosinase and (c) chicken lysozyme. Both proteins co-localize with DsRed2-SKL in human peroxisomes as demonstrated by the fluorescence overlay. Owing to the evolutionary conservation of PEX5 within the metazoans \[7,13,33\], a chicken protein (lysozyme) can be assayed in a human cell line and the species barrier is not an issue in this study.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Targeting of GFP-MRP7 and GFP-GSA to peroxisomes in yeast cells. Fluorescence of CB80 yeast cells expressing **(a)**GFP and DsRed-SKL; **(b)**GFP-SKL and DsRed-SKL; **(c)**GFP-MRP7 and DsRed-SKL; or **(d)**GFP-GSA and DsRed-SKL. Transformed cells were cultured on oleate and observed live for fluorescence. Control experiments (a) show that GFP co-localizes with Ds-Red-SKL only when the sequence -SKL is appended at its extreme carboxyl terminus (b). The figures reveal a punctuate fluorescence pattern for GFP fused to the yeast mitochondrial ribosomal protein L2 encoded by *MRP7*(c) or to the bacterial enzyme glutamate-1-semialdehyde 2,1-aminomutase (GSA) (d). Both fusion proteins co-localize with DsRed-SKL in yeast peroxisomes. GFP fused to GSA without its carboxy-terminal -AKL gave rise to a diffuse (cytosolic) fluorescence pattern (data not shown).
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Multiple alignment of lysozyme carboxyl termini. A screen of the SWISS-PROT database \[15\] for proteins that harbour PTS1 signals produced a set of lyosozymes, well characterized secreted enzymes that are not usually found in peroxisomes. Rather than occurring sporadically, a large fraction of the known sequences from this family was obtained using the PTS1 prediction tool \[7\]. Moreover, these hits could not be rationalized as false positives as they did not deviate from the PTS1 sequence motif \[11-13\]. The multiple alignment shows intact vertebrate lysozyme carboxy-terminal 20-mers (with accession number and species name) retrieved from the SWISS-PROT database. From a total of 76 entries, 23 have predicted PTS1s (score \> 0; at the top, marked with \'+\'), seven are in the twilight zone (-10 \< score \< 0; in the middle, marked with \'\#\') and 46 are not predicted (score \< -10; at the bottom, marked with \'-\'). There appears to be an overlap between the PTS1 motif and sequence variability within the lysozyme family. For example, the absolutely conserved cysteine near the carboxyl terminus is needed for the formation of a disulfide bridge in the mature protein \[21\]. This cysteine also meets the requirement for a small residue at the antepenultimate position of the PTS1 sequence.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Results of the yeast-two hybrid interaction assays with PEX5
:::
Yeast PEX5 Human PEX5
---------------------------- ------------ ------------ ------------ ---- ------- ----- ---- ------------------ ---------------------------------------------
*Canis familiaris* P81708 \- \- \- 0.17 25 2 HCKGKDLSKYLASCNL Lysozyme
*Drosophila melanogaster* P13368 \- \- \- 6.70 29 11 PLKDKQLYANEGVSRL Sevenless protein
*Gallus gallus* P00698 \- \- \- 2.02 73 4 RCKGTDVQAWIRGCRL Lysozyme
*Rana nigromaculata* Q04604 \- \- \- 0.13 91 15 LLMEAEDYQATYQSNL Tyrosinase
*Homo sapiens* P14679 \- \- \- 4.01 242 10 LLMEKEDYHSLYQSHL Tyrosinase
*Bos taurus* P80209 \- \- \- 7.04 310 58 FDRDQNRVGLAEAARL Cathepsin D
*Saccharomyces cerevisiae* P12687 2.72 482 37 \- \- \- KVEVIARSRRAFLSKL Mitochondrial ribosomal protein L2, or MRP7
Synthetic construct DHFR-SKL 11.51 195 45 \- \- \- EKGIKYKFEVYEKSKL DHFR-SKL
*Escherichia coli* P23893 4.81 270 26 11.35 473 57 DINNTIDAARRVFAKL Glutamate-1-semialdehyde 2,1-aminomutase
*E. coli* P78258 -9.46 164 31 5.59 566 70 FAVDQRKLEDLLAAKL Transaldolase A
*Methanopyrus kandleri* NP\_613646 6.08 45 8 10.41 358 46 GMGRREGHPDVGPARL Riboflavin synthase
*Archaeoglobus fulgidus* NP\_070998 7.57 206 19 -1.36 0 NA EEVIRKIAEGLNKAKF 2-nitropropane dioxygenase
All eukaryotic target sequences (characterized by species, SWISS-PROT or NCBI-Refseq accession number, score from the PTS1 predictor \[7\], carboxy-terminal sequence and description) were tested for interaction with the tetratricopeptide (TPR) repeat domain of human PEX5, except for P12687 and DHFR-SKL where the corresponding TPR domains were derived from yeast PEX5. The prokaryotic proteins were assayed using PEX5 from both yeast and human. As the estimated length of the PTS1 signal is 12 carboxy-terminal residues \[13\], we chose the carboxy-terminal 16-mers to be sure that we have included the complete motif-carrying segment. \*A PTS1 prediction score above zero is considered predictive of a functional PTS1 signal; a score between -10 and 0 is considered a \'twilight zone\' prediction. It should be noted that the negative score for the DHFR-SKL carboxyl terminus in its context is generated by the PTS1 predictor \[7\] solely by terms that evaluate its potential accessibility for PEX5. ^†^A yeast-two hybrid assay is considered positive if the measured β-galactosidase activity is clearly greater than zero. Experience from previous test series suggests a lower limit of around 10 Miller Units per mg protein \[12\] for the detection of a productive interaction. The measured β-galactosidase activities (including standard deviations) range from weak (P81708, P13368) to strong (P80209, P12687).
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Oligonucleotides used for the amplification of the GFP-constructs
:::
Construct Forward primer Reverse primer
------------------------- -------------------------------------- -----------------------------------------------
EGFP-tyrosinase GAATTCAATGCTCCTGGCTGTTTTGTACTG GGATCCTTATAAATGGCTCTGATACAAGCTG
EGFP-lysozyme GAATTCCATGAGGTCTTTGCTAATCTTGGT GGATCCGGCAGCTCCTCACAGCCG
GFP-MRP7 CGGGATCCAATGTGGAATCCTATTTTACTAGATAC GGGCATGCTCAAAGCTTGCTCAAAAAAGCCCG
GFP-GSA CGGGATCCAATGAGGAAGTCTGAAAATCTTTACCAG GGGCATGCTCACAACTTCGCAAACACCCGACG
DsRed2-SKL (COS7 cells) CGGCTAGCGCTACCGGTCGCCACCATGGCC CGTCTCGAGTTATAATTTGGACAGGTGGTGGCGGCC
DsRed-SKL (yeast cells) AGATCTATGGTGAGGTCTTCCAAG CTGCAGTTATAATTTGGATAGGATCCCAAGGAACAGATGGTGGCG
:::
|
PubMed Central
|
2024-06-05T03:55:51.899929
|
2004-11-30
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545800/",
"journal": "Genome Biol. 2004 Nov 30; 5(12):R97",
"authors": [
{
"first": "Georg",
"last": "Neuberger"
},
{
"first": "Markus",
"last": "Kunze"
},
{
"first": "Frank",
"last": "Eisenhaber"
},
{
"first": "Johannes",
"last": "Berger"
},
{
"first": "Andreas",
"last": "Hartig"
},
{
"first": "Cecile",
"last": "Brocard"
}
]
}
|
PMC545801
|
Background
==========
Different types of genomic features have characteristic patterns of evolution that, when sequences from closely related organisms are available, can be exploited to annotate genomes \[[@B1]\]. Methods for comparative sequence analysis that exploit variation in rates and patterns of nucleotide evolution can identify coding exons \[[@B1],[@B2]\], noncoding sequences involved in the regulation of transcription \[[@B3],[@B4]\] and various types of RNAs \[[@B5]-[@B7]\]. While most of these methods have been developed for and applied to pairwise comparisons, sequence data are increasingly available for multiple closely related species \[[@B8]\]. It is therefore of considerable importance to develop sequence-analysis methods that optimally exploit evolutionary information, and to explore the dependence of these methods on the evolutionary relationships of the species in comparison.
Sequence-specific DNA-binding proteins involved in transcriptional regulation (transcription factors) play a central role in many biological processes. Despite extensive biochemical and molecular analysis, it remains exceedingly difficult to predict where on the genome a given factor will bind. Transcription factors bind to degenerate families of short (6-20 base-pairs (bp)) sequences that occur frequently in the genome, yet only a small fraction of these sequences are actually *bona fide*targets of the transcription factor \[[@B9]\]. A major challenge in understanding the regulation of transcription is to be able to distinguish real transcription factor binding sites (TFBSs) from sequences that simply match a factor\'s binding specificity. Because the evolutionary properties of TFBSs are expected to be different from their nonfunctional counterparts, comparative analyses hold great promise in helping to address this challenge.
In the past few years, several methods have been introduced to identify conserved (and presumably functional) TFBSs for a factor of known specificity (in contrast to the larger set of methods that use comparative data in motif discovery or to otherwise identify sequences likely to be involved in *cis*-regulation). Each of these methods explicitly or implicitly adopts one of several distinct definition of a conserved TFBS. These include a binding site in a reference genome that is perfectly or highly conserved \[[@B8],[@B10]-[@B12]\]; a binding site in a reference genome that lies in a highly conserved region \[[@B4]\]; or a position at which the binding model predicts a binding site in all species \[[@B13]-[@B18]\].
In a previous study we characterized the evolution of experimentally validated TFBSs in the *Saccharomyces cerevisiae*genome, finding that functional TFBSs evolve more slowly than flanking intergenic regions, and more strikingly, that there is considerable position-specific variation in evolutionary rates within TFBSs \[[@B19]\]. We further showed that evolutionary rate at each position is a function of the selectivity of the factor for bases at that position.
Our goal here is to incorporate these specific evolutionary properties of TFBSs into the search for conserved TFBSs. Or, more precisely, to develop a method that, given the specificity of a transcription factor, identifies conserved binding sites in multiple alignments by taking into account the sequence specificity and patterns of evolution expected for TFBSs, while still fully exploiting the phylogenetic relationships of the species being compared.
In addition to developing new methods, there are several hypotheses regarding the comparative annotation of TFBSs that we are interested in testing. It has been noted that the effectiveness of such analyses will depend critically on the evolutionary distance separating the species used. At very close distances TFBSs will appear conserved because there has been insufficient time for substitutions to occur. As distance increases, and substitutions occur most rapidly at nonfunctional positions, our ability to detect constrained binding sites should improve until we are no longer able to reliably assign orthology based on sequence alignment. To overcome this problem of divergence distances exceeding what can be aligned, the sequences of multiple closely related species can be used to span the same evolutionary distances (and presumably provide the same discriminatory power) as fewer more distantly related ones. However, aside from these qualitative expectations, the dependence of the ability to identify conserved TFBSs on evolutionary distance and tree topology has not been rigorously investigated. Because the software MONKEY can be applied to multiple alignments of varying numbers of species and produces scores that can be meaningfully compared across different sets of species, we are now able to address these issues.
Results
=======
Overview
--------
We developed an approach to identify conserved TFBSs that combines probabilistic models of binding-site specificity \[[@B20]-[@B22]\] with probabilistic models of evolution \[[@B23],[@B24]\]. Starting with an alignment of sequences from multiple related species, we use the known sequence specificity for a transcription factor to compare the likelihood of the sequences under two evolutionary models - one for background and one for TFBSs. The central feature of this method that underlies its ability to identify conserved TFBSs is that it uses a specific probabilistic evolutionary model for the binding sites of each transcription factor. The evolutionary model we use for TFBSs \[[@B25]\] assumes that sites were under selection to remain binding sites throughout the evolutionary history of the species being studied. This model uses the sequence specificity of the factor to predict patterns and rates of evolution that recapitulate the patterns and rates observed in real TFBSs \[[@B19]\].
MONKEY: scanning alignments to identify conserved transcription factor binding sites
------------------------------------------------------------------------------------
MONKEY, our tool for identifying conserved TFBSs, takes as input a multiple sequence alignment, a tree describing the relationship of the aligned species, a model of a transcription factor\'s binding specificity and a model for background noncoding DNA. It returns, for each position in the alignment, a likelihood ratio comparing the probability that the position is a conserved binding site for the selected factor compared to the probability that the position is background.
Extending matrix searches to multiple sequence alignments
---------------------------------------------------------
For the model of binding specificity, we use a traditional frequency matrix \[[@B20]-[@B22]\]. The values in the matrix - *f*~*ib*~- represent the probability of observing the base *b*(A, C, G or T) at the *i*th position in a binding site of width *w*. For the model of the background, we use a single set of base frequencies *g*~*b*~.
A widely used statistic for scoring the similarity of a single sequence to a frequency matrix is the log likelihood ratio comparing the probability of having observed a sequence *X*of width *w*under the motif model (a frequency matrix, designated as *motif*) to the probability of having observed *X*under the background model (designated by *bg*), which can be easily reduced to:
where *X*~*ib*~is an indicator variable which equals 1 if base *b*is observed at position *i*, and zero otherwise.
This classifier can be motivated by the approximation that the data are distributed as a two-component mixture of sequences matching the frequency matrix and sequences drawn from a uniform background. In practice, we compute this score using a position-specific scoring matrix (PSSM) with entries, *M*~*ib*~= log(*f*~*ib*~/*g*~*b*~), and find *S*for a particular *w*-mer by adding up the entries that correspond to the bases in the query sequence.
In extending this to a pair of aligned sequences *X*and *Y*, we want to perform the same calculation on their common ancestor *A*. Since *A*is not observed, we consider all possible ancestral sequences by summing over them, weighting each by their probability given the data (*X*and *Y*), the phylogenetic tree (*T*) that relates the sequences, and a probabilistic evolutionary model \[[@B23]\].
We can write a new score representing the log-likelihood ratio that compares the hypothesis that *X*and *Y*are a conserved example of the binding site represented by the frequency matrix to the hypothesis that they have been drawn from the background:
where *R*~*motif*~and *R*~*bg*~are rate matrices describing the substitution process of the binding site and background respectively. Using the conditional independence of the sequences *X*and *Y*on the ancestor, *A*, and writing *T*~*AX*~for the evolutionary distance separating sequence *X*from *A*, this becomes:
The class of evolutionary models used by MONKEY define a substitution matrix, *p*(*X*~*i*~\|*A*~*i*~, *t*) = *e*^*Rt*^, that represents the probability of observing each base at position *i*in the extant sequence (*X*) given each base in the ancestral sequence (*A*) after *t*units of evolutionary time or distance, given some rate matrix, *R*\[[@B23]\]. Since these models retain positional independence, we can rewrite this as:
This can be extended to more than two sequences, that is, (*X*, *Y*, \..., *Z*), by replacing the probabilities of *X*and *Y*with the probability with the left and right branches of the tree below, and performing the calculation at the root. The probabilities of the left and right branches of the tree can be calculated recursively as has been described previously \[[@B23]\].
Once again, for practical purposes we can convert these scores to a PSSM, whose entries are given for the pairwise case by:
where at each position we now index by the bases *a*and *b*in the two sequences. For multiple alignments of *n*species, each position requires 4^*n*^entries.
Evolutionary models
-------------------
The use of evolutionary models is critical to the function of MONKEY. Myriad of such models exist, and in principle all can be used in MONKEY. For the background, it is natural to use a model appropriate for sites with no particular constraint, such as the average intergenic or synonymous rates. MONKEY allows the use of the JC \[[@B26]\] or HKY \[[@B27]\] models, and here we use the latter with the base frequencies, rates and transition-transversion rate-ratio estimated from noncoding alignments assuming a single model of evolution over the noncoding regions (see details in Materials and methods). It is also possible to estimate the evolutionary model separately for each intergenic alignment, although the small size of yeast intergenic regions leads to variable estimates.
In principle, the JC and HKY models can also be used for the motif, with rates set according to our expectation of the overall rate of evolution in functional binding sites, which has been estimated as two to three times slower than the average intergenic rate \[[@B19]\]. However, we have previously shown that there is position-specific variation in evolutionary rates within functional transcription factor binding sites \[[@B19]\] and that positions in a motif with low degeneracy in the binding-site model evolve more slowly than positions with high degeneracy; this relationship between the equilibrium frequencies and the position-specific evolutionary rates is accurately predicted by an evolutionary model from Halpern and Bruno (HB model) \[[@B25]\].
In using this model, we assume that sequences evolve under constant purifying selection to maintain a particular set of equilibrium base frequencies. The use of this model corresponds to a definition of a conserved TFBS as a sequence position where there has always been a binding site for the transcription factor. Although the model does not strictly require that a binding site be present in each of the observed species, positions lacking such sites will have lower probabilities as they require the use of less probable substitutions. The rate of change from residue a to b at position *i*in the motif is given by:
where *Q*is the (position independent) underlying mutation matrix, which we set equal to the background model (*Q*= *R*~*bg*~), and *f*is the frequency matrix describing the specificity of the factor. Thus, for each position in the motif, the HB model predicts the rates of each type of substitution as a function of the frequency matrix, and the background model.
Comparing hits for different factors and evolutionary distances: computing the null distribution
------------------------------------------------------------------------------------------------
To compare scores from different evolutionary distances and different factors, it is critical that we are able to assign significance to a particular value of the score. To do so, we need to compute the distribution of the score under the null hypothesis that the sequence is part of the background. Calculating a *p*-value for a score *S*in a single sequence requires the enumeration of all possible *w*-mers that have a score *S*or greater under the background model. For *n*aligned sequences this requires the enumeration all 4^*wn*^possible sets of aligned *w*-mers with scores *S*or greater under the background model. While the number of possible alignments of *n w*-mers can be unmanageably large for even small values of *n*and *w*, because we treat each position independently we can enumerate these possibilities efficiently using an algorithm developed for matrix searches of single sequences \[[@B28],[@B29]\].
Every observed score is a sum of *w*numbers, one from each column of the matrix. The probability of observing exactly score *S*is the number of paths through the matrix whose entries add up to *S*, weighted by the probability of the path. By converting the matrix to integers, we can compute this probability for all values of *S*recursively. We initialize *P*~*i*~(*S*) (the probability of observing score *S*after *i*columns in the matrix) by setting *P*~*0*~(*S*) = 1 for *S*= 0, and *P*~*0*~(*S*) = 0 for *S*≠ 0. We then compute the values of the function for *i*= \[1, *w*\] as follows:
For aligned sequences, *c*represents a column in the alignment, and the sum is over all 4^*n*^possible columns an alignment of *n*sequences. The probability distribution function (PDF) of scores is *P*~*w*~(*S*), and from this the cumulative distribution function (CDF), the probability of observing a score of *S*or greater, can be directly computed. Although in principle we can compute the probabilities to arbitrary precision, because the time complexity increases with the number of possible scores, we limit the precision to within approximately 0.01 bits.
Figure [1](#F1){ref-type="fig"} compares empirical *p*-values from 5,000 pairs of sequences evolved in a simulation (see Materials and methods) with those computed by this method, and shows that they agree closely. We have used this method to compute the CDFs for alignments of up to six species, and therefore can apply our method to most comparative genomics applications. We note, in addition, that the likelihood ratio scores are approximately Gaussian (data not shown). As the means and variance of the scores under each model can be computed efficiently (see Materials and methods) we can estimate *p*-values using a Gaussian approximation (Figure [1](#F1){ref-type="fig"}) when the number of sequences in the alignment is large.
Heuristics for alignments with gaps
-----------------------------------
The treatment of alignment gaps in identifying conserved TFBSs is somewhat problematic. One the one hand, nonfunctional sequences may be inserted and deleted over evolution more rapidly than functional elements \[[@B30]-[@B32]\], and thus the presence of a gap aligned to a predicted binding site could indicate that it is nonfunctional. On the other hand, alignment algorithms are imperfect, and must often make arbitrary decisions about the placement of gaps. We sought to design a heuristic that accommodated both these aspects of genomic sequence data by locally optimizing alignments for the purpose of comparative annotation of regulatory elements.
The idea is to assign a poor score to regions of the alignment with a large number of gaps, but to locally realign regions with a small number of gaps to identify conserved but misaligned binding sites. To do this, we scan along the ungapped version of one of the aligned sequences - the \'reference\' sequence. For each position in the reference sequence *p*~*r*~, we define a window in each other sequence around *p*~*s*~, the position in sequence *s*aligned to position *p*~*r*~. The window runs from *p*~*s*~- (*a*+ *b*) to *p*~*s*~+ *w*+ (*a*+ *b*), where *a*and *b*are the number of gaps in the aligned versions of sequences *r*and s in position *p*to *p*+ *w*, where *p*is the position in the alignment of *p*~*r*~. For each subsequence of length *w*in the window, we calculate the percent identity to the reference sequence, and create an alignment of *p*~*r*~to *p*~*r*~+ *w*(in the reference sequence) to the most similar word in the window of each other sequence. This locally optimized alignment is then scored. Note that if *a*and *b*are zero (meaning there are no gaps in the aligned sequence), no optimization is done. If *a*is too large (in most contexts greater than five) we exclude that region of the alignment from further. This heuristic encapsulates the idea that too many gaps are indicative of lack of constraint, but conservatively allows for a few gaps due to alignment or sequence imperfections.
Application to *Saccharomyces*
------------------------------
The genome sequences of several species closely related to the budding yeast *Saccharomyces cerevisiae*have recently been published and become models for the comparative identification of transcription factor binding sites \[[@B8],[@B11]\]. We aligned the intergenic regions of *S. cerevisiae*genes to their orthologs in *S. paradoxus*, *S. mikatae*, *S. bayanus*and *S. kudriavzevii*genomes using CLUSTALW (see Materials and methods) and sought to evaluate the effectiveness of MONKEY under different evolutionary models and distances.
Ideally, we would use several diverse transcription factors with known binding specificity, where the set of matches to the factor\'s matrix in the *S. cerevisiae*genome could be divided into two reasonably sized sets: those known to be bound by the factor (positives) and those known not to be bound by the factor (negatives). Unfortunately, even in yeast, the number of such cases is limited. For many factors we can identify true positives by combining high- and low-throughput experimental data that supports the hypothesis that a particular position in the genome is bound by a given factor. A true negative set, however, must be constructed on the basis of lack of evidence that a sequence is functional, as the interpretation of negative results almost always is ambiguous. In the case of transcription factor binding sites this is particularly problematic, because DNA-binding proteins have overlapping specificity, and we may therefore observe conservation of a binding site because it is bound by another factor with similar specificity. After evaluating all factors with binding specificity in *Saccharomyces cerevisiae*Promoter Database (SCPD) \[[@B33]\], we focus on Gal4p and Rpn4p for further analysis (see Table [1](#T1){ref-type="table"} for properties of these factors, and Materials and methods for a description of the selection of positive and negative sets).
The effects of evolutionary models on the discrimination of functional binding sites
------------------------------------------------------------------------------------
To evaluate the performance of our evolutionary method in correctly identifying *bona fide*binding sites, we calculated the *p*-values of the positive and negative sites for each factor, using MONKEY on alignments of all five genomes for Rpn4p and four species (with *S. kudriavzevii*excluded because too few sequences were available) for Gal4p. We compared the performance of MONKEY with the HB model to scores from *S. cerevisiae*alone and to a \'simple\' score (equal to the average of the single sequence log likelihood ratios) that utilizes all the comparative data without an evolutionary model.
The results are summarized in Table [2](#T2){ref-type="table"}. An ideal scoring method would assign low *p*-values to real sites (positives) and high *p*-values to spurious sites (negatives), and we therefore compared the *p*-values assigned by monkey based on the HB model to those based on the \'simple\' score. Not surprisingly, both methods were a great improvement over searching in *S. cerevisiae*alone. Overall, when compared to each other, the HB score assigned lower *p*-values to the binding sites more often in the positive sets (90% for Gal4p and 80% for Rpn4p) and less often in the negative sets (20% for Gal4p and 25% for Rpn4p) than did the simple score. We note that some of the supposedly functional Rpn4p sites were assigned higher *p*-values in *S. cerevisiae*alone, suggesting that they are not in fact conserved; these will be discussed below.
The effect of evolutionary distance on the discrimination of functional binding sites
-------------------------------------------------------------------------------------
As evolutionary distance increases, we expect fewer matches to the matrix to be conserved by chance, which implies that the probability of observing matches as highly conserved as the functional sites should decrease. Similarly, we expect the nonfunctional sites to show many substitutions and their *p*-values to increase over evolution. To explore the change in *p*-values over evolutionary distance, we scored the functional and nonfunctional sets of binding sites at a variety of evolutionary distances by creating alignments of different combinations of species (see Materials and methods). The median *p*-value of the positive set of TFBSs decreases monotonically with evolutionary distance, with the rate of decrease an approximately constant function of evolutionary distance (see Figure [2](#F2){ref-type="fig"}). The median *p*-value for the binding sites in the negative set increases with evolutionary distance, although somewhat erratically. This demonstrates that MONKEY effectively exploits evolutionary distance, and confirms our intuition that as evolutionary distance increases, functional elements should be increasingly easy to distinguish from spurious predictions.
To test this hypothesis on a more quantitative level we sought to compare the observed scores with the expected scores assuming that binding sites evolved precisely according to the evolutionary models used by MONKEY. Briefly, given a binding-site model and a phylogenetic tree, we assume we have observed a binding site in the reference genome, and that this site evolves along the tree under either the motif model (HB) or background model (HKY), representing functional and nonfunctional binding sites, respectively (see Materials and methods for details). The expected *p*-values associated with the functional binding sites (Figure [2](#F2){ref-type="fig"}, solid lines) showed reasonable agreement with the models, consistent with previous observations that they are evolving under constraint that is well modeled by the purifying selection on the base frequencies in the specificity matrix \[[@B19]\].
Pairwise versus multi-species comparisons
-----------------------------------------
The comparisons at the different evolutionary distances used in Figure [2](#F2){ref-type="fig"} employed variable numbers of species, with the shorter distances representing primarily pairwise comparisons and the longer distances comparisons of three or more species. While we expect the variation in *p*-values with different combinations of species to be primarily a function of the evolutionary distance spanned by these species, there will also be effects related to the number of species and the topology of the three. For example, in the limit of very long branch lengths, the evolutionary *p*-values are on the order of the power of the number of species and are independent of evolutionary distance. In contrast, in the limit of very short branch lengths, the evolutionary *p*-values depend only on the distance spanned by the comparison, as most of the information provided by additional species is redundant. However, because most comparisons that are actually carried out are far from either of these extremes, we sought to evaluate the effects of species numbers and tree topology for the *Saccharomyces*species analyzed here.
First, we recomputed the expected *p*-values for all the distances analyzed in Figure [2](#F2){ref-type="fig"}, except that instead of using the real tree topology, we used a single pairwise comparison at the same evolutionary distance (Figure [2](#F2){ref-type="fig"}, dotted lines). For example, for the Rpn4p analyses using all five species we assumed a pairwise comparison at an evolutionary distance of around 1.1 substitutions per site. Note that this is considerably more distant than any of the pairwise comparisons available among these species. The predictions for the pairwise and multi-species comparisons are very similar, suggesting that at the evolutionary distances spanned by these species there is little difference in using multiple species alignments relative to a pairwise alignment that spans the same evolutionary distance. Only at the longest distances considered (greater than 0.8 substitutions per site) does the power of the pairwise comparison begin to level off, although there are other reasons that multiple species comparisons might still be preferred (see Discussion).
To complement this theoretical analysis, we were interested in using empirical data to compare pairwise and multi-species analyses. Fortuitously, the evolutionary distance between *S. cerevisiae*and *S. kudriavzevii*is almost exactly equal to the evolutionary distance spanned by *S. cerevisiae*, *S. paradoxus*and *S. mikatae*(median tree length approximately 0.5 substitutions per site; see Figure [3a](#F3){ref-type="fig"}). Because our models predict that we are in a regime where evolutionary distance is the primary determinant of the *p*-values, we expect searches using these different sets of species to yield similar results. We tested this hypothesis by calculating the *p*-values associated with the Rpn4p-binding sites using the sequences from these two comparisons. The median *p*-values in both the positive and negative sets are very similar (Figure [3b](#F3){ref-type="fig"}), confirming that at these relatively short evolutionary distances, the power of the comparative method is independent of the number of species considered (see Discussion).
Taken together, these results strongly support the idea that when appropriate methods are used, data from multiple species can be combined effectively to span larger evolutionary distances. Note that this in no way implies that the addition of extra species to an existing pairwise comparisons is not useful - such additions will always increase the evolutionary distance spanned by the species and thus will increase the power of the comparison.
Testing the power of comparative annotation of transcription factor binding sites
---------------------------------------------------------------------------------
At the distances spanned by all available sequence data, the *p*-values are so small that we no longer expect to find matches of the quality of those in the positive set by chance, especially for Rpn4p. To test this further, we scanned both strands of all the available alignments of all five *sensu stricto*species (around 2.7 Mb) to identify our most confident predictions of conserved matches to the Rpn4p matrix. We chose the *p*-value cutoff of 1.85 × 10^-8^, which corresponds to a probability of 0.05 of observing one match at that level over the entire search (using a Bonferroni correction for multiple testing). After excluding divergently transcribed genes, there were 56 genes that contained putative binding sites at that *p*-value. Of 32 genes in our positive set that had sequence available for all five species, 30 had binding sites below this *p*-value. Of the 28 genes in the negative set for which sequences were available, only three had binding sites below this cutoff. In this (nearly ideal) case we have ruled out nearly 90% of the negative set at the expense of less than 10% of the positives.
Examining the expression patterns of these genes (Figure [4a](#F4){ref-type="fig"}) allows them to be divided into three major classes. The first is a group (indicated by a blue bar) containing 30 genes (28 of which were in our original positive set and two other genes) that show a very similar pattern over the entire set of conditions. The second group (indicated by a green bar) contains 11 genes (of which only one was in our original positive set) that show uncoordinated gene expression changes in some conditions in addition to the stereotypical Rpn4p expression pattern. It is possible that these genes\' regulation is controlled by multiple mechanisms under different conditions \[[@B34]\], and regulation by Rpn4p is one contribution to their overall pattern of expression. Further supporting this hypothesis, only one of these genes (*UFD1*) is annotated as involved in protein degradation, and three (*YBR062C*, *YOR052C*and *YER163C*) have unknown functions.
Finally, and most surprising from the perspective of comparative annotation, is a third set of 14 genes, including one from our original positive set and three from our negative set, most of which show no evidence of the proteasomal expression pattern associated with Rpn4p (Figure [4b](#F4){ref-type="fig"}). It is extremely unlikely that these sequences have been conserved by chance, and we suggest that they represent matches that are conserved for reasons other than binding by Rpn4p (see Discussion).
Nonconserved binding sites in regulated genes
---------------------------------------------
Having identified examples of conserved binding sites whose nearby genes showed no evidence of function, we decided to examine the converse: binding sites near regulated genes, and therefore presumably functional, that are not conserved. Figure [5](#F5){ref-type="fig"} shows the *p*-values of individual positive Rpn4p sites at different evolutionary distances. While most of the sites follow the trajectory predicted for sites evolving under the HB model, the *p*-values for four of the positive sites seem to be well-modeled by the \'background\' or unconstrained model. This is surprising because we expect these binding sites to be functional, and therefore under purifying selection. One explanation is that some of these sites may have been misannotated as functional. For example, in addition to a nonconserved positive site, the upstream region of *REH1*contains another binding site that is a weaker match to the Rpn4p matrix (Figure [5b](#F5){ref-type="fig"}) and did not pass our threshold for inclusion in the positive set (see Materials and methods). This weaker match is more highly conserved and may represent the functional site in this promoter. In the case of *PTC3*, however, we can find no other candidate binding sites nearby (Figure [5c](#F5){ref-type="fig"}). This represents a possible example of binding-site gain, a proposed mechanism of regulatory evolution at the molecular level (see Discussion).
Different factors have different relationships between significance and evolutionary distance
---------------------------------------------------------------------------------------------
The optimal selection of species for comparative sequence analysis remains an open question. To analyze this question for transcription factor binding sites, we examined the relationship between evolutionary distance and the MONKEY *p*-values for several *S. cerevisiae*transcription factors (Figure [6](#F6){ref-type="fig"}) for which sufficient characterized binding sites were available in SCPD \[[@B33]\]. We find that while all factors show the tendency for *p*-values to decrease with evolutionary distance, the *p*-values for each factor remain very different. For example, with alignments of four species spanning about 0.8 substitutions per site, we expect a conserved match to the Gcn4p matrix as good as the median functional binding site (Figure [6a](#F6){ref-type="fig"}, red triangles) approximately every million bases of aligned sequence. This in contrast to Rpn4p, for which in the same alignments we expect such a match (Figure [6a](#F6){ref-type="fig"}, violet crosses) only once in about 1 billion base pairs. Thus, the evolutionary distance required to achieve a desired *p*-value is different for different factors. Understanding the relationship between a frequency matrix and the behavior of its *p*-values is an area for further theoretical exploration. We note that, once again, we can predict the behavior of these *p*-values (Figure [6b](#F6){ref-type="fig"}), and that while our predictions agree qualitatively, there is considerable variability.
Software
--------
MONKEY is implemented in C++. It is available for download under the GPL and can be accessed over the web at \[[@B35]\].
Discussion
==========
By formulating the problem of identifying conserved TFBSs in a probabilistic evolutionary framework, we have both created a useful tool (MONKEY) for comparative sequence analysis capable of functioning on relatively large numbers of related species, and enabled the examination of several important questions in comparative genomics. While most previous approaches to this problem have used heuristics to define conserved and nonconserved TFBSs, with the probabilistic scores and *p*-value estimates presented here the assumptions underlying our approach can be made explicit, and where those assumptions hold we can be assured the reliability of our method. In addition, the probabilistic framework allows us to estimate the amount of evolutionary distance required to achieve a certain level of significance.
Evolutionary models
-------------------
The score based on the evolutionary model proposed by Halpern and Bruno \[[@B25]\] effectively discriminated the functional and nonfunctional Gal4p- and Rpn4p-binding sites in *S. cerevisiae*(Table [2](#T2){ref-type="table"}). We believe the success of the HB model in predicting position-specific rates of evolution \[[@B19]\] and identifying conserved TFBSs reflects its encapsulation of a model of binding sites evolving under constant purifying selection. Although not every functional binding site will remain under purifying selection, as a result of either functional change or binding-site turnover (see below), a large subset of functional binding sites do remain under purifying selection, and for these, the \'HB\' score performs better than the \'simple\' score. It is interesting to note, however, that the simple score, which is not based on an evolutionary model and does not take into account the relationships of the species used in the comparison, still shows great improvement over one genome alone, highlighting the value of comparative sequence data even when used suboptimally.
Effects of evolutionary distance
--------------------------------
An important hypothesis of the comparative genomics paradigm is that as evolutionary distance increases, observing a match with a given level of conservation should become less and less likely by chance - the *p*-values for functional sites that are conserved are expected to decrease. We confirm this hypothesis for a small number of factors from *S. cerevisiae*. In addition, our probabilistic models allow us to quantify this relationship. We can directly measure the confidence that a specific site is a conserved binding site, and we can predict the evolutionary distance needed to achieve a desired level of significance.
Typical *p*-values for functional binding sites scored by matching a matrix to a single genome are on the order of 10^-4^to 10^-6^. Even in a relatively small genome like yeast, with roughly 12 million bases, we expect many matches at this significance level to occur by chance. Adding four closely related species that span a total evolutionary distance of approximately one substitution per site reduces these *p*-values by approximately three orders of magnitude to the range 10^-7^to 10^-9^. In the yeast genome we expect few, if any, matches to occur at this level of significance by chance. When we search the alignments of these species with the Rpn4p matrix with a low enough *p*-value that we expect a match at that significance to occur only once in a random 50 Mb genome, we recover nearly the entire positive set of Rpn4p-binding sites while excluding most of the negative set, highlighting the utility of MONKEY and the statistics we have developed. As a measure of the improvement over searching a single genome alone, we note that even the best possible match to the Rpn4p matrix in one genome does not meet this significance criterion.
The expected relationship between evolutionary distance and *p*-value can, in principle, be used to guide to choice of species to be sequenced for comparative analyses. However, the dependence of *p*-values on evolutionary distance is not the same for all factors (Figure [6](#F6){ref-type="fig"}). This suggests that our ability to annotate functional sequences by comparative methods will depend on the type of sequences that we are trying to annotate, and that there is no single evolutionary distance sweet-spot for identifying TFBSs.
Pairwise versus multiple species comparisons
--------------------------------------------
In theory, for a given reference genome it should be possible to pick a single comparison species at an evolutionary distance sufficient to identify any conserved feature of interest. Our results suggest that at distances of up to approximately 0.6 substitutions per site, pairwise alignments provide essentially the same amount of resolving power as multiple comparisons spanning the same evolutionary distance. We showed that *S. cerevisiae*and *S. kudriavzevii*span almost exactly the same evolutionary distance as *S. cerevisiae*, *S. paradoxus*and *S. mikatae*, and that that distance is well below 0.6 substitutions per site. Consistent with this, MONKEY produces nearly identical *p*-values for conserved binding sites from these two sets of species. Thus, our results suggest that from a theoretical perspective, if the goal of comparative analysis is to identify conserved binding sites for factors like the ones considered here, it is not necessary to sequence species much more closely related than this limit.
We note, however, that there are myriad practical reasons other than evolutionary resolving power (the only factor considered in our models) for sequencing multiple closely related sequences. First, there may simply be no extant species at the exact evolutionary distance desired. Second, the quality of DNA alignments is expected to be much higher for multiple closely related species than for more distant pairwise alignments - if alignment errors prevent correct assignment of orthology, conserved binding sites will not be identified. For the factors considered here, the pairwise comparison performed nearly as well as the multiple species comparison well beyond the evolutionary distances at which pairwise alignments are reliable \[[@B36]\], suggesting that the necessity of alignment will limit the maximum distance between species. Finally, and perhaps most important, is the assumption that our models make about constant functional constraint over evolution. To illustrate this, consider the binding sites for Gal4p used in the analysis in Figure [2a](#F2){ref-type="fig"}. These binding sites could not be included in Figure [3](#F3){ref-type="fig"} because *S. kudriavzevii*orthologs for these genes were not available in SGD, apparently because of the degeneration of the galactose-utilization pathway in this species \[[@B37]\]. Sequencing multiple closely related species provides insurance against such functional changes, because they are less likely to have occurred in all the lineages.
Conserved sites and binding-site turnover
-----------------------------------------
MONKEY was very effective in identifying functional Rpn4p-binding sites from the alignment of five *Saccharomyces*species. In our search, 41 of 56 (73%) predicted sites were found near genes showing the expected expression pattern, and are therefore likely to be functional. Even at this level of stringency, however, there are highly conserved sequences that match the matrix, but do not appear to be near genes that are regulated by Rpn4p. It is very unlikely that these sites are conserved by chance. One possible explanation for this high degree of conservation is that these are functional sites, but that the expression of these genes is not accurately detected in high-throughput assays, or their function has not been accurately determined. A more likely possibility is that these sites are conserved because they perform other, unknown functions. Consistent with this hypothesis is the fact that many of these matches fall near other highly conserved sequences (Figure [4b](#F4){ref-type="fig"}), suggesting that they may be parts of larger conserved features.
In addition to the conserved sequences that are unlikely to represent *bona fide*binding sites, we also found examples of binding sites associated with properly regulated genes that do not seem to be conserved (Figure [5](#F5){ref-type="fig"}). Once again there are several possible explanations for this observation. First, these binding sites may not actually be functional and may have been included in our positive set erroneously. While this is a possible explanation for the case of the Rpn4p-binding sites shown in Figure [5](#F5){ref-type="fig"} (and may be likely in the case of *REH1*, where we could identify another apparently conserved binding site in the region) we have also found nonconserved examples among the TFBSs in the SCPD database (approximately 20% of TFBSs we examined, see Additional data file 1), all of which have at least some direct experimental support.
Another potential explanation is that these binding sites are actually conserved, but were not aligned correctly. While this is difficult to rule out in general, in the few nonconserved cases for Rpn4p at least we could not find (by eye) errors in the alignments. Most interesting, of course, would be the situation where these nonconserved binding sites are not due to some error on our part, but rather represent a biological change in the functional constraints on these sequences, possibly resulting in a change in the regulation of the expression of these genes. Our results represent an upper bound on the number of TFBSs for which this has occurred. *Cis*-regulatory changes have been proposed to be an important source of genetic variation \[[@B32]\]. Gains and losses of functional binding sites represent an important class of these changes \[[@B38],[@B39]\], and an important area for future computational and experimental analysis, particularly as the genome sequences of closely related metazoans become available. We expect MONKEY to be a useful tool in the comparative analysis of these genomes, and we have found comparable increases in the significance of functional binding sites in alignments of *Drosophila melanogster*and *D. pseudoobscura*(see Additional data file 2).
Conclusions
===========
We have developed a method to identify conserved TFBSs in sequence alignments from multiple related species that provides a quantitative framework for evaluating results. The method - implemented in the open-source software MONKEY - extends probabilistic models of binding specificity to multiple species with probabilistic models of evolution. We have found that a probabilistic evolutionary model \[[@B25]\] that assumes binding sites are under constant purifying selection performs effectively in discriminating functional binding sites. We have developed methods to assess the significance of hits, and have shown that the significance of functional matches increases while the significance of spurious matches decreases over increasing evolutionary distance. We can explicitly model the relationship between the significance of a hit and evolutionary distance, allowing the assessment of the potential of any collection of genomes for identifying conserved binding sites. Applying MONKEY to a collection of related yeast species we find that most functional binding sites are highly significantly conserved, but also find evidence for conserved sites that are not functional and *vice versa*. Our results suggest that development of methods that model the evolutionary relationships between species and the evolution of the genomic features of interest yield insight into the challenges for comparative genomics.
Materials and methods
=====================
Simulating pairs of sequences
-----------------------------
To generate the empirical *p*-values shown in Figure [1](#F1){ref-type="fig"}, random sequences of length *w*were generated according to the average intergenic base frequencies of the *S. cerevisiae*genome. These were then evolved according to the Jukes-Cantor substitution model, to a specified evolutionary distance. Likelihood ratio scores and *p*-values were then calculated for each of the pairs of sequences using the method implemented in MONKEY. Finally, all pairs of sequences were ranked by their scores, and the rank divided by the total number of pairs was taken as the empirical *p*-value.
Preparation of alignments for different groups of species
---------------------------------------------------------
We aligned the upstream regions of all *S. cerevisiae*genes to their orthologs in *S. paradoxus*, *S. mikatae*, *S. bayanus*and *S. kudriavzevii*by taking the 1,000 bp upstream of each gene, identifying the corresponding region from the other species using data in the *Saccharomyces*Genome Database \[[@B40]\], aligning them with CLUSTAL W \[[@B41]\] and trimming them to remove regions corresponding to *S. cerevisiae*coding sequence. We used this strategy rather than simply aligning intergenic regions to control for differences in alignments that might arise from the use of variably sized regions.
To obtain estimates of the evolutionary distance spanned by each comparison, we ran PAML \[[@B24]\] on the entire set of intergenic alignments, using the HKY model \[[@B27]\], with constant rates across sites. We used the median PAML estimate of kappa (the transition-transversion rate ratio) of 3.8, the *S. cerevisiae*background frequencies (ACGT) = (0.3, 0.2, 0.2, 0.3) and the median of the branch lengths estimates as the \'background\' evolutionary model. The trees with these branch lengths were used as input to MONKEY to calculate *p*-values. The distances in Figure [4](#F4){ref-type="fig"} represent the sum of the median branch lengths in each comparison. The subsets (with evolutionary distances in parentheses) were as follows: *S. cerevisiae*and *S. paradoxus*(0.194); *S. cerevisiae*and *S. mikatae*(0.403); *S. cerevisiae*, *S. paradoxus S. mikatae*(0.477); *S. cerevisiae*and *S. bayanus*(0.559); *S. cerevisiae*, *S. paradoxus, S. mikatae*and *S. bayanus*(0.816); *S. cerevisiae*, *S. paradoxus, S. mikatae, S. bayanus*and *S. kudriavzevii*(1.090).
Definition of Rpn4p and Gal4p matrices and positive and negative sets
---------------------------------------------------------------------
**Rpn4p:**we used Rpn4p sites in proteasomal genes \[[@B42],[@B43]\] to build an Rpn4p specificity matrix (using a pseudocount of 1 per base per position). To identify additional likely targets, we obtained expression data from public sources \[[@B30],[@B31]\] and compared the expression patterns of all genes to the average expression pattern of proteasomal genes using the following metric:
where *θ*is the \'uncentered correlation\', a commonly used distance metric for gene-expression data \[[@B44]\]. Our score adds a correction for the number of datapoints, *n*, that are available for each gene. All matches to the Rpn4p matrix (*S. cerevisiae*likelihood ratio score \> 9) in the upstream region of a gene that matched the proteasomal expression pattern (*t*\> 8) were considered to be true Rnp4p sites. The negative set consists of all sites that matched the Rpn4p matrix with a score greater than 9, and excluded sites in genes with even weak similarity to the proteasomal expression pattern (*t*\> 0) or that were annotated \[[@B40]\] as involved in protein processing or degradation.
**Gal4p:**we used the matrix from SCPD \[[@B33]\] (with a pseudo count of 1 per base per position). To define a positive set we used the binding sites in SCPD and systematic studies of this Gal4p regulatory system \[[@B45],[@B46]\], and used matches near additional genes that we identified in these studies with scores above the lowest score in the SCPD set. To define a negative set, we again scanned the *S. cerevisiae*genome with a cutoff equal to the lowest score in the positive set and then eliminated any binding sites near genes that showed evidence for regulation in the systematic studies.
It is important to note that our categorization of sequences as positive and negative is done independently of the comparative sequence data, thus avoiding potential circularity.
Calculations of expected scores
-------------------------------
Because our methods employ explicit probabilistic models for the evolution of noncoding DNA, it is possible to compute the expected scores under various assumptions. The expectation of the log likelihood ratio for examples of the motif is the \'information content\' and its calculation has been addressed \[[@B47]\]. We can extend this to calculation to our evolutionary scores, as follows. Using the fact that all the scores treat the positions of the matrix independently, and the linearity of the expectation, we write:
where *E*\[*x*\] denotes the expectation of the random variable *x*, *m*denotes a frequency matrix and a corresponding evolutionary model, either {*motif*, *R*~*motif*~} or {*bg*, *R*~*bg*~}. *p*(*X*~*i*~, *Y*~*i*~\|*m*, *T*) is calculated as above, and we define:
We can write a similar expression for the variance, *V*:
In order to predict the scores for the genes in our positive and negative sets, we are interested in the case were we have observed a match to the motif in one species, but the constraints on its evolution are either those of the background or the motif. We can compute the expected scores under these assumptions as follows:
where *p*(*X*~*i*~\|*motif*) is the single species probability of observing the base *X*~*i*~at position *i*in the specificity matrix (*f*), and using Bayes\' theorem:
This calculation can be extended to the multiple species case, by replacing the distributions *p*(*X*~*i*~, *Y*~*i*~) and *p*(*Y*~*i*~\|*X*~*i*~) with *p*(*X*~*i*~, *Y*~*i*~, \..., *Z*~*i*~) and *p*(*Y*~*i*~, \..., *Z*~*i*~\|*X*~*i*~) and changing the sum over *Y*~*i*~to a sum over all the other leaves in the tree except the reference, in this case, *X*~*i*~. For the functional set, we assumed the binding sites were evolving under the HB model \[[@B25]\], and for the nonfunctional set we assumed evolution under the HKY background model described above. To model the sequence-specificity matrices most accurately, we reduced the pseudocount (equal to the background probability of observing each base).
Additional data files
=====================
Additional data file [1](#s1){ref-type="supplementary-material"} shows the fraction of binding sites that are not conserved for several different *S. cerevisiae*transcription factors. Additional data file [2](#s2){ref-type="supplementary-material"} shows the conservation p-values of predicted binding sites in high-density binding site clusters in the *Drosophila*melanogaster genome, with the binding sites grouped according to whether the cluster has regulatory activity.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
The fraction of binding sites that are not conserved for several different *S. cerevisiae*transcription factors
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
The conservation p-values of predicted binding sites in high-density binding site clusters in the *Drosophila*melanogaster genome, with the binding sites grouped according to whether the cluster has regulatory activity
:::
::: {.caption}
######
Click here for additional data file
:::
Acknowledgements
================
We thank Dan Rokhsar for suggesting simple heuristics for gapped sequences, and Audrey Gasch, Casey Bergman, Ewan Birney and Bill Bruno for comments on the manuscript.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Accuracy of *p*-value estimations. To examine the accuracy of our *p*-value estimates, we compared the empirical *p*-value (computed from the observed distribution of scores) to *p*-values computed using either the exact method described above (black points) or Gaussian approximation (gray points). The scores represent the simple score at a distance of 0.1 substitutions per site calculated using the Gcn4p matrix from SCPD \[33\]. Other models and matrices produce similar results.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Significance of matches increases with evolutionary distance. Median *p*-values for the positive (black squares) and negative (white triangles or white triangle points) sets of binding sites for **(a)**Gal4p and **(b)**Rpn4p at different evolutionary distances represented by comparing *S. cerevisiae*to different subsets of the available species. For both factors, as evolutionary distance increases, the median *p*-value of the functional matches decreases, indicating that they are less likely to have appeared by chance. Conversely, the median *p*-value of the nonfunctional matches (negative set, white symbols) increases. These observations agree with our predictions for the behavior of the *p*-values (solid traces) under either the HB evolution for the motif or HKY evolution for the background. There is little difference between these predictions and similar ones that assume that all the comparisons were pairwise (dotted traces).
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Significance of binding sites in pairwise or three-way comparisons at similar evolutionary distance. **(a)**Histogram of the percent identities of all aligned noncoding regions of *S. cerevisiae*and *S. kudriavzevii*(open squares) and *S. cerevisiae*, *S. paradoxus*and *S. mikatae*(filled squares). **(b)**Median *p*-values of functional matches (positive set, gray bars) and the nonfunctional matches (negative set, open bars) for *S. cerevisiae*and *S. kudriavzevii*alignments (left) and *S. cerevisiae*, *S. paradoxus*and *S. mikatae*alignments (right). The similarity of these *p*-values supports the idea that multiple similar genomes can be used to span longer evolutionary distances, but at these close evolutionary distances provide little additional power.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Relationship between conserved Rpn4p-binding sites and expression. **(a)**We identified 56 Rpn4p-binding sites with *p*-values below 1.85 × 10^-8^using all five species and the HB model. The expression patterns of these genes (clustered and displayed as in \[44\]) fall into two major groups: the \'stereotypical\' proteasomal pattern (indicated by a blue bar at the right), and a second group expressed in these and additional conditions (indicated by the green bar). The orange bars above the expression data correspond to (left to right) temperature changes, treatment with H~2~O~2~, treatment with the superoxide generating drug menadione, treatment with the sulfhydryl oxidant diamide, deletions of *YAP1*and *MSN2/4*, treatment with the DNA damaging agent methylmethanesulfonate (MMS), and heat shock in deletions of *MEC1*and *DUN1* \[48,49\]. **(b)**Examples of conserved Rpn4p sites (boxed) that do not fall in either expression group (neither blue nor green bar).
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Some apparently functional Rpn4p-binding sites are not conserved. **(a)**The MONKEY *p*-values (points) of all putatively functional Rpn4p-binding sites at varying evolutionary distances, along with the expected values under the HB and HKY models (solid traces). The majority of sites behave as expected for conserved binding sites (lower trace). Several, however, behave as expected for unconstrained sites (upper trace). **(b)**The predicted binding site (indicated by a box) in *REH1*, which encodes a protein of unknown function in *S. cerevisiae*, is not conserved, whereas a binding site with a lower score is conserved (indicated by a black bar). **(c)**A very poorly conserved match upstream of *PTC3*; in this case no other sites can be found in the region.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
The evolutionary distance required to confidently identify conserved binding sites varies among transcription factors. **(a)**Median *p*-values for functional binding sites for various factors at different evolutionary distances. The evolutionary distance needed to obtain a desired significance varies between factors. **(b)**Predicted dependence of the *p*-values on evolutionary distance. Specificity data and functional binding sites were obtained from the SCPD.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Definition of positive and negative sets of matrix matches
:::
Criterion Gal4p Rpn4p
------------------------------------ --------------------------------------------------------------------- ------------------------------------------------------------------------------------------------
Unique specificity Spacer of 11 bp \[50\] Atypical zinc finger \[42\]
Well characterized specificity Protein-DNA co-crystal \[51\] Large number of binding sites, low degeneracy \[42\]
Well characterized target gene set Classic genetic system \[52\] and high-throughput studies \[45,46\] Targets include almost all proteasomal subunits \[42\] stereotypical expression pattern \[48\]
Criteria used to define positive and negative sets to use in this study. It is important to avoid factors whose specificity overlaps with other factors, because binding sites that are not occupied by one factor may be constrained because of binding by another, and to choose factors with characterized specificity because our methods rely on the assumption that the specificity is known.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Performance of different scores in recognizing functional and nonfunctional sites
:::
Percent of binding sites assigned a lower *p*-value Gal4p Rpn4p
----------------------------------------------------- ------- ------- ----- -----
Halpern-Bruno vs simple 90% 30% 80% 25%
Halpern-Bruno vs *S. cerevisiae*alone 100% 40% 87% 34%
Simple vs *S. cerevisiae*alone 100% 30% 90% 48%
The score based on the Halpern-Bruno (HB) model assigns lower *p*-values to functional binding sites and higher *p*-values to nonfunctional binding sites than the simple score, defined as the average of the single species scores in at that position in the alignment. Both methods are far superior to *p*-values from *S. cerevisiae*alone. See text for details.
:::
|
PubMed Central
|
2024-06-05T03:55:51.903045
|
2004-11-30
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545801/",
"journal": "Genome Biol. 2004 Nov 30; 5(12):R98",
"authors": [
{
"first": "Alan M",
"last": "Moses"
},
{
"first": "Derek Y",
"last": "Chiang"
},
{
"first": "Daniel A",
"last": "Pollard"
},
{
"first": "Venky N",
"last": "Iyer"
},
{
"first": "Michael B",
"last": "Eisen"
}
]
}
|
PMC545802
|
Background
==========
Genome-wide analyses of cellular mRNA, protein or metabolite complements have become workhorses in biological research that produce unprecedented amounts of data on cellular network composition. In contrast to such compositional information, molecular fluxes through intact metabolic networks link genes and proteins to higher-level functions that result from biochemical and regulatory interactions between the components \[[@B1]\]. As such, quantitative knowledge of *in vivo*molecular fluxes is highly relevant to functional genomics, metabolic engineering and systems biology \[[@B2],[@B3]\]. Intracellular fluxes, or *in vivo*reaction rates, can be assessed by methods of metabolic flux analysis that are based on stable isotopic tracer experiments \[[@B4],[@B5]\], which have successfully unraveled novel biochemical pathways \[[@B6],[@B7]\] and gene functions \[[@B8],[@B9]\]. The presently tedious and limited methodologies, however, hamper broader application to a large range of environmental conditions, isotopic tracers and higher biological systems \[[@B4]\].
We set out to overcome a principal bottleneck in metabolism-wide flux (fluxome \[[@B10]\]) analysis: the requirement for mathematical frameworks to interpret the isotopic tracer data from nuclear magnetic resonance (NMR) or mass spectrometric (MS) analyses within a detailed metabolic model \[[@B4],[@B5]\]. Constructing such models requires *a priori*knowledge on possible distributions of the tracer used within the network, and, more importantly, extensive labeling and physiological data to resolve all fluxes within a given model. The lack of such structural knowledge on metabolic pathways and the technical difficulty of acquiring sufficient data hamper studies of metabolism, in particular in higher cells with complex nutrient requirements and for exotic tracer molecules. Hence, fluxome analysis is largely restricted to few ^13^C-labeled carbon sources in microbes or plants cultivated in minimal medium \[[@B7],[@B11]-[@B16]\].
Here we discriminate mutants/conditions and assess their metabolic impact directly from \'raw\' mass-isotope data by unsupervised multivariate statistics without *a priori*knowledge of the biochemical reaction network. To illustrate the applicability of this conceptually novel profiling method, we focused on the reactions of central metabolism in the model bacterium *Bacillus subtilis*, for which detailed flux data were available to validate the results \[[@B9],[@B11],[@B14]\].
Results
=======
^2^H and ^13^C tracer experiments
---------------------------------
Environmental and genetic modifications were used to perturb intracellular metabolic activities in *B. subtilis*. In particular, we chose 10 knockout mutants \[[@B17]\] that were affected in metabolic genes or transcriptional regulators linked to central metabolism (Table [1](#T1){ref-type="table"} and Figure [1](#F1){ref-type="fig"}). These mutants were grown in 1-ml batch cultures \[[@B18]\] with six combinations of the carbon sources \[U-^13^C\] or \[U-^2^H\]glucose, \[U-^13^C\]sorbitol or \[3-^13^C\]pyruvate and the nitrogen sources ammonium or casein amino acids (CAA). As a proof of concept, we detected the isotopic labeling patterns in proteinogenic amino acids by gas chromatography MS (GC-MS), which provides direct access to several metabolic nodes in the network \[[@B6],[@B7],[@B19]\] (Figure [1](#F1){ref-type="fig"}). The raw mass isotope data of all mutants under each of the six experimental conditions are given in Additional data file 2.
In media supplemented with amino acids, cell protein was only partly synthesized from the isotopically labeled substrate. In such cases, current flux-analysis methods such as isotopomer balancing or flux ratio analysis are not applicable \[[@B4],[@B5]\] because they do not account for variations in the labeling patterns due to amino-acid uptake and catabolism. Practically, we tackled here a worst-case scenario: growth in a medium enriched with unlabeled amino acids and profiling of the labeling pattern from tracers in the proteinogenic amino acids, which may potentially originate entirely from the medium. Nevertheless, a sufficiently high fraction of all analyzed amino acids was synthesized *de novo*from the labeled substrates to obtain relevant MS signals, indicating that information on pathway activities was recorded in the labeling patterns (Figure [2](#F2){ref-type="fig"}). To capture the impact of genetic or environmental modifications, we analyzed the 260-330 raw mass isotope data points for each mutant and condition. This is essentially a table of mass-distribution vectors for all detected amino-acid fragments upon correction for naturally occurring stable isotopes, that is, the list of the relative frequencies of all possible isotope isomers for each detected analyte.
Identification of metabolic determinants for altered flux profiles
------------------------------------------------------------------
For the visualization of metabolic effects, the corrected MS signals of the wild type were subtracted from those of the mutants (Figures [3](#F3){ref-type="fig"} and [4](#F4){ref-type="fig"}). Some mutations, such as *pps*, were silent under the conditions tested and exhibited only noise in the wild-type-normalized data. In other mutants, characteristic profiles of strongly affected amino acids were readily apparent. One example was the almost identical signature of serine (S) fragments in the profiles of the *glcP*and *cggR*mutants during growth on sorbitol with CAA; that is, high fractions of masses *m*~0~and *m*~3~and low fractions of *m*~1~and *m*~2~(where the subscripts denote the number of ^13^C atoms in each amino-acid fragment). While the S signature of the *mdh*mutant on sorbitol with CAA was also distinct, it was different from that in the above two mutants with low *m*~1,~*m*~2,~and *m*~3~fractions (Figure [3](#F3){ref-type="fig"}). These characteristic labeling profiles are biochemically very informative and may be linked to precise metabolic causes. For the above examples, the high fraction of uncleaved serine molecules with intact C~3~backbones (that is, *m*~0~and *m*~3~) in *glcP*and *cggR*is evidence of a lower exchange with the glycine pool, when compared with the wild type \[[@B19],[@B20]\]. In the *mdh*mutant, the high fraction of uncleaved but unlabeled S (*m*~0~) reveals high incorporation of unlabeled serine from the CAA supplement, and thus low *de novo*biosynthesis from ^13^C-labeled sorbitol.
As well as consistency with the data in the literature, the analysis also revealed new information on pathway activity and regulation that was not previously accessible. One example is the pronounced signatures of the *sdhC*mutant on glucose and sorbitol. Because the *sdhC*mutation disrupts the tricarboxylic acid (TCA) cycle, the wild-type flux through the cycle must be similar on these substrates, both with and without CAA (Figure [3](#F3){ref-type="fig"}). The *sdhC*signatures of the TCA cycle-derived amino acids aspartate (D) and glutamate (E) were also present in the CAA profiles of the other TCA cycle mutant *mdh*. Their absence on ammonium indicates activity of the malic enzyme-based pyruvate bypass \[[@B11]\] in the *mdh*mutant.
While such a level of detailed biochemical insight is possible, it requires considerable expertise and time to retrieve. Alternatively, metabolic impacts in new mutants can be identified by comparison of the mass fingerprints in mutants with known metabolic lesions. During growth on sorbitol and pyruvate in minimal media but not with CAA, the CggR repressor of the glycolytic *gapA*operon, for example, appears to affect TCA cycle fluxes because the mutant profile matches those of the TCA cycle mutants *sdhC*and *mdh*(Figure [3](#F3){ref-type="fig"}). In contrast to glucose, sorbitol does not elicit catabolite repression; hence, comparison of sorbitol and glucose profiles can identify repression-dependent effects. Examples are the signatures of the oxaloacetate-derived amino acids isoleucine (I), threonine (T) and aspartate in the *cggR*profile that reveal, by the similarity to the *sdhC*and *mdh*mutants, a TCA cycle flux-promoting effect of CggR on sorbitol but not on glucose. This is consistent with the repression of *cggR*on glucose \[[@B21]\], and the TCA cycle effect is probably indirect, through the repression of glycolytic genes \[[@B22]\].
A significant extension beyond the canonical ^13^C-tracer methods is the applicability to any isotope, which broadens the observable metabolic processes. Here we used fully deuterated \[U-^2^H\]glucose that allows us to monitor dehydrogenase activities and water release. The ^2^H-label was present exclusively in the variable side chains, because the α-carbon hydrogen was lost in the transaminase reaction. Thus, glycine contains no label and the acidic aspartate and glutamate lose the label proximal to the carboxyl group as a result of exchange with water at the low pH during hydrolysis. The remaining amino acids provided a stable and informative ^2^H-pattern (see Additional data file 1). An illustrative example is the *cggR*mutant signatures for the pyruvate-derived amino acids valine (V), leucine (L) and, partially, alanine (A) (Figure [3](#F3){ref-type="fig"}) In all three cases, reduced *m*~2~and increased *m*~0~fractions revealed a double loss of ^2^H-label in their common precursor pyruvate at position C-3. This loss of ^2^H indicates increased exchange of ^2^H with water at the C-3 position of pyruvate (or any upstream triose), which is fully consistent with increased transcription of the glycolytic enolase in the *cggR*mutant on glucose \[[@B23]\] that could catalyze this exchange. As the enolase activity does not affect the carbon backbone, the corresponding patterns cannot be identified in ^13^C experiments
Independent component analysis (ICA)
------------------------------------
For large-scale profiling studies, automated mutant classification based on metabolic function without user supervision would be desirable. Initially, we used principal component analysis (PCA), which is often used for graphical representation of multidimensional variables from profiling experiments \[[@B24],[@B25]\], as was recently described for pretreated (summed fractional labels) mass isotope data \[[@B26]\]. From the raw mass isotope data, the first two PCs discriminated, under most conditions, mutants with extreme labeling patterns (see Additional data file 1). The differences become smaller with increasing PCs, and only the initial three to four PCs allowed reliable discrimination. In the present data, PCA tended to discriminate extreme singular labeling patterns in few fragments or, more frequently, combinations of altered patterns in the fragments of many amino acids, as was expected from the variance maximization of PCA. Unfortunately, the resulting complex PCs are difficult to interpret metabolically, and thus are of limited biochemical relevance.
Consequently we used independent component analysis (ICA) for unsupervised, automatic recognition of conserved labeling patterns that are biochemically relevant. The underlying assumption is that these patterns result from the superposition of independent metabolic activities. Each activity causes a specific shift in the mass distributions of one or more intermediates. ICA seeks to separate the observed variables into non-gaussian components that are statistically as independent as possible \[[@B27]\]. Generally, ICA clearly discriminated mutants and conditions from the corrected (non-normalized) MS data (see Additional data file 1). While the weights in PCs were more broadly distributed among the input variables, ICs were dominated by fewer, sharper peaks (Figure [4](#F4){ref-type="fig"}).
For the particular example of the \[U-^13^C\]sorbitol with ammonium experiment, we explored the ICA results in more detail (Figure [5](#F5){ref-type="fig"}). The first, striking, observation was that the second IC contains the biochemically redundant signals of *m*~2~T, *m*~2~D, and *m*~1~and *m*~3~E (highlighted in red in Figure [5a](#F5){ref-type="fig"}) that arise from acetyl-CoA units in the TCA cycle \[[@B19]\]. This shows that ICA automatically provides insights into the biosynthetic linkage between amino acids with a resolution that eclipses visual comparison of the normalized signatures. For amino acids, this information was of course previously available, but statistical identification of biochemical relations could potentially also be obtained for less well-characterized compounds. Second, ICA often clustered biosynthetically related signals in the same component (Figure [5](#F5){ref-type="fig"}): IC7 grouped the similar signatures of phenylalanine (F) and tyrosine (Y) together; IC1 reports labeling shifts in glycine (G) and partially serine; and IC4 concentrated high weights in signals of the pyruvate derivatives alanine, valine and leucine (highlighted in blue in Figure [5](#F5){ref-type="fig"}). While isoleucine is also synthesized from pyruvate, it had only a marginal weight in IC4 because of interference from its second precursor oxaloacetate. Third, specific signatures of proline (P), leucine and serine are clearly recognized in IC3, IC8 (highlighted in green in Figure [5a](#F5){ref-type="fig"}), and IC10, respectively. These signatures reflect those previously identified in the normalized profiles (Figures [3](#F3){ref-type="fig"} and [5c](#F5){ref-type="fig"}). Among the remaining components, IC5 and IC6 emphasize outliers in the *cggR*and *ytsJ*MS data, respectively, whereas the noisy IC9 profile indicates that the identified ICs in our small dataset approach a limit.
Akin to PCA, ICA allowed us to discriminate mutants from the corrected MS data (Figure [5b](#F5){ref-type="fig"} and Additional data file 1). On sorbitol, mutants such as *pgi*, *yqjI*, *pps, glcP*and *glcR*were mostly silent, and typically projected in proximity to the parent strain. In contrast to PCA, ICs classified the mutants on the basis of specific metabolic effects. In some cases (IC2 or IC4 in Figure [5b](#F5){ref-type="fig"}), the IC defined well-separated clusters of mutants, usually two groups, reflecting a binary (on-off) effect. In the majority of the components, however, the even distribution between the extremes reveals progressive metabolic responses (for example, IC3, IC7 or IC10). Overall, the ICs correlated favorably with the signatures of wild-type-normalized profiles (Figure [5](#F5){ref-type="fig"} and Additional data file 1). Thus, ICA clearly outperformed PCA by its capacity for unsupervised recognition of metabolic responses and its ability to correlate biochemically redundant information in the data.
Comparison of PCA and ICA with analytically determined flux ratios
------------------------------------------------------------------
For most experimental conditions tested, mathematical frameworks for numerical flux analysis such as isotopomer balancing or flux-ratio analysis \[[@B4],[@B5]\] were not available. Only the \[U-^13^C\]glucose minimal medium experiments allowed a direct comparison of fluxome profiles with flux ratios. Therefore, we examined whether any of the statistically identified PCs and ICs was linearly correlated with eight analytically determined flux ratios \[[@B9],[@B19]\] that were obtained from the same MS data (Figure [6](#F6){ref-type="fig"}). For PCs, the correlation coefficients decreased with increasing component number, and singular correlations could not be detected between individual PC-flux ratio pairs. Generally, the ICs were much better correlated with the flux ratios, for particular pairs with coefficients close to 0.90. This indicates that the identified ICs define signatures in the mass distribution of the analytes that bear high metabolic relevance, similarly to analytically derived flux ratios.
Notably, IC6 was almost perfectly correlated with the flux ratio of oxaloacetate derived through the TCA cycle (Figure [6](#F6){ref-type="fig"}). This IC contained high weights in TCA-cycle-derived amino acids signals that are linked to the incorporation of C~2~units from acetyl-CoA (Figure [4](#F4){ref-type="fig"}). As shown above, the projection of a data point on the axis defined by a component reflects the presence of the fluxome signature in its labeling patterns, and hence directly quantifies the occurrence of a particular metabolic activity. When plotting the projection versus the numerical values, the IC6-derived data exhibited a highly linear correlation, while the correlation coefficient was almost halved for PC3, the closest relative to IC6 (Figure [7](#F7){ref-type="fig"}). This confirms numerically the enhanced capacity of ICA to capture essential and independent information for a complex metabolic trait such as the TCA cycle activity. The extraordinarily high correlation coefficient of 0.99 demonstrates that IC6 represents very closely the analytically deduced TCA-cycle flux ratio. This is surprising because IC6 was statistically identified from 265 masses, whereas the flux ratio was calculated on the basis of a large body of biochemical background information \[[@B19],[@B20]\].
Discussion
==========
For the example of central and amino-acid metabolism in *B. subtilis*, we show that fluxome profiling by multivariate statistics from mass isotopomer distribution analysis is meaningful for the discrimination of mutants or conditions on the basis of their metabolic behavior, and applicable to conditions that are inaccessible to previous flux analysis. In sharp contrast to metabolome concentration data \[[@B24],[@B25]\], fluxome profiles contain functional information on the operation of fully assembled networks \[[@B1],[@B4]\]. As shown here by ICA, this approach enables us to distill the essential signatures of independent metabolic activities, and supports the identification of the underlying biochemical causality. Because no model or *a priori*knowledge on the investigated system is required, the metabolic imprints of any tracer atom and molecule can be followed in virtually any biological system, including multicellular organisms in complex multisubstrate media.
Similarly, *a priori*knowledge of the number of ICs to be computed is not a prerequisite. As a matter of fact, the optimal number depends primarily on the labeling patterns and can hardly be estimated from the dataset dimensions. An underestimate will generally leave some relevant signatures unrecognized, whereas an overestimate will lead to an increased fraction of components reflecting measurement or biological noise. Although statistical significance can be assessed with duplicates, this becomes prohibitive with large datasets (that is, hundreds of mutants or analytes) or reduced availability of replicas. The bottleneck resides in the stochastic approach of most ICA algorithms, for which independent runs result in different ICs or ordering thereof. Instead, algorithmic and statistical reliability of the ICs can be evaluated by repeating the estimation several times either with randomly chosen initial guesses or by slightly varying the dataset (bootstrapping \[[@B28]\]), respectively, and then clustering all results to identify robust ICs \[[@B29]\].
Two factors directly affect the results that can be obtained by comparative fluxome profiling: the detected analytes and the choice of isotopic tracer. As well as polymer-based analytes such as the proteinogenic amino acids monitored here, fluxome profiles can be detected in any set of intra- or extracellular metabolites, thereby widening the observable metabolic processes The choice of tracer depends, to some extent, on the metabolic subsystem of interest. Uniformly labeled substrates provide a more global perspective because they allow assessment of the scrambling of any carbon backbone and, in the case of experiments performed in rich media, also allow quantification of the fraction of *de novo*biosynthesis from the tracer relative to the uptake of a medium component. Similarly, uniformly deuterated substrates or ^2^H~2~O are valuable for simultaneously capturing a wide number of ICs that are affected by the release, binding and exchange of water or protons. Substrates that are labeled at specific positions, in contrast, enable deeper interrogation of particular sub-networks, for example, \[1-^13^C\]hexoses for the initial catabolic reactions \[[@B8],[@B19]\] or \[1-^13^C\]aspartate to assess urea cycle activity.
The results also revealed new biological information on pathway activity, function or regulation. First, both glycolysis and the pentose phosphate pathway actively catabolized glucose in the presence of CAA, because the *pgi*and *yqjI*mutant signatures were different from the wild type and from each other. On sorbitol, in contrast, the same mutants were very similar to the wild type, suggesting that both reactions are only marginally involved in catabolism of this sugar. Second, the Krebs cycle flux was similar on glucose and sorbitol (with and without CAA), as deduced from the similarly pronounced signatures of the *sdhC*mutant. Third, absence of the *sdhC*signatures in the Krebs cycle-derived amino acids aspartate and glutamate of the *mdh*mutant when grown with ammonium (but not CAA) indicates activity of the malic enzyme-based pyruvate bypass \[[@B30]\]. Fourth, activity of the NADP-dependent malic enzyme appears to be independent of catabolite repression because pronounced signatures of the *ytsJ*mutant were seen on all substrates. The gluconeogenic phosphoenolpyruvate synthetase Pps, in contrast, was inactive in the presence of the repressing glucose but active on pyruvate or sorbitol. Fifth, as discussed above the data reveal a Krebs cycle-promoting effect of the repressor CggR on sorbitol but not on glucose, most likely through the repression of glycolytic genes \[[@B22]\].
The comparative fluxome profiling presented here complements traditional flux analysis because it enables potentially rapid and automated identification of relevant mutants or conditions from large-scale datasets, for example from entire mutant libraries. The approach is quantitative in terms of the relative difference between variants, but qualitative with respect to the *in vivo*flux. Interesting variants are then subjected to deeper interrogation of the specific metabolic phenomenon identified. Besides mere data mining, fluxome profiling also has the potential to identify complex functional traits in higher cells where current flux methods fail, and possibly even identify the underlying biochemical mechanism of discriminant mass isotope signatures.
Materials and methods
=====================
Strains and growth conditions
-----------------------------
Wild-type *B. subtilis*168 (*trpC2*) \[[@B31]\] and knockout mutants containing an antibiotic marker in single genes \[[@B17]\] were grown in M9 minimal medium \[[@B9]\] at pH 7.0 with 50 mg tryptophan. Six different combinations of ^2^H- or ^13^C-labeled isotopic tracers (3 g/l) and nitrogen sources were used: (i + ii) uniformly ^13^C-labeled \[U-^13^C\]glucose with either 0.5 g/l CAA (Sigma) or 1 g/l NH~4~Cl; (iii + iv) \[U-^13^C\]sorbitol with either 0.5 g/l CAA or 1 g/l NH~4~Cl; (v) \[U-^2^H\]glucose (\[1,2,3,4,5,6,6-^2^H\]glucose) with 1 g/l NH~4~Cl; and (vi) \[3-^13^C\]pyruvate with 1 g/l NH~4~Cl and twofold higher concentrations of phosphate to ensure pH buffering. \[U-^13^C\]glucose (Martek Biosciences), \[U-^13^C\]sorbitol (Omicron Biochemicals), and \[1,2,3,4,5,6,6-^2^H\]glucose (Euriso-Top) were supplemented as 50:50 mixtures of labeled and unlabeled isotopomers. Pyruvate was supplied entirely as the \[3-^13^C\] isotopomer (Euriso-Top).
Aerobic batch cultures were grown in silicone-covered, deep-well microtiter plates at 37°C and 300 rpm in a 5-cm orbital shaker \[[@B18]\]. Frozen stocks were used to inoculate 1 ml LB medium with selective antibiotics. After 10 h of incubation, 10 μl were used to inoculate 1 ml M9 medium with 5 g/l glucose and selective antibiotics, incubated for 12 h, and 10 μl of these precultures were used to inoculate 1.2 ml of M9 medium with isotopic tracers. Cultures were harvested upon entry into stationary phase (assessed by visual evaluation). Because the length of batch growth varied, cultures with CAA, with NH~4~Cl, and with pyruvate were harvested after 10, 14 and 24 h, respectively. Labeling patterns in the analyzed proteinogenic amino acids are rather stable \[[@B10],[@B19]\]; hence differences of a few hours in growth phase at harvest were irrelevant. This was also confirmed in separate (data not shown) and duplicate experiments for each combination of strain and medium that was independently started from culture stocks.
GC-MS analysis and data preprocessing
-------------------------------------
Cell harvest, protein hydrolysis and GC-MS analysis of amino acids were done exactly as described before \[[@B19],[@B32]\]. Amino-acid mass distributions were derived from the spectra after correction for the natural abundance of stable isotopes \[[@B19]\]. Since amino acids are fragmented during electron impact ionization in the MS, we obtained three to five fragments with partially redundant information for each amino acid. For each fragment, a normalized vector *m*~0~, *m*~1~, \..., *m*~n~, expresses the fraction of molecules that are labeled at 0,1, \...,*n*positions, depending on the total number *n*of carbon or hydrogen atoms present. Considering all corrected fragment vectors obtained per sample, a complete dataset typically consisted of about 260 and 330 single mass values from ^13^C and ^2^H experiments, respectively, depending on the quality of the MS measurement.
Multivariate data analysis
--------------------------
To obtain a new representation of the multivariate MS data and to make their essential structure accessible, we applied PCA to the corrected fragment vectors. This approach projects the input variables in an orthogonal space that is spanned by the PCs. Among the infinite number of possibilities, each successive PC is selected to maximize the variance of the projected data and to be orthonormal to the previous ones \[[@B33]\]. Consequently, PCA concentrates the maximum and nonredundant information of the entire dataset in the minimal number of dimensions, and thus is best suited for data compression \[[@B27]\]. The computation was performed with Matlab (The Mathworks) using the *princomp*function of the Statistics toolbox 4.0. No input vectors were eliminated from the dataset to filter outliers in PCA, because this operation affected only PCs with higher order but only marginally PC1 and PC2.
To reveal hidden information in the labeling patterns, the corrected MS vectors were subjected to ICA \[[@B27]\], which is frequently used in the neurosciences \[[@B34],[@B35]\] and in gene-expression studies \[[@B36],[@B37]\]. For ICA, we assume that independent metabolic processes such as reactions or pathways produce characteristic fingerprints in the labeling pattern. These metabolic fingerprints are defined by *m*fundamental components *S*= (*s*~1~, \..., *s*~*m*~)^T^, each of which is represented by a vector of *p*MS-signals. We assumed that the experimental data *X*= (*x*~1~, \..., *x*~n~)^T^, with *n*vectors of *p*corrected MS signals for each mutant/condition, result from a linear combination of the *m*fundamental processes, given by *x*~*i*~= *a*~*i*1~*s*~1~+\...+ *a*~*i*m~*s*~m~. In matrix notation, this leads to *X*~*p*×*n*~= *A*~*p*×*m*~*S*~*m*×*n*~, with *A*as the mixing or loading matrix. ICA seeks to estimate the unknown terms *A*and *S*from the observed values *X*but has different objectives from PCA. Briefly, ICA identifies statistically ICs by selecting those with maximum non-gaussianity \[[@B27]\]. Hence, ICs are nonlinearly decorrelated and assumed to have non-gaussian distributions. Because of the central limit theorem, which states that the sum of non-gaussian random variables is closer to gaussianity than the original ones, ICs are identified by selecting the linear combinations of the observed variables that have maximum non-gaussianity \[[@B27]\]. In particular, we used the publicly available FastICA 2.1 algorithm \[[@B38]\] to estimate the number of components that were equal to the number of strains in the dataset, excluding duplicates. The data dimension was not reduced (by PCA) before IC computation.
Additional data files
=====================
The following additional data is available with the online version of this paper. Additional data file [1](#s1){ref-type="supplementary-material"} contains three figures (Additional Figure 1 shows the mass distribution in the ^2^H experiment; Additional Figure 2 shows mutant discrimination by PCA (less relevant than by ICA); Additional Figure 3 is a complete representation of the 660 ICs (10 ICs in 6 experiments for 11 strains). All the raw data is contained in six Excel tables in Additional data file [2](#s2){ref-type="supplementary-material"}.
Supplementary Material
======================
::: {.caption}
###### Additional data file 1
Three additional figures (Additional Figure 1 shows the mass distribution in the ^2^H experiment; Additional Figure 2 shows mutant discrimination by PCA (less relevant than by ICA); Additional Figure 3 is a complete representation of the 660 ICs (10 ICs in 6 experiments for 11 strains)
:::
::: {.caption}
######
Click here for additional data file
:::
::: {.caption}
###### Additional data file 2
All the raw data contained in six Excel tables
:::
::: {.caption}
######
Click here for additional data file
:::
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Simplified biochemical reaction network of *Bacillus subtilis*central carbon metabolism. Gray arrows outline the biosynthesis of precursor amino acids that are indicated by their one-letter code. Amino acids in square brackets were not detected. Black dashed arrows illustrate the uptake of substrates. Black boxes highlight pathways or reactions that are affected in the mutants used (see also Table 1). G6P, glucose 6-phosphate; F6P, fructose 6-phosphate; T3P, triose phosphate; PGA, phosphoglycerate; PEP, phosphoenolpyruvate; PYR, pyruvate; OAA, oxaloacetic acid; MAL, malic acid; OGA, 2-oxoglutarate.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Fraction of amino acids that were synthesized *de novo*from \[U-^13^C\]glucose (white bars) and sorbitol (gray bars) in batch experiments supplemented with 0.5 g/l casein hydrolysate. Amino acids are given in the one-letter code.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Comparison of labeling profiles in amino acids of *B. subtilis*mutants that were normalized by subtraction with the wild-type values obtained under the same condition, as obtained from five different medium compositions. The line deviates above (or below) the null line when an amino acid (represented by their one letter code at the top of the first panel) mass is more (or less) abundant in the mutant than in the parent. For each amino acid, the available data points are in the order of their total mass fragment. Gray areas represent the deviation of the normalized values, based on duplicate analyses of mutant and wild type. To reduce the dimension of the data for visual comparison, we excluded those values that, on average, accounted for less than 5% of the fragment pool in all mutants under a given condition.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Weights of input variables. Weights of input variables in the first eight components obtained by **(a)**PCA and **(b)**ICA from the corrected MS data of the \[U-^13^C\]glucose experiment with ammonium.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Fluxome profiling by independent component analysis of *B. subtilis*mutants grown on a 50:50 mixture of \[U-^13^C\]- and naturally labeled sorbitol with ammonium. **(a)** Weights of input variables (amino-acid mass-distribution vectors) in the mixing matrix of 10 ICs. **(b)** Projections (on *x*-axis) of samples on the IC shown in (a). The vertical line is drawn to intersect the average of the wild-type values. **(c)** Wild-type-normalized labeling profiles. Colors are used to highlight those aspects of the amino-acid profiles that were identified by ICA as relevant for the discrimination of the samples (b) along selected components.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
Correlation between analytically derived metabolic flux ratios (on the *y*-axis) \[19\] and the projections of the data on the first eight components obtained by PCA and ICA for the \[U-^13^C\]glucose experiment with ammonium. The brightness reflects the correlation coefficient, with black and white corresponding to values of 0 and 1, respectively. For coefficients higher than 0.8, the numerical value is reported. ub, upper bound; lb, lower bound.
:::
:::
::: {#F7 .fig}
Figure 7
::: {.caption}
######
Weights of input variables in the component that is linked to TCA cycle activity, identified by either **(a)**PCA or **(b)**ICA from the \[U-^13^C\]glucose experiment with ammonium. In **(c)**and **(d)**, the projections of the mutant data on the component shown in (a) and (b), respectively, were plotted versus the analytically derived fraction of oxaloacetate (OAA) originating from TCA cycle \[19\]. The correlation coefficients are for linear fits.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
*B. subtilis*strains used
:::
Strain Description of deleted gene
----------- --------------------------------------
Wild-type 168 *trpC2*
*pgi* P-glucoisomerase
*yqjI* 6-P-gluconate dehydrogenase
*sdhC* Succinyl-CoA dehydrogenase component
*ytsJ* Malic enzyme
*mdh* Malate dehydrogenase
*pps* PEP synthetase
*ccpA* Main carbon catabolite repressor
*cggR* Repressor of the *gapA*operon
*glcP* Hexose/H^+^symporter
*glcR* Repressor of PTS system
Strains were provided by S. Aymerich (INRA, CNRS, Thiverval-Grignon, France) and K. Kobayashi (Nara Institute of Science and Technology, Nara, Japan) \[17\].
:::
|
PubMed Central
|
2024-06-05T03:55:51.908146
|
2004-11-16
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545802/",
"journal": "Genome Biol. 2004 Nov 16; 5(12):R99",
"authors": [
{
"first": "Nicola",
"last": "Zamboni"
},
{
"first": "Uwe",
"last": "Sauer"
}
]
}
|
PMC545934
|
Background
==========
Increased airway narrowing in response to nonspecific stimuli is a characteristic feature of human obstructive diseases, including bronchial asthma. This abnormality is an important symptom of the disease, although the pathophysiological variations leading to the hyperresponsiveness are unclear now. Several mechanisms have been suggested to explain the airway hyperresponsiveness (AHR), such as alterations in the neural control of airway smooth muscle \[[@B1]\], increased mucosal secretions \[[@B2]\], and mechanical factors related to remodeling of the airways \[[@B3]\]. In addition, it has also been suggested that one of the factors that contribute to the exaggerated airway narrowing in asthmatics is an abnormality of the nature of airway smooth muscle \[[@B4],[@B5]\]. Rapid relief from airway limitation in asthmatic patients by β-stimulant inhalation may also suggest an involvement of augmented airway smooth muscle contraction in the airway obstruction. Thus, it may be important for development of asthma therapy to understand changes in the contractile signaling of airway smooth muscle cells associated with the disease.
Smooth muscle contraction including airways is mainly regulated by an increase in cytosolic Ca^2+^concentration in myocytes. Recently, additional mechanisms have been suggested in agonist-induced smooth muscle contraction by studies in which the simultaneous measurements of force development and intracellular Ca^2+^concentration, and chemically permeabilized preparations in various types of smooth muscles were used. It has been demonstrated that agonist stimulation increases myofilament Ca^2+^sensitivity in permeabilized smooth muscles of the rat coronary artery \[[@B6]\], guinea pig vas deferens \[[@B7]\], canine trachea \[[@B8]\] and rat bronchus \[[@B9]\]. Although the detailed mechanism is not fully understood, a participation of RhoA, a monomeric GTP binding protein, in the agonist-induced Ca^2+^sensitization has been suggested by many investigators \[[@B10]\]. Moreover, an augmented RhoA-mediated Ca^2+^sensitization in smooth muscle contraction has been reported in experimental animal models of diseases such as hypertension \[[@B11]-[@B13]\], coronary \[[@B14]-[@B16]\] and cerebral \[[@B17]-[@B19]\] vasospasm. It is thus possible that RhoA-mediated signaling is the key for understanding the abnormal contraction of diseased smooth muscles.
Here, we show an increased acetylcholine (ACh)-induced contraction of bronchial smooth muscle (BSM) isolated from repeatedly ovalbumin (OA)-challenged BALB/c mice, which have been reported to have in vivo AHR \[[@B20]\]. A participation of RhoA-mediated Ca^2+^sensitization in the augmented ACh-induced contraction of BSM was demonstrated in this animal model of AHR.
Methods
=======
Sensitization and antigenic challenge
-------------------------------------
Male BALB/c mice (6-week old, specific pathogen-free; Charles River Japan, Inc., Kanagawa, Japan) were used. All experiments were approved by the Animal Care Committee at the Hoshi University (Tokyo, Japan). Preparation of a murine model of allergic bronchial asthma, which has in vivo airway hyperresponsiveness (AHR), was performed as described by Kato *et al*. \[[@B20]\]. In brief, mice were actively sensitized by intraperitoneal injections of 8 μg ovalbumin (OA; Seikagaku Co., Tokyo, Japan) with 2 mg Imject Alum (Pierce Biotechnology, Inc., Rockfold, IL, USA) on day 0 and day 5. The sensitized mice were challenged with aerosolized OA-saline solution (5 mg/ml) for 30 min on days 12, 16 and 20. A control group of mice received the same immunization procedure but inhaled saline aerosol instead of OA challenge. The aerosol was generated with an ultrasonic nebulizer (Nihon Kohden, Tokyo, Japan) and introduced to a Plexiglas chamber box (130 × 200 mm, 100 mm height) in which the mice were placed.
Determination of intact bronchial smooth muscle (BSM) responsiveness
--------------------------------------------------------------------
Twenty-four h after the last antigen challenge, the mice were sacrificed by exsanguination from abdominal aorta under urethane (1.6 g/kg, *i.p*.) anesthesia. Then the airway tissues under the larynx to lungs were immediately removed. About 3 mm length of the left main bronchus (about 0.5 mm diameter) was isolated and epithelium was removed by gently rubbing with keen-edged tweezers \[[@B21]\]. The resultant tissue ring preparation was then suspended in a 5 ml-organ bath by two stainless-steel wires (0.2 mm diameter) passed through the lumen. For all tissues, one end was fixed to the bottom of the organ bath while the other was connected to a force-displacement transducer (TB-612T, Nihon Kohden) for the measurement of isometric force. A resting tension of 0.5 g was applied. The buffer solution contained modified Krebs-Henseleit solution with the following composition (mM); NaCl 118.0, KCl 4.7, CaCl~2~2.5, MgSO~4~1.2, NaHCO~3~25.0, KH~2~PO~4~1.2 and glucose 10.0. The buffer solution was maintained at 37°C and oxygenated with 95% O~2~-5% CO~2~. The BSM responsiveness to exogenously applied Ca^2+^in acetylcholine (ACh)-stimulated or high K^+^-depolarized muscle was determined as previously \[[@B22]\]. In brief, after an equilibration period, the organ bath solution was replaced with Ca^2+^-free solution containing 10^-6^M nicardipine with the following composition (mM); NaCl 122.4, KCl 4.7, MgSO~4~1.2, NaHCO~3~25.0, KH~2~PO~4~1.2, glucose 10.0 and EGTA 0.05. Fifteen min later, 1 mM ACh was added and, after attainment of a plateau (almost baseline level) response to ACh, a cumulative concentration-response curve for Ca^2+^(0.1--6.0 mM) was made. A higher concentration of Ca^2+^was added after the response to the previous concentration reached a plateau. In another series of experiments, bronchial smooth muscles were depolarized with 60 mM K^+^, instead of ACh, in the presence of 10^-6^M atropine and in the absence of nicardipine in the Ca^2+^-free solution. All these functional studies were performed in the presence of 10^-6^M indomethacin. The concentration of indomethacin had no effect both on baseline tension and on the ACh- and high K^+^-induced constrictions of BSMs (data not shown).
BSM permeabilized fiber experiments
-----------------------------------
To determine the change in Ca^2+^sensitization of BSM contraction, permeabilized BSMs were prepared as described previously \[[@B21]\] with minor modification. In brief, 24 h after the last antigen challenge, the left main bronchus was isolated as described above and cut into ring strips (about 200 μm width, 500 μm diameter). The epithelium was removed by gently rubbing with keen-edged tweezers. The ring strips were then permeabilized by a 30-min treatment with 83.3 μg/ml α-toxin (Sigma, St. Louis, MO, USA) in the presence of Ca^2+^ionophore A23187 (10 μM, Sigma) at room temperature in relaxing solution. Relaxing solution contained: 20 mM PIPES, 7.1 mM Mg^2+^-dimethanesulfonate, 108 mM K^+^-methanesulfonate, 2 mM EGTA, 5.875 mM Na~2~ATP, 2 mM creatine phosphate, 4 U/ml creatine phosphokinase, 1 μM carbonyl cyanide p-trifluoromethoxyphenylhydrazone (FCCP) and 1 μg/ml E-64 (pH 6.8) containing 10 μM A23187. Free Ca^2+^concentration was changed by adding an appropriate amount of CaCl~2~. The apparent binding constant of EGTA for Ca^2+^was considered to be 10^6^M^-1^\[[@B23]\]. The permeabilized muscle strip was then suspended in a 400-μL organ bath at room temperature. The contractile force developed was measured by an isometric transducer (T7-8-240; Orientec, Tokyo, Japan) under a resting tension of 50 mg. To determine the involvement of RhoA in the ACh-induced myofilament Ca^2+^sensitization, the α-toxin-permeabilized muscle strips were treated with *Clostridium botulinum*C3 exoenzyme (10 μg/ml; Calbiochem-Novabiochem Corp., La Jolla, CA) in the presence of 100 μM NAD for 20 min at room temperature.
Determination of RhoA protein level in BSM
------------------------------------------
Protein samples of BSMs were prepared as previously \[[@B21]\]. In breif, the airway tissues below the main bronchi to lungs were removed and immediately soaked in ice-cold, oxygenated Krebs-Henseleit solution. The airways were carefully cleaned of adhering connective tissues, blood vessels and lung parenchyma under a stereomicroscopy. The epithelium was removed as much as possible by gently rubbing with keen-edged tweezers \[[@B21]\]. Then the bronchial tissue (containing the main and intrapulmonary bronchi) segments were quickly frozen with liquid nitrogen, and the tissue was crushed to pieces by CryopressTM (CP-100W; Microtec, Co. Ltd., Chiba, Japan: 15 sec × 3). The tissue powder was homogenized in ice-cold tris(hydroxymethyl)aminomethane (Tris, 10 mM; pH 7.5) buffer containing 5 mM MgCl~2~, 2 mM EGTA, 250 mM sucrose, 1 mM dithiothreitol, 1 mM 4-(2-aminoethyl)benzenesulfonyl fluoride, 20 μg/ml leupeptin, 20 μg/ml aprotinin, 1% Triton X-100 and 1% sodium cholate. The tissue homogenate was then centrifuged (3,000 g, 4°C for 15 min) and the resultant supernatant was stored at -85°C until use. To determine the level of RhoA protein in BSMs, the samples (10 μg of total protein per lane) were subjected to 15% SDS-PAGE and the proteins were then electrophoretically transferred to a PVDF membrane. After blocking with 3% gelatin, the PVDF membrane was incubated with polyclonal rabbit anti-RhoA antibody (1:3,000; Santa Cruz Biotechnology, Inc., Santa Cruz, CA, USA). Then the membrane was incubated with horseradish peroxidase-conjugated goat anti-rabbit IgG (1:2,500 dilution; Amersham Biosciences, Co., Piscataway, NJ, USA), detected by an enhanced chemiluminescent system (Amersham Biosciences, Co.) and analyzed by a densitometry system. Thereafter, the primary and secondary antibodies were stripped and the membrane was reprobed by using monoclonal mouse anti-glyceraldehyde-3-phosphate dehydrogenase (GAPDH; 1:3,000 dilution; Chemicon International, Inc., Temecula, CA, USA) to confirm the same amount of proteins loaded.
Determination of active form of RhoA in BSM
-------------------------------------------
The active form of RhoA, GTP-bound RhoA, in BSMs was measured by RhoA pull down assay. In brief, bronchial tissues containing the main and intrapulmonary bronchi were isolated as described above. The isolated bronchial tissues were equilibrated in oxygenated Krebs-Henseleit solution at 37°C for 1 hr. After the equilibration period, the tissues were stimulated by ACh (10^-3^M for 10 min) and were quickly frozen with liquid nitrogen. The tissues were then lysed in lysis buffer with the following composition (mM); HEPES 25.0 (pH 7.5), NaCl 150, IGEPAL CA-630 1%, MgCl~2~10.0, EDTA 1.0, glycerol 10%, NaF 25.0, sodium orthovanadate 1.0 and peptidase inhibitors. Active RhoA in tissue lysates (200 μg protein) was precipitated with 25 μg GST-tagged Rho binding domain (amino acids residues 7--89 of mouse rhotekin; Upstate, Lake Placid, NY, USA), which was expressed in *Escherichia coli*and bound to glutathione-agarose beads. The precipitates were washed three times in lysis buffer, and after adding the SDS loading buffer and boiling for 5 min, the bound proteins were resolved in 15% polyacrylamide gels, transferred to nitrocellulose membranes, and immunoblotted with anti-RhoA antibody as described above.
Determination of phosphorylation of myosin phosphatase and myosin light chain in BSM
------------------------------------------------------------------------------------
Phosphorylated proteins were detected by using the fluorescent Pro-Q-Diamond dye (Molecular Probes, Eugene, OR, USA), which can directly detect phosphate groups attached to tyrosine, serine or threonine residues in gels \[[@B24]\]. In brief, bronchial tissue lysates (50 μg protein) with SDS loading buffer prepared as described above were resolved in 10 -- 20% gradient polyacrylamid gels (Atto Co., Tokyo, Japan). Proteins were fluorescently stained by fixing the gels in 50% methanol and 10% acetic acid for 1 h. The gels were washed with deionised water for 20 min, stained with Pro-Q-Diamond for 1.5 h and destained by three washes in 4% acetonitrile in 50 mM sodium acetate, pH 4.0, for 2 h. Gels were scanned with a fluorimager, a Typhoon 9410 laser scanner (Amersham Biosciences, Co.), with excitation at 532 nm and a 580 nm band pass emission filter for Pro-Q-diamond dye detection. Phosphorylated proteins were quantified densitometrically with the ImageQuant software (Amersham Biosciences, Co.). After scanning, the gels were washed with deionised water for 30 min and incubated in 0.7% glycine-0.2% SDS in 0.3% Tris buffer for 15 min. The proteins were then electrophoretically transferred to a PVDF membrane and immunoblottings for myosin phosphatase target subunit 1 (MYPT1; polyclonal goat anti-MYPT1 antibody; 1:1000; Santa Cruz Biotechnology, Inc.), GAPDH and myosin light chain (MLC; polyclonal rabbit anti-MLC2 antibody; 1:3000; Santa Cruz Biotechnology, Inc) were performed as described above.
Statistical analyses
--------------------
All the data are expressed as the mean ± S.E. Statistical significance of difference was determined by unpaired Student\'s *t*-test, Bonferroni/Dunn\'s test or two-way analysis of variance (ANOVA).
Results
=======
Contractile response of intact BSM preparations
-----------------------------------------------
Under Ca^2+^-free condition (in the presence of 10^-6^M nicardipine and 0.05 mM EGTA), ACh (10^-3^M) generated a transient phasic contraction in all BSM preparations used. The generated tension of BSM from the repeatedly OA-challenged mice (69 ± 12 mg, N = 6) was significantly greater than that from the sensitized control animals (20 ± 12 mg, N = 6; P \< 0.05). The concentration of nicardipine used in the present study completely blocked high K^+^(10--90 mM)-induced BSM contraction in Ca^2+^(2.5 mM) containing normal Krebs-Henseleit solution (data not shown), indicating that voltage-dependent Ca^2+^channels were completely blocked in this condition.
The tension returned to baseline level within 5 min after the ACh application, and then the contraction induced by cumulatively administered Ca^2+^was measured. Figure [1A](#F1){ref-type="fig"} shows the concentration-response curves to Ca^2+^of murine BSMs that were preincubated with nicardipine (10^-6^M) and ACh (10^-3^M) under Ca^2+^-free (0.05 mM EGTA) condition. Addition of Ca^2+^induced a concentration-dependent BSM contraction in both the sensitized control and OA-challenged groups. The contractile response to Ca^2+^of the ACh-stimulated BSMs from the repeatedly OA-challenged mice was markedly augmented as compared to that from the sensitized control animals. By contrast, no significant difference in the response to Ca^2+^of BSMs depolarized with 60 mM K^+^(in the absence of nicardipine and presence of 10^-6^M atropine) was observed between groups (Fig. [1B](#F1){ref-type="fig"}). Likewise, the ACh (10^-7^--10^-3^M) concentration-response curve determined in normal Krebs-Henseleit solution (2.5 mM Ca^2+^) was significantly shifted upward in BSMs from the OA-challenged mice as compared with that from the sensitized control animals, whereas no significant difference in the contractile response induced by isotonic high K^+^(10--90 mM) was observed between groups (data not shown).
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Cumulative concentration-response curves to Ca^2+^of bronchial rings obtained from sensitized control (Control; *open circles*) and repeatedly ovalbumin-challenged (OA-challenged; *closed circles*) mice. Bronchial rings were preincubated with 10^-3^M acetylcholine (ACh) in the presence of 10^-6^M nicardipine (*A*) or isotonic 60 mM K^+^in the presence of 10^-6^M atropine (*B*) in Ca^2+^-free, 0.05 mM EGTA solution. Each point represents the mean ± S.E. from 6 experiments. The Ca^2+^-induced contraction of the ACh-stimulated bronchial smooth muscles was significantly augmented in the OA-challenged group (*A*; P \< 0.05 by ANOVA), whereas no significant change in the Ca^2+^-induced contraction of the high K^+^-depolarized muscles was observed between groups (*B*).
:::
:::
Contractile response of α-toxin-permeabilized BSM preparations
--------------------------------------------------------------
The BSM contractility was also determined by using α-toxin-permeabilized BSM preparations. In all BSM preparations treated with a-toxin, application of free Ca^2+^(pCa = 6.5, 6.3, 6.0, 5.5 and 5.0) induced a concentration-dependent reproducible contractile response, indicating successful permeabilization. In the α-toxin-permeabilized BSM, no significant difference in the Ca^2+^responsiveness or the maximal contractile response induced by pCa 5.0 (Emax) was observed between the sensitized control (pEC~50~\[Ca^2+^(M)\] = 5.67 ± 0.04, Emax = 26.7 ± 1.2 mg; N = 6) and repeatedly OA-challenged (pEC~50~\[Ca^2+^(M)\] = 5.78 ± 0.15, Emax = 22.8 ± 5.9 mg; N = 6) groups. In both groups, when the Ca^2+^concentration was clamped at pCa 6.0, application of ACh (10^-5^--10^-3^M) in the presence of GTP (10^-4^M) caused a further contraction, *i.e*., ACh-induced Ca^2+^sensitization, in an ACh concentration-dependent manner (Fig. [2](#F2){ref-type="fig"}). The ACh-induced Ca^2+^sensitization was significantly greater in the repeatedly OA-challenged group (Fig. [2B](#F2){ref-type="fig"}).
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Acetylcholine (ACh)-induced Ca^2+^sensitization of murine bronchial smooth muscle. (*A*) A typical recording of contraction induced by Ca^2+^(pCa 6.0 and 5.0) and ACh (10^-5^--10^-3^M) with guanosine triphosphate (GTP; 10^-4^M) in α-toxin-permeabilized bronchial smooth muscle isolated from a sensitized control mouse. In the presence of GTP, ACh induced further contractions even in the constant Ca^2+^concentration at pCa 6.0, *i.e*., ACh-induced Ca^2+^sensitization, in an ACh-concentration-dependent manner. (*B*) Concentration-response curves for ACh (10^-5^--10^-3^M)-induced Ca^2+^sensitization in α-toxin-permeabilized bronchial smooth muscle isolated from sensitized control (Control; *open circles*) and repeatedly ovalbumin-challenged (OA-challenged; *closed circles*) mice. The data are expressed as percentage increase in tension induced by ACh (10^-5^--10^-3^M) in the presence of Ca^2+^(pCa 6.0) and GTP (10^-4^M) from the sustained contraction induced by pCa 6.0. Each point represents the mean ± S.E. from 6 experiments. The ACh-induced Ca^2+^sensitization of bronchial smooth muscle contraction was significantly augmented in the OA-challenged mice (\*P \< 0.05 vs. Control group by unpaired Student\'s *t*-test).
:::
:::
To determine an involvement of RhoA protein in the ACh-induced Ca^2+^sensitization, the effect of pretreatment with C3 exoenzyme on the contractile response of the α-toxin-permeabilized BMS was also investigated. The C3 treatment alone had no significant effect on the Ca^2+^responsiveness of α-toxin-permeabilized BSMs in any groups (data not shown). However, the ACh (10^-3^M, in the presence of 10^-4^M GTP)-induced Ca^2+^sensitizing effect was inhibited by treatment with C3 in both the sensitized control and OA-challenged groups (Fig. [3](#F3){ref-type="fig"}). Interestingly, the remaining C3-insensitive component of the ACh-induced Ca^2+^sensitization was the same level between groups, whereas the Ca^2+^sensitization before treatment with C3 was significantly greater in BSMs of the OA-challenged mice (Fig. [3B](#F3){ref-type="fig"}). These findings indicate that the C3-sensitive Ca^2+^sensitization, probably mediated by RhoA \[[@B25],[@B26]\], might be augmented in BSMs of the OA-challenged AHR mice.
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Effect of *Clostridium botulinum*C3 exoenzyme, an inhibitor of RhoA protein, on the acetylcholine (ACh)-induced Ca^2+^sensitization of the α-toxin-permeabilized bronchial smooth muscle of mice. (*A*) Typical recordings of contraction induced by Ca^2+^(pCa 6.0 and 5.0) and ACh (10^-3^M) with guanosine triphosphate (GTP; 10^-4^M) in α-toxin-permeabilized bronchial smooth muscle isolated from a sensitized control mouse. In the presence of GTP, ACh induced a further contraction even in the constant Ca^2+^concentration at pCa 6.0, *i.e*., ACh-induced Ca^2+^sensitization (*a*). The ACh-induced Ca^2+^sensitization was re-estimated after treatment with C3 exoenzyme (10 μg/mL, for 20 min; *b*). (*B*) Summary of the effects of C3 exoenzyme on the ACh-induced Ca^2+^sensitization of bronchial smooth muscle contraction in the sensitized control (Control) and repeatedly ovalbumin (OA)-challenged (OA-challenged) mice. The data are expressed as percentage increase in tension induced by ACh (in the presence of Ca^2+^and GTP) from the sustained contraction induced by pCa 6.0. Each column represents the mean ± S.E. from 6 experiments. \*P \< 0.05 vs. Control group (Before C3) and \#P \< 0.05 vs. respective Before C3 group by Bonferroni/Dunn\'s test.
:::
:::
Upregulation of RhoA protein in BSMs of OA-challenged mice
----------------------------------------------------------
The expression of RhoA protein in BSM homogenates was assessed by using immunoblotting. As shown in Fig. [4A](#F4){ref-type="fig"}, immunoblotting with the antibody against RhoA gave a single 21 kD band, indicating the expression of RhoA protein in murine BSM. The level of RhoA protein in samples of the OA-challenged mice was significantly increased as compared with that of the sensitized control animals. Moreover, the GTP-bound active form of RhoA in ACh-stimulated BSMs was markedly increased in the OA-challenged mice (Fig. [5](#F5){ref-type="fig"}).
::: {#F4 .fig}
Figure 4
::: {.caption}
######
The levels of RhoA protein in the bronchi obtained from the sensitized control (Control) and repeatedly ovalbumin (OA)-challenged (OA-challenged) mice. (*A*) Typical immunoblots. *Lane 1*; Control, *lane 2*; OA-challenged, and GAPDH; glyceraldehyde-3-phosphate dehydrogenase. The bands were analyzed by a densitometer and normalized by the intensity of corresponding GAPDH band, and the data are summarized in *B*. Each column represents the mean ± S.E. from 5 experiments. The expression level of RhoA protein in the bronchi was significantly increased in the OA-challenged group (\*P \< 0.001 vs. Control group by unpaired Student\'s *t*-test).
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Representative immunoblots showing activation of RhoA in acetylcholine (ACh)-stimulated bronchi obtained from the sensitized control (Control) and repeatedly ovalbumin (OA)-challenged (Challenged) mice. Isolated bronchial tissues were incubated for 10 min in the absence (-) or presence (+) of 10^-3^M ACh (*see Methods*). Tissues were then rapidly lysed, GTP-bound active form of RhoA was pulled down with GST-tagged Rho binding domain of rhotekin, and RhoA was visualized by Western blotting. The respective blot of total RhoA in each sample is also shown. The GTP-bound RhoA in ACh-stimulated bronchi was markedly increased in the OA-challenged mice.
:::
:::
Augmented ACh-induced phosphorylation of MLC in BSMs of OA-challenged mice
--------------------------------------------------------------------------
Figure [6](#F6){ref-type="fig"} shows the levels of total and phosphorylated MLCs in BSMs determined by immunoblotting and Pro-Q Diamond dye staining, respectively. Immunoblotting with the antibody against MLC protein revealed a single 20 kD band, which contains both phosphorylated and non-phosphorylated MLC proteins (total MLC). The levels of total MLC were the same between groups (Fig. [6](#F6){ref-type="fig"}, *middle panel*). In the Pro-Q Diamond dye-stained gels, there were several positive bands, *i.e*., phosphorylated proteins \[[@B24]\], in the ACh-stimulated BSM samples. Among them, a 20 kD band corresponding to MLC was distinctly found (Fig. [6](#F6){ref-type="fig"}, *bottom panel*). The ACh-induced phosphorylation of MLC in BSMs of OA-challenged mice was markedly augmented as compared with those of control animals. A Pro-Q Diamond dye-positive 140 kD band probably corresponding to MYPT1, *i.e*., phosphorylated MYPT1, was also found in the ACh-stimulated BSM samples and was increased in the OA-challenged group (data not shown).
::: {#F6 .fig}
Figure 6
::: {.caption}
######
Representative photographs showing phosphorylation of myosin light chain (MLC) in acetylcholine (ACh)-stimulated bronchi obtained from the sensitized control (Cont) and repeatedly ovalbumin-challenged (OA) mice. Isolated bronchial tissues were incubated for 10 min in the absence (non-stimulated; NS) or presence of 10^-3^M ACh (*see Methods*). The electrophoretically separated proteins on gels were stained by Pro-Q Diamond dye, which can detect phosphorylated proteins specifically and quantitatively. After detection of phosphorylated proteins, immunoblotting for MLC was performed to detect total (phosphorylated and non-phosphorylated) MLC. The respective Pro-Q Diamond dye-positive band (*bottom panel*), which has same molecular weight with MLC visualized by immunoblotting (*middle panel*), in each sample was determined as phosphorylated MLC (p-MLC). The ACh-induced phosphorylation of MLC was augmented in the OA-challenged mice whereas the total MLC levels were equal to the control.
:::
:::
Discussion
==========
An *in vivo*AHR accompanied by increased IgE production and pulmonary eosinophilia has been demonstrated in the actively sensitized and repeatedly OA-challenged BALB/c strain of mice \[[@B20]\]. By using the same sensitization and challenge protocol in BALB/c mice, the current study demonstrated an increased BSM contractility in ACh-stimulated, but not in high K^+^-depolarized (without receptors stimulation), intact muscle strips of the repeatedly OA-challenged mice (Fig. [1](#F1){ref-type="fig"}). Likewise, the ACh-induced, C3-sensitive Ca^2+^sensitization of BSM contraction was significantly augmented in α-toxin-permeabilized BSMs of the OA-challenged mice (Figs. [2](#F2){ref-type="fig"} and [3](#F3){ref-type="fig"}), whereas the contraction induced by Ca^2+^itself was the same as the control level (see Results section). These findings suggest that the C3-sensitive, RhoA-mediated Ca^2+^sensitization might be augmented in BSMs of the OA-challenged AHR mice. Indeed, the current study also demonstrated a marked increase in the expression and activation of RhoA protein in BSMs of the AHR mice (Fig. [4](#F4){ref-type="fig"} and [5](#F5){ref-type="fig"}).
In the present study, no significant difference in the Ca^2+^-induced contraction (in the absence of ACh and GTP) of α-toxin-permeabilized BSMs was observed between groups (see Result section), indicating that the contents of typical contractile elements such as calmodulin, myosin light chain (MLC; Fig. [6](#F6){ref-type="fig"}) and SM α-actin might be the same as control even in the BSMs of the OA-challenged mice. Moreover, the results also indicate that the downstream signaling activated by Ca^2+^-calmodulin complex, including phosphorylation of MLC via activation of MLC kinase, might be in an analogous fashion between groups. The results that the contractile response of intact (non-permeabilized) BSMs induced by high K^+^depolarization was not changed after OA challenge also support our speculation. Thus, the baseline Ca^2+^sensitivity of contractile elements themselves in BSM cells is unlikely to change in AHR.
By contrast with the contraction induced by Ca^2+^itself, the ACh-stimulated contraction of intact BSM strips from the OA-challenged mice was significantly augmented as compared to that from the sensitized control animals (Fig. [1](#F1){ref-type="fig"}). BSMs are predominantly innervated by vagal efferent nerves, which release ACh when stimulated leading to an activation of muscarinic ACh receptors. The activation of muscarinic receptors existing on BSM, which are mainly thought to be of the M~3~subtype \[[@B27]\], results in BSM contraction by increasing intracellular Ca^2+^concentration through Ca^2+^release from sarcoplasmic reticulum and Ca^2+^influx from voltage-dependent (nicardipine-sensitive) and receptor-operated (nicardipine-insensitive) Ca^2+^channels \[[@B28]\]. Therefore, one possible explanation for the increased response to ACh of OA-challenged BSMs may be attributable to an enhanced Ca^2+^mobilization in BSM cells. However, the possibility might be denied by the current result that the ACh-induced contraction of α-toxin-permeabilized BSMs from the OA-challenged mice was significantly augmented as compared with that from the control animals even at a constant Ca^2+^concentration (pCa 6.0; Fig. [2B](#F2){ref-type="fig"}). Moreover, it has also been reported that there is no difference between normal and antigen-induced AHR animals in ACh-induced increase in intracellular Ca^2+^concentration in BSMs, irrespective of a great difference in ACh-induced BSM contraction \[[@B29],[@B30]\].
In addition to the classical Ca^2+^-mediated contractile signaling in smooth muscle, it has been demonstrated that agonist stimulation increases myofilament Ca^2+^sensitivity in various types of smooth muscles including airways \[[@B8],[@B10],[@B21],[@B31]\]. Recent studies suggest a participation of RhoA in the agonist-induced Ca^2+^sensitization of smooth muscle contraction \[[@B10]\]. Hirata *et al*. \[[@B32]\] firstly reported an involvement of RhoA in the mechanism for the increase in Ca^2+^sensitization in smooth muscle. It was then shown that RhoA is responsible for the inhibition of MLC phosphatase through the activation of Rho-associated kinases \[[@B33]\]. The present study demonstrated an ACh-induced Ca^2+^sensitization in murine BSM contraction (Fig. [2](#F2){ref-type="fig"}),which is sensitive to C3 exoenzyme (Fig. [3](#F3){ref-type="fig"}), in the α-toxin-permeabilized BSMs. Furthermore, western blot analysis clearly demonstrated a distinct expression of RhoA protein in BSMs of mice (Fig. [4](#F4){ref-type="fig"}). Collectively, these findings firstly demonstrated a participation of RhoA-mediated Ca^2+^sensitization in ACh-induced BSM contraction in mice.
One of the important findings in the present study is that the C3-sensitive, RhoA-mediated Ca^2+^sensitization in ACh-induced contraction was significantly augmented in BSMs of the repeatedly OA-challenged AHR mice (Figs. [2](#F2){ref-type="fig"} and [3](#F3){ref-type="fig"}). Moreover, the protein level of RhoA in BSMs of the AHR mice was significantly increased (Fig. [4](#F4){ref-type="fig"}). Thus, the current study demonstrated an augmentation of ACh-induced, RhoA-mediated Ca^2+^sensitization of BSM contraction, which coincides with enhanced protein expression of RhoA, in antigen-induced AHR. Although the mechanism(s) of up-regulation of RhoA in OA-challenged BSMs is not known here, inflammatory cytokines such as tumor necrosis factor-α \[[@B34]\], which is also demonstrated in airways of this murine model of asthma (unpublished data), may be involved in. On the other hand, it has been reported that an introduction of active forms of RhoA to permeabilized smooth muscle induced contractile response \[[@B32],[@B35]\]. It is thus likely that ACh stimulation activates the upregulated RhoA (Fig. [5](#F5){ref-type="fig"}), resulting in a greater phosphorylation of MLC (Fig. [6](#F6){ref-type="fig"}) and contraction of BSMs in AHR mice.
An increase in responsiveness to muscarinic agonists of airway smooth muscle has been demonstrated in animal models of AHR \[[@B21],[@B22],[@B36],[@B37]\] and asthmatic patients \[[@B38]\], although no change in the levels of plasma membrane receptors was observed \[[@B36],[@B37],[@B39]\]. Moreover, the agonist-induced increase in cytosolic Ca^2+^level was within normal level even in the hyperresponsive BSMs \[[@B29],[@B30]\]. Taken together with our current findings, it is likely that the enhanced contractility to agonists reflects, at least in part, the augmentation of muscarinic receptor- and RhoA-mediated Ca^2+^sensitization, although the mechanism(s) for activation of RhoA by ACh is still unclear. If RhoA proteins are activated by receptors other than muscarinic receptor, it might account for the \'non-specific\' AHR, which is a common feature of allergic asthmatics. Indeed, the BSMs of the OA-challenged mice also have hyperresponsiveness to endothelin-1 \[[@B40]\], which has been reported to activate RhoA via its own receptors \[[@B41]\].
An upregulation of RhoA/Rho-kinase associated with the augmented smooth muscle contractility has also been reported in rat myomertium during pregnancy \[[@B42],[@B43]\], arterial smooth muscle of spontaneously hypertensive rats \[[@B12]\], coronary vasospasm in pigs \[[@B16]\], dog femoral artery in heart failure \[[@B44]\], and BSMs in rat experimental asthma \[[@B21]\]. Thus, the upregulation of RhoA might be widely involved in the enhanced contraction of the diseased smooth muscles including the BSMs in AHR over species.
Conclusions
===========
In conclusion, the current study demonstrated an ACh-induced, RhoA-mediated Ca^2+^sensitization in murine BSM contraction. An augmentation of the Ca^2+^sensitizing effect, probably by the upregulation of RhoA protein, might be involved in the enhanced BSM contraction observed in the antigen-induced AHR in mice.
Authors\' contributions
=======================
YC conceived of the study, participated in its design and coordination, and drafted the manuscript. AU carried out the intact smooth muscle studies. KS, HT and HS carried out the skinned fiber studies and immunoblot analysis. SN carried out the analysis of active RhoA. MM participated in the direction of the study as well as writing and preparing the manuscript. All authors read and approved the final manuscript.
Acknowledgements
================
This work was partly supported by a Grant-in-Aid for Encouragement of Young Scientists from the Ministry of Education, Science, Sports and Culture of Japan.
|
PubMed Central
|
2024-06-05T03:55:51.911577
|
2005-1-8
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545934/",
"journal": "Respir Res. 2005 Jan 8; 6(1):4",
"authors": [
{
"first": "Yoshihiko",
"last": "Chiba"
},
{
"first": "Ayako",
"last": "Ueno"
},
{
"first": "Koji",
"last": "Shinozaki"
},
{
"first": "Hisao",
"last": "Takeyama"
},
{
"first": "Shuji",
"last": "Nakazawa"
},
{
"first": "Hiroyasu",
"last": "Sakai"
},
{
"first": "Miwa",
"last": "Misawa"
}
]
}
|
PMC545935
|
Background
==========
The adipocyte plays a central role in overall metabolic regulation serving as a storage depot for fatty acids and as an endocrine cell to regulate energy utilization and feeding behavior \[[@B1],[@B2]\]. The mass of adipose tissue is maintained by a well-controlled balance of cell proliferation (hyperplasia) and increase in fat cell size (hypertrophy). Contributing to adipocyte hypertrophy is the assimilation of fatty acids into cytosolic triacylglycerol-rich lipid droplets. Fatty acids enter the adipocyte through the plasma membrane, are converted to their acyl-CoA derivatives and transported through the cytosol with the assistance of fatty acid binding proteins due to the lipophilic nature of the fatty acid hydrocarbon chain \[[@B3],[@B4]\]. They are then reassembled into triacylglycerol units by acyltransferases. The intracellular lipid droplet that forms from the coalescence of triacylglycerols has recently been shown to associate with regulators of membrane trafficking in addition to enzymes needed for fatty acid storage and utilization, suggesting a complex and dynamic role deserving of the name adiposome \[[@B5]\].
Extracellular fatty acids that are available for adipocyte uptake are either 1) associated with circulating albumin, 2) hydrolyzed from triacylglycerol-rich lipoprotein particles by lipoprotein lipase, or 3) in the form of VLDL particles which can be directly internalized by adipocyte lipoprotein receptors. In the circulation, VLDL represents the major source of fatty acids for peripheral tissues in the form of triacylglycerols and provides a concentrated source of esterified fatty acids. It is interesting that in light of the well studied processes of cytosolic transport and assimilation of free fatty acids into triacylglycerol-rich storage droplets, the mechanism of transport of fatty acids across the adipocyte plasma membrane remains controversial. Two mechanisms, which are not mutually exclusive, have been proposed: one involves passive diffusion across the plasma membrane \[[@B6],[@B7]\], the other requires protein-mediated transport \[[@B8],[@B9]\]. Passive diffusion, which requires protonation of the fatty acid prior to entering the bilayer, has long been regarded as the major pathway for uptake of fatty acids by cells. However, recent kinetic data suggest that passive diffusion, while sufficient for cells with relatively low metabolic rates, is likely to be insufficient for cells with high fatty acid utilization such as skeletal muscle and adipose \[[@B10]-[@B12]\]. Moreover, the role of fatty acid-albumin complexes as a significant source of diffusible free fatty acids has recently been questioned, as evidence indicates that a significant transfer of fatty acids from albumin occurs only at very high and non-physiological fatty acid to albumin ratios \[[@B13],[@B14]\]. Protein-mediated transport of fatty acids has been investigated using fatty acid binding and uptake studies \[[@B15],[@B16]\]. These results show that fatty acid permeation demonstrates concentration-dependent, nonlinear saturation kinetics with a Km of transport of \~7 nM \[[@B9]\]. Moreover, uptake of long-chain fatty acids (\>18 carbons) was competable \[[@B17],[@B18]\], further suggesting a receptor-mediated process.
Several cell surface proteins are expressed by adipocytes which potentially contribute to receptor-mediated uptake of extracellular fatty acids; these include CD36, fatty acid transport protein-1 (FATP1), very low density lipoprotein receptor (VLDL-R), low density lipoprotein receptor-related protein (LRP), and heparan sulfate proteoglycans (HSPG). CD36 is a cell surface glycoprotein that binds long-chain fatty acids with high affinity and demonstrates a subcellular distribution that is consistent with a role in fatty acid transport across the plasma membrane \[[@B19]-[@B23]\]. Following the induction of pre-adipocyte 3T3-L1 cells, CD36 expression increases \[[@B24]\] with a concomitant increase in fatty acid uptake. *In vivo*studies corroborate these observations as uptake of long-chain fatty acids is impaired in CD36 knock-out mice \[[@B25]\] or when CD36 expression is reduced by anti-sense RNA treatment \[[@B22]\]. Although these data indicate a significant role for CD36 in long-chain fatty acid uptake, the data further suggest that it is not sufficient to account for the entirety of fatty acid uptake by adipocytes since the absence of CD36 reduces uptake by only 50% \[[@B25]\]. FATP1 is a 71 kD transmembrane protein and the major FATP family member expressed in adipocytes \[[@B26]\]. In adipocytes, insulin is known to induce the translocation of FATP1 from a perinuclear compartment to the plasma membrane \[[@B27]\]. Similar to CD36, FATP1 expression enhances the uptake of fatty acids \[[@B26],[@B28]\].
Studies using a VLDL-R knock-out animal model have suggested a role for this receptor in the accumulation of fatty acids by adipose tissue; mice lacking VLDL-R demonstrate only modest weight gain and a reduction in adipose tissue mass when placed on a high-fat diet \[[@B29]\]. VLDL-R mRNA is known to increase \~3--5-fold during adipocyte differentiation \[[@B30]\], suggesting that it may play a critical role in adipocyte physiology. The LRP is also expressed by adipocytes \[[@B31]-[@B33]\] and like VLDL-R can bind and internalize apoE-enriched VLDL particles \[[@B34],[@B35]\]. However, its role in lipid accumulation by adipocytes has not been investigated. HSPG have been well characterized in their ability to bind and internalize apoE-enriched VLDL \[[@B36]-[@B38]\], as well as localize lipases to the cell surface through high affinity binding \[[@B39],[@B40]\]. HSPG have also been shown to play a significant role in the hepatic clearance of lipoprotein, however, like LRP, no studies to date have investigated their role in intracellular lipid accumulation by adipocytes.
To better understand the process of fatty acid transport across the plasma membrane of adipocytes, we have examined the expression levels and functional contributions of lipoprotein receptors, VLDL-R, LRP and HSPG toward intracellular lipid accumulation. Our findings suggest that cell surface HSPG play an essential role in fatty acid uptake and intracellular lipid accumulation in adipocytes.
Results
=======
VLDL-R and LRP protein expression increases during adipocyte differentiation
----------------------------------------------------------------------------
Since members of the LDL receptor family are known to play a predominant role in the uptake of lipoproteins, we chose to examine the protein expression levels of VLDL-R and LRP during adipocyte differentiation using 3T3-L1 cells \[[@B41]\], which are a well-established *in vitro*model for adipocyte differentiation \[[@B42],[@B43]\]. 3T3-L1 cells were incubated with or without differentiation agents (dexamethasone, 3-isobutyl-1-methylxanthine, and insulin) and total protein extracts were immunoblotted with antibodies specific for VLDL-R, LRP, or the LDL receptor family-specific chaperone, receptor associated protein (RAP) (Fig. [1](#F1){ref-type="fig"}). Densitometric analysis of the resulting immunoblots indicated that VLDL-R expression increased by \~5-fold over non-differentiated, control cells. LRP expression demonstrated a more modest increase of \~2-fold over control cells. RAP showed no difference in expression between treated and non-treated cells, thus serving as a useful internal loading control. These results are consistent with previous studies which have shown an increase in mRNA levels during adipocyte differentiation for VLDL-R (3--5-fold) \[[@B30]\] and LRP (1.5 to 2-fold) \[[@B31]\].
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Expression of VLDL-R and LRP increases following adipocyte differentiation.**Pre-adipocyte 3T3-L1 cells were incubated for 4 d in the presence (+) or absence (-) of 100 ng/ml dexamethasone, 100 μg/ml 3-isobutyl-1-methylxanthine, and 1 μg/ml insulin, followed by 1 μg/ml insulin for an additional 4--8 d. Total cellular protein was obtained by detergent extraction, equal amounts of protein (20 μg/lane) were separated by SDS-PAGE and immunoblotted with anti-VLDL-R polyclonal IgG (4 μg/ml), or anti-LRP (1:2000) or anti-RAP (1:2000) antisera. The chemiluminescence image was quantitated by densitometry. Values obtained for non-treated cells were assigned as 100% for comparison with treated cells. Data shown is representative of 3 separate experiments.
:::
:::
RAP, besides being an exocytic chaperone for members of the LDL receptor family, is also a high affinity ligand for most, if not all, members of the family \[[@B44],[@B45]\]. This property is useful for evaluating specific receptor binding and internalization activities by cells. Cell surface expression of LDL receptor family members was compared between control, non-treated 3T3-L1 cells and differentiated cells by incubating with ^125^I-RAP at 4°C (Fig. [2A](#F2){ref-type="fig"}). Differentiated cells were found to bind \~7-fold more ^125^I-RAP than non-treated control cells. This increase in receptor expression, as measured by ligand binding, is quantitatively consistent with that measured by immunoblotting. When ^125^I-RAP was incubated with cells at 37°C to evaluate receptor internalization, we found that differentiated adipocytes internalized \~2-fold more ^125^I-RAP than control 3T3-L1 cells (Fig. [2B](#F2){ref-type="fig"}). Together, these results indicate that levels of both VLDL-R and LRP are increased at the cell surface resulting in increased receptor internalization activity in differentiated adipocytes.
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Differentiated adipocytes bind and internalize more RAP than pre-adipocytes.**Pre-adipocyte 3T3-L1 cells and differentiated adipocytes were incubated with ^125^I-RAP (500 ng/ml) in the presence (hatched bars) or absence (solid bars) of unlabeled RAP (20 μg/ml) at 4°C for 3 h (A) or 37°C for 3 h (B). For 4°C incubations, bound ligand was solubilized and directly quantitated by scintillation counting. For 37°C incubations, culture media was processed for trichloroacetic acid (TCA) precipitation and soluble radioactivity, representing degraded ligand, was quantitated by scintillation counting.
:::
:::
Sulfated proteoglycan levels are increased in differentiated adipocytes
-----------------------------------------------------------------------
Although HSPG are known to play a critical role in binding of lipoproteins to the cell surface \[[@B46],[@B47]\] and additionally serve as a primary interaction site for LPL \[[@B48],[@B49]\], little evidence has been reported as to their function in lipid accumulation by adipocytes. To begin to characterize their role in adipocyte growth, we first examined changes in overall proteoglycan synthesis during adipocyte differentiation. 3T3-L1 preadipocytes and differentiated adipocytes were incubated with ^35^SO~4~to label glycosaminoglycan moieties. Cell lysates were prepared by extracting with urea and radiolabeled proteoglycans were analyzed by anion exchange chromatography. As shown in Fig. [3A](#F3){ref-type="fig"}, the amount of ^35^SO~4~-labeled proteoglycan recovered from mature adipocytes was \~2.5-fold greater than that extracted from pre-adipocytes indicating that proteoglycan synthesis is augmented during differentiation. When the chromatographic fractions were separated by SDS-PAGE, interestingly, we found primarily high molecular weight species of proteoglycans in extracts from either pre-adipocytes or mature adipocytes (Fig. [3B](#F3){ref-type="fig"}).
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Adipocyte differentiation results in increased synthesis of a high molecular weight proteoglycan that consists of a mixture of heparan and chondroitin sulfate glycosaminoglycan moieties.**Pre-adipocyte 3T3-L1 cells and differentiated adipocytes were incubated for 20 h with ^35^SO~4~(125 μCi/ml). (A), cells were lysed with 8 M urea and proteins were fractionated by anion exchange chromatography as described in Methods. Radioactivity in each fraction was determined by scintillation counting. (B), fractions from (A) were separated by 7% SDS-PAGE and subjected to phosphorimager analysis (upper panel, pre-adipocytes; lower panel, adipocytes). (C), ^35^SO~4~-labeled proteoglycans, enriched by anion exchange chromatography, were incubated with or without heparinase I (5 Units/ml), chondroitinase ABC (5 Units/ml), or both enzymes for 20 h at 37°C. Labeled material was then fractionated by size exclusion chromatography and fractions were assessed by scintillation counting.
:::
:::
To analyze the composition of the high molecular weight proteoglycan species synthesized by mature adipocytes, we incubated the chromatographically enriched material with heparinase I, chondroitinase ABC, or a mixture of both enzymes followed by size fractionation (Fig. [3C](#F3){ref-type="fig"}). With no enzymatic digestion, the material eluted at the exclusion limit of the column. Treatment with heparinase I reduced peak fractions 6--8 by \~15% and yielded a lower molecular weight product that eluted between fractions 10--14. By contrast, chondroitinase ABC treatment reduced the high molecular weight, peak fractions 6--8 by \~65--70% and generated mostly low molecular weight products eluting between fractions 11--15. The high molecular weight material remaining after chondroitinase ABC treatment likely consists primarily of heparan sulfate moieties. This was confirmed by treating the enriched proteoglycans with both heparinase I and chondroitinase ABC which resulted in essentially a quantitative shift of labeled material from high molecular weight fractions to fractions consisting of primarily lower molecular weight products. These data provide us with two important observations; 1) major proteoglycan species synthesized by mature adipocytes are of very high molecular weight, and 2) this proteoglycan structure consists mostly of chondroitin sulfate glycosaminoglycans with a smaller percentage of heparan sulfate moieties.
HSPG participate in the uptake of DiI-labeled apolipoprotein E-enriched VLDL by adipocytes
------------------------------------------------------------------------------------------
The extent to which extracellular lipoproteins contribute to lipid accumulation by adipocytes has not been fully established. A recent study suggests that uptake of VLDL by adipocytes stimulates intracellular lipid accumulation \[[@B50]\]. To identify if either HSPG or LDL receptor family members play a role in this process, we incubated mature adipocytes with DiI-labeled apoE-enriched VLDL in the presence or absence of either heparin (to inhibit HSPG activity) or RAP (to inhibit VLDL-R and LRP function) and visualized the cells by fluorescence microscopy (Fig. [4](#F4){ref-type="fig"}). In the absence of potential competitors, mature adipocytes readily internalized DiI-apoE-VLDL into small cytosolic vesicles (upper panel). Incubation with RAP showed little or no effect on DiI-apoE-VLDL uptake (lower panel). However, incubation with heparin significantly inhibited intracellular accumulation of DiI-apoE-VLDL (middle panel) indicating that HSPG, rather than VLDL-R or LRP, play a major role in apoE-VLDL internalization.
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Heparin, but not RAP, competes for DiI-apoE-VLDL uptake by adipocytes.**Differentiated adipocytes were cultured on glass coverslips and incubated with DiI-labeled apoE-VLDL (4 μg/ml) in the presence or absence (upper panel) of either heparin (500 μg/ml, middle panel) or RAP-GST (50 μg/ml, lower panel) at 37°C for 3 h. Cells were then fixed and processed for fluorescence microscopy. Left panels, phase contrast image; right panels, rhodamine filter set (550 nm excitation-573 nm emission). Magnification, 630×.
:::
:::
HSPG are necessary for intracellular lipid accumulation in mature adipocytes
----------------------------------------------------------------------------
4-methylumbelliferyl-β-D-xylopyranoside (4-MUmb) and p-nitrophenyl-β-D-xylopyranoside (pNP-Xyl) are reagents that can serve as alternative acceptors within cells for heparan sulfate moieties and are thus able to compete for heparan sulfate chain addition to proteoglycan core proteins \[[@B51],[@B52]\]. The bare proteoglycan core proteins that result from this treatment traverse to the plasma membrane but are devoid of any heparan sulfate glycosaminoglycan modifications. This in effect removes heparan sulfate proteoglycan functionality from the surface of treated cells and permits an examination of HSPG contributions to cell function. To ensure that these reagents do not interfere with the events of adipocyte differentiation, we treated 3T3-L1 pre-adipocytes with 4-MUmb or pNP-Xyl concurrently with the addition of differentiation reagents and examined the expression levels of proteins known to increase during adipocyte formation, namely VLDL-R, LRP, CD36/FAT \[[@B21]\], and leptin \[[@B53]\]. By immunoblot analysis, we found that the expression levels of these proteins were comparable between treated and non-treated cells (Fig. [5A](#F5){ref-type="fig"}). Since lipoprotein lipase (LPL) synthesis is also known to increase following 3T3-L1 differentiation \[[@B54],[@B55]\], we measured enzymatic activity of LPL in the culture supernatants of induced and non-induced 3T3-L1 cells and compared these results with those obtained from induced 3T3-L1 cells treated with either 4-MUmb or pNP-Xyl (Fig. [5B](#F5){ref-type="fig"}). As expected, conversion of 3T3-L1 pre-adipocytes to mature adipocytes resulted in \~5--6-fold increase in LPL activity. Importantly, no significant difference in LPL activity was noted between normal mature adipocytes and those treated with either 4-MUmb or pNP-Xyl. These data indicate that expression of these proteins is unaffected by inhibited proteoglycan maturation and, importantly for our purposes, that addition of 4-MUmb or pNP-Xyl to cells does not prevent adipocyte differentiation.
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Inhibitors of HSPG maturation have no effect on adipocyte differentiation.**3T3-L1 pre-adipocytes were incubated with or without 3 mM pNP-Xyl or 4-MUmb concurrently with differentiation reagents for 4 d. Media was then changed to include just insulin with or without 3 mM pNP-Xyl or 4-MUmb for an additional 4--8 d. (A), cell lysates were immunoblotted with anti-VLDL-R mAb (1:50, culture supernatant), anti-LRP antisera (1:2000), or anti-CD36 IgG (5 μg/ml). For anti-leptin, proteins in culture supernatant were concentrated 10-fold by acetone precipitation prior to SDS-PAGE fractionation and immunoblotted with anti-leptin IgG (2 μg/ml). (B), following differentiation and xyloside treatment, cells were cultured for 16 h in serum-free, phenol red-free media. Culture supernatants were then assayed for LPL enzymatic activity as described in Methods. No significant difference was found in LPL levels between adipocytes treated with xylosides and untreated adipocytes (asterisk, p-value \<0.01).
:::
:::
To determine if HSPG are involved in intracellular lipid accumulation by mature adipocytes, we assessed cytosolic lipid droplet formation in induced 3T3-L1 cells following treatment with either 4-MUmb or pNP-Xyl. Again 3T3-L1 pre-adipocytes were incubated with 4-MUmb or pNP-Xyl concurrently with the addition of differentiation reagents and intracellular lipid accumulation was assessed by Oil Red O staining. Upon Oil Red staining, untreated mature adipocytes demonstrated large intracellular droplets; the characteristic morphologic feature of cytosolic lipid accumulation (Fig. [6A](#F6){ref-type="fig"}, left panel). By contrast, adipocytes treated with either 4-MUmb or pNP-Xyl showed little staining by the lipophilic dye (Fig. [6A](#F6){ref-type="fig"}, middle and right panels, respectively). To quantitate this effect, cell-associated lipophilic dye was extracted and measured by spectrophotometry (Fig. [6B](#F6){ref-type="fig"}). Treatment of adipocytes with 4-MUmb or pNP-Xyl resulted in \~6--7-fold decrease in lipid accumulation. Moreover, adipocytes treated with varying concentrations of 4-MUmb or pNP-Xyl demonstrated a concentration-dependent effect of these HSPG inhibitors on intracellular lipid accumulation (Fig. [7](#F7){ref-type="fig"}). Together, these data indicate that HSPG play an essential role in lipid accumulation by adipocytes.
::: {#F6 .fig}
Figure 6
::: {.caption}
######
**Reduction of cell surface heparan sulfate glycosaminoglycans significantly reduces intracellular lipid accumulation in adipocytes.**3 mM 4-MUmb or pNP-Xyl was added to 3T3-L1 cells concurrently with differentiation reagents as in Fig. 5. Cells were then fixed with 10% formaldehyde, stained with 0.1% Oil Red O and photographed with phase contrast optics (A; magnification, 400×). (B), after image capture, cell-associated Oil Red O was extracted with 0.1 N HCl and dye was quantitated by spectrophotometry. Asterisk indicates a statistically significant difference compared to untreated adipocytes (p \< 0.01).
:::
:::
::: {#F7 .fig}
Figure 7
::: {.caption}
######
**Reduction in intracellular lipid accumulation in adipocytes by heparan sulfate inhibitors is concentration-dependent.**4-MUmb (solid bars) or pNP-Xyl (hatched bars) was added to 3T3-L1 cells at the indicated concentrations concurrently with differentiation reagents as in Fig. 5. Cells were then fixed with 10% formaldehyde, stained with 0.1% Oil Red O, and cell-associated lipophilic dye was extracted with 0.1 N HCl and quantitated by spectrophotometry. Data shown represents an average of three experiments.
:::
:::
Discussion
==========
Intracellular lipid accumulation is a hallmark event of adipocyte development and the major factor in adipocyte hypertrophy. This prompted us to investigate the molecular mechanism by which adipocytes take up extracellular lipid components. In the present study, we provide new information on the expression levels and functional activities of certain cell surface receptors in mature adipocytes that are known to play primary roles in lipoprotein processing and clearance. These receptors have been extensively studied in liver, the major organ for dietary lipoprotein clearance \[[@B56]\], however little information is available regarding their function in lipid accumulation by adipocytes. Using 3T3-L1 cells, which is a well established model of adipocyte differentiation, we show that protein expression of the lipoprotein receptors, VLDL-R and LRP, is increased during adipocyte differentiation, 5-fold and 2-fold, respectively. This increase is consistent with previous studies reporting increases in mRNA levels for these proteins \[[@B30],[@B31]\]. The comparable increase between mRNA and protein levels also suggests that the overall increase in receptor expression following differentiation is primarily due to an increase in transcription with little or no post-transcriptional regulation. We also demonstrate that both pre-adipocytes and mature adipocytes express one major form of sulfated proteoglycan with a relative molecular mass of \>500 kD. Expression of this species of proteoglycan increases \~2.5-fold during differentiation and appears to contain both heparan and chondroitin sulfate glycosaminoglycans. Furthermore, we found that uptake of DiI-labeled VLDL by adipocytes is inhibited by heparin, but not RAP, suggesting that in adipocytes HSPG play a greater role in lipoprotein uptake than do members of the LDL receptor family. Since VLDL is a major carrier of fatty acids in the form of triacylglycerols, these data suggest to us that HSPG may provide a mechanism for facilitating the uptake of fatty acids and significantly contribute to intracellular lipid accumulation. To address this hypothesis, we treated adipocytes with 4-methylumbelliferyl-β-D-xylopyranoside and p-nitrophenyl-β-D-xylopyranoside, which are competitive inhibitors of heparan sulfate chain addition, to prevent the synthesis of functional cell surface HSPG molecules. We found this treatment to effectively block intracellular lipid accumulation in adipocytes without affecting adipocyte differentiation.
These findings offer a novel observation that cell surface HSPG appear to be essential for lipid accumulation and lipid droplet formation in adipocytes. However, this observation also raises the question as to the exact role of HSPG in the transport of fatty acids across the adipocyte plasma membrane. Based on the results presented in this study and those reported elsewhere, we have drawn a testable model to define how HSPG contribute to intracellular lipid accumulation by adipocytes (Fig. [8](#F8){ref-type="fig"}). HSPG are known to serve as binding sites for apoE-enriched lipoproteins \[[@B36]-[@B38]\]. In liver, HSPG are thought to either directly internalize bound lipoproteins by hepatocytes or, alternatively, localize lipoproteins to the cell surface and subsequently transfer particles to LDL receptor family members for endocytosis \[[@B57]\]. In the present study, we show that heparin effectively competes for uptake of DiI-labeled apoE-VLDL while RAP has little or no effect suggesting that, in adipocytes, HSPG are able to internalize apoE-VLDL independent of LDL receptor family members. A direct internalization function by HSPG is also supported by recent findings from our laboratory \[[@B47]\] and others \[[@B36],[@B37]\]. Alternatively, HSPG may serve as a reaction center for triacylglycerol hydrolysis. In addition to apoE-rich lipoproteins, HSPG are also capable of binding LPL \[[@B39],[@B40]\]. The binding function of HSPG for both apoE-rich lipoproteins and LPL can serve to co-localize enzyme and substrate and facilitate release of fatty acids in proximity of the adipocyte cell surface. Uptake of these liberated fatty acids can then be accomplished by fatty acid transport proteins such as CD36 \[[@B58]\] or FATP1 \[[@B59]\].
::: {#F8 .fig}
Figure 8
::: {.caption}
######
**Model for HSPG activity contributing to fatty acid uptake and intracellular lipid accumulation in adipocytes.**Cell surface heparan sulfate proteoglycans (HSPG) serve as primary binding sites for apoE-enriched VLDL (apoE-VLDL) and lipoprotein lipase (LPL) on adipocytes. Localization of apoE-VLDL and LPL to the cell surface can create a focal reaction center for triacylglycerol hydrolysis thereby releasing fatty acids for cellular uptake by fatty acid transporters such as FATP1 or CD36. Alternatively, similar to that proposed for hepatic clearance of apoE-VLDL and chylomicron remnants \[57\], initial binding to HSPG serves to concentrate lipoprotein particles at the cell surface and their uptake is mediated either by direct HSPG internalization or following their transfer to VLDL-R or LRP.
:::
:::
The identity of this HSPG is currently unknown; however, its large relative molecular mass and composition containing both heparan and chondroitin sulfates are consistent with the structural properties of the syndecan \[[@B60],[@B61]\] and perlecan \[[@B62]\] families of proteoglycans. There are four members to the syndecan family, all of which are type I transmembrane cell surface molecules. Syndecan-1 is found primarily in epithelial and plasma cells, syndecan-2 is found in endothelial cells and fibroblasts, syndecan-3 is expressed in cells of neural crest origin, and syndecan-4 demonstrates a more ubiquitous distribution, including adipose tissue \[[@B63]\]. bFGF treatment increases expression of syndecan-1 in 3T3 cells \[[@B64]\]; however its presence in adipocytes has not been investigated. Although the core proteins of the syndecan family are of modest molecular weight, they are typically modified with the addition of heparan sulfate glycosaminoglycan chains near their N-termini and chondroitin sulfate moieties attached more proximal to their membrane spanning domains, which results in very high molecular masses when resolved by SDS-PAGE, often \>500 kD. Perlecan has a large complex modular core protein with a molecular weight of \~400 kD \[[@B62],[@B65]\]. It typically contains three heparan sulfate chains near its N-terminus, but can also have chondroitin sulfate substitutions. Perlecan is a secreted proteoglycan that demonstrates a widespread distribution as a basement membrane component \[[@B66]\]. As both syndecan and perlecan contain heparan sulfate glycosaminoglycan chains, they are able to bind LPL with high affinity \[[@B67]\] and localize its enzymatic activity near the cell surface thereby serving as a reaction center for triacylglycerol hydrolysis. Moreover, studies have shown that they can also bind and internalize lipoprotein particles independent of the classical lipoprotein receptors \[[@B37],[@B38]\]. We are currently in the process of accurately quantitating the amounts of heparan and chondroitin sulfate moieties present on these high molecular weight proteoglycans and identifying their core protein structure. Once this information is available, we will be able to confirm their role in fatty acid transport and adipocyte hypertrophy using expression inhibition procedures and obtain additional information on their role in adipocyte physiology.
The availability of circulating triacylglycerol-rich lipoproteins for adipocyte utilization and storage relies on transport of these particles across the endothelial barrier. Transcytosis of albumin across endothelium has been shown to occur by a vesicular transport pathway \[[@B68]-[@B70]\] and is a likely mechanism for transport of fatty acid-albumin complexes. More recently, members of the LDL receptor family, including VLDL-R \[[@B71]\] and megalin \[[@B72],[@B73]\], have also been shown to undergo transcytosis across endothelium and thus provide a mechanism for transport of triacylglycerol-rich lipoproteins into the tissues. Transport of VLDL across the endothelium is also likely to be assisted by hydrostatic and osmotic pressures within the capillary lumen. VLDL and related remnant lipoprotein particles represent the richest source of triacylglycerols in the body. The high metabolic requirements of adipocytes for fatty acids for both storage and utilization make these triacylglycerol-rich particles a suitable source of available fatty acids. The presence of a transendothelial transport mechanism via lipoprotein receptors makes VLDL particles a logical source of fatty acids for adipose growth.
How HSPG might coordinate its activity with CD36 or FATP1 for fatty acid uptake or possibly with members of the LDL receptor family for assisted internalization is under current investigation. Presently, our results provide novel findings indicating that cell surface HSPG activity is necessary for lipid accumulation by adipocytes. It is anticipated that these observations will aid in our understanding of the complex mechanism leading to adipose hypertrophy and obesity, and provide a novel avenue to explore for targeted reduction of intracellular lipid accumulation
Methods
=======
Materials
---------
Anti-LRP polyclonal antibody was raised against an 18 amino acid peptide from the cytoplasmic tail of human LRP \[[@B74]\] and anti-RAP polyclonal antibody was raised against a recombinant RAP-GST fusion protein as described \[[@B75]\]. Anti-VLDL-R polyclonal antibody was a generous gift from Dr. Dudley Strickland (School of Medicine, University of Maryland, Baltimore). Anti-VLDL-R monoclonal antibody-producing hybridoma was purchased from American Type Culture Collection (Manassas, VA) anti-leptin polyclonal antibody was purchased from Chemicon (Temecula, CA), and anti-CD36 polyclonal antibody (H-300) was obtained from Santa Cruz Biotechnology (Santa Cruz, CA). Dexamethasone, 3-isobutyl-1-methylxanthine and insulin were obtained from Sigma-Aldrich (St. Louis, MO). ^35^SO~4~were purchased from MP Biomedicals (Irvine, CA). p-nitrophenyl-β-D-xylopyranoside and 4-methylumbelliferyl-β-D-xylopyranoside were from Calbiochem (La Jolla, CA). Heparin, heparinase I, chondroitinase ABC, p-nitrophenylbutyrate, lipoprotein lipase, and Oil Red O were purchased from Sigma-Aldrich. Optiprep was from Greiner Bio-One (Longwood, FL). DiI (1,1\'-dictadecyl-3,3,3\',3\'-tetramethylindocarbocyanine perchlorate) was purchased from Molecular Probes (Eugene, OR). Apolipoprotein E was obtained from Calbiochem. RAP-GST fusion protein was purified as previously described \[[@B76]\]. Tissue culture plastics were purchased from Corning (Corning, NY) or Greiner Bio-One. Buffers, salts, and detergents were obtained from either Sigma-Aldrich or Calbiochem.
Cell culture and adipocyte differentiation
------------------------------------------
3T3-L1 cells were obtained from American Type Culture Collection (Manassas, VA) and grown in Dulbecco\'s modified Eagle\'s medium (DMEM) (Invitrogen, Carlsbad, CA) supplemented with 10% (v/v) fetal calf serum (Irvine Scientific, Santa Ana, CA), 1 mM sodium pyruvate, 100 μg/ml streptomycin sulfate, and 100 units/ml penicillin. Cells were cultured at 37°C with 5% CO~2~and passaged twice weekly. To differentiate 3T3-L1 cells into adipocytes, cells were incubated with 100 ng/ml dexamethasone, 100 μg/ml 3-isobutyl-1-methylxanthine, and 1 μg/ml insulin for 4 days, followed by 1 μg/ml insulin for an additional 4--8 days.
Immunoblotting
--------------
Cell lysates were prepared with 20 mM Tris pH 7.4, 150 mM NaCl (TBS) containing 1% (v/v) Triton-X100. For leptin, culture supernatants were concentrated by mixing with 3 volumes ice cold acetone, incubating at -20°C for 1 h, followed by centrifugation at 10,000 × g at 4°C for 15 min. Protein pellets were resuspended with SDS-PAGE sample buffer supplemented with 2% (v/v) β-mercaptoethanol. Proteins were then separated by SDS-PAGE and transferred to Immobilon-P (Millipore, Billerica, MA) using a wet tank transfer system (BioRad, Hercules, CA). Membranes were blocked with TBS, 0.1% (v/v) Tween-20, 5% (w/v) non-fat dry milk for 20 minutes at 23°C and incubated with the indicated antibody for 2 h at 23°C. Membranes were washed three times (10 min each) with TBS, 0.1% (v/v) Tween-20, and bound antibodies were detected with species-specific HRP-conjugated secondary antibodies (1:3000, BioRad) followed by chemiluminescence detection according to the manufacturer\'s instructions (Pierce, Rockford, IL). Images were captured using a Syngene GeneGnome system equipped with a Peltier-cooled 16-bit CCD camera and saturation detection. Densitometric analysis was performed using Scion Image, version 4.0.2.
^125^I-RAP cell surface 4°C binding assay
-----------------------------------------
RAP-GST was purified \[[@B76]\] and labeled with Na^125^I as previously described \[[@B77]\]. Differentiated 3T3-L1 cells were grown on tissue culture plates precoated with 1% (w/v) gelatin and incubated with ^125^I-RAP-GST (500 ng/ml) diluted into 20 mM Hepes, pH 7.4, 150 mM NaCl, 2 mM CaCl~2~, 1% (w/v) bovine serum albumin (buffer A) at 4°C for 3 h in the presence or absence of a 50-fold molar excess of unlabeled RAP-GST. Unbound ^125^I-RAP-GST was removed by rinsing cells three times with cold buffer A after which cells with bound ligand were solubilized with 0.1 N NaOH. Solubilized proteins were added to EcoLume (ICN Biomedicals, Costa Mesa, CA) and subjected to scintillation counting (Packard Tri-Carb 1600CA, 73% efficiency for ^125^I). Results were normalized to total cellular protein (BCA Protein Assay, Pierce, Rockford, IL). Specificity was determined as the difference between total binding (without competition) and non-specific binding (non-competable) \[[@B78]\]. The actual amount of ligand bound to cells was calculated as cpm ÷ the specific activity of ^125^I-labeled ligand. All data points represent averages of duplicates or triplicates with standard errors of \<5%.
37°C ligand degradation assay
-----------------------------
Differentiated 3T3-L1 cells were incubated at 37°C/5% CO~2~for 3 h with 500 ng/ml ^125^I-RAP-GST diluted into DMEM containing 1% (w/v) bovine serum albumin in the presence or absence of a 50-fold molar excess of unlabeled RAP-GST. Media was then removed and processed for trichloroacetic acid (TCA) precipitation \[[@B75]\]. TCA-soluble material was added to EcoLume and subjected to scintillation counting. Degradation was calculated as TCA-soluble cpm ÷ specific activity of the radioiodinated ligand.
^35^SO~4~incorporation and proteoglycan analysis
------------------------------------------------
3T3-L1 pre-adipocytes and differentiated adipocytes grown in 100 mm tissue culture dishes were incubated with sulfate-free DME containing 10% (v/v) fetal calf serum and 125 μCi/ml Na~2~^35^SO~4~for 20 h at 37°C/5% CO~2~. Cells were detached by incubating with phosphate buffered saline containing 5 mM EDTA for 20 min at 23°C. Following centrifugation to pellet cells (1000 × g, 4°C, 10 min), they were lysed by incubation with solubilization buffer (100 mM Tris, pH 7.5, 150 mM NaCl, 8 M urea) for 1 h at 4°C. Insoluble material was removed by centrifugation (10,000 × g, 4°C, 5 min) and supernatant was batch adsorbed onto 0.5 ml Macro-Prep DEAE Support (BioRad) (pre-washed with solubilization buffer) with gentle agitation for 1 h at 4°C. DEAE resin was applied to a 1.0 cm × 5.0 cm column and washed with 10 column volumes solubilization buffer followed by 10 column volumes 100 mM Tris, pH 7.5, 150 mM NaCl. Bound proteins were eluted with 100 mM Tris, pH 7.5, 1.0 M NaCl and 120 μl fractions were collected. Fifteen μl from each fraction was mixed with 3 ml EcoLume and subjected to scintillation counting. Peak fractions from DEAE enrichment of proteoglycans obtained from labeled differentiated adipocytes (fraction no. 3 and 4, Fig. [3A](#F3){ref-type="fig"}) were pooled and diluted with dH~2~O to adjust to 150 mM NaCl. This preparation was then incubated with or without either heparinase I (5 units/ml) or chondroitinase ABC (5 units/ml) or both for 20 h at 37°C. The material was applied to a 1.0 cm × 12 cm Sephacryl S-200 HR (Amersham Biosciences, Piscataway, NJ) gel filtration column and eluted with 20 mM Tris, pH 7.4, 150 mM NaCl. 350 μl fractions were collected of which 200 μl was mixed with 3 ml EcoLume and analyzed by scintillation counting.
Preparation of DiI-labeled apoE-VLDL
------------------------------------
New Zealand White rabbits were placed on a high-fat chow diet (10% peanut oil/1% cholesterol) for a minimum of 4 d, blood was drawn into in 1 mM EDTA and centrifuged at 2000 × g, 15 min to remove cells. Chylomicrons were floated by centrifuging plasma at 100,000 × g for 10 min and removed by pipetting. Plasma was then mixed with OptiPrep™ (12% iodixanol final concentration) and centrifuged at 350,000 × g for 3 h (SW55Ti rotor) with slow acceleration and deceleration. VLDL particles (density of 1.006 g/ml) were removed from the top of the gradient by pipetting. Purified VLDL was analyzed by SDS-PAGE and Coomassie R staining to confirm the presence of apoB100 (515 kD) and apoE (35 kD). Animal protocol (\#2032) was approved by the University of New Mexico, Health Sciences Center Laboratory Animal Care and Use Committee. For DiI labeling, a working stock of 3 mg/ml was made in dimethylsulfoxide and 0.15 mg was slowly added to 1.67 mg VLDL (in 1.9 ml) with vortexing to rapidly mix. The mixture was then wrapped in foil and incubated for 8 h at 37°C. Unbound DiI was removed from DiI-labeled VLDL by OptiPrep gradient centrifugation as described above.
DiI-labeled apoE-VLDL uptake assay
----------------------------------
3T3-L1 cells were plated on glass coverslips and incubated with differentiation reagents as described above. After conversion to mature adipocytes, cells were rinsed twice with DMEM and incubated with DiI-VLDL (4 μg/ml) and apolipoprotein E (3 μg/ml) diluted into DMEM in the presence or absence of either heparin (500 μg/ml) or RAP-GST (50 μg/ml). After 3 h at 37°C, cells were rinsed with phosphate buffered saline, fixed with 1.5% (w/v) paraformaldehyde for 30 min, and mounted in Gelvatol (Air Products, Allentown, PA) containing 1 mg/ml p-phenylenediamine. Cells were observed with a Zeiss Axioskop microscope equipped for epifluorescence. Images were capture with a Hamamatsu digital/video camera and AxioVision software.
LPL activity
------------
Treated and non-treated 3T3-L1 cells (as indicated in figure legends) were cultured overnight in serum-free DMEM without phenol red. Media was removed and combined with an equal volume of 100 mM sodium phosphate buffer, pH 7.2, 150 mM NaCl, 0.5% (v/v) Triton-X100. p-nitrophenyl butyrate in acetonitrile was added to a final concentration of 0.5 mM and absorbance at 400 nm was recorded every 10 s for 5 min using a Genesys UV Spectrophotometer. A standard curve for LPL activity was generated by plotting absorbance values obtained with varying concentration of purified LPL (from bovine milk). Quantitation of LPL activity in the individual media samples was determined from the standard curve.
Oil Red O staining and quantitation
-----------------------------------
Cells were fixed with 10% (v/v) formaldehyde in phosphate buffered saline for 1 h at 23°C, rinsed twice with water, then stained for 2 h at 23°C with 0.1% (w/v) Oil Red O in 75% (v/v) isopropanol, followed by rinsing twice with water to remove unincorporated dye. Stained cells were viewed and photographed using a Zeiss AxioVert microscope with phase contrast optics and Hamamatsu digital/video camera. For quantitation, cells were dried for 2 h at 37°C, followed by incubation with 100% isopropanol for 15 m at 23°C to extract bound dye. Solubilized dye was then quantitated by spectrophotometry in a BioRad Model 680 Microplate Reader by measuring absorbance at 510 nm. Statistical significance was determined by performing a paired t-test.
List of abbreviations
=====================
ApoE = apolipoprotein E
HSPG = heparan sulfate proteoglycan
VLDL = very low density lipoprotein
VLDL-R = VLDL receptor
LRP = low density lipoprotein receptor-related protein
Syn-1 = Syndecan-1
FATP1 = fatty acid transport protein-1
RAP = receptor associated protein
LPL = lipoprotein lipase
DiI = 1,1\'-dictadecyl-3,3,3\',3\'-tetramethylindocarbocyanine perchlorate
4-MUmb = 4-methylumbelliferyl-β-D-xylopyranoside
pNP-Xyl = p-nitrophenyl-β-D-xylopyranoside
bFGF = basic fibroblast growth factor
Authors\' contributions
=======================
LCW carried out the majority of studies and drafted the manuscript. SC performed the xyloside titration study and DN performed the LPL assays. RAO provided the original conceptual framework for the study, carried out pilot experiments, participated in the experimental design and finalized the manuscript for submission. All authors read and approved the final version.
Acknowledgements
================
We thank Dr. Robert H. Glew (Department of Biochemistry and Molecular Biology, University of New Mexico, School of Medicine) for his valuable input and critical evaluation of this work. We also express our gratitude to Dr. Dudley Strickland (American Red Cross, Holland Branch) for supplying us with anti-VLDL-R polyclonal antibody. This work was supported by the National Institutes of Health Grant HL63291 (to R.A.O.). The images presented in this study were generated in the Fluorescence Microscopy Facility which is supported by NCRR 1 S10 RR14668, NSF MCB9982161, NCRR P20 RR11830, NCI R24 CA88339, NCRR S10 RR19287, NCRR S10 RR016918, the University of New Mexico Health Sciences Center, and the University of New Mexico Cancer Center.
|
PubMed Central
|
2024-06-05T03:55:51.914421
|
2005-1-6
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545935/",
"journal": "Lipids Health Dis. 2005 Jan 6; 4:2",
"authors": [
{
"first": "Larissa C",
"last": "Wilsie"
},
{
"first": "Shree",
"last": "Chanchani"
},
{
"first": "Deepti",
"last": "Navaratna"
},
{
"first": "Robert A",
"last": "Orlando"
}
]
}
|
PMC545936
|
Background
==========
Cough is the most common symptom presenting to medical practitioners in Australia, the UK and USA \[[@B1]-[@B3]\]. Cough quality, specifically dry versus we t\[[@B4]\] or productive cough, is often used in epidemiological \[[@B5]-[@B7]\] and clinical research \[[@B8],[@B9]\]. Clinically, physicians also often differentiate between dry and wet cough \[[@B10]-[@B12]\] but there are no studies that have evaluated if these are reproducible descriptors. In adults, productive cough is usually obvious but children however often swallow their sputum and hence a \'wet cough\' is used inter-changeably with \'productive cough\' to describe cough quality in young children who are unable to expectorate \[[@B10],[@B13]\]. It is known that nocturnal cough is unreliably reported in both children \[[@B14]\] and adults \[[@B15]\] but there is no data on cough quality. Wet and dry cough are determined subjectively as there are no \'gold standards\'. To date there are no human studies that have identified the objective relationship of the cough descriptors to mucus secretory states.
The sound of a cough is due to vibration of larger airways and laryngeal structures during turbulent flow in expiration \[[@B16],[@B17]\]. It is not known which generation of the airways is involved when the human ear identifies a wet cough and currently there are no validated human models that allow measurement of increased airway mucus. Mucus hypersecretory states in human diseases can occur from a variety of mechanisms which include; hypersecretion of stored mucin, hypertrophy or hyperplasia of goblet cells and/or increased synthesis from over-expression of mucin genes \[[@B18]\]. In disease states, it is not known which mechanism or site of production is the most important but in smokers with chronic bronchitis, a common cause of productive cough in adults, the larger bronchi (bronchi of diameter \>4 mm ie segmental bronchi and above) \[[@B19]\] are the site of greatest inflammation \[[@B18]\]. Flexible bronchoscopy allows an *in-vivo*visual assessment of larger airways usually to the 3^rd^(lobar bronchi) or 4^th^generation (segmental bronchi) in young children.
The study aims were to compare (1) cough quality (wet vs dry and brassy vs non-brassy) with bronchoscopic findings of secretions and tracheomalacia respectively and, (2) parent(s) vs clinician\'s evaluation of the cough quality (wet and dry). We hypothesised that clinical assessment of cough is good compared to bronchoscopic findings and that a wet cough is related to presence of airway secretions.
Methods
=======
Children electively admitted for bronchoscopy without a known underlying respiratory diagnosis were seen by a member of the research team 0.5--3 hours prior to bronchoscopy. The clinician\'s assessment of cough quality (wet or dry) was recorded on a standardised sheet (based on the cough present on the day of the bronchoscopy), before the parent(s) independently evaluated the current (the morning of, or last 12 hours) cough quality (wet or dry) of their child. For clinician\'s assessment of wet/dry cough, when no spontaneous cough was heard or if child was too young to elicit a cough, cough quality (wet or dry) was deemed \'non-assessable\'. Clinicians also rated cough as \'brassy\' or \'non-brassy\' based on coughs heard anytime before bronchoscopy. For assessment of reliability of cough quality (wet/dry and brassy/non-brassy), 21 cooperative children had their coughs digitally recorded (Acer Pocket PC n11, Taiwan) using music compact disc quality format (44.1 kHz, 16 bit) on the morning of their bronchoscopy. These stored cough sounds were later replayed (using headphones 30--10,000 Hz, Lanier, Japan) from a computer and re-scored in a blinded manner (blinded to the child\'s name and cough quality assigned earlier) for wet/dry and brassy/non-brassy qualities. Written consent was obtained from a parent and the study approved by the hospital\'s ethics committee on human research.
Bronchoscopy and quantification of secretions seen during bronchoscopy
----------------------------------------------------------------------
Flexible bronchoscopy was performed under general anaesthesia as previously described \[[@B20]-[@B22]\]. Briefly, anaesthesia was induced with sevoflurane in 100% oxygen administered through a Jackson Rees T piece circuit, the vocal cords and upper trachea then sprayed (4 mg/kg lignocaine via a Cass needle). Atropine was given intravenously to most children aged \<12 months. In all children a video flexible bronchoscope (BF 3C160, Olympus, Tokyo, Japan) entered the circuit via the port of a swivel right angle connector attached to a facemask. Images were projected onto a monitor (Sony Trinitron, Tokyo, Japan).
A respiratory consultant (ABC or IBM) blinded to the child\'s history and cough quality scored the bronchoscopy sheet quantifying the amount of secretions at the time of the bronchoscopy in real time. When no scorer was available, the session was videotaped and played back. A secretion quantification card (figure [1](#F1){ref-type="fig"}) was visible to the scorer at all times. Secretions were quantified according to amount of mucus in the airways in relation to lumen size (fig [1](#F1){ref-type="fig"}) and scored from the trachea to the level of lobar bronchi (total of 9; trachea, right main stem, right upper lobe, right middle lobe, right lower lobe, left main stem, left upper lobe, left lingula, left lower lobe). When segmental bronchi were seen, the worst segment (ie segment with most secretions) was scored. These scores were used to obtain a final grade of bronchoscopic secretions (BS) from grades 1 to 6;
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Bronchoscopic secretion quantification card.
:::
:::
BS Grade 1 = Nil secretions
BS Grade 2 = Near dry = Bubbles only in \< half total number of bronchi involved
BS Grade 3 = Minimal = Bubbles found in \> half total number of bronchi involved or Secretion type-I in \< half total number of bronchi involved
BS Grade 4 = Mild = Secretion type-I, \> half total number of bronchi involved or Secretion type-II, \< half total number of bronchi involved
BS Grade 5 = Mod = Secretion type-II, \> half total number of bronchi involved or Secretion type-III, \< half total number of bronchi involved
BS Grade 6 = Large = Secretion type-III, \> half total number of bronchi involved Inter-rater reliability of BS grading was assessed by replaying the videotapes of the recorded bronchoscopy of 20 children, with the 2^nd^assessor blinded to the child\'s condition.
BAL was obtained from the macroscopically most abnormal lobe; when changes were generalised, BAL was obtained from the right middle lobe. Cell count was performed on the cell suspension, cytocentrifuge slides were prepared and stained (modified Wright\'s stain) for cell differential profile. All cellular examinations were performed by cytologists blinded to the children\'s medical history.
Statistics
----------
Data were not normally distributed and thus non parametric analyses were used; medians and inter-quartile range (IQR) were used for all descriptive data and Kruskal Wallis for comparisons between groups. Cohen\'s kappa (K) with 95%CI was utilised for inter and intra-observer reliability and graded from \'poor\' (K\<0.2) to \'very good\' (K = 0.81--1.0)\[[@B23]\]. For calculation of sensitivity and specificity, negative and positive predictive values (NPV, PPV); cough quality was assigned to dry when a history of cough was absent and bronchoscopy findings at two cut offs (grades 3 and 4) of BS grades were taken as the \'gold standard\' eg for cut-off at BS grade 3, BS grades 1--2 were defined as no secretions and BS grades = 3 defined as secretions present. To determine if cough quality (wet/dry) was predictive of amount of secretions found during bronchoscopy, a receiver operating characteristic (ROC) curve was generated \[[@B24]\] where cough quality wet/dry was considered the true positive/negative and the bronchoscopic secretion scoring (1 to 6) as the ordinal rating scale. Two tailed p value of \< 0.05 was considered significant. SPSS ver 11.1 was utilised for most statistical calculation.
Results
=======
Median age of the 106 children (62 boys, 44 girls) enrolled was 2.6 years (IQR 5.7). Indications for bronchoscopy were chronic cough (n = 44, 41.5%), wheeze (n = 21, 19.8%), stridor (n = 16, 15.4%), investigation of persistent radiological changes (n = 14, 13.5%), recurrent pneumonia (n = 6, 5.8%), suspicion of aspiration lung disease (n = 3, 2.9%), BAL and suspected foreign body (n = 1 each, 2%). In four children, BS grades were not obtained (session was inadvertently not recorded and \'blinded\' clinician not present at bronchoscopy). Scores of BS were done in real time in all but 9 children.
In 30 children, cough was non-assessable. Agreement between clinicians and parents assessment of cough quality (wet/dry) was good (K = 0.75, 95%CI 0.58, 0.93). For cough quality of \'wet/dry\', cough assessed by clinicians had the highest specificity, sensitivity, NPV, PPV and positive likelihood ratio for both BS cut-offs (tables [1](#T1){ref-type="table"} and [2](#T2){ref-type="table"}). Parent(s) assessment were less precise but only marginally so. The area under the fitted ROC curve (figure [2](#F2){ref-type="fig"}) was 0.85 95%CI 0.77, 0.92. The specificity, NPV and likelihood ratio for brassy cough assessed against gold standard bronchoscopic finding of tracheomalacia was good (table [1](#T1){ref-type="table"}) but less than that for cough quality of wet/dry.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Assessment of cough quality vs bronchoscopic findings with BS cut off at grade 3\*
:::
**Assessment type (clinical vs bronchoscopic findings)** **Sensitivity** **Specificity** **NPV** **PPV** **Positive LR**
---------------------------------------------------------- ----------------- ----------------- --------- --------- -----------------
**Clinician** 1.00 0.55 1 0.64 2.21
Cough quality (wet/dry)
assessed by clinician (n = 96)
**Parent(s)** 0.95 0.54 0.93 0.61 2.06
Cough quality (wet/dry)
assessed by parents (n = 92)
**Combined**\*(n = 100) 0.98 0.54 0.97 0.62 2.10
**Tracheomalacia**(n = 81)\# 0.57 0.81 0.84 0.52 3.12
\*Cough quality (wet/dry) assessed by clinicians combined with parents. When cough was non-assessable by clinician and child has current cough, parental assessment of the cough (wet or dry) was taken. If child has no history of current cough, cough was assigned \'dry\'.
LR = likelihood ratio.
Specificity, sensitivity of dry and wet cough was assessed against bronchoscopic findings as the gold standard where \*BS grades ≥ 3 were considered abnormal (secretions present) and ≤ 2 considered normal (no secretions). \#That for tracheomalacia was assessed using clinicians record of presence/absence of brassy cough with bronchoscopic findings of tracheomalacia.\[21\]
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Assessment of cough quality vs bronchoscopic findings with BS cut off at grade 4\*
:::
**Assessment type (clinical vs bronchoscopic findings)** **Sensitivity** **Specificity** **NPV** **PPV** **Positive LR**
---------------------------------------------------------- ----------------- ----------------- --------- --------- -----------------
**Clinician** 0.79 0.75 0.82 0.72 3.22
Cough quality (wet or dry)
assessed by clinician (n = 96)
**Parent(s)** 0.78 0.71 0.80 0.67 2.69
Cough quality (wet or dry)
assessed by parents (n = 92)
**Combined**\* (n = 100) 0.77 0.73 0.80 0.69 2.88
\*Cough quality (wet/dry) assessed by clinicians combined with parents. When cough was non-assessable by clinician and child has current cough, parental assessment of the cough (wet or dry) was taken. If child has no history of current cough, cough was assigned \'dry\'.
LR = likelihood ratio.
Specificity, sensitivity of dry and wet cough was assessed against bronchoscopic findings as the gold standard where BS grades ≥ 4 were considered abnormal (secretions present) and ≤ 3 considered normal (no secretions).
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
ROC curve with 95%CI relating cough quality (wet/dry) to bronchoscopic secretion (BS) grades from 1--6.
:::
:::
There was little difference in sensitivity and specificity between children grouped by indication for bronchoscopy (cough or other indications). Values were marginally better in older children (tables 4 and 5 in supplementary data additional file [2](#S2){ref-type="supplementary-material"}). Area under the fitted ROC curve was similar for both age groups (aged ≤ 2 years = 0.811, 95%CI 0.79, 0.84; age \>2 = 0.84, 95%CI 0.74, 0.95). Agreement for clinicians vs parents cough quality (dry/wet) was better in children aged ≤ 2 years (K = 0.85, 95%CI 0.57, 1.0; n = 42 but 18 non-assessable) than that for those age \>2 years (K = 0.70, 95%CI 0.49, 0.92; n = 64, but 12 non-assessable) (see additional file [1](#S1){ref-type="supplementary-material"}).
Using recorded coughs, kappa scores were \'very good\' for both intra-observer and inter-clinician agreement for wet and dry cough (K = 1.0 and 0.88 \[95%CI 0.82--0.94\] respectively). There was only one disagreement in wet and dry cough between clinicians and in this child the cough was mildly wet (BS grade of 3). Kappa scores for intra-observer and inter-observer clinician agreement for brassy cough was good, K in both was 0.79, 95%CI 0.73, 0.86. Inter-rater agreement for BS grades was \'very good\' (weighted K = 0.95, 95%CI 0.87--1).
Cellularity for total cell count, percentages of neutrophils and macrophages were significantly different between children grouped by BS grade cut-offs of 3 and 4 as well as wet/dry cough (table [3](#T3){ref-type="table"}).
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Cellular differential profile in BALs
:::
**Median** **TCC (IQR)** **% M IQR)** **% N (IQR)** **% Lym (IQR)** **% Eos (IQR)**
-------------------------- --------------- -------------- --------------- ----------------- -----------------
**BS cutoff at grade 3**
≤2 (n = 31) 195 (290) 82.0 (15.8) 5.0 (7) 13.5 (15.8) 0 (0)
≥3 (n = 70) 334.0 (425) 66.0 (45) 12.0 (38) 11.0 (16.0) 0 (0)
p value\^ 0.038 0.001 0.006 0.605 0.758
**BS cutoff at grade 4**
≤3 (n = 52) 176 (257) 81.0 (17.0) 6.0 (8.0) 13.0 (16.0) 0 (0)
≥4 (n = 49) 368 (574) 51.5 (59.8) 20.0 (47.0) 11.0 (15.0) 0 (5)
p value\^ 0.0001 0.0001 0.0001 0.445 0.613
**Cough quality**\*
Wet (n = 45) 365 (522) 51.5 (49.8) 25.0 (43) 13.0 (16) 0 0
Dry (n = 25) 176 (315) 80.5 (24.8)) 5.5 (13.0) 1.8 (16.0) 0 (0)
No history (n = 28) 80 (310) 15 (16.5) 1 (7.5) 1 (11.5) 0 (0)
310 16.5 7.5 11.5 0
p value\^ 0.017 0.0001 0.001 0.242 0.769
\^p value = examined using Kruskal Wallis test.
\*assessed by clinician
TCC = total cell count; N = neutrophils, M = macrophages, L = lymphocytes, Eos = eosinophils,
:::
Discussion
==========
We have shown that clinical assessment of cough quality of wet/dry cough generally relates to bronchoscopic secretions determined using a standardised scoring system (BS grades). When cough is wet, secretions were always present; when cough was dry secretions if present, were usually minimal or mild. Clinicians were marginally better than parents at assessing wet/dry cough and agreement between the 2 groups was good. When clinicians detected presence of a brassy cough, tracheomalacia was usually present. Inter-rater clinician agreement for cough qualities of dry/wet and brassy/non brassy was good.
Accuracy and reliability of symptoms are important in clinical and research settings. Cane and colleagues \[[@B25],[@B26]\]. found that parental reports of wheeze and stridor are often not accurately reported in a clinic setting. There is no data on the validity of cough quality in spite its use in management and diagnostic guidelines \[[@B11],[@B27],[@B28]\] and cough being the most common symptom seen by general practitioners \[[@B1]-[@B3]\]. The level of agreement recommended for symptoms and signs to be used in clinical prediction rules is kappa value of ≥ 0.6 \[[@B29]\]. The kappa values we obtained in this study well exceeded 0.6. Specifically, intra and inter-clinician evaluation was very good and parental reporting of cough quality (wet/dry) also related well to clinicians\' evaluation.
When compared to bronchoscopic findings, this study showed that a wet cough is always associated with BS grades of 3 or more. Dry cough is less valid; the presence of dry cough does not necessary indicate absence of secretions. However BS grades are less in dry cough as shown in the ROC curve. The generation of cough sounds and some factors that influence cough sounds have been examined in the laboratory \[[@B16],[@B30]\]. Using cough sound analysis (spectrogram and time-expanded waveform), productive and non-productive cough can be differentiated in the laboratory \[[@B30]\]. However to date there is no data on its clinical reliability and its relationship to quantification of airway secretions. In humans, it is not known how much mucus is required and where it has to be located for the human ear to detect presence of a moist cough. It is likely that mucus in the large airways is required for detectable difference in cough quality as the sound of cough is generated from vibration of larger airways and laryngeal structures during turbulent flow in expiration \[[@B16],[@B17]\]. Laminar airflow, which occurs in smaller airways, is inaudible \[[@B31]\]. In an animal model, Korpas and colleagues showed that a certain amount of mucus is required to alter cough sound; 0.5 ml of mucus instilled into the trachea of cats altered cough sound, too little mucin had no effect on cough quality whilst too much mucin impaired breathing \[[@B32]\]. Our study findings support this and it is not surprising that when the cough is dry, BS grades were less. The rheological properties of airway mucus also influence cough sound \[[@B17]\]. It is not known how airway secretions in the more peripheral airways influences the sound of cough.
One possible limiting factor of our study is the choice of cut offs for BS grades in determining presence or absence of significant secretions. We chose to use a cut off of 3 as a minor amount of bubbles in the airways can be present from trickling of lignocaine into the airways or spillage from the upper airways. BS cut-off at grade 4 resulted in improved specificity but decreased sensitivity. Children grouped by both BS cut-offs (3 and 4) had significantly different airway cellular profile. The clinical significance of minimal BS grades and appropriate cut-offs can only be determined in a prospective follow-up study which is not an aim of this study. This study did determine that our BS scoring method was easy to use (most done in real time) and had very good inter-rater agreement. The clinical outcomes of wet and dry cough were not the aims of this study and thus cannot be determined here. To relate clinical outcomes to cough descriptors would ideally require a randomised controlled trial with dry and wet cough as entry criteria. A follow-up cohort study with strict clinical diagnostic categories would be useful and we have shown in a preliminary study that dry cough was significantly more likely to naturally resolve than wet cough \[[@B33]\].
In addition to the limitation of quantifying airway secretions using a bronchoscopic method, this study is also limited by a number of factors. Firstly, clinical repeatability or agreement of cough sounds was assessed by doctors in a tertiary setting. Whether or not these findings can be extrapolated to the secondary and primary setting can only be speculated. Hay and colleagues showed that inter-observer agreement for clinical signs of fever, tachypnoea and chest signs were poor to fair (kappa of 0.12--0.39) in the primary care setting but these signs are known to have good agreement in secondary care settings \[[@B34]\]. However as parents were almost as good as clinicians in our study and are \'untrained\' compared to medical practitioners, we would expect that this data can be extrapolated to most primary and secondary settings. Secondly, anaesthesia and atropine could possibly influence mucus quantity and properties. However this influence, if any, is likely to be small as both bronchoscopists (ABC, IBM) are experienced (our recorded average total theatre time is relatively short at 22 mins) \[[@B22]\], and atropine is given just immediately prior to commencement of bronchoscopy.
Determining the validity of cough quality in children is important not only because of the commonality of the clinical problem of cough but also its use in guidelines and research studies \[[@B11],[@B27],[@B28]\]. A particularly important finding is the presence of small amounts of secretions in children with dry cough which may have implications in the management of suppurative lung disease; a dry cough may represent early disease process where only a small amount of mucous is present.
Conclusion
==========
We conclude that the description of a cough as wet or dry cough as determined by clinicians and parents has good clinical validity as it has good agreement with, and relates to, quantification of airway secretions. However as minimal amount of secretions may be present in children with dry cough, clinicians should be cognisant that a dry cough may eventually become wet if airway secretions increase. Thus it should not be assumed that airway secretions are absent in children with chronic dry cough and cough quality in these children should be reviewed. We also conclude that the brassy cough determined by respiratory physicians is highly specific for presence of tracheomalacia.
List of Abbreviations
=====================
BAL Bronchoalveolar lavage
BS Bronchoscopic secretion
K Kappa
NPV Negative predictive value
PPV Positive predictive value
ROC receiver operating characteristic
Authors\' contributions
=======================
AC conceived the idea, designed the study, performed the data analysis and drafted the manuscript. JG participated in data acquisition and coordination of project. ME participated in electronic acquisition of data and software for sound recordings. JF and NC designed the microbiology and cytological components respectively and both helped draft the manuscript. IBM helped in formulation of overall study design, data acquisition and drafting of the manuscript. All authors read and approved the manuscript.
Supplementary Material
======================
::: {.caption}
###### Additional File 1
**Figure 3: ROC curve**ROC curve with 95%CI relating cough quality (wet/dry) to bronchoscopic secretion (BS) grades from 1--6 in children grouped according into age (a) ≤ 2 years and (b) \> 2 years.
:::
::: {.caption}
######
Click here for file
:::
::: {.caption}
###### Additional File 2
**Table 4: Assessment of cough quality vs bronchoscopic findings in children grouped by indication for bronchoscopy**4a: Assessment of cough quality vs bronchoscopic findings in children whose indication for bronchoscopy was cough 4b: Assessment of cough quality vs bronchoscopic findings in children whose indication for bronchoscopy was others (ie not cough) **Table 5: Assessment of cough quality vs bronchoscopic findings in children grouped by age**5a: Assessment of cough quality vs bronchoscopic findings in children aged ≤ 2 years 5b: Assessment of cough quality vs bronchoscopic findings in children aged \> 2 years
:::
::: {.caption}
######
Click here for file
:::
Acknowledgment
==============
We thank members of the anaesthetic department, Royal Children\'s Hospital for their help, in particular Dr. L Patterson and Dr. J Wuth. We also thank Dr M McElrea for proof reading the manuscript and Barry Dean for providing the images used in the bronchoscopic secretion card (figure [1](#F1){ref-type="fig"}). ABC is supported by the National Health and Medical Research Council and the Royal Children\'s Hospital Foundation.
|
PubMed Central
|
2024-06-05T03:55:51.917781
|
2005-1-8
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545936/",
"journal": "Respir Res. 2005 Jan 8; 6(1):3",
"authors": [
{
"first": "Anne Bernadette",
"last": "Chang"
},
{
"first": "Justin Thomas",
"last": "Gaffney"
},
{
"first": "Matthew Michael",
"last": "Eastburn"
},
{
"first": "Joan",
"last": "Faoagali"
},
{
"first": "Nancy C",
"last": "Cox"
},
{
"first": "Ian Brent",
"last": "Masters"
}
]
}
|
PMC545937
|
Background
==========
Sexually transmitted diseases (STDs) have become a major public health concern in the UK during recent years. The rates of STDs in England have been rising steadily since the mid 1990s. In 2003, the number of STDs in England rose by 4% compared to 2002. Overall, 672,718 people were diagnosed with an STD in England in 2003, and around one third of those cases were diagnosed in the London area alone \[[@B1],[@B2]\].
The House of Commons Hansard Written Answers for 15 October 2004 included an answer by Melanie Johnson MP, Minister for Public Health at the Department of Health, to a question by Sarah Teather MP on \"how many cases of diagnosed STDs there were in each Primary Care Trust (PCT) in London in each year since 1997\". The answer was provided in the form of a long table showing the figures for 25 PCTs in London (Table [1](#T1){ref-type="table"} -- \[[@B2]\]).
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Diagnoses of STDs by PCT in London, 1997 -- 2003
:::
**PCT Name** **1997** **1998** **1999** **2000** **2001** **2002** **2003**
------------------------------------- ---------- ---------- ---------- ---------- ---------- ---------- ----------
Barking and Dagenham 2,454 2,797 3,020 2,764 2,457 3,461 3,292
Barnet 1,542 1,591 1,546 1,993 2,001 1,899 923
Brent 8,753 8,851 8,724 9,258 9,479 10,156 10,113
Bromley 1,386 1,899 2,287 2,468 3,296 3,683 3,575
Camden 15,714 17,117 17,478 18,750 20,962 22,257 26,987
City and Hackney 14,998 16,595 16,601 16,169 19,171 22,673 21,072
Croydon 4,089 5,546 5,707 6,968 7,805 7,995 7,587
Ealing 1,514 585 1,518 1,687 2,062 2,751 1,832
Enfield 581 1,513 1,168 1,481 1,957 2,129 1,659
Greenwich 4,361 4,314 4,901 6,012 5,170 5,385 6,607
Hammersmith and Fulham 7,838 7,889 8,281 9,556 9,992 4,996 5,723
Haringey 5,029 5,229 5,568 6,344 7,183 6,617 6,379
Hillingdon 1,931 2,387 3,038 3,096 4,131 4,000 3,110
Hounslow 3,270 3,185 3,029 4,101 4,637 5,861 6,473
Islington 5,339 5,552 6,259 5,781 5,309 4,988 6,186
Kensington and Chelsea 11,995 11,636 11,040 13,149 13,301 11,739 12,243
Kingston 1,926 2,443 2,646 3,129 3,474 3,872 4,481
Lambeth 17,865 19,998 21,327 19,971 19,754 22,003 22,209
Lewisham 412 353 346 482 21 \- \-
Newham 8,948 9,033 11,214 11,023 12,769 13,200 15,236
Southwark 16,481 17,073 15,232 13,836 16,699 19,618 19,249
Sutton and Merton 10,870 9,110 11,569 13,464 14,229 16,589 17,784
Walthamstow, Leyton and Leytonstone 1,157 1,355 1,726 1,713 2,311 2,433 2,988
Wandsworth 2,629 2,712 3,072 3,004 3,519 4,151 4,199
Westminster 18,639 21,490 21,359 20,503 20,870 18,462 19,657
Diagnoses of sexually transmitted diseases by Primary Care Trust in London, 1997 -- 2003 (Source: \[2\]). Overall figures may be lower than stated for the London region in the annual report because the data presented here have not been imputed. One clinic in each of these PCTs (Barnet; City and Hackney) did not submit all the KC60 returns for 2003. The Alexis Clinic of Lewisham PCT closed in June 2001; no GUM clinics currently open in this PCT.
:::
Though the table presents all the requested data, it remains very difficult for the reader to fully appreciate the patterns and trends buried in them, or make quick and effective comparisons between the figures for different PCTs or between the seven data sets for the years from 1997 to 2003.
Such data patterns, trends and comparisons derived from this Hansard table and other sources, e.g., demographic, deprivation/social exclusion, transport and existing GUM (Genito-Urinary Medicine) clinic data sets, are crucial for the decision maker wanting, for example, to:
\- improve access to GUM clinics and make decisions regarding the expansion or closure of existing clinics, or the creation of new ones;
\- channel resources and target STD prevention programmes to areas with the most need, or scale such programmes according to the magnitude of the problem in different areas (this is especially important in a climate of finite resources); and/or
\- monitor the impact of such programmes in a given area over time.
In this paper, we describe a much better way of presenting the same Hansard table data in the form of interactive Web maps in Scalable Vector Graphics (SVG) format to further support health planners and decision makers in their planning and management tasks.
Results
=======
Using GeoReveal, a tool from Graphical Data Capture Ltd (<http://www.graphdata.co.uk/product.asp?product_id=GeoReveal> -- see \'Methods\' section below), we produced Web-based interactive choropleth maps of diagnoses of STDs by PCT in London for the years from 1997 to 2003, which readers can browse at <http://healthcybermap.org/PCT/STDs/> (Figure [1](#F1){ref-type="fig"}). These maps and companion bar chart (\'Chart Panel\') are based on the data in Table [1](#T1){ref-type="table"}, and show steadily rising rates of STDs in London over the covered seven-year period. Also, one can clearly see on the maps that PCTs located in central London, e.g., Camden PCT and Lambeth PCT, had the highest numbers of STD diagnoses throughout the seven years from 1997 to 2003.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Screenshot from our interactive Web maps of STD diagnoses by London PCT, 1997 -- 2003.**Screenshot from our Web-based interactive choropleth maps of diagnoses of STDs by PCT in London for the years from 1997 to 2003 <http://healthcybermap.org/PCT/STDs/>. The map shown in this screenshot is for the year 2002, with Camden PCT highlighted in yellow. The bar chart (\'Chart Panel\') on the left shows steadily rising STD rates in Camden PCT over the covered seven-year period. Camden\'s rates are well above the average for all 25 mapped PCTs over the same period (the purple portions of the bars represent the average for all PCTs). The maps require the free Adobe SVG Viewer <http://www.adobe.com/svg/viewer/install/main.html>. Visitors will be automatically prompted to download it on their first visit to the site, if they don\'t already have it installed on their machine. Scripting must also be enabled in Internet Explorer.
:::
:::
As the mouse cursor is moved around the main map window, the \'Chart Panel\' changes to display statistical data about the currently highlighted PCT. Each row of the bar chart represents data for one year and has two bars; a red bar that shows the value for the highlighted PCT, and a transparent blue bar which shows a mean value of this piece of data (or year) for all 25 mapped PCTs (when both bars overlap, a purple colour is produced -- Figure [1](#F1){ref-type="fig"}).
Additionally, clicking a PCT area on the map will display an information box with all available data for that PCT (Figure [2](#F2){ref-type="fig"}).
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Screenshots of two information boxes from our interactive Web map interface.**Upper information box: clicking a PCT area on the map (\'Camden PCT\' for this screenshot) will display this pop-up box with all available data for that PCT. Lower information box: clicking the \'Info\' button (shown in Figure 1 -- bottom right \'Navigation Panel\') will open this pop-up box with extra information about the maps, links to the Web sites of London PCTs, and detailed help about the map interface.
:::
:::
An \'Options Panel\' controls the data that are shown in the map window. Users can tick the \'Background Map\' box to add a raster background map of London to the main map. They can use the \'Choose Classification Method\' list to select how the data are mapped; with equal ranges, equal counts or by highlighting the highest values. A \'Number of Colours\' list allows users to select how many classes or ranges are used in the choropleth map (two to five ranges). Using the \'Select Map Topic\' list, users can select the topic that is shown on the main map (a year from 1997 to 2003). Finally, users can select from the \'Colour of Theme\' list the colour theme that is used in the main map (orange, green, or blue).
Map zooming, panning, MapTips (displaying PCT names), and a dynamic legend are available. An overview map shows a miniature version of the full extent map. When the user zooms in, a rectangle on the overview map highlights the area that is currently being displayed in the main map window. The user can click and drag this rectangle to change the view in the main map window. After zooming into the main map, users can use the \'Reset\' button to return to the full extent of the main map.
An \'Info\' button is also available. Clicking this button will open a pop-up window with extra information about the maps, links to the Web sites of London PCTs, and detailed help about the map interface (Figure [2](#F2){ref-type="fig"}).
Discussion
==========
From complex raw data to valuable decision support information
--------------------------------------------------------------
Turning raw tabular data into much more useful and accessible visual information in the form of interactive Web maps is much needed to support and empower decision makers, and even members of the general public. Such maps help us understand the relationships, patterns and trends buried in the original data sets and also enable instant visual comparisons to be made between different geographical areas and over time (when data sets for successive periods of time are available) \[[@B3]\].
We believe this transformation of raw data into valuable decision support information is very evident in the London STD example described in this paper. Readers only have to compare the original Hansard table (Table [1](#T1){ref-type="table"}) with the corresponding interactive Web maps we have produced <http://healthcybermap.org/PCT/STDs/> to see the difference for themselves and appreciate the value of interactive maps.
SVG: an ideal format for interactive Web maps
---------------------------------------------
SVG is a non-proprietary language for describing rich, stylable two-dimensional graphics and graphical applications in XML (eXtensible Markup Language). SVG is fully endorsed by the W3C (World Wide Web Consortium -- <http://www.w3.org/Graphics/SVG/>). It is rapidly becoming a popular choice for delivering interactive Web maps, being designed to work effectively across platforms, output resolutions, colour spaces, and a range of available bandwidths. It offers a rich modern graphics format providing the ability for better map display, and advanced graphical features such as transparency, arbitrary geometry, filter effects (shadows, lighting effects, etc.), scripting, and animation \[[@B4]\].
All these features have made SVG a direct competitor to the proprietary Macromedia^®^Flash format \[[@B5]\]. Vector-based images (describing shapes and paths), such as those in SVG and SWF (Macromedia^®^Shockwave/Flash File) formats, will keep their sharp character when enlarged, while raster-based images (storing information about each and every pixel in the image), such as those saved in GIF (Graphics Interchange Format) or JPEG (Joint Photographic Experts Group) formats, will show jagged edges.
A free SVG Web browser plug-in is available from Adobe for different platforms (Adobe SVG Viewer -- <http://www.adobe.com/svg/viewer/install/main.html>), in the same way the free Adobe Reader software is available for rendering PDF (Portable Document Format) files.
Besides the example described in this paper, other examples of SVG interactive Web maps in the health arena include the Office for National Statistics\' England and Wales 2001 Census Key Statistics maps <http://www.statistics.gov.uk/census2001/censusmaps/index_new.html> and 2001 Area Classification for Health Areas maps <http://www.statistics.gov.uk/about/methodology_by_theme/area_classification/ha/maps.asp>, and Leeds Interactive Health Atlas <http://www.leeds.nhs.uk/professional/healthatlas/>.
Tools for producing interactive SVG and Flash maps from desktop GIS projects
----------------------------------------------------------------------------
Besides GeoReveal <http://www.graphdata.co.uk/product.asp?product_id=GeoReveal>, the tool we have used to produce the maps described in this paper (see \'Methods\' section below), other SVG/Flash mapping tools available today for publishing maps created in desktop GIS (Geographic Information Systems) include GéoClip <http://www.geoclip.net/an/>, SVGMapMaker <http://www.tetrad.com/svgmapmaker/svgmapmaker.html>, MapViewSVG <http://www.uismedia.de/mapview/eng/>, and SVGMapper <http://www.svgmapper.com/>. The latter two tools (MapViewSVG and SVGMapper) are specific to ESRI ArcView GIS. Table [2](#T2){ref-type="table"} provides an overview of the features of GeoReveal, GéoClip, and SVGMapMaker.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Overview of the features of GeoReveal, GéoClip, and SVGMapMaker
:::
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
**GeoReveal** **GéoClip** **SVGMapMaker**
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
GeoReveal is a powerful .NET application that combines geography and statistics to produce revealing interactive graphics for the Web. It produces fully interactive geo-statistical presentations. There is no requirement for a GIS. GeoReveal uses MID/MIF files that can be exported from practically any GIS.\ GéoClip is a MapBasic program that takes TAB file information and turns it into interactive Macromedia^®^Flash presentations. There is also a version for ESRI ArcGIS. GéoClip does not currently have an SVG component.\ SVGMapMaker is a MapBasic program that takes TAB file information and uses it to replicate limited GIS functionality in a browser window. It cannot be considered as a geo-statistical tool, as the level of interactivity is limited. For example, there are no interactive charts as provided in GeoReveal.\
Other information:\ Other information:\ Other information:\
- Unlimited data fields and includes a comprehensive set of pre-defined templates\ - Written in MapBasic and therefore dependent on having a MapInfo Pro licence\ - Written in MapBasic and therefore dependent on having a MapInfo Pro licence\
- Features a simple wizard driven interface and an advanced set of menus for users who need to fine tune the finished output\ - Does not offer a very intuitive interface\ - Limited interface -- no ability to change map topic and no statistical delivery\
- Generates SVG data, overlay boundaries, overview and background maps. For UK Local Government users, SVG is also an e-GIF (e-Goverment Interoperability Framework) compliant format\ - Limited to ten data fields and to only one page layout\ - Map display cannot be changed after generation, i.e., no ability to change map colour, number of ranges, and classification method
- Presentations can include interactive charts that change to display statistics about the currently selected map region or area\ - Users cannot remove the GéoClip logo/branding from the map page\
- Map display can be changed after generation, i.e., ability to change map topic, map colour, number of ranges, and classification method\ - Can only produce Flash (proprietary format, not e-GIF compliant) and not SVG (open format)\
- Includes pre-prepared UK census boundary data and sample projects - Some of its unique features like adjustable ranges and multiple themes are planned for the next release of GeoReveal
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
A quick comparison of three tools available today for producing interactive SVG and Flash maps from desktop GIS projects: GeoReveal (from Graphical Data Capture Ltd, London, UK -- <http://www.graphdata.co.uk/product.asp?product_id=GeoReveal>), GéoClip (from eMc3, a young company based in Toulouse, France -- <http://www.geoclip.net/an/>), and SVGMapMaker (from TETRAD Computer Applications Inc, Bellingham, WA, USA -- <http://www.tetrad.com/svgmapmaker/svgmapmaker.html>).
:::
Conclusions
===========
Using GeoReveal v1.1 for Windows, we produced Web-based interactive choropleth maps of diagnoses of STDs by PCT in London for the years from 1997 to 2003 <http://healthcybermap.org/PCT/STDs/>. These maps are in SVG format and require a freely available Adobe SVG browser plug-in to be displayed. They are based on data obtained from the House of Commons Hansard Written Answers for 15 October 2004. They show steadily rising rates of STDs in London over the covered seven-year period. Also, one can clearly see on the maps that PCTs located in central London had the highest numbers of STD diagnoses throughout the mapped seven years. A companion bar chart allows users to instantly compare the STD figure of a given PCT for a given year against the average figure for all 25 mapped PCTs for the same year, and also compare those figures across all seven years. The maps offer users a rich set of useful features and functions, including the ability to change the classification method in use, the number of ranges in the map, and the colour theme, among others.
We also presented a quick review of some of the tools available today for creating interactive vector graphics maps from desktop GIS projects.
Wizard-driven tools like GeoReveal have made it very easy to transform complex raw data into valuable decision support information products (interactive Web maps) in very little time and without requiring much expertise. The resultant interactive maps have the potential of further supporting health planners and decision makers in their planning and management tasks by allowing them to graphically interrogate data, instantly spot trends, and make quick and effective visual comparisons of geographically differentiated phenomena between different geographical areas and over time.
SVG makes an ideal format for such maps. SVG is a W3C non-proprietary, XML-based vector graphics format, and is an extremely powerful alternative to Macromedia^®^Flash and bitmap graphics.
Methods
=======
We used GeoReveal v1.1 for Windows to create the interactive SVG maps described in this paper. GeoReveal runs under Windows 98/NT/2000/XP and requires Microsoft^®^.NET framework v1.1 to be installed on the production machine.
We started by extracting London PCT boundaries from a larger data set of all England (2001 Census PCT -- post April 2002 change), which is the copyright of the Crown/Ordnance Survey <http://www.ordsvy.gov.uk/>, and is freely available in both ArcView and MapInfo formats to the UK academic community from EDINA UKBORDERS service with the support of the ESRC and JISC <http://edina.ac.uk/ukborders/>. We also prepared a spreadsheet containing data about the number of STDs recorded in each London PCT between 1997 and 2003 using data from \[[@B2]\]. The two files were merged using MapInfo Professional v7.5 <http://www.mapinfo.com/>, creating a MapInfo .TAB file. This file was exported to MID/MIF format, the format that files need to be in, in order to be used by GeoReveal.
We created our presentation (the interactive SVG Web maps) using the GeoReveal Wizard; an eight-step process that allows users to create a fully interactive SVG page.
Introduction: setting the output directory
------------------------------------------
The first Wizard dialog that must be completed is the \'Introduction\' dialog, which is used to specify the directory to which the final GeoReveal output files will be saved (Figure [3](#F3){ref-type="fig"}).
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Screenshot of the \'Introduction\' dialog.**Screenshot of the \'Introduction\' dialog in GeoReveal Wizard. For a detailed description of the functions available in this dialog, please refer to the \'Methods\' section \> \'Introduction: setting the output directory\'.
:::
:::
To select the output directory, we clicked the \'Browse\' button and in the resulting file browse dialog, we selected the required directory and clicked \'OK\'. To move to the next step, we clicked \'Next\'.
Wizard step 1: choosing the template
------------------------------------
The first Wizard step is the \'Choose Template\' dialog (Figure [4](#F4){ref-type="fig"}). This dialog is used to specify general settings for a GeoReveal presentation.
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Screenshot of the \'Choose Template\' dialog.**Screenshot of the \'Choose Template\' dialog in GeoReveal Wizard (Wizard Step 1 of 8). For a detailed description of the functions available in this dialog, please refer to the \'Methods\' section \> \'Wizard step 1: choosing the template\'.
:::
:::
We selected the template to be used. Ten templates are provided; four that include a bar chart, four that include a pie chart and two that have no chart. With our London PCT STD data, the most appropriate is a bar chart template.
We selected a logo image to be added to the top-left corner of our GeoReveal page. We clicked the \'Browse\' button and selected the required file -- in this instance, a University of Bath logo has been used.
We then specified the title of the GeoReveal page. This title is placed next to the logo at the top of the output page (see the logo and title at <http://healthcybermap.org/PCT/STDs/>.
Finally, we selected the title colour and page background colour. To move to the next step, we clicked \'Next\'.
Wizard step 2: selecting the main map data
------------------------------------------
The second step is the \'Select Main Map Data\' dialog (Figure [5](#F5){ref-type="fig"}). This dialog is used to specify settings for the main GeoReveal map.
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Screenshot of the \'Select Main Map Data\' dialog.**Screenshot of the \'Select Main Map Data\' dialog in GeoReveal Wizard (Wizard Step 2 of 8). For a detailed description of the functions available in this dialog, please refer to the \'Methods\' section \> \'Wizard step 2: selecting the main map data\'.
:::
:::
We selected the MID/MIF file that contains the statistics that will form the main map in our GeoReveal page by clicking the \'Browse\' button and choosing the required file.
We selected the field from this MID/MIF file that will be used for ToolTips. A ToolTip (or MapTip) is the piece of text that is displayed when the mouse cursor is hovered over a region in the map. In this instance, the PCT Name field has been selected.
Users can also choose to turn the information box on or off. When it is turned on, a user can click a region in the map to display a dialog containing all information held within the MID/MIF file about that region (Figure [2](#F2){ref-type="fig"}).
Finally, we selected the background colour and highlight colour for the map. The highlight colour is the colour a map region (an individual PCT area in our case) will be displayed in when the mouse cursor hovers over it. To move to the next step, we clicked \'Next\'.
Wizard step 3: selecting the overview map data
----------------------------------------------
The third step is the \'Select Overview Map Data\' dialog (Figure [6](#F6){ref-type="fig"}). This dialog is used to specify settings for the GeoReveal overview map.
::: {#F6 .fig}
Figure 6
::: {.caption}
######
**Screenshot of the \'Select Overview Map Data\' dialog.**Screenshot of the \'Select Overview Map Data\' dialog in GeoReveal Wizard (Wizard Step 3 of 8). For a detailed description of the functions available in this dialog, please refer to the \'Methods\' section \> \'Wizard step 3: selecting the overview map data\'.
:::
:::
We selected the MID/MIF file to be used for the overview map by clicking the \'Browse\' button and selecting the required file. It is possible to use the same MID/MIF file for both the main and overview maps.
We selected the overview map colour and the overview rectangle colour (this rectangle shows a user where they are currently zoomed in on the main map). Additionally, users can select the background colour and border colour for the overview map panel. To move to the next step, we clicked \'Next\'.
Wizard step 4: selecting options controls
-----------------------------------------
The fourth step is the \'Select Options Controls\' dialog (Figure [7](#F7){ref-type="fig"}). This dialog is used to specify settings for the \'Options Panel\' in the presentation.
::: {#F7 .fig}
Figure 7
::: {.caption}
######
**Screenshot of the \'Select Options Controls\' dialog.**Screenshot of the \'Select Options Controls\' dialog in GeoReveal Wizard (Wizard Step 4 of 8). For a detailed description of the functions available in this dialog, please refer to the \'Methods\' section \> \'Wizard step 4: selecting options controls\'.
:::
:::
First, we selected the background map image that is to be used. We ticked the top \'Checkbox Visible\' box to activate the background map option and then clicked the top \'Browse\' button to select the file that will be used. This file must be a JPEG or GIF image and information is needed about its width and height in real terms, and also the bounding coordinates.
Once the required file is selected, the \'Background Image Settings\' dialog is displayed (Figure [8](#F8){ref-type="fig"}). In this dialog, we entered the bounding coordinates of our selected image, the height and width of the image in real terms, and specified how opaque the background map will be in the presentation. We then clicked \'OK\' to confirm and return to the previous dialog (Figure [7](#F7){ref-type="fig"}).
::: {#F8 .fig}
Figure 8
::: {.caption}
######
**Screenshot of the \'Background Image Settings\' dialog.**Screenshot of the \'Background Image Settings\' dialog in GeoReveal Wizard. For a detailed description of the functions available in this dialog, please refer to the \'Methods\' section \> \'Wizard step 4: selecting options controls\'.
:::
:::
In the \'Select Options Controls\' dialog, it is also possible to:
\- enable a vector map option; and
\- select the text colour, border colour, background colour and control colour (the colour in which the fields are rendered) for the \'Options Panel\'.
To move to the next step, we clicked \'Next\'.
Wizard step 5: setting the bar chart
------------------------------------
The fifth step is the \'Set Bar Chart\' dialog (Figure [9](#F9){ref-type="fig"}). This dialog is used to specify settings for the bar chart in the presentation. This dialog will differ depending on the template that was selected. In this dialog, one can:
::: {#F9 .fig}
Figure 9
::: {.caption}
######
**Screenshot of the \'Set Bar Chart\' dialog.**Screenshot of the \'Set Bar Chart\' dialog in GeoReveal Wizard (Wizard Step 5 of 8). For a detailed description of the functions available in this dialog, please refer to the \'Methods\' section \> \'Wizard step 5: setting the bar chart\'.
:::
:::
\- specify the bar colour, grid colour (if the grid is enabled), text colour, background colour and border colour;
\- add or remove a % sign to the figures that are shown at the end of each bar. As the data being used in this presentation are absolute, the \'Show %\' box should be unchecked; and
\- specify the opacity of the bars in the bar chart.
After specifying our settings for the bar chart, we clicked \'Next\' to move to the next step.
Wizard step 6: setting the legend
---------------------------------
The sixth step is the \'Set Legend\' dialog (Figure [10](#F1){ref-type="fig"}). This dialog is used to specify settings for the legend panel in the presentation. Users can:
::: {#F10 .fig}
Figure 10
::: {.caption}
######
**Screenshot of the \'Set Legend\' dialog.**Screenshot of the \'Set Legend\' dialog in GeoReveal Wizard (Wizard Step 6 of 8). For a detailed description of the functions available in this dialog, please refer to the \'Methods\' section \> \'Wizard step 6: setting the legend\'.
:::
:::
\- specify a title and subtitle for the legend. Rather than specifying a main title, we checked the \'Use Dynamic Legend\' box. When this box is checked, the title will be determined by the topic that the GeoReveal map is based on (a year from 1997 to 2003 in our case);
\- enable or disable the \'Highlight Range\' option. If this option is enabled (as is the case in our presentation), when the mouse is hovered over a map region (PCT), the legend range that this region falls into will be highlighted; and
\- specify the text colour, background colour and border colour for the legend panel.
After specifying our settings for the legend, we clicked \'Next\' to move to the next step.
Wizard step 7: setting the \'Information Panel\'
------------------------------------------------
The seventh step is the \'Set Information Panel\' dialog (Figure [11](#F11){ref-type="fig"}). This dialog is used to specify settings for the \'Information Panel\' in the presentation. In the presentation, this panel displays information about the region in the map (PCT) that the mouse cursor is currently over. Users can:
::: {#F11 .fig}
Figure 11
::: {.caption}
######
**Screenshot of the \'Set Information Panel\' dialog.**Screenshot of the \'Set Information Panel\' dialog in GeoReveal Wizard (Wizard Step 7 of 8). For a detailed description of the functions available in this dialog, please refer to the \'Methods\' section \> \'Wizard step 7: setting the \'Information Panel\".
:::
:::
\- enable or disable each of the information panels and select the fields from the MID/MIF file that will be used to populate each panel. A title can also be entered for each panel (in our instance, Panel 1 was enabled, has been given the title \'PCT Name\' and will display the PCT Name field from the MID/MIF file); and
\- specify the text colour, background colour and border colour for the \'Information Panel\'.
We then clicked \'Next\' to move to the next step.
Wizard step 8: setting the \'Navigation Panel\'
-----------------------------------------------
The eighth step is the \'Set Navigation Panel\' dialog (Figure [12](#F12){ref-type="fig"}). This dialog is used to specify settings for the \'Navigation Panel\' in the presentation: the text colour, background colour, border colour and button colour.
::: {#F12 .fig}
Figure 12
::: {.caption}
######
**Screenshot of the \'Set Navigation Panel\' dialog.**Screenshot of the \'Set Navigation Panel\' dialog in GeoReveal Wizard (Wizard Step 8 of 8). For a detailed description of the functions available in this dialog, please refer to the \'Methods\' section \> \'Wizard step 8: setting the \'Navigation Panel\".
:::
:::
We then clicked \'Next\' to move to the final step.
Generating the presentation
---------------------------
Lastly, we saved the settings for our presentation and generated an initial presentation page (Figure [13](#F13){ref-type="fig"}). The \'Wizard Final Step\' dialog enables users to:
::: {#F13 .fig}
Figure 13
::: {.caption}
######
**Screenshots of the \'Wizard Final Step\' dialog and final \'Build Successful\' message box.**Screenshots of the \'Wizard Final Step\' dialog and final \'Build Successful\' message box in GeoReveal Wizard. For a detailed description of the functions available in this dialog, please refer to the \'Methods\' section \> \'Generating the presentation\'.
:::
:::
\- save the settings they have just made by clicking \'Save Settings\', then in the resulting dialog, browsing to the required directory and clicking \'Save\'. The saved settings file enables users to restore their settings and edit the presentation at a later date; and
\- generate the presentation by clicking \'Finish\'. A message is displayed to confirm that the presentation has been successfully generated (Figure [13](#F13){ref-type="fig"}). Clicking \'Yes\' on the message box will open the page in Internet Explorer.
When the presentation is created, the \'Advanced View\' window is also opened. This can be used to add the finishing touches to the presentation. It contains all the options that the wizard does, along with a few advanced options.
\'Advanced View\': \'Legend\' tab
---------------------------------
The \'Legend\' tab can be used to select the legend colour schemes that will be available (Figure [14](#F14){ref-type="fig"}). In this tab, users can add and remove colour schemes using the \'Add\' and \'Remove\' buttons, and ensure that the legend to be first loaded is at the top. Legend schemes are provided with the GeoReveal installation, and additional ones can be created using this tab. Moreover, the number of decimal places used in the legend can be changed. For this presentation, it has been set to 0.
::: {#F14 .fig}
Figure 14
::: {.caption}
######
**Screenshot of the \'Legend\' tab in the \'Advanced View\' window.**Screenshot of the \'Legend\' tab in the \'Advanced View\' window in GeoReveal Wizard. For a detailed description of the functions available in this tab, please refer to the \'Methods\' section \> \"Advanced View\': \'Legend\' tab\'.
:::
:::
\'Advanced View\': \'Bar Chart\' tab
------------------------------------
The \'Bar Chart\' tab can be used to edit the bar chart (Figure [15](#F15){ref-type="fig"}). It is possible to add average value bars to the bar chart. When enabled, a second bar is added to each row of the chart which shows the mean value of all the data in that field, as opposed to the first set of bars showing values for the currently selected map region (PCT) alone. To enable this:
::: {#F15 .fig}
Figure 15
::: {.caption}
######
**Screenshot of the \'Bar Chart\' tab in the \'Advanced View\' window.**Screenshot of the \'Bar Chart\' tab in the \'Advanced View\' window in GeoReveal Wizard. For a detailed description of the functions available in this tab, please refer to the \'Methods\' section \> \"Advanced View\': \'Bar Chart\' tab\'.
:::
:::
\- we ticked the \'Show Average Value Bars\' box to add average value bars to the bar chart; and
\- set the average bar colour and specified the opacity of the bars.
\'Advanced View\': \'Navigation Toolbar\' tab
---------------------------------------------
Finally, we used the \'Navigation Toolbar\' tab to edit the \'Navigation Panel\' (Figure [16](#F16){ref-type="fig"}). In this tab, it is possible to add a button to the \'Navigation Panel\', which when clicked, will open a simple HTML (HyperText Markup Language) page that contains information on how the GeoReveal presentation can be used and additional information about what it shows. To do this:
::: {#F16 .fig}
Figure 16
::: {.caption}
######
**Screenshot of the \'Navigation Toolbar\' tab in the \'Advanced View\' window.**Screenshot of the \'Navigation Toolbar\' tab in the \'Advanced View\' window in GeoReveal Wizard. For a detailed description of the functions available in this tab, please refer to the \'Methods\' section \> \"Advanced View\': \'Navigation Toolbar\' tab\'.
:::
:::
\- we enabled the help button by ticking the \'Display Help Button\' box;
\- entered the text that will be shown on the button. In this instance, it will show \'Info\'; and
\- selected the help file that will be used by clicking the \'Browse\' button and selecting the required file. In this instance, an HTML document was created that contains hyperlinks to the PCT Web sites, and information about using the presentation (see lower information box in Figure [2](#F2){ref-type="fig"}).
Regenerating the presentation
-----------------------------
We then resaved our settings by going to \'File\' \> \'Save Settings\'. In the resulting dialog, we browsed to the required directory and clicked \'Save\'. The presentation was regenerated by clicking \'File\' \> \'Generate SVG Page\'. Finally, we uploaded all the generated final page files to our Web server (<http://healthcybermap.org/PCT/STDs/> -- Figure [1](#F1){ref-type="fig"}).
Competing interests
===================
CR and MS work for GDC (Graphical Data Capture Ltd), the company that produces GeoReveal.
Authors\' contributions
=======================
All three authors contributed equally to this paper.
|
PubMed Central
|
2024-06-05T03:55:51.920580
|
2005-1-18
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545937/",
"journal": "Int J Health Geogr. 2005 Jan 18; 4:4",
"authors": [
{
"first": "Maged N Kamel",
"last": "Boulos"
},
{
"first": "Chris",
"last": "Russell"
},
{
"first": "Michael",
"last": "Smith"
}
]
}
|
PMC545938
|
Background
==========
The inclusion of patients\' opinions in the assessment of interventions has gained greater prominence over the last decades. Regulator agencies now call for the inclusion of patient-reported outcomes (PRO) in clinical trials evaluating pharmaceuticals interventions \[[@B1]-[@B4]\]. PRO of interest include health-related quality of life (HRQL), symptom assessment, and more recently, treatment satisfaction, in gastroesophageal reflux disease (GERD).
Whereas HRQL measures the patient\'s physical, psychological, and social level of function, treatment satisfaction assesses the patient\'s attitude towards the treatment, or the extent to which the patient is satisfied or not with the results of the treatment. Thus, treatment satisfaction focuses on the interaction of expectations and preferences for treatments and is defined as the individual\'s rating of important attributes of the process and outcomes of the treatment experience \[[@B5]\]. Coyne and co-workers \[[@B6]\] have summarized a number of patient important domains that describe satisfaction with treatment including symptom relief, flexibility with dosing, and treatment expectations. Treatment satisfaction is also associated with prescription regimens that involve less invasive dosing regimens \[[@B5],[@B7]-[@B10]\], such as daily versus twice daily use \[[@B11]\].
Evaluating treatment satisfaction may assist healthcare providers in understanding the issues that influence adherence with therapeutic interventions. In addition, treatment satisfaction can be a useful PRO when treatments show similar efficacy because differences in satisfaction could lead to patient preferences for one treatment over another and greater adherence with various treatment regimens.
Demographic variables such as age, ethnicity, and gender may influence satisfaction \[[@B12]\]. Older people tend to be more satisfied with medical care than younger people \[[@B13]-[@B15]\], and Caucasian people on the whole are more satisfied than non-Caucasians \[[@B16]\]. In contrast, gender does not appear to influence treatment satisfaction \[[@B17]\].
The objectives of this study were to assess correlates of treatment satisfaction, including demographic factors, symptoms, and HRQL, as well as change scores in PRO instruments in patients with moderate to severe GERD receiving a proton pump inhibitor, esomeprazole.
Methods
=======
Participants
------------
No statistical determination of sample size has been done since the study is of exploratory nature. We enrolled 249 patients with GERD in 13 gastroenterology practices and four general practices across Canada between March 2002 and March 2003.
Included patients were 18 years of age or older and had a diagnosis of moderate to severe GERD and presence of symptoms for three months or longer \[[@B18]\]. Prior to inclusion all patients gave written informed consent in accordance with the Helsinki declaration. Of 249 patients, 217 (87%) completed the study. We excluded twelve patients because upon review they did not meet the initial inclusion criteria. Of the 20 patients who withdrew after the baseline visit, 4 withdrew because of adverse events, 2 were unwilling to continue, 4 were lost to follow-up and 10 were excluded because of improper administration or completion of the questionnaires at one visit. Figure [1](#F1){ref-type="fig"} shows the flow of patients through the study. The final group of 217 completed patients received four weeks of therapy with esomeprazole 40 mg once daily, in the morning.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Flow chart
:::
:::
Procedure
---------
Patients completed PRO instruments at the clinic before and approximately 28 days after treatment. The completed PRO instruments included the Quality of Life in Reflux and Dyspepsia (QOLRAD) \[[@B19]\], the Feeling Thermometer (FT) \[[@B20]\], a four symptoms scale, the Standard Gamble (SG) \[[@B21]\], and an upper gastrointestinal (GI) symptom severity scale at baseline and follow-up. Patients completed the Health Utilities Index Mark 2 and 3 (HUI2 and HUI3) \[[@B22]\], and the Medical Outcomes Short-Form 36 (SF-36) \[[@B23]\] at baseline only; and the treatment satisfaction item at follow-up only. We describe these instruments below. In addition, trained research assistants collected information concerning demographic data and clinical data. Each visit lasted approximately 80 minutes.
Treatment satisfaction
----------------------
Patients rated their satisfaction with treatment on a seven point scale responding to the question: \'How satisfied are you with the study treatment you received?\' with the response options: completely satisfied, very satisfied, quite satisfied, no change, dissatisfied, very dissatisfied, and completely dissatisfied.
PRO instruments
---------------
### QOLRAD
The QOLRAD is a 25-item disease-specific self-administered instrument asking about the impact of heartburn and acid regurgitation on the patient\'s HRQL during the previous week. The QOLRAD includes questions related to 5 domains; emotional distress, sleep disturbance, problems with food and drink, limitations in physical and social functioning, and lack of vitality. Patients respond to each question on a seven-point scale on which a higher score indicates better HRQL. The psychometric properties concerning validity, reliability, and responsiveness to change are reported elsewhere \[[@B19],[@B24]\]. The minimal important difference (MID) that patients perceive as important is approximately 0.5 on the 1 -- 7 scale \[[@B25]\].
### FT
The FT is a visual analogue scale that resembles a thermometer. It is divided into 100 segments with a mark to represent each segment. Its anchors are dead (0) and full health (100) \[[@B21]\]. Patients mark their own health state and/or that of hypothetical patient scenarios or clinical marker states. In this study, three patient scenarios represented mild, moderate, and severe GERD. We developed and tested the clincal marker states with patients and clinicians \[[@B26]\]. The MID of the FT is approximately 6 on the 0 to 100 scale \[[@B27]\].
### HUI
This is a 15 item questionnaire designed to quantify HRQL \[[@B22]\]. Each item has 4--6 response options. There are 8 attributes in the HUI3 classification system: vision, hearing, speech, ambulation, dexterity, emotion, cognition, and pain. In the HUI2 there are 7 attributes: sensation, mobility, emotion, cognition, self-care, pain, and fertility.
### SF-36
The SF-36 contains 36 items that measure 8 dimensions: physical functioning, role limitations due to physical health problems, bodily pain, general health perceptions, vitality, social functioning, role limitations due to emotional problems, and general mental health. This questionnaire has been extensively tested for validation and reliability \[[@B23]\]. Each domain is scored on a 0 to 100 scale where higher scores indicate better HRQL. Scores on the SF-36 can also be expressed as two summary measures, the physical component score and the mental component score, which provide a measure of the overall effect of physical and mental impairment on HRQL.
### Rating of four symptoms
To assess common symptoms in GERD, patients evaluated their heartburn, acid reflux, stomach pain, and belching for the past week using a seven-point scale ranging from no discomfort to very severe discomfort.
### SG
The SG involves decision in the face of uncertainty, where in the standard administration the uncertainty involves a risk of death. The SG offers the patients two alternatives from which a choice must be made: Choice A is a hypothetical treatment with two possible outcomes: 1) returning to full health (probability p) for t years, at the end of which they die, or 2) immediate death (probability 1 -- p). The alternative (choice B) is a certain outcome that he or she will stay in a health state (their own health state, or a patient scenario) for t years until death. t varies depending on the patient\'s age. The interviewer used a change board with the ping-pong approach varying the probability p in steps of 0.05 to find the value p where the respondent considered choice A = choice B. This value of p is the utility value for the health state in choice A in the interval from dead (0) to full health (1). The greater a patient\'s willingness to accept the risk of a worse outcome (e.g. dead) to avoid the health state in choice A, then the lower is the utility of the state in choice A to them.
### Rating of upper GI symptom severity
Patients documented the severity of overall upper GI symptom on a seven-graded scale (1 = no symptoms; 7 = severe symptoms) over the past seven days. At baseline, patients who had no, minimal or mild symptoms were not included in this study.
Statistical analyses
--------------------
We calculated the mean and standard deviation of the basic demographic variables. Our multiple linear regression analysis focused on the outcome variable treatment satisfaction, which we treated as a continuous outcome variable. Evaluation of the data with polynomial regression yielded similar results. Potential correlates were demographic variables and baseline scores, as well as change scores for the PRO instruments described in the previous section. We first modelled these variables univariately as correlates of treatment satisfaction and only those that were significant at p \< 0.1 entered into the multiple regression model. After having entered the multiple regression model, only those significant at p \< 0.05 remained in the final model.
Results
=======
Table [1](#T1){ref-type="table"} shows the baseline demographic characteristics and frequencies of the included patients. The mean age was 50 years, and approximately 50% of the patients were female. The mean number of months since diagnosis was 86 months. Approximately 70% were full-time or part-time employed, and 88% were Caucasians.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Demographic characteristics and frequencies at baseline for the study sample (N = 217).
:::
Frequency Percentage
---------------------------------------------------- ------------- ------------
Gender
Male 103 47.5
Female 114 52.5
Age
Mean (SD) 49.7 (13.7)
Range 20--82
Months since diagnosis
Mean (SD) 86.3 (99.4)
Range 1--504
Smoking history
Never 94 43.5
Yes 38 17.6
Previous 84 38.9
Living alone 23 10.6
Employed: full-time and part-time 149 68.7
Ethnicity
Caucasian 191 88.0
Other 26 12.0
Severity of gastroesophageal reflux disease (GERD)
Moderate problem 112 51.6
Moderate severe problem 74 34.1
Severe problem 27 12.5
Very severe problem 4 1.8
:::
Table [2](#T2){ref-type="table"} depicts the mean baseline scores for the QOLRAD, the four symptoms scale, the FT, the SG, the HUI, and the SF-36. The mean QOLRAD scores at baseline were lowest for the food/drink domain, indicating worse HRQL for this domain, and the mean scores at baseline for the four symptoms show that patients had most problems with heartburn. Furthermore, the mean SF-36 scores at baseline were lowest (worse) for the bodily pain dimension, and highest (best) for the social functioning domain. Figure [2](#F2){ref-type="fig"} shows the distribution of the treatment satisfaction scores. Approximately 50% of the patients were completely satisfied, 25% were very satisfied, and approximately 15% were quite satisfied. About 7% reported no change or dissatisfaction of different severity.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Baseline scores for Quality of Life in Reflux and Dyspepsia (QOLRAD), four symptoms, Feeling Thermometer (FT), Standard Gamble (SG), Health Utilities Index Mark 2 and 3 (HUI), and Medical Outcomes Short Form-36 (SF-36).
:::
Mean SD
------------------------------ ------ ------
**QOLRAD dimensions**
Emotional distress 4.5 1.4
Sleep disturbance 4.5 1.5
Food/drink problem 3.8 1.2
Physical/social functioning 5.4 1.4
Vitality 4.3 1.3
**Four symptoms**
Stomach pain 3.9 1.5
Heartburn 4.5 1.2
Belching 3.6 1.6
Acid reflux 4.1 1.6
**FT** 0.7 0.2
**SG** 0.8 0.2
**HUI2** 0.8 0.2
**HUI3** 0.8 0.2
**SF-36**
Physical functioning 46.6 9.0
Role-physcial 45.5 11.4
Bodily pain 42.8 9.4
General health 46.2 9.7
Vitality 45.9 9.8
Social functioning 47.7 10.3
Role-emotional 46.5 12.0
Mental health 46.9 10.4
Physial component 45.1 8.7
Mental component 47.6 11.0
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Distribution of treatment satisfaction scores
:::
:::
Table [3](#T3){ref-type="table"} portrays the results from the multiple linear regression analysis. Ethnicity, baseline QOLRAD vitality, baseline heartburn from the four symptoms scale, and QOLRAD vitality change score remained as independent variable when all variables had entered the model. Caucasian patients were more likely to be satisfied with the treatment than patients of other ethnicity. Higher baseline QOLRAD vitality scores, higher levels of heartburn and larger change on the QOLRAD vitality score were associated with greater treatment satisfaction.
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Results from the multiple linear regression analysis with treatment satisfaction as outcome variable.
:::
Correlate variables Parameter estimate (β) SE P-value
--------------------------------- ------------------------ ------- ---------
Ethnicity (Caucasian vs. other) -0.570 0.190 0.003
QOLRAD Vitality baseline -0.628 0.068 \<0.001
Four symptoms Heartburn -0.195 0.055 \<0.001
QOLRAD Vitality change -0.593 0.071 \<0.001
Note. R^2^= 0.34 includes ethnicity, QOLRAD Vitality, heartburn, and QOLRAD Vitality change score
:::
Discussion
==========
The objective of this study was to assess correlates of treatment satisfaction in patients with moderate to severe GERD receiving esomeprazole. We found that Caucasian ethnicity, greater vitality and more severe heartburn at baseline, correlates with treatment satisfaction. Furthermore, the greater the improvement on vitality change score, the more likely the patient is to be satisfied with the treatment.
The strengths of this study include the detailed assessment of a number of demographic characteristics, HRQL and symptoms. However, this study has two important limitations. First, we did not perform a placebo controlled trial limiting our ability to assess satisfaction as a true treatment result versus other reasons for satisfaction. Second, investigators have not conducted a thorough psychometric assessment of the treatment satisfaction instrument we used in this study.
Nevertheless, the present study yields four important results. First, in this sample of GERD patients without prior endoscopic evaluation of their symptoms, Caucasian ethnicity was positively associated with treatment satisfaction. Ethnic origin is perhaps one of the most complex demographic characteristics \[[@B12]\] and it has previously been reported that Caucasian people on the whole are more satisfied than non-Caucasians \[[@B16]\].
Second, higher vitality scores, as assessed by the QOLRAD, were associated with higher treatment satisfaction. A patient\'s health status prior to receiving treatment may cause the patient to be either more or less satisfied with treatment. Clearly and McNeil \[[@B28]\] reported positive correlations between health status and satisfaction. However, it is unclear if satisfaction was correlated with health status before intervention or with health status after intervention. A possible interpretation of the positive association between QOLRAD vitality and treatment satisfaction in our study might be that patients with a high vitality score at baseline are less distressed by their disease, and therefore tend to be more satisfied. The association in our study between higher vitality scores, as assessed by the QOLRAD, and higher treatment satisfaction is in line with Revicki and co-workers \[[@B29]\] who found that patients reporting greater severity in heartburn symptoms were more likely to report psychological distress and impaired well-being compared with those who reported no or mild symptoms. However, Revicki et al measured HRQL with a generic instrument while we used a disease-specific instrument.
Third, higher scores for heartburn, assessed with the four symptoms scale, were related to higher treatment satisfaction. Thus, in our study population, patients with high discomfort from heartburn at baseline perceived a high satisfaction with treatment.
Fourth, the higher the improvement on the QOLRAD vitality (change score), the more likely the person is to be satisfied with the treatment.
Patients\' age is regarded as the most consistent determinant characteristic of satisfaction \[[@B13]-[@B15]\]. The results from this study did not reveal that treatment satisfaction was related to age. However, Fitzpatrick \[[@B30],[@B31]\] and Fox and Storms \[[@B32]\] highlight the lack of consistency of the effect of age in satisfaction studies. Since satisfaction studies focused on a variety of concepts, such as satisfaction with medical care, satisfaction with hospital management, satisfaction with health services, and satisfaction with treatment, it might be that the association between age and satisfaction is dependent on the concept assessed. The lack of an association to age reveals also the possible that our study population was too homogenous with regard to age.
Although some studies have reported that patient gender affects satisfaction values \[[@B33],[@B34]\], other studies did not find such association \[[@B17],[@B35]\]. In line with this, in our study population treatment satisfaction was not associated with gender.
The current results may be unique to the study sample since no placebo control group was included in the study and, therefore, we were unable to evaluate whether the factors related to treatment satisfaction are related to real treatment effects or patients\' need to please and placebo effects. The efficacy, tolerability and safety of esomeprazole versus other proton pump inhibitors has been shown in other studies \[[@B36]-[@B40]\]. In this study, patients had moderate to severe symptoms of GERD and some patients had received proton pump inhibitors prior to this study. The latter indicates that our study population is selected with regard to symptom severity, and mixed with regard to previous medication, which might limit generalizability of the findings. Treatment satisfaction in patients with mild GERD symptoms and with no previous experience of proton pump inhibitors remains unknown.
Investigators often use several PRO instruments, each with many dimensions and single items that are more or less correlated in clinical studies. This can lead to a large number of statistical tests being carried out and an increased risk of statistically significant findings occurring by chance in the absence of adjustment of P-values. In the present report we did not carry out adjustments for multiple comparisons for two main reasons. Firstly, the analysis of correlations was intended to be exploratory rather than confirmatory. Secondly, there is no consensus on how to adjust in analyses of the nature we conducted in this study. A simple adjustment according to Bonferroni would be too conservative, in part because many of the PRO variables are closely correlated.
Different drug therapies may elicit unwanted side-effects, which could compromise the patients\' HRQL, and adherence with the treatment. Thus, a challenge in the management of GERD is to achieve as high adherence as possible. In addition, treatment satisfaction can be of use when different drug therapies show similar efficacy since it can lead to a preference for one drug over another and greater adherence.
Our study also supports the need for validated treatment satisfaction instruments because the available instruments vary widely in clinical trials \[[@B41]\] and the majority of studies rely on single items. There is a need for developing and improving psychometric documentation of instruments measuring treatment satisfaction \[[@B42]\].
Conclusions
===========
We examined correlates of treatment satisfaction, including demographic factors, symptoms, and HRQL, as well as change scores in HRQL, in patients with moderate to severe GERD who were not investigated by endoscopy. We observed that Caucasian ethnicity was positively related to treatment satisfaction. Furthermore, higher vitality and more severe heartburn were associated with treatment satisfaction. Finally, the higher the improvement on the QOLRAD vitality (change score), the more likely the patient is to be satisfied with the treatment.
Authors\' contributions
=======================
Alessio Degl\' Innocenti was the project leader for this manuscript, edited the clinical protocol of the study, interpreted the data, and wrote the final manuscript as well as early versions. Gordon H Guyatt and Holger Schünemann were the principal investigators of the study, wrote the clinical protocol and grant application, are responsible for the study protocol, interpreted data and participated in writing the final as well as early versions of this manuscript. Ingela Wiklund contributed to the study protocol, interpreted the data and edited the manuscript. Diane Heels-Ansdell was responsible for the statistical analysis and edited the final manuscript as well as early versions. David Armstrong was co-principal investigator of the study, revised the clinical protocol, assessed patients, interpreted data and edited the final manuscript as well as early versions. Carlo A Fallone and Sander Veldhuyzen van Zanten revised the clinical protocol, assessed patients, interpreted data and edited the final manuscript as well as early versions. Samer El-Dika, Alan N Barkun, and Peggy Austin revised the clinical protocol, interpreted data and edited the final manuscript as well as early versions. Peggy Austin also co-ordinated the study. Lisa Tanser contributed to co-ordination of the study. All authors read and approved the final manuscript. The AstraZeneca global publications group approved the manuscript.
Acknowledgments
===============
Participating investigators and affiliation: Dr. Iain Murray, Quest Clinical Trials, Markham, Ontario; Dr. Daniel Sadowski, Hys Medical Centre, Edmonton, Alberta; Dr. Alan Barkun & Dr. Serge Mayrand, Montreal General Hospital, Montreal, Quebec; Dr. Ford Bursey, St. John General Hospital, St. John\'s, Newfoundland; Dr. Naoki Chiba, Surrey GI Research/Clinic, Guelph, Ontario; Dr. Lawrence Cohen, Sunnybrook & Women\'s College, Toronto, Ontario; Dr. Carlo Fallone, Royal Victoria Hospital, Montreal, Quebec; Dr. Francis Joanes, Port Arthur Clinic, Thunder Bay, Ontario; Dr. David Morgan, Hamilton Health Sciences Centre, Hamilton, Ontario; Dr. Marc Bradette, L\'Hotel-Dieu de Quebec, Quebec; Dr. David Armstrong, McMaster University Medical Centre, Hamilton, Ontario; Dr. Sander Veldhuyzen van Zanten, Queen Elizabeth II Health Sciences Centre, Halifax, Nova Scotia; Dr. Pierre Pare, Hospital St. Sacrement, Quebec, Quebec; Dr. W. Olsheski, Albany Medical Clinic, Toronto, Ontario; Dr. Ivor Teitelbaum, Yorkview Medical Centre, North York, Ontario; Dr. Subodh Kanani, Lakeshore West Medical Professional Centre, Toronto, Ontario; Dr. Paul Braude, Markham Research, Thornhill, Ontario.
|
PubMed Central
|
2024-06-05T03:55:51.924111
|
2005-1-13
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545938/",
"journal": "Health Qual Life Outcomes. 2005 Jan 13; 3:4",
"authors": [
{
"first": "Alessio",
"last": "Degl' Innocenti"
},
{
"first": "Gordon H",
"last": "Guyatt"
},
{
"first": "Ingela",
"last": "Wiklund"
},
{
"first": "Diane",
"last": "Heels-Ansdell"
},
{
"first": "David",
"last": "Armstrong"
},
{
"first": "Carlo A",
"last": "Fallone"
},
{
"first": "Lisa",
"last": "Tanser"
},
{
"first": "Sander Veldhuyzen",
"last": "van Zanten"
},
{
"first": "Samer",
"last": "El-Dika"
},
{
"first": "Naoki",
"last": "Chiba"
},
{
"first": "Alan N",
"last": "Barkun"
},
{
"first": "Peggy",
"last": "Austin"
},
{
"first": "Holger J",
"last": "Schünemann"
}
]
}
|
PMC545939
|
Background
==========
The Kosovo crisis resulted in the largest population displacement in Europe since World War II. During this time, over 300,000 Albanian Kosovars were expelled from Kosovo, \[[@B1]\] resulting in a Complex Humanitarian Emergency. (CHE). The United Nations High Commission for Refugees (UNHCR) managed a Humanitarian Evacuation Program (HEP) to move the Kosovars to new international destinations, including Canada.
Canada received approximately 5,500 refugees in 1999, with approximately 500 arriving to the city of Hamilton, in southern Ontario. Some Kosovars stayed temporarily in military bases and arrived after sponsorship housing had been arranged, while others arrived directly to temporary reception \"houses\" established at two Hamilton hotels. The reception in Hamilton included the organized provision of health services by physicians, nurses, dentists and optometrists, and settlement services from the Settlement and Integration Services Organization (SISO).
This HEP was unique for many countries, including Canada, and involved a complex coordinated effort on the part of international, national, and local organizations traditionally involved with immigrants and refugees, settlement, and health. This event was an important opportunity for health and settlement organizations operating at international, national, and local levels to critically examine and reflect on their efforts to organize and deliver health and settlement services to new immigrants and refugees.
The goal of this study was to explore the main challenges and successes of the Kosovar arrival, from international, national, and local perspectives and to develop recommendations to advise and guide planning of the future complex humanitarian emergencies, using interviews with key informants.
Methods
=======
A qualitative case study methodology was used with semi-structured interviews of key informants who were involved in the humanitarian evacuation at the international, national and local levels in Hamilton, Canada. This method was deemed most appropriate to provide a framework for understanding the challenges and successes associated with Complex Humanitarian Emergencies and generate recommendations for future efforts. The background of research team members was diverse. Medicine, nursing, anthropology, sociology and settlement were represented. Some members had been directly involved with the Kosovar settlement process at the local level (NF, SR, MJ).
Sampling
--------
A purposive sampling strategy was used, in which appropriate key informants known to the researchers were approached for participation first. Snowball sampling, whereby additional participants were identified by the initial respondents, was used to increase the diversity of respondents. These sampling approaches are used when certain known individuals are likely to have in-depth information on a topic \[[@B2]\]. Individuals from agencies involved with the 1999 Kosovar HEP were included. There was a very specific attempt to sample individuals from the international to the local levels even if they had no direct association with the Hamilton group. The sampling included three Health Canada officials who were known to be involved at the international and national levels and two Citizenship and Immigration Canada (CIC) officials involved at the provincial and local levels. Also sampled were two representatives from the Department of Social and Public Health Services in Hamilton, five local healthcare professionals (one family physician, two nurse practitioners, one dentist, one optometrist), and five local settlement providers. All of the CIC and the local key informants were referred and identified by the local settlement agency SISO as groups that had been integral to the Kosovo efforts in Hamilton. The international and national key informants were identified through Health Canada and internationally as people who were involved in the effort. An initial group of nine key informants identified eight additional participants. All 17 key informants approached, agreed to participate.
Interviews
----------
Interviews were conducted in English between April and July 2001. The interviewer (EM) took extensive field notes to supplement the taped and transcribed interviews. Of the 17 key informant interviews, the three international and national ones were conducted by telephone, and 14 were in person at the workplace of the interviewee. Six participants were male and 11 were female. Interviews were approximately 30 minutes in length. The interviews were based on an interview guide and the main part of the interview focused on the organization and delivery of health services to the Kosovars. Questions were further refined during the study. Specific probes were used to follow up on open-ended questions, where appropriate.
Consultation with the local settlement agency (SISO) took place throughout the research process. Key informants were told that the purpose of the study was to gather information that would help in planning for similar events in the future. All participants provided verbal consent and were assured of confidentiality of responses. All interviews were tape recorded with the participants\' permission and all but one (due to technical difficulty) was transcribed. Ethics approval for the study was received from McMaster University Research Ethics Board.
Questions pertained to the participants\' involvement in the process, challenges, flow of information and communication, and involvement in health care services and settlement.
Data Analysis
-------------
Two research team members (EM and NF) initially independently reviewed the transcripts, coded categories and themes, and then compared results. A third researcher with expertise in qualitative research, who was not involved in the project, also reviewed the transcripts using NVivo software version 2.0 (QSR International Pty Ltd, Melbourne, Australia) and compared results. The interviewer\'s field notes were also used for comparison with the data. Ambiguities were resolved and themes were developed from categories through discussion among the research group members and re-reading of transcripts. This process continued until no new themes emerged from the interviews. It was felt that saturation was achieved after 17 interviews, and data collection was ended. The process was iterative, thus the later interviews probed new emerging themes identified in previous interviews.
Results
=======
Six core themes emerged from the data analysis.
Theme 1: A Sense of Being Overwhelmed
-------------------------------------
There was agreement among the participants that the Kosovo crisis in 1999 was very unusual in nature. Participants stressed that there was a lot of \"scrambling\" to prepare the infrastructure to receive a large number of people in a very short period of time. Respondents/informants were surprised by the magnitude, immediacy, scope and scale of the response required. Most described feeling overwhelmed.
Although the local settlement agency (SISO) had begun preparations by meeting with various local agencies and organizations weeks in advance, they were given only three days notice by national immigration authorities that the Kosovars were arriving locally. One Hamilton Social and Public Health Services representative noted \"We had many community people come together to talk about how we could plan for this big influx of refugees\... we knew were in dire straits\". One settlement worker called it \"organized chaos\". One local health care professional explained the triage process and it was evident that a major organizational effort was required to process the arrivals.
\"We had to do our own triage at the hotel to find out whether there was anybody needing medications, anyone with heart conditions, anyone with diabetes \... we had to scan the place to find out, and even then we didn\'t have medical records that came with them\... We just went around asking: Do you know about anybody who is pregnant? Do you know about anybody who has heart conditions? Do you know about anybody who is diabetic? \... And we tried to identify those people or they presented in the little clinic\".
Theme 2: A Multitude of Health Issues
-------------------------------------
Participants described facing a multitude of challenges in responding to the health needs of Kosovar refugees. Women\'s health services, including pregnancy care and abortion services were required. Dental care, mental health services and general curative services were also required. Participants were struck by the poor oral health of the Kosovars. Many Kosovars had been deprived of any curative or preventative care in Kosovo for a number of years prior to the 1999 crisis. One local health care professional observed,
\"I noticed with the Kosovars\...that a lot of them had not had any health care for a long time so they had many of the things that we take for granted. Immunizations being up to date and those sorts of things had not been done\.... They only saw doctors if they were absolutely on their deathbed. So sometimes, you sort of had a lot of catching up to do in terms of getting their health up to date\"
Chronic and poorly managed conditions such as diabetes, hypertension and renal failure were common. An interviewee with an international perspective noted,
\"I think the world, from a health point of view, went into Kosovo thinking that all refugees were like the Great Lakes and Rwanda\...The challenges \[in Kosovo\] were more of chronic diseases, diseases of socialization, hypertension, diabetes, renal failure\...I think that people have learned to be a bit more comprehensive in their approach to conflicts and emergencies.\"
Theme 3: Critical Challenges in Providing Health Care -- Lack of information at the local level
-----------------------------------------------------------------------------------------------
The participants expressed three main challenges which made their work more difficult: (i) tuberculosis screening, (ii) the lack of medical records and tuberculosis test results and (iii) mental health issues.
Several participants noted that the tuberculosis screening process was not optimally organized. One local healthcare professional lamented,
\"I was reassured, but without any documentation to back it, that everyone had been screened for TB\...When I contacted Public Health, they had not received any notification \...Certainly I didn\'t know where or how to get that information, and it seemed to me that Public Health didn\'t either\".
Communication to local health care providers about test results (tuberculosis and radiographs) was often described as sub-optimal. Immigration officials in Europe attempted to keep medical information flowing. One interviewee from an international perspective explained that,
\"We provided by fax and e-mail a summary of the medical conditions on the aircraft, so that they can be dealt with appropriately on arrival in Canada\...We also used colour-coded cards and things so that (when) people who got off the plane (who needed care) they did receive expedient care\".
However, many local respondents expressed concern over the location of medical records and the inability to access this information in a timely manner. A local healthcare professional expressed,
\"At no point did I receive any medical documents about people with serious or chronic medical conditions that required care\...if that information was available, I never got it\".
It was clear among the participants that mental health concerns were prevalent and service providers struggled with the delivery of appropriate mental health services. One local healthcare professional explained,
\"\...things like headaches presenting when really the underlying condition was one of stress and anxiety, distress. Most of them came under the umbrella of what we might call, presenting with trivial complaints, but really what was beneath it all was stress and anxiety..\".
Theme 4: Access to Health and Settlement Services
-------------------------------------------------
The rapid settlement of the Kosovars resulted in the local collaboration and coordination of health and settlement services at a single geographic site. This multi-agency collaboration was thought to enhance access and provision of services. One local healthcare professional described,
\"\[The Kosovars\] had, I thought, very well organized access\...The SISO organization provided chauffeurs and translators, and administrators to pull all those areas together. So\...if a refugee needed to go to a lab, they were driven, they were translated for, and they were brought back\...\".
Since translators organized through SISO were available at the local health care sites, language was not perceived as a significant challenge by local health care providers.
There was an awareness that the care provided was transitional in nature. This created some hesitancy in initiating treatment, as frequently there was uncertainty about the medical follow-up arrangements. Mental health issues and the lack of opportunity for identification and communication with future care providers were identified as concerns.
While the Kosovars received medical coverage through the Interim Federal Health (IFH) program, local health care providers felt that the amount and nature of coverage was inadequate for services such as home care, dental care and optometry. One local healthcare professional described the issues surrounding home care,
\"One access issue that came up was the fact that under their IFH coverage, IFH does not cover home care, and without special arrangements to be in place\...they don\'t qualify for it\...There were a lot of people\...who had walked miles and miles\...and I saw \[some\] who had really bad foot conditions, infections and ulcers requiring daily treatment, soakings, dressings, bandage changes\... Normally, those are things in the community that we would involve home care in\...\".
In several instances, both the optometrist and dentist interviewed provided services that were not covered, free of charge, but in general, there was much confusion and uncertainty regarding the payment scheme for professionals. A dentist explained,
\"..we were not informed as a health professional that if we were approached by people from Kosovo that plans were available for treatment\...we knew absolutely nothing until we started asking the questions..\".
Theme 5: Overall successes
--------------------------
Specific examples of successes included provision of comprehensive onsite health care integrating both health and settlement services, the use of nurse practitioners to allow physicians to focus on more complex cases, the policy to keep families intact, and the positive media coverage that contributed to an atmosphere of acceptance. A local settlement provider observed,
\"The Kosovars was something that the whole nation took on\...We saw what kind of role the media can play in making the host community aware what other people are going through, explaining that refugees are not to be seen as invaders but as people who are in need of welcoming, and the Kosovars received one of the best, I think at least in my experience, one of the warmest welcomes\...\".
Services available to the Kosovar refugees were deemed better compared to that provided to other refugee groups. This was a concern to many participants. Some stated that all refugees should be offered the same standard of high quality settlement services as those made available to the Kosovars. Another local settlement provider noted,
\"Refugees, regardless from where, should be treated the same way because it creates resentment not just in other refugee communities that have come to Canada, but \[it\] creates resentment among the workers that are providing services and sometimes struggling to get resources for a group of people\...The Kosovars should be an example of how refugees should be treated in general\".
Theme 6: Need for a Coordinated Approach in Migration Health
------------------------------------------------------------
A number of participants suggested that we should learn from this experience with the Kosovars and prepare to put a contingency plan in place for similar events in the future.
Better communication and organization were repeatedly stressed. One participant used the term \'Migration Health\' to describe such an overall coordinated approach. Several participants commented on the importance of the coordinated approach as a necessary societal investment. One interviewee from an international perspective noted,
\...a lot of people who are working on the receiving end are simply following the process for receiving refugees\.... there may not be the resources to look at the longer term issues: primary health care, health education, explaining how the health systems work, looking at some of the parameters that may influence longer term mortality and morbidity: dietary counselling, smoking cessation counselling, primary preventive health care procedures that we do in Canada that may not take place in the developing world.\"
Most participants suggested that health care and settlement providers need to enhance their cultural sensitivity and cultural competence and better understand the health conditions of the displaced individuals in their country of origin.
Discussion
==========
Despite organized attempts to coordinate efforts at different stages of the migration process, communication gaps and the sheer size of the influx resulted in challenges at the local level. Officials and service organizers at the international/national levels were unaware of these local gaps at the time of the evacuation. The flow of medical information and health records is an example. Primary care health workers needed to have easy access to targeted health information about the Kosovars, however this information was not available. If this information had been more easily available, health care services may have been more streamlined, and unnecessary duplication of lab tests and radiographs could have been avoided.
Health care providers did not know what to \'expect\'. Mental health, dental care and communicable diseases (specifically tuberculosis) were identified as requiring further specialized planning. This finding is similar to other refugee experiences in Australia \[[@B3]\] and Canada \[[@B4]\]. Our study was also consistent with other refugee literature suggesting that there is often gap in addressing refugee women\'s health services \[[@B5]-[@B7]\].
There was a consensus among participants that this international evacuation represented an improved approach and a good foundation on which to organize refugee health and settlement in the future. Informants\' concerns about local preparedness and the need for future advanced planning was consistent with recent United Nations High Commission for Refugees (UNHCR) findings \[[@B8]\]. Participants wanted to see a contingency plan developed for the future with enhanced communication and better organization. These wishes have been incorporated into the detailed recommendations.
A potential limitation of this study is the time elapsed between the refugees\' arrival and the interviews. Asking people to recall events that took place approximately 18 months previously may have influenced the nature and detail of the information obtained. However, it may also have enabled informants to recover from the initial emotional reaction and to see these events from a different perspective.
It must be stated that the \'success\' as defined by key informants may be very different than \'success\' from the refugees\' perspective. This study examined the perceptions of those involved with the Kosovars. The majority (but not all) of the local health professionals and settlement workers interviewed had worked in the local \'settlement houses\' site where there was a shorter notice of arrival of the refugees. This may have contributed to the sense of \'chaos\' echoed in many of the comments.
Many of the participants in this study made observations about the diverse nature of complex emergencies. The importance of logistics and planning has also been described in other studies \[[@B8]-[@B10]\]. Some countries and jurisdictions are starting to develop \'rapid response\' protocols for similar situations. After the Kosovar influx, Australia developed a surveillance, triage, clinical and database system (*\'Operation Safe Haven\'*). Triage questionnaires for primary health care based on International guidelines were developed \[[@B11]\]. *Operation Safe Haven*produced a template for refugee \"acute health response\" system. Using the Australian template combined with findings from this study, we propose the following attributes for an *International Rapid Response System*(IRRS):
1\. Establish lead organizations at the different levels (international, national, provincial, local)
2\. Clearly define roles for different organizations involved
3\. Establish communication linkages between lead organizations at different levels
4\. Identify strategies for flow of information (\"situation reports\") from authorities to local organizations
5\. Establish a protocol for triage/rapid assessment of health, settlement needs and cultural preferences
6\. Establish a system of medical charts that follow individual refugees though the process
7\. Establish links to primary and specialized health care especially for urgent communicable disease and mental health issues
8\. Identify a plan for the provision of urgent dental and eye care
9\. Establish a surveillance/data-tracking system to collect essential health information, tack service use, and provide the ability to conduct quality assurance assessments
10\. Use information technology -- key internet links; background data; briefing; high quality background and cultural information about refugee groups
11\. Introduce better training of professionals who will be dealing with refugees. Includes increasing cultural competence of health care and settlement providers, women\'s health, mental health, chronic illness, dental care and current health coverage for refugees. Global health issues should be introduced into health school curricula.
12\. Local health care professionals need access to better information regarding the background, circumstances and organizational arrangements relating to refugees. Regular, ongoing information sharing sessions with health professionals involved with refugees, public health and government would facilitate communication when the next refugee crisis occurs. There must be political commitment at all levels.
Conclusions
===========
For those involved in the Kosovar Humanitarian Evacuation Program, the experience was both overwhelming and rewarding. Many perceived that a superior effort was made for the Kosovars compared to other groups of refugees and that positive media coverage contributed to a warm and effectively organized reception. Interviewees\' major concerns were the need for a more comprehensive and coordinated approach to the flow of information and handling of specific health problems.
List of Abbreviations Used
==========================
Complex Humanitarian Emergency: CHE
United Nations High Commission for Refugees: UNHCR
Humanitarian Evacuation Program: HEP
Settlement and Integration Services Organization: SISO
Tuberculosis: TB
Interim Federal Health: IFH
International Rapid Response System: IRRS
Competing Interests
===================
The author(s) declare that they have no competing interests.
Authors\' Contributions
=======================
NF and LRC designed and implemented the study, analysed the data and critically revised the manuscript. SR, MJ, MH, JK, SR critically revised the manuscript, EM conducted the interviews, assisted with study design and implementation, data analysis, and drafted the manuscript.
Acknowledgements
================
This work was supported by the Hamilton Community Foundation and Citizenship and Immigration Canada.
|
PubMed Central
|
2024-06-05T03:55:51.926665
|
2005-1-12
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545939/",
"journal": "Int J Equity Health. 2005 Jan 12; 4:1",
"authors": [
{
"first": "Nancy",
"last": "Fowler"
},
{
"first": "Lynda",
"last": "Redwood-Campbell"
},
{
"first": "Elizabeth",
"last": "Molinaro"
},
{
"first": "Michelle",
"last": "Howard"
},
{
"first": "Janusz",
"last": "Kaczorowski"
},
{
"first": "Morteza",
"last": "Jafarpour"
},
{
"first": "Susan",
"last": "Robinson"
}
]
}
|
PMC545940
|
Commentary
==========
Cancer is being increasingly recognized as a very heterogeneous disease, both within an individual tumor and within a tumor type and among tumor types. This heterogeneity is manifested both at the genetic and phenotypic level and determines the progression of disease and response to therapy. It is possible to see the heterogeneity in examples of differential disease progression and response to therapy of the same tumor type, as morphology does not always reveal underlying biology. The diagnosis of tumors by histopathological and morphological criteria cannot fully account for the variability seen in prognosis and therapy outcome. A classic example of this heterogeneity is diffuse large B cell lymphoma (DLBCL) morphologically defined as one tumor type but only 40% of patients respond to treatment suggesting there are at least two distinct tumor groups. Alizadeh et. al. performed microarrays on DLCBL to assess gene expression profiles and identified several genetically distinct groups that correlated with differential survival rates \[\[[@B1]\], reviewed in \[[@B2]\]\]. Genome wide screening technology such as microarray offer the potential to diagnose, prognose and develop new therapeutic strategies for cancers based on grouping by genetic signature.
Whereas previously, it was possible only to study one or a few genes at a time, microarray technology allows the simultaneous assessment of the expression of thousands of genes within a cell population at a single time. By looking at the full spectrum of the genetic contribution within a tumor, microarray technology has furthered our understanding of the complexity in terms of tumor subclassification. The advantage of a global gene expression analysis is that it assesses many genes within a sample at a given timepoint and allows comparison to a myriad of other samples. This has resulted in the improved classification of tumors, identification of potential new biomarkers, and detection of possible therapeutic targets as in DLBCL. In addition, the same gene expression data can be reanalyzed according to a user defined phenotype or without bias while looking for patterns in the data that correlate with a phenotype such as progression, prognosis, or treatment outcome. As data analysis and data mining become more sophisticated, the information acquired will provide scientists and clinicians with a significant improvement in correlating patient data with tumor diagnosis and enabling us to better select patient groups who will respond (or not respond) to therapy. Microarray technology has become the best hope in developing a global and accurate assessment of the tumor type and all its complexity. However, the road to achieve this goal will be long and hard because we have to learn to ask the right questions, select the appropriate patients, collect their material and then verify the initial results.
Clinical pathological analysis cannot predict clinical outcome or metastatic potential of melanoma, a very heterogeneous cancer with an unpredictable progression rate. In a previous issue, Wang et. al. performed gene expression analysis on RNA from a number of human solid tumor lesions, including melanoma \[[@B3]\]. In their comparison they showed that it is possible to identify tumor specific gene expression profiles which can rapidly aid in tumor identification and classification. In addition, the study identified commonly expressed genes between melanoma and Renal Cell carcinoma, both known to be responsive clinically to IL-2 treatment, allowing for comparison of immunologically related genes to identify common response pathways. Gene expression profiles of melanoma lesions can also be used for prognosis by stratifying patients based on risk and thus identifying subtypes. For example, an early study by Clark et. al. assessed the gene expression differences of metastatic versus non-metastatic melanoma cell lines, identifying a metastatic profile which was linked to the small GTPase RhoC \[[@B4]\]. Bittner et. al. performed a more comprehensive array analysis of 31 cutaneous melanomas and identified a major cluster of melanoma samples \[[@B5]\]. Further, the authors were able to verify the validity of the cluster by correlating the melanomas within the cluster to reduced motility, invasive ability, and vasculogenic mimicry potential *in vitro*. This showed that lesions can be stratified into subtypes by gene expression analysis. Further, microarray gene expression data can be used to define responders and non-responders to known anti-cancer treatments prospectively or retrospectively. By combining clinical data with microarray data, it will be possible to predict patient response based on gene expression profile or biomarkers, which may allow for better, more targeted therapies to be selected. This new information will lead to improved treatments and prolonged survival for cancer patients.
DNA microarray technology may help us understand the complex pathogenesis of melanoma and will allow us to determine the role of the different genetic profiles in determining different disease outcomes. From this we will be able to identify new biomarkers, leading to the development of more pathologically relevant models. To achieve better prediction for optimal treatment strategies, microarray studies as presented here are only the beginning of a long road, in which we need to drastically par down the markers to be tested. We need to verify and validate biomarker candidates in ways that go beyond the capacity of individual laboratories. Instead we need to establish consortia of scientists from bioinformatics and computational biology, who team up with oncologists, pathologists, and immunobiologists. Any selected biomarker requires validation in independent multi-center analyses. Once the appropriate tools and infrastructures are on hand, we can select better new treatment modalities and may realize that previously unsuccessful regimens would have shown more success, if we would have know how to select most appropriate patients. We have to start now to develop the groundwork for such multidisciplinary, multi-institutional work that will challenge us in the years to come.
|
PubMed Central
|
2024-06-05T03:55:51.928885
|
2005-1-13
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545940/",
"journal": "J Transl Med. 2005 Jan 13; 3:2",
"authors": [
{
"first": "Patricia",
"last": "Brafford"
},
{
"first": "Meenhard",
"last": "Herlyn"
}
]
}
|
PMC545941
|
Introduction
============
Alzheimer\'s disease (AD) is the most common form of dementia in the elderly. Its main pathological features include extracellular amyloid beta (Aβ) deposition in plaques, neurofibrillary tangles (composed of hyperphosphorylated tau protein) in neurons, progressive loss of synapses and cortical/hippocampal neurons, and upregulation of inflammatory components including activated microglia and astrocytes and complement activation \[[@B1]\]. Although the contribution of abnormal phosphorylation and assembly of tau to AD dementia remains a focus of investigation, therapies that interfere with Aβ production, enhance its degradation, or cause its clearance from the central nervous system (CNS) have been the center of many studies in search of a cure for this disease.
Microglial cells, when activated, are believed to be responsible for much of the Aβ clearance through receptor-mediated phagocytosis \[[@B2],[@B3]\]. Upon activation, microglia acquire features more characteristic of macrophages, including high phagocytic activity, increased expression of leukocyte common antigen (CD45), major histocompatibility complex (MHC) class II and costimulatory molecules B7, and secretion of proinflammatory substances \[[@B4]\]. In addition, phagocytic microglia also participate in the removal of degenerating neurons and synapses as well as Aβ deposits (\[[@B5]\], and reviewed in \[[@B6]\]). Thus, while some microglial functions are beneficial, the destructive effects of the production of toxins (such as nitric oxide, superoxide) and proinflammatory cytokines by activated microglia apparently overcome the protective functions in the chronic stage of neuroinflammation \[[@B7],[@B8]\]. *In vitro*studies have shown both protection and toxicity contributed by microglia in response to Aβ depending on the state of activation of microglia \[[@B9],[@B10]\]. Correlative studies on AD patients and animal models of AD strongly suggest that accumulation of reactive microglia at sites of Aβ deposition contributes significantly to neuronal degeneration \[[@B3],[@B11]\], although decreased microglia have been reported to be associated with both lowered and enhanced neurodegeneration in transgenic animals \[[@B12],[@B13]\]. Aβ itself is believed to initiate the accumulation and activation of microglia. However, recent reports provide evidence for neuron-microglial interactions in regulating CNS inflammation \[[@B14]\]. Nevertheless, the molecular mechanisms responsible for activation and regulation of microglia remain to be defined.
Complement proteins have been shown to be associated with Aβ plaques in AD brains, specifically those plaques containing the fibrillar form of the Aβ peptide \[[@B11]\]. Complement proteins are elevated in neurodegenerative diseases like AD, Parkinson\'s disease, and Huntington\'s disease as well as more restricted degenerative diseases such macular degeneration and prion disease \[[@B11],[@B15]-[@B18]\]. Microglia, astrocytes, and neurons in the CNS can produce most of the complement proteins upon stimulation. C1q, a subcomponent of C1, can directly bind to fibrillar Aβ and activate complement pathways \[[@B19]\], contributing to CNS inflammation \[[@B13]\]. In addition, C1q has been reported to be synthesized by neurons in several neurodegenerative diseases and animal injury models, generally as an early response to injury \[[@B20]-[@B23]\], possibly prior to the synthesis of other complement components.
Interestingly, C1q and, upon complement activation, C3 also can bind to apoptotic cells and blebs and promote ingestion of those dying cells \[[@B24]-[@B26]\]. Elevated levels of apoptotic markers are present in AD brain tissue suggesting that many neurons undergo apoptosis in AD \[[@B27]-[@B29]\]. Excess glutamate, an excitatory neurotransmitter released from injured neurons and synapses, is one of the major factors that perturb calcium homeostasis and induce apoptosis in neurons \[[@B30]\]. Thus, it is reasonable to hypothesize that neuronal expression of C1q, as an early injury response, may serve a potentially beneficial role of facilitating the removal of apoptotic neurons or neuronal blebs \[[@B31]\] in diseases thereby preventing excess glutamate release, excitotoxicity, and the subsequent additional apoptosis.
We have previously reported that in rat hippocampal slice cultures treated with exogenous Aβ42, C1q expression was detected in pyramidal neurons following the internalization of Aβ peptide. This upregulation of neuronal C1q could be a response to injury from Aβ that would facilitate removal of dying cells. Concurrently, microglial activation was prominent upon Aβ treatment. In the present study, the relationship of Aβ-induced neuronal C1q production to microglia activation and Aβ uptake in slice cultures was investigated.
Materials and methods
=====================
Materials
---------
Aβ 1--42, obtained from Dr. C. Glabe (UC, Irvine), was synthesized as previously described \[[@B32]\]. Aβ 10--20 was purchased from California Peptide Research (Napa, CA). Lyophilized (in 10 mM HCl) Aβ peptides were solubilized in H~2~O and subsequently N-2-hydroxyethylpiperazine-N\'-2-ethanesulfonic acid (HEPES) was added to make a final concentration of 10 mM HEPES, 500 μM peptide. This solution was immediately diluted in serum-free medium and added to slices. Glycine-arginine-glycine-aspartic acid-serine-proline (RGD) peptide was purchased from Calbiochem (San Diego, CA). D-(-)-2-amino-5-phosphonovaleric acid (APV) was purchased from Sigma (St. Louis, MO). Both compounds were dissolved in sterile Hanks\' balanced salt solution (HBSS) without glucose at 0.2 M and 5 mM, respectively, before diluted in serum-free medium. Antibodies used in experiments are listed in Table [1](#T1){ref-type="table"}; RT-PCR primers, synthesized by Integrated DNA Technologies (Coralville, IA), are listed in Table [2](#T2){ref-type="table"}. All other reagents were from Sigma unless otherwise noted.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Antibodies used in immunohistochemistry.
:::
**antibody/antigen** **concentration** **source**
---------------------- ------------------- --------------------------------------
anti-rat C1q 2 μg/ml M. Wing, Cambridge, UK
OX-42 (CD11b/c) 5 μg/ml BD/PharMingen, San Diego, CA
ED-1 3 μg/ml Chemicon, Temecula, CA
anti-CD45 0.5 μg/ml Serotec Inc, Raleigh, NC
4G8 (Aβ) 1 μg/ml Signet Pathology Systems, Dedham, MA
6E10 (Aβ) 0.5 μg/ml Signet Pathology Systems
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
PCR primers and cycling conditions for RT-PCR assay.
:::
-------------------------------------------------------------------------------------------------------------------
Gene Primer sequences Denaturation Annealing Extension cycle Ref
--------- ----------------------------------- -------------- ------------------ ------------------ ------- --------
C1qB 5\'-cgactatgcccaaaacacct-3\'\ 94°C 1 min 60°C 1 min30 sec 72°C 2 min 35 \[61\]
5\'-ggaaaagcagaaagccagtg-3\'
MCSF 5\'-ccgttgacagaggtgaacc-3\'\ 92°C 30 sec 58°C 1 min 72°C 1 min30 sec 35 \[62\]
5\'-tccacttgtagaacaggaggc-3\'
CD40 5\'-cgctatggggctgcttgttgacag-3\'\ 94°C 30 sec 58°C30 sec 72°C 1 min 30 \[63\]
5\'-gacggtatcagtggtctcagtggc-3\'
β-actin 5\'-ggaaatcgtgcgtgacatta-3\'\ 94°C 30 sec 60°C30 sec 72°C 1 min 25 \[61\]
5\'-gatagagccaccaatccaca-3\'
IL-8 5\'-gactgttgtggcccgtgag-3\'\ 94°C 1 min 56°C 1 min 72°C 1 min 39 \[64\]
5\'-ccgtcaagctctggatgttct-3\'
-------------------------------------------------------------------------------------------------------------------
:::
Slice cultures
--------------
Hippocampal slice cultures were prepared according to the method of Stoppini et al \[[@B33]\] and as described in Fan and Tenner \[[@B34]\]. All experimental procedures were carried out under protocols approved by the University of California Irvine Institutional Animal Care and Use Committee. Slices prepared from hippocampi dissected from 10d-old Sprague Dawley rat pups (Charles River Laboratories, Inc., Wilmington, MA) were kept in culture for 10 to 11 days before treatment started. All reagents were added to serum-free medium (with 100 mg/L transferrin and 500 mg/L heat-treated bovine serum albumin) which was equilibrated at 37°C, 5% CO~2~before addition to the slices. Aβ 1--42 or Aβ 10--20 was added to slice cultures as described previously \[[@B34]\]. Briefly, peptide was added to cultures in serum-free medium at 10 or 30 μM. After 7 hours, the peptide was diluted with the addition of an equal amount of medium containing 20% heat-inactivated horse serum. Fresh peptide was applied for each day of treatment. Controls were treated the same way except without peptide. RGD or APV was added to the slice cultures at the same time as Aβ 42.
Immunohistochemistry
--------------------
At the end of the treatment period, media was removed, the slices were washed with serum-free media and subjected to trypsinization as previously described \[[@B34]\] for 15 minutes at 4°C to remove cell surface associated, but not internalized, Aβ. After washing, slices were fixed and cut into 20 μm sections for immunohistochemistry or extracted for protein or RNA analysis as described in Fan and Tenner \[[@B34]\]. Primary antibodies (anti-Aβ antibody 4G8 or 6E10; rabbit anti rat C1q antibody; CD45 (leukocyte common antigen, microglia), OX42 (CD11b/c, microglia), or ED1 (rat microglia/macrophage marker), or their corresponding control IgGs were applied at concentrations listed in Table [1](#T1){ref-type="table"}, followed by biotinylated secondary antibody (Vector Labs, Burlingame, CA) and finally FITC- or Cy3-conjugated streptavidin (Jackson ImmunoResearch Laboratories, West Grove, PA). Slides were examined on an Axiovert 200 inverted microscope (Carl Zeiss Light Microscopy, Göttingen, Germany) with AxioCam (Zeiss) digital camera controlled by AxioVision program (Zeiss). Images (of the entire CA1-CA2 region of hippocampus) were analyzed with KS 300 analysis program (Zeiss) to obtain the percentage area occupied by positive immunostaining in a given field.
ELISA
-----
Slices were homogenized in ice-cold extraction buffer (10 mM triethanolamine, pH 7.4, 1 mM CaCl~2~, 1 mM MgCl~2~, 0.15 M NaCl, 0.3% NP-40) containing protease inhibitors pepstatin (2 μg/ml), leupeptin (10 μg/ml), aprotinin (10 μg/ml), and PMSF (1 mM). Protein concentration was determined by BCA assay (Pierce, Rockford, IL) using BSA provided for the standard curve.
An ELISA for rat C1q was adapted from Tenner and Volkin \[[@B35]\] with some modifications as previously described \[[@B34]\].
RNA preparation and RT-PCR
--------------------------
Total RNA from cultures was isolated using the Trizol reagent (Life Technologies, Grand Island, NY) according to the manufacturer\'s instructions. RNA was treated with RNase-free DNase (Fisher, Pittsburgh, PA) to remove genomic DNA contamination. Each RNA sample was extracted from 3 to 5 hippocampal organotypic slices in the same culture insert. The reverse transcription (RT) reaction conditions were 42°C for 50 min, 70°C for 15 min. Tubes were then centrifuged briefly and held at 4°C. Primer sequences and PCR conditions are listed in Table [2](#T2){ref-type="table"}. PCR products were electrophoresed in 2% agarose gel in TAE buffer and visualized with ethidium bromide luminescence. To test for differences in total RNA concentration among samples, mRNA level for rat β-actin were also determined by RT-PCR. Results were quantified using NIH image software \[[@B36]\] by measuring DNA band intensity from digital images taken on GelDoc (BIO-RAD) with Quantity One program.
Results
=======
NMDA receptor antagonist APV inhibits Aβ42 uptake and Aβ42-induced microglial activation and neuronal C1q production
--------------------------------------------------------------------------------------------------------------------
We have previously reported that C1q was detected in cells positive for neuronal markers and that microglial cells were activated in slices following Aβ42 ingestion \[[@B34]\]. Lynch and colleagues have shown that APV, a specific NMDA glutamate receptor antagonist, was able to block Aβ42 uptake by hippocampal neurons in slice cultures \[[@B37]\]. This provided a mechanism to down-modulate the Aβ42 internalization and test the effect on induction of C1q synthesis in neurons. Slices were treated with no peptide, 50 μM APV, 30 μM Aβ42, or 30 μM Aβ42 + 50 μM APV for 3 days with fresh reagents added daily. Cultures were collected and processed as described in Materials and Methods. Similar to reported previously, addition of exogenous Aβ42 resulted in Aβ uptake by hippocampal neurons, induction of C1q synthesis in neurons, and activation of microglial cells (Figure [1d, e, f](#F1){ref-type="fig"} compared with [1a, b, c](#F1){ref-type="fig"}). As anticipated, Aβ42 uptake in neurons detected by both 4G8 (Figure [1g](#F1){ref-type="fig"}) and 6E10 (data not shown) was inhibited by APV co-treatment. Neuronal C1q immunoreactivity was also inhibited when APV was added to Aβ42 treated slices (Figure [1h](#F1){ref-type="fig"}). Aβ42-triggered microglial activation, assessed by upregulation of antigens detected by anti-CD45 (Figure [1i](#F1){ref-type="fig"} vs. [1f](#F1){ref-type="fig"}), OX42 and ED1 (data not shown) was also fully diminished by APV. To quantify the immunohistochemistry results, images were taken from the entire CA1-CA2 region of each immunostained hippocampal section and averaged. Image analysis further substantiated the reduction in Aβ uptake, C1q synthesis and microglial activation (Figure [1j](#F1){ref-type="fig"}). C1q gene expression at mRNA and protein levels was also assessed by RT-PCR and ELISA, respectively. Results showed decrease of C1q mRNA and protein in slice extracts treated with 30 μM Aβ42 + APV, compared to 30 μM Aβ42 alone (Figure [2a](#F2){ref-type="fig"} and [2b](#F2){ref-type="fig"}, n = 2).
::: {#F1 .fig}
Figure 1
::: {.caption}
######
APV inhibited Aβ uptake, neuronal C1q production, and microglial activation. Slices were treated with no peptide (a, b, c), 30 μM Aβ 42 (d, e, f), or 30 μM Aβ 42 + 50 μM APV (g, h, i) for 3 days with fresh reagents added daily. Immunohistochemistry for Aβ (4G8, a, d, g), C1q (anti-rat C1q, b, e, h), and microglia (CD45, c, f, i) was performed on fixed and sectioned slices. Scale bar = 50 μm. Results are representative of three separately performed experiments. j. Immunoreactivity of Aβ (open bar), C1q (black bar), or CD45 (striped bar) was quantified as described in Materials and Methods. Values are the mean ± SD (error bars) from images taken from 8 slices (2 sections per slice) in 3 independent experiments (\* p \< 0.0001 compared to Aβ, Anova single factor test).
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Inhibition of Aβ-induced C1q synthesis by APV. a. C1q and β-actin mRNAs were assessed by RT-PCR in slices after 3 days of no peptide, 30 μM Aβ, or 30 μM Aβ + 50 μM APV treatment. Results are from one experiment representative of two independent experiments. b. Slices were treated with no peptide (open bar), 30 μM Aβ (black bar), or 30 μM Aβ + 50 μM APV (striped bar) daily for 3 days. 3 or 4 slices that had received same treatment were pooled, extracted and proteins analyzed by ELISA. Data are presented as percentage of control in ng C1q/mg total protein (mean ± SD of three independent experiments, \*\*p = 0.01 compared to Aβ, one-tailed paired t-test).
:::
:::
Integrin receptor antagonist GRGDSP (RGD) peptide enhances Aβ42 uptake and Aβ42-induced microglial activation and neuronal C1q expression
-----------------------------------------------------------------------------------------------------------------------------------------
It has been shown that an integrin receptor antagonist peptide, GRGDSP (RGD), can enhance Aβ ingestion by neurons in hippocampal slice cultures \[[@B37]\]. Therefore, we adopted this experimental manipulation as an alternative approach to modulate the level of Aβ uptake in neurons and assess the correlation between Aβ ingestion and neuronal C1q expression. Slices were treated with no peptide, 2 mM RGD, 10 μM Aβ42, or 10 μM Aβ42 + 2 mM RGD for 3 days with fresh peptides added daily. At the end of treatments, slices were collected and processed.
Addition of RGD peptide by itself did not result in neuronal C1q induction or microglial activation (CD45) compared to no treatment control, as assessed by immunostaining (data not shown). While greater ingestion was seen at 30 μM (Figure [1d, e, f](#F1){ref-type="fig"}), addition of 10 μM Aβ shows detectable Aβ ingestion, C1q expression, and microglial activation (Figure [3d, e, f](#F3){ref-type="fig"} compared with [3a, b, c](#F3){ref-type="fig"}). The lower concentration of Aβ was chosen for these experiments to ensure the detection of potentiation of uptake (vs. a saturation of uptake at higher Aβ42 concentrations). When RGD was provided in addition to 10 μM Aβ42, Aβ immunoreactivity in neurons with antibody 4G8 (Figure [3g](#F3){ref-type="fig"} vs. [3d](#F3){ref-type="fig"}) and 6E10 (similar results, data not shown), neuronal C1q expression (Figure [3h](#F3){ref-type="fig"} vs. [3e](#F3){ref-type="fig"}), and CD45 (Figure [3i](#F3){ref-type="fig"} vs. [3f](#F3){ref-type="fig"}) upregulation in microglia triggered by Aβ42, were significantly enhanced. Enhanced microglial activation was also detected with OX42 and ED1 antibodies (data not shown). Quantification by image analysis (Figure [3j](#F3){ref-type="fig"}) definitively demonstrated that the increased accumulation of Aβ in neurons, microglial activation, and induction of neuronal C1q synthesis in the presence of RGD. RT-PCR (Figure [4a](#F4){ref-type="fig"}) and ELISA (Figure [4b](#F4){ref-type="fig"}) further demonstrated that both mRNA and protein expression of C1q was enhanced by RGD. Thus, under the conditions tested, both neuronal C1q synthesis and microglial activation are coordinately affected when the internalization of Aβ is modulated negatively by APV or positively by RGD.
::: {#F3 .fig}
Figure 3
::: {.caption}
######
RGD enhanced Aβ uptake, neuronal C1q expression, and microglial activation. Hippocampal slices were treated with no peptide (a, b, c), 10 μM Aβ 42 (d, e, f), or 10 μM Aβ 42 + 2 mM RGD (g, h, i) for 3 days with fresh peptides added daily. Immunohistochemistry for Aβ (4G8, a, d, g), C1q (anti-rat C1q, b, e, h), and microglia (CD45, c, f, i) was performed on fixed slice sections. Scale bar = 50 μm. Results are representative of three separately performed experiments. j. Immunoreactivities of Aβ (open bar), C1q (black bar), or CD45 (striped bar) were quantified as described in Materials and Methods. Values are the mean ± SD (error bars) from images taken from 8 slices (2 sections per slice) in 3 independent experiments (\* p \< 0.0001, compared to Aβ, Anova single factor test).
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Enhancement of Aβ-induced C1q synthesis by RGD. a. C1q and β-actin mRNAs were assessed by RT-PCR in slices after 3 days of no peptide, 10 μM Aβ, or 10 μM Aβ + 2 mM RGD treatment. Results are from one experiment representative of two independent experiments. b. Slices were treated with no peptide (open bar), 10 μM Aβ (black bar), or 10 μM Aβ + 2 mM RGD (striped bar) daily for 3 days. 3 or 4 slices that had received same treatment were pooled, extracted and proteins analyzed by ELISA. Data are presented as percentage of control in ng C1q/mg total protein (mean ± SD of three independent experiments, \*\*p = 0.06 compared to Aβ, one-tailed paired t-test).
:::
:::
Aβ10--20 blocks Aβ42 induced microglial activation but triggers C1q synthesis in hippocampal neurons
----------------------------------------------------------------------------------------------------
Data reported by Giulian et al suggests that residues 13--16, the HHQK domain in human Aβ peptide, mediate Aβ-microglia interaction \[[@B38]\]. To investigate the effect of HHQK peptides in this slice culture system, rat hippocampal slices were treated with no peptide, 10 μM Aβ42, 10 μM Aβ42 + 30 μM Aβ10--20, or 30 μM Aβ10--20 for 3 days with fresh peptides added daily. Sections were immunostained for Aβ, C1q, and microglia. Aβ immunoreactivity was significantly reduced in the Aβ42 +Aβ10--20 treated tissues compared to the Aβ42 alone treatment (Figure [5g](#F5){ref-type="fig"} vs. [5d](#F5){ref-type="fig"}). Aβ10--20 alone-treated slices lacked detectable immunopositive cells with either 4G8 or 6E10 anti-Aβ antibody (Figure [5j](#F5){ref-type="fig"} and data not shown). Furthermore, as anticipated \[[@B38]\], when Aβ10--20 was present, microglial activation by Aβ42 as assessed by level of CD45, OX42, and ED1, was significantly reduced (Figure [5i](#F5){ref-type="fig"} vs. [5f](#F5){ref-type="fig"} and data not shown). Image analysis confirmed the inhibition of Aβ uptake (Figure [5m](#F5){ref-type="fig"}, open bars) and microglial activation (Figure [5m](#F5){ref-type="fig"}, striped bars) by the HHQK-containing Aβ10--20 peptide. However, production of C1q in neurons treated with Aβ42 was not inhibited by Aβ10--20 (Figure [5h](#F5){ref-type="fig"} vs. [5e](#F5){ref-type="fig"}). In fact, with Aβ10--20 alone, neurons were induced to express C1q to a similar level as Aβ42 (Figure [5k](#F5){ref-type="fig"}). The sustained C1q induction by Aβ10--20 was confirmed by RT-PCR for C1q with mRNAs extracted from slices (Figure [6a](#F6){ref-type="fig"}).
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Aβ10--20 blocked Aβ42 uptake, microglial activation, but not neuronal C1q induction. Slices were treated with no peptide (a, b, c), 10 μM Aβ 42 (d, e, f), 10 μM Aβ 42 + 30 μM Aβ 10--20 (g, h, i) or 30 μM Aβ 10--20 (j, k, l) for 3 days with fresh peptides added daily. Immunohistochemistry for Aβ (4G8, a, d, g, j), C1q (anti-rat C1q, b, e, h, k), and microglia (CD45, c, f, i, l) was performed on fixed and sectioned slices. Results are representative of three independent experiments. Scale bar = 50 μm. m. Immunoreactivities of Aβ (open bar), C1q (black bar), or CD45 (striped bar) were quantified as described in Materials and Methods. Values are the mean ± SD (error bars) from images taken from 8 slices (2 sections per slice) in 3 independent experiments. Microglial activation by Aβ42 was significantly inhibited by Aβ10--20 (\* p \< 0.0001, compared to either Aβ42 + Aβ10--20 or Aβ10--20, Anova single factor test).
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
a\. Aβ10--20 inhibited Aβ42-induced C1q and CD40 mRNA elevation, but not that of MCSF. C1q, MCSF, CD40, and β-actin mRNAs were assessed by RT-PCR in slices treated for 3 days with no peptide, 10 μM Aβ 42, 30 μM Aβ 10--20, or 10 μM Aβ 42 + 30 μM Aβ 10--20. Results are from one experiment representative of two independent experiments. b. APV blocked MCSF, CD40, and IL-8 mRNA induction triggered by Aβ42. RT-PCR for MCSF, CD40, IL-8, and β-actin were performed on RNA extracted from slices treated with no peptide (control), 30 μM Aβ 42, or 30 μM Aβ42 + 50 μM APV for 3 days. Results are from one experiment representative of two separate experiments.
:::
:::
CD40, IL-8, and MCSF mRNAs are induced by Aβ42 and differentially regulated by Aβ10--20 and APV
-----------------------------------------------------------------------------------------------
It is known that activated microglia cells can produce pro-inflammatory cytokines, chemokines, and nitric oxide, as well as higher expression of co-stimulatory molecules like CD40 and B7 \[[@B39]\]. Many of those proteins have been shown to be upregulated in microglia stimulated by Aβ in cell culture and *in vivo*\[[@B40]\]. Semi-quantitative reverse transcriptase PCR technique was used to determine how certain inducible activation products were modified in slice cultures stimulated with exogenous Aβ42 and in the presence of Aβ10--20 or APV. Rat slices were treated with 30 μM Aβ42 +/-APV or 10 μM Aβ42 +/- 30 μM Aβ10--20 for 3 days before mRNAs were extracted from tissues. LPS, was added at 150 ng/ml for 24 hr, served as positive control, with positive detection for all molecules tested (data not shown).
RT-PCR revealed that mRNAs for CD40 and IL-8 were enhanced in Aβ treated slice cultures relative to the control after 3 days (Figure [6a](#F6){ref-type="fig"} and [6b](#F6){ref-type="fig"}). Both Aβ10--20 and APV inhibited Aβ42-triggered upregulation of CD40 (Figure [6a](#F6){ref-type="fig"} and [6b](#F6){ref-type="fig"}), consistent with the inhibition of microglial activation by both Aβ10--20 and APV assessed by immunohistochemistry. APV also blocked Aβ42-induced IL-8 expression (Figure [6b](#F6){ref-type="fig"}), as did Aβ10--20 (data not shown).
Macrophage-colony stimulating factor (MCSF), a proinflammatory mediator for microglial proliferation and activation, has been shown to be expressed by neurons upon Aβ stimulation \[[@B41]\]. The expression of MCSF was induced in slice culture by Aβ treatment by Day 3 (Figure [6a](#F6){ref-type="fig"} and [6b](#F6){ref-type="fig"}) and this increase was blocked by the presence of APV (Figure [6b](#F6){ref-type="fig"}). In contrast, Aβ10--20 did not alter the Aβ42-triggered MCSF induction (Figure [6a](#F6){ref-type="fig"}), suggesting that MCSF may be required for microglial activation, but alone is not sufficient to induce that activation.
Discussion
==========
Previously, it has been shown that Aβ is taken up by pyramidal neurons in hippocampal slice culture and that the synthesis of complement protein C1q is induced in neurons \[[@B34]\]. Here we demonstrate that blocking of Aβ42 accumulation in neurons by NMDA receptor antagonist APV and increasing Aβ42 ingestion by integrin antagonist RGD is accompanied by inhibition and elevation in neuronal C1q expression, respectively. However, Aβ10--20, which markedly inhibits Aβ42 accumulation in pyramidal neurons, does not have any inhibitory effect on neuronal C1q expression. Thus, intraneuronal accumulation of Aβ is not necessary for Aβ-mediated induction of neuronal C1q synthesis.
Since Aβ10--20 alone can induce a level of C1q expression in neurons comparable to Aβ42, it is hypothesized that amino acids 10--20 in Aβ peptide contain the sequence that is recognized by at least one Aβ receptor. It was reported by Giulian et al. that the HHQK domain (residues 13--16) in Aβ is critical for Aβ-microglia interaction and activation of microglia, as they demonstrated that small peptides containing HHQK suppress microglial activation and Aβ-induced microglial mediated neurotoxicity \[[@B38]\]. We have previously reported that rat Aβ42, which differs in 3 amino acids from human Aβ42, including 2 in the 10--20 region and 1 in the HHQK domain, was internalized and accumulated in neurons but failed to induce neuronal C1q expression \[[@B34]\]. This is consistent with the hypothesis that a specific Aβ interaction (either neuronal or microglial), presumably via the HHQK region of the Aβ peptide, but not intracellular Aβ accumulation, can lead to neuronal C1q induction in hippocampal neurons.
Neurons are the major type of cells that accumulate exogenous Aβ in slice cultures. Microglial activation, as assessed by CD45, OX42, and ED1, was increased with enhanced neuronal Aβ42 uptake and inhibited when Aβ42 uptake was blocked by APV or Aβ10--20 in this slice culture system. These data would be consistent with a model in which neurons, upon internalization of Aβ peptide, secrete molecules to modulate microglial activation \[[@B14],[@B41],[@B42]\] (Figure [7](#F7){ref-type="fig"}, large arrows). Synthesis and release of those molecules may require the intracellular accumulation of Aβ since blocking intraneuronal Aβ accumulation always blocked microglial activation. The finding that treatment with Aβ10--20 alone did not result in intraneuronal Aβ immunoreactivity or microglial activation, while rat Aβ42, which did accumulate within neurons, induced activation of microglial cells, is consistent with this hypothesis. It should be noted that an absence of Aβ immunoreactivity in Aβ10--20 treated slices does not exclude the possibility that Aβ10--20 was ingested but soon degraded by cells, and thus accumulation of Aβ rather than ingestion alone may be necessary to induce secretion of microglia activating molecules from neurons. Giulian et al. reported that the HHQK region alone was not able to activate microglia \[[@B38]\]. Thus, Aβ10--20 might block microglial activation by competing with Aβ42 for direct microglial binding, as well as by blocking uptake and accumulation of Aβ in neurons.
::: {#F7 .fig}
Figure 7
::: {.caption}
######
Model of Aβ interaction with neurons and microglia in slice cultures. Exogenous Aβ peptide interacts with neuronal receptors leads to at least two separate consequences, in one of which C1q expression is upregulated in neurons. A second receptor mediates the secretion of certain modulatory molecules, which lead to microglial activation involving the expression of CD45, CR3, CD40, and IL-8. This does not exclude the direct interactions of Aβ with receptor(s) on microglia that may also contribute to microglial activation.
:::
:::
Activated glial cells, especially microglia, are major players in the neuroinflammation seen in of Alzheimer\'s disease \[[@B43]\]. Microglial cells can be activated by Aβ and produce proinflammatory cytokines, nitric oxide, superoxide, and other potentially neurotoxic substances *in vitro*, although the state of differentiation/ activation of microglia and the presence of other modulating molecules is known to influence this stimulation \[[@B7],[@B9],[@B43]\]. \"Activated\" microglia also become more phagocytic and can partially ingest and degrade amyloid deposits in brain. This leads many to hypothesize that there are multiple subsets of \"activated\" microglia, each primed to function in a specific but distinct way \[[@B5],[@B43]\].
In hippocampal slice cultures, we and others have shown that Aβ42 triggered microglial activation as assessed by immunohistochemical detection of CR3 (OX42), and cathepsin D \[[@B34],[@B37]\]. Several chemokines, including macrophage inflammatory protein-1 (MIP-1α, MIP-1β), monocyte chemotactic protein (MCP-1), and interleukin 8 (IL-8), have been reported to increase in Alzheimer\'s disease patients or cell cultures treated with Aβ \[[@B44],[@B45]\]. CD40, a co-stimulatory molecule, is also upregulated in Aβ-treated microglia \[[@B10]\]. In this study, similar to reports of cultured microglia, immunoreactivity of CD45 was found increased on microglia in Aβ42 treated slice cultures, and CD40 and IL-8 messenger RNAs were elevated after Aβ42 exposure. As expected, CD40 and IL-8 mRNA induction was blocked whenever immunohistochemistry analysis showed the inhibition of microglial activation. \[We did not observe change in MIP-1α, 1β mRNAs in slice culture with Aβ42 treatment, and MCP-1 was too low to be detected with or without Aβ stimulation although it was detectable in LPS treated slices (data not shown).\]
The data presented thus far suggest the hypothesis that neurons, upon uptake and accumulation of Aβ, release certain substances that activate microglia. One possible candidate of those neuron-produced substances is MCSF, which has been reported to be induced in neuronal cultures upon Aβ stimulation \[[@B41],[@B46]\], and is known to be able to trigger microglial activation \[[@B47]\]. Indeed, MCSF mRNA was found to increase after 3 days of Aβ treatment (Figure [6a](#F6){ref-type="fig"} and [6b](#F6){ref-type="fig"}). The diminished MCSF signal with the addition of APV and coordinate lack of microglial activation is consistent with a proposed role of activating microglia by MCSF produced by stimulated neurons. However, in the presence of Aβ10--20, MCSF induction was unaltered, though microglial activation was inhibited. Thus, MCSF alone does not lead to the upregulation of the above-mentioned microglial activation markers.
In this organotypic slice culture, no significant neuronal damage was observed in 3 day treatment with Aβ at concentrations that have been reported to cause neurotoxicity in cell cultures. One possible explanation is that the peptide has to penetrate the astrocyte layer surrounding the tissue to reach the multiple layers of neurons. Thus, the effective concentration of Aβ on neurons is certainly much lower than the added concentration. Aβ failing to induce neurotoxicity in slices to the same extent as in cell cultures may also indicate the loss of certain protective mechanisms in isolated cells. A distinct advantage of the slice culture model is that the tissue contains all of the cell types present in brain, the cells are all at the same developmental stage, and cells may communicate in similar fashion as *in vivo*.
Our data demonstrating distinct pathways for the induction of neuronal C1q and the activation of microglial by amyloid peptides suggest the involvement of multiple Aβ receptors on multiple cell types in response to Aβ (Figure [7](#F7){ref-type="fig"}, model) and possibly in Alzheimer\'s disease progression. This multiple-receptor mechanism is supported by reports suggesting many proteins/complexes can mediate the Aβ interaction with cells \[[@B48]\]. These include, but not limited to, the alpha7nicotinic acetylcholine receptor (alpha7nAChR), the P75 neurotrophin receptor (P75NTR) on neurons, the scavenger receptors and heparan sulfate proteoglycans on microglia, as well as receptor for advanced glycosylation end-products (RAGE) and integrins on both neurons and microglia (Figure [7](#F7){ref-type="fig"}). Several signaling pathways have been implicated in specific Aβ-receptor interactions \[[@B49]-[@B51]\]. However, it is not known which receptors are required for induction of C1q in neurons. In addition, as of yet the function of neuronal C1q has not been determined. Previous reports from our lab have shown that C1q is associated with hippocampal neurons in AD cases but not normal brain \[[@B52]\], and the fact that it is synthesized by the neurons has been documented by others \[[@B23],[@B53]\]. In addition, C1q was prominently expressed in a preclinical case of AD (significant diffuse amyloid deposits, with no plaque associated C1q, and no obvious cognitive disorder) and is expressed in other situations of \"stress\" or injury in the brain \[[@B54]-[@B58]\]. Indeed, overexpression of human cyclooxygenase-2 in mice leads to C1q synthesis in neurons and inhibition of COX-2 activity abrogates C1q induction. These data suggest that in addition to the facilitation of phagocytosis by microglia \[[@B59],[@B60]\] (particularly of dead cells or neuronal blebs), the induction of C1q may be an early response of neurons to injury or regulation of an inflammatory response, consistent with a role in the progression of neurodegeneration in AD. Whether and how the neuronal C1q production affects the survival of neurons is still under investigation. Identifying the receptors responsible for neuronal C1q induction may be informative in understanding the role of C1q in neurons in injury and disease.
Conclusions
===========
In summary, induction of C1q expression in hippocampal neurons by exogenous Aβ42 is dependent upon specific cellular interactions with Aβ peptide that require HHQK region-containing sequence, but does not require intraneuronal accumulation of Aβ or microglial activation. Thus, induction of neuronal C1q synthesis may be an early response to injury to facilitate clearance of damaged cells, while modulating inflammation and perhaps facilitating repair. Microglial activation in slice culture involves the induction of CD45, CD40, CR3, and IL-8, which correlates with intraneuronal accumulation of Aβ, indicating contribution of factors released by neurons upon Aβ exposure. MCSF may be one of those stimulatory factors, though by itself MCSF cannot fully activate microglia.
Removal of Aβ to prevent deposition and of cellular debris to avoid excitotoxicity would be a beneficial role of microglial activation in AD. However, activated microglia also produce substances that are neurotoxic. Therefore, the goal of modulating the inflammatory response in neurodegenerative diseases like AD is to enhance the phagocytic function of glial cells and inhibit the production of proinflammatory molecules. Being able to distinguish in the slice system C1q expression (which has been shown to facilitate phagocytosis of apoptotic cells in other systems \[[@B24]\]) from microglial activation suggests a plausible approach to reach that goal *in vivo*.
List of abbreviations
=====================
Aβ: amyloid beta; AD: Alzheimer\'s disease; APV: D-(-)-2-amino-5-phosphonovaleric acid; BSA: bovine serum albumin; GRGDSP (RGD): glycine-arginine-glycine-aspartic acid-serine-proline; HBSS: Hanks\' balanced salt solution; HEPES: N-2-hydroxyethylpiperazine-N\'-2-ethanesulfonic acid; MCSF: macrophage colony stimulating factor; NMDA: N-methyl-D-aspartic acid; PMSF: phenylmethylsulfonylfluoride; TAE: triethanolamine.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
RF cultured and processed the tissue, performed all experiments (immunohistochemistry, ELISA, PCR and others), analyzed the data, and drafted the manuscript. AJT contributed to the design of the study, guided data interpretation and presentation and edited the manuscript.
Acknowledgments
===============
This work is supported by NIH NS 35144 and P50 AG16573. The authors thank Dr. Saskia Milton and Dr. Charles Glabe for providing the synthetic human Aβ peptide, and Dr. Maria Fonseca, Dr. Ming Li, and Karntipa Pisalyaput for their review of this manuscript.
|
PubMed Central
|
2024-06-05T03:55:51.930260
|
2005-1-10
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545941/",
"journal": "J Neuroinflammation. 2005 Jan 10; 2:1",
"authors": [
{
"first": "Rong",
"last": "Fan"
},
{
"first": "Andrea J",
"last": "Tenner"
}
]
}
|
PMC545942
|
Background
==========
Dyslipidemia in type 2 diabetes are most frequently characterized by elevation of total serum triglycerides, of very low density lipoprotein-triglyceride (VLDL-TG) and low level of high density lipoprotein-cholesterol (HDL-C) \[[@B1]\]. Hypertriglyceridemia is an independent risk factor for coronary artery disease (CAD) in type 2 diabetes \[[@B2]\]. Triglyceridemia is modulated by environmental and genetic factors. A new identified gene associated with triglyceride level was the gene encoding for apo A-V located at the chromosome 11 (11q23), in the vicinity of apoA-I/C-III/A-IV cluster \[[@B3]\]. Studies on transgenic mice overexpressing human apo A-V showed a decreased level of triglyceride, whereas knock-out mice showed an increased level of triglyceride \[[@B3]\]. These results prove the regulator effect of apo A-V on triglyceride metabolism. Moreover apo A-V regulates levels of circulating triglyceride and cholesterol \[[@B4]\]. Four neighboring single nucleotide polymorphisms (SNP1 to SNP4) within apo A-V were identified by Pennachio et al \[[@B3]\]. The first three of the SNPs (SNPs1--3) were in significant linkage disequilibrium suggesting the existence of a common haplotype in apo A-V gene. The minor allele of each SNP was associated with high triglyceride level. In other study, the SNP3 (T/C polymorphism) was also associated with HDL-C concentration \[[@B5]\]. This suggests that genetic variability of the apo A-V gene is likely to also have an impact on the lipid profile of type 2 diabetic patients, but reports on the subjects are few \[[@B6]\].
We have addressed the issue to examine the interaction between SNP3 and lipid profile and coronary artery disease (CAD) in type 2 diabetic patients compared to controls in Tunisian population.
Results
=======
Description of the participating groups
---------------------------------------
Type 2 diabetic patients have a BMI values more important than non diabetic subjects. Whereas the WHR in lower in controls. Males and smokers are more frequent in patients with CAD (Table [1](#T1){ref-type="table"}). To compare lipid parameters, we take into consideration the lipid lowering drugs use, then patients who are taking lipid lowering drugs were excluded (Table [2](#T2){ref-type="table"}). Plasma total cholesterol and total triglyceride concentrations were significantly higher in type 2 diabetic patients than in the non diabetic patients. Subjects with CAD had lower concentration of HDL-C and higher concentration of triglyceride as compared to those without CAD (Table [2](#T2){ref-type="table"}).
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Clinical characteristics of the four studied groups
:::
**Diabetes- CAD-** **Diabetes+ CAD-** **Diabetes+ CAD+** **Diabetes- CAD+** **Diabetes vs non Diabetes** **CAD vs non CAD**
----------------------------- -------------------- -------------------- -------------------- -------------------- ------------------------------ --------------------
Number 99 78 74 57
Sex (% of men) 50.5 49.4 75.4 84.2 ns \< 0.001
Age (years) 52.9 ± 8.7 51.5 ± 7.3 55.7 ± 6.8 59.0 ± 6.8 ns \< 0.001
BMI (Kg/m^2^) 27.6 ± 4.0 28.8 ± 4.6 28.1 ± 3.8 26.6 ± 4.8 0.029 ns
Smokers (%) 42 53 86 80 0.003 \< 0.001
SBP (mmHg) 12.5 ± 1.3 13.2 ± 1.4 13.0 ± 1.8 12.4 ± 2.1 0.001 ns
DBP (mmHg) 7.5 ± 0.7 7.8 ± 0.9 7.6 ± 1.0 7.1 ± 1.1 0.006 0.018
WHR 0.82 ± 0.09 0.92 ± 0.07 0.95 ± 0.07 0.92 ± 0.06 ns 0.016
Glucose (mmol/l) 5.2 ± 0.4 10.4 ± 3.4 11.7 ± 3.5 6.8 ± 3.6 \< 0.001 \< 0.001
HbA1c (%) 4.5 ± 0.9 10.4 ± 3.4 9.4 ± 2.4 11.3 ± 4.1 0.009 \< 0.001
Diabetes duration (years) 6.8 ± 5.2 7.9 ± 6.1
Lipid lowering drug use (%) 11.1 0 16.2 70.2
SBP : systolic blood pressure
DBP : diastolic blood pressure
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
lipid profiles of the four studied groups
:::
**Diabetes- CAD-** **Diabetes+ CAD-** **Diabetes+ CAD+** **Diabetes- CAD+** **Diabetes vs non Diabetes** **CAD vs non CAD**
---------------------- -------------------- -------------------- -------------------- -------------------- ------------------------------ --------------------
Number 88 78 62 17
Cholesterol (mmol/l) 4.70 ± 1.22 5.16 ± 1.14 5.13 ± 1.12 4.06 ± 1.22 \< 0.001 ns
HDL-C (mmol/l) 0.94 ± 0.24 0.96 ± 0.38 0.79 ± 0.22 0.72 ± 0.21 ns \< 0.001
Total TG (mmol/l) 1.39 ± 0.75 2.11 ± 1.48 2.27 ± 1.77 2.81 ± 2.53 \< 0.001 \< 0.001
:::
Heterozygous genotype had the high triglyceride level
-----------------------------------------------------
According to SNP3 of the apo A-V gene, variation of lipid parameters in diabetic or non diabetic patients are shown in Table [3](#T3){ref-type="table"}. Non diabetic subjects having the heterozygous genotype (T/C) showed an increased triglyceride level and decreased HDL-C concentration. However these variations were not significant. In type 2 diabetic patients, triglyceride level increased significantly in C/T genotype in association with non significant decrease in HDL-C concentration. The genotype frequencies of T/T, T/C and C/C were 0.74, 0.23 and 0.03 respectively in non diabetic subjects, 0.71, 0.25 and 0.04 respectively in type 2 diabetic patients. The SNP3 was shown to be in Hardy-Weinberg equilibrium. It is clear that there was no difference in genotype distribution between diabetic and non diabetic subjects. The type 2 diabetic population was classified further into those with high and low triglyceride concentration (cut point 2.2 mmol/l which was more than 90 percentile level in the healthy population). The SNP3 frequencies for T/T, T/C and C/C genotypes in the low triglyceride groups were 77.1 %, 18.8 % and 4.2 % respectively, and those in the high triglyceride group were 59.1 %, 40.9 % and 0 % respectively. The difference in genotype frequencies between low and high triglyceride groups were significant (p = 0.011) (Table [4](#T4){ref-type="table"}).
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Clinical Characteristics and lipid profile according to SNP3 of apo A-V in type 2 diabetic and in non diabetic subjects
:::
**Type 2 Diabetes** **Non Diabetes**
---------------------- --------------------- ------------------ ------------- -------------
Characteristics **Genotypes**
**TT** **TC** **TT** **TC**
Number 100 36 76 25
Sex (% men) 61 50 55.8 60
Smokers (%) 67.3 81.3 44.4 46.2
Age (years) 52.8 ± 7.6 54.8 ± 6.2 53.3 ± 8.8 51.5 ± 7.5
BMI (Kg/m^2^) 28.5 ± 4.3 28.7 ± 3.9 27.3 ± 3.0 28.1 ± 5.0
WHR 0.94 ± 0.06 0.93 ± 0.08 0.93 ± 0.08 0.9 ± 0.08
Cholesterol (mmol/l) 5.17 ± 1.15 5.17 ± 1.07 4.61 ± 1.32 4.53 ± 0.99
HDL-C (mmol/l) 0.90 ± 0.32 0.83 ± 0.34 0.92 ± 0.25 0.83 ± 0.26
Total TG (mmol/l) 2.05 ± 1.61 2.62 ± 1.59\* 1.6 ± 1.42 1.75 ± 1.07
\* significant difference between TT and TC genotypes (p = 0.016).
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Distribution of different apo A-V genotypes in type 2 diabetes between low and high triglyceride groups
:::
**Low triglyceride group ≤ 2.2 mmol/l** **High triglyceride group \> 2.2 mmol/l**
-------------- ----------------------------------------- ------------------------------------------- ------------ -------
Polymorphism **Number** **%** **Number** **%**
T/T 74 77.1 26 59.1
T/C 18 18.8 18 40.9
C/C 4 4.2 nd nd
Total 96 100 44 100
nd : not detected
X2 = 8.962, p = 0.011, degree of freedom = 2.
:::
No association between SNP3 and CAD
-----------------------------------
To investigate the relation of SNP3 polymorphism with coronary artery disease we studied the association in all subjects (those who are taking lipid lowering drugs are included). There was no association between coronary artery disease and SNP3 either in non diabetic subjects or in type 2 diabetic patients (Table [5](#T5){ref-type="table"}).
::: {#T5 .table-wrap}
Table 5
::: {.caption}
######
Association between SNP3 of apo A-V and risk for coronary artery disease in diabetic and non diabetic subjects
:::
**Genotypes** **CAD-** **CAD +**
----------- --------------- ----------- ----------- ----------------------------------
Diabetes- T/T 74 (74.7) 41 (71.9) P = 0.514 OR = 1.29 (0.6--2.77)
T/C 21 (21.1) 15 (26.3)
C/C 4 (4.0) 1 (1.8)
Diabetes+ T/T 56 (71.8) 52 (70.3) P = 0.717 OR = 0.87 (0.41--1.83)
T/C 21 (26.9) 17 (23.0)
C/C 1 (1.3) 5 (6.7)
OR was calculated using T/T and T/C genotypes only.
:::
Discussion
==========
Elevated serum lipid levels are an important risk factor for atherosclerosis. Both environmental and genetic factors contribute to variability in serum lipid levels \[[@B5]\]. In this study, we choose the apo A-V gene as a genetic factor that predispose to elevated triglyceride level. The dynamic interfacial properties of apo A-V are consistent with the hypothesis that apo A-V impedes triglyceride particle assembly \[[@B7]\]. Thus, the effect of apo A-V on triglyceride level can be attributed to the intracellularly function of apo A-V to modulate hepatic VLDL synthesis and/or secretion. Also apo A-V can lower plasma triglyceride by activating the lipoprotein lipase (LPL)\[[@B8]\]. In our study we showed that SNP3 was significantly associated with hypertriglyceridemia especially in type 2 diabetic patients. This finding confirmed the idea that the effect of this SNP on triglyceride metabolism was not influenced by ethnic background \[[@B9]\]. However, it has been reported that the major and minor allele frequencies differed between populations such as : 0.06, 0.09 and 0.34 for the C allele in UK, Caucasian and Japanese respectively \[[@B10],[@B3],[@B9]\]. In our population, the minor allele frequency (0.13) is almost the same as in Caucasian \[[@B3]\] but lower than in Japanese population \[[@B9]\]. However, the triglyceride level is higher in Tunisian than in Japanese population. This paradoxical observation confirms that triglyceride level is influenced by environmental factors and other genetic factors (apo CIII\...).
In type 2 diabetes the triglyceride level is increased, which is due to multiple factors related to insulin and carbohydrate metabolism, LPL activity, CETP activity\... The absence of an unusual allele frequency of SNP3 in type 2 diabetic patients compared to non diabetic subjects, in spite of triglyceride level variation, shows that there is no association between apo A-V SNP3 polymorphism and the presence or absence of diabetes. Diabetic homozygous for the major allele are more frequent in low triglyceride group, showing that SNP3 is associated with triglyceride variation in type 2 diabetic patients. Contrary to Esteve et al., who have reported no significant difference in triglyceride concentration with the apo A-V polymorphism in type 2 diabetic patients \[[@B6]\], our results showed, then, that SNP3 is associated with triglyceride level.
The relationship between SNP3 and coronary artery disease is under discussion. The LOCAT study showed that there is no association between SNP3 and progression of coronary heart disease \[[@B11]\]. Our results find no association between SNP3 of apo A-V and development of coronary artery disease either in diabetic patients or in non diabetic subjects. This suggests that the high triglyceride level in T/C genotype alone was not a good discriminator of coronary heart disease. In contrast ; Szalai et al. showed an association between SNP3 and an increased risk for severe coronary artery disease \[[@B12]\]. Identifying genetic and environmental factors that influence plasma lipid levels represents a key step towards developing strategies for preventing and treating CAD. Usually, in the case of type 2 diabetes, patients are taking lipid lowering drugs in order to ameliorate their lipid profile, namely lower triglyceride level and increase HDL-C concentration. Fibrates represent a commonly used therapy for lowering plasma triglyceride, its mechanism of action involves the activation of the nuclear receptor peroxisome proliferator activated receptor alpha (PPARα). The apo A-V is a highly responsive PPAR target gene \[[@B13]\]. While SNP3 is located in promoter region, it should be interesting to study the interaction between this polymorphism and lipid lowering drugs response in different population.
Conclusion
==========
In summary, the apo A-V SNP3 is associated with triglyceride level in Tunisian type 2 diabetic patients. However, this SNP is unlikely to be associated with the presence of diabetes. Although SNP3 is associated with hypertrigyceridemia, there was no relationship between this polymorphism and coronary artery disease. Further investigations were needed to determine the effect of SNP3 on lipid lowering drug response.
Methods
=======
Subjects
--------
Three hundred and eight subjects, aged 45--70 years, participated in this study. They belong to four groups. The first group contained 74 type 2 diabetic patients with CAD, among whom 12 are taking lipid lowering drugs. The second group contained 78 type 2 diabetic patients without CAD, all of them did not take lipid lowering drugs. The third group contained 57 patients with CAD and without diabetes, 70% of the patients are taking lipid lowering drugs. The last group contained 99 controls without diabetes nor CAD, 11 subjects are taking lipid lowering drugs. The clinical and biological characteristics of each group are summarized in Table [1](#T1){ref-type="table"}.
All participants were recruited in the departments of internal medicine and cardiology in Monastir university hospital. Written or verbal informed consent were obtained from all patients and controls before the study. All written informed consent were not possible because the majority of eligible subjects in our study were illiterate. The diagnosis of diabetes was based on a previous history of diabetes according to American Diabetes Association criteria \[[@B14]\]. The patients with CAD were defined as clinical history of stable angina pectoris, previous acute coronary syndromes with or without ST segment elevation. This CAD was confirmed by coronary angiography. Exclusion criteria were taking insulin, having a renal or liver failure or thyroid disease, alcohol consumption before 3 days or less, Body Mass Index (BMI) more than 35 Kg/m^2^and glycosylated hemoglobin (Hb A1c) more than 12%. Post menopausal women had no hormone replacement therapy. BMI was calculated using the formula : weight (Kg)/height^2^(m^2^). Obesity was defined as BMI\>30 kg/m^2^. Waist-to-hip ratio (WHR) was calculated from measurements of the waist circumference taken at the mid point between umbilicus and xiphoid and hip circumference, at the widest point around the hips, respectively. Blood samples were drawn after subjects had fasted overnight (12 hours) into tubes containing EDTA. Plasma was immediately separated by centrifugation.
Laboratory Analysis
-------------------
After DNA extraction, the single nucleotide polymorphism SNP3 within the apo A-V gene (-1311 T/C) was determined by PCR-RFLP analysis using MseI restriction endonuclease as described previously \[[@B3]\]. Plasma glucose, glycosylated haemoglobin (HbA1c), lipids and lipoproteins were determined as described by Smaoui et al. \[[@B15]\].
Statistical Analyses
--------------------
Data management and statistical analysis were performed using SPSS 10.0 software. Results are summarized as mean ± SD. Since triglyceride levels were not normally distributed, logarithmic transformation of triglyceride concentration was performed before the statistical analysis. Student\'s test was used to compare continuous variables and Chi square (χ^2^) test was used to examine distribution of categorical variables. A value of p \< 0.05 was considered significant.
List of abbreviations
=====================
Apo : apolipoprotein
BMI : Body Mass Index
CAD : Coronary Artery Disease
HDL : High Density Lipoproteins
HDL-C : HDL cholesterol
PCR-RFLP : Polymerase Chain Reaction-Restriction Fragment Length Polymorphism
SNP : Single Nucleotide Polymorphism
TG : Triglyceride
VLDL : Very Low Density Lipoproteins
WHR : Waist-to-Hip Ratio
Authors\' contributions
=======================
R.Ch and N.A: carried out the molecular studies, participated in the design of the study and drafted the manuscript ; M.S : carried out the biochemical essay ; S.H and Sy.M : interested in the clinical aspect ; M.H : conceived of the study, and participated in its design and coordination and helped to draft the manuscript and revised it critically for important intellectual content and have given final approval of the version to be published; MS.M : revised the article critically for important intellectual content and have given final approval of the version to be published.
Acknowledgements
================
This study was supported by grants from the \"Ministère de l\'Enseignement Supérieur et de la recherche Scientifique et Technologique\" (DGRST -- \"UR Nutrition Humaine et Désordres Métaboliques\").
|
PubMed Central
|
2024-06-05T03:55:51.933126
|
2005-1-6
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545942/",
"journal": "Lipids Health Dis. 2005 Jan 6; 4:1",
"authors": [
{
"first": "Raja",
"last": "Chaaba"
},
{
"first": "Nebil",
"last": "Attia"
},
{
"first": "Sonia",
"last": "Hammami"
},
{
"first": "Maha",
"last": "Smaoui"
},
{
"first": "Sylvia",
"last": "Mahjoub"
},
{
"first": "Mohamed",
"last": "Hammami"
},
{
"first": "Ahmed Slaheddine",
"last": "Masmoudi"
}
]
}
|
PMC545943
|
Introduction
============
Alzheimer\'s disease (AD) is a progressive degeneration of neural structure and function that arises in the cerebral cortex. Behaviorally, affected individuals usually present with semantic difficulty, followed by a deficiency in episodic memory, spatial disorientation, sleep disturbances, depression, agitation, loss of longer memories, general difficulty with the activities of daily living, and eventually, death. Neuropathological findings include a relatively high number of extracellular deposits of the amyloid β-peptide (Aβ), argyrophillic cytoskeletal aggregates in neurons, accumulation of α-synuclein, loss of synapses, loss of cholinergic and adrenergic fibers, loss of pyramidal neurons, and cerebral amyloid angiopathy (CAA) -- deposition of Aβ around blood vessels.
Most of the AD correlates above have been connected in some way to inflammation. For instance, the plaques -- comprised primarily of aggregated amyloid β-peptide (Aβ) -- are inundated with microglia that show profiles of morphology and gene expression consistent with inflammation. Indeed, if one characterizes any activity by microglia as a sign of \"neuroinflammation,\" it can be said that inflammatory responses have been evident in AD for at least 40 years \[[@B1]\]. But, it was not until the late 1980s that investigators were willing to express the hypothesis that inflammatory events were causal or otherwise contributing to the progression of the disease. Recognition of the powerful impact of a cytokine like interleukin-1 (IL-1), elevated in AD microglia, permitted such speculation \[[@B2]\]. Similarly, research accrued showing that primary inflammation could lead to many of the aberrations found in AD, fueling the consideration that inflammatory events were seminal \[[@B3]-[@B5]\]. Many of the individual molecules produced by activated microglia and astrocytes are conditional neurotoxins: hydrogen peroxide, glutamate and other agonists of glutamate receptors, complement components, prostanoids. (Nitric oxide from inducible nitric oxide synthase, produced abundantly in rodent glia, may be less important in human tissues.) Retrospective epidemiological studies showed protection against AD -- either in age of onset or rate of progression -- by nonsteroidal antiinflammatory drugs (NSAIDs); such correlations have now been born out in a prospective study \[[@B6]\]. Perhaps most compelling, polymorphisms in the genes for proinflammatory cytokines are indicative of risk for AD \[[@B7]\].
Despite these indications, there are reasons to believe that the changes observed in glia and inflammatory cytokines constitute a compensatory response in AD. Indeed, some investigators have been reluctant to apply the term \"inflammation\" to the constellation of events related to AD pathology. Some of the cytokines and other gene products expressed in peripheral sites of inflammation are present in the AD brain, but there is no apparent vasodilation or extravasation of neutrophils. In general, there seems to be less of the molecular and cellular behavior that is responsible for bystander tissue damage in peripheral inflammation. This journal was founded partially out of recognition that \"neuroinflammation\" is distinct. In essence, the concept reflects a compromise befitting the difficult line that must be maintained between effective cell-mediated immune responses and damage to the precious components of the CNS. Microglia elevate their expression of neurotrophic factors under many of the same conditions in which they show inflammation-related responses such as phagocytosis, retraction of processes, release of excitotoxins, and production of IL-1β and IL-6 and tumor necrosis factor \[[@B8]\]; in fact, the latter cytokines can have neurotrophic effects themselves \[[@B9],[@B10]\]. Astrocytes deposit proteoglycans around the Aβ deposits destined to become plaques \[[@B11]\], perhaps sequestering this neurotoxic peptide from doing its harm. Even the apparent benefits of NSAIDs can be parsed from their presumed mechanism of inhibiting cyclooxygenase-2 \[[@B12],[@B13]\](and references therein).
Discussion
==========
Recent experiments with anti-Aβ immunization have highlighted another beneficial effect of \"activated\" microglia: removal of Aβ. It has long been recognized that microglia can efficiently phagocytose and at least partially degrade Aβ both in vitro and in vivo. But the persistence of amyloid plaques suggests that microglia are stymied in this regard during the development of AD or in the deposition of Aβ in mice transgenically engineered to produce large amounts of the peptide. Introduction of antibodies recognizing Aβ, either by active vaccination or by passive immunization (injection of antibodies, typically monoclonal), results in removal of some Aβ deposits and/or prevention of their formation. Although the phenomenon has been studied most rigorously in the transgenic mouse models, similar clearance of parenchymal plaques seems to have occurred in two human subjects that participated in an Aβ-vaccine trial \[[@B14],[@B15]\]. And microglia appear to contribute; Aβ can be readily detected in microglia of immunized mice \[[@B16]\] and was also abundant in some microglia and related syncitia in the AD trial subjects \[[@B14],[@B15]\]. However, the only reason we are privy to the effects of the vaccination paradigm in humans is because these two individuals died after complications of meningeal encephalitis -- rampant cranial inflammation brought on by the immunization. This iatrogenic event occurred in about five percent of the human subjects vaccinated against Aβ, prompting discontinuation. One interesting finding from both autopsies is that while parenchymal Aβ deposits were substantially lower than to be expected in AD victims, both individuals had relatively high levels of vascular deposition. This CAA was accompanied by microhemorrhage in at least one of the subjects \[[@B15]\], consistent with the majority of advanced cases of CAA \[[@B17]\].
Wilcock *et al*. \[[@B18]\] have now produced evidence that the appearance of CAA after immunization may represent an actual increase in this parameter triggered by anti-Aβ antibodies. Furthermore, the investigators also found that the CAA was accompanied by an increase in hemorrhages -- similar to a previous report \[[@B19]\] -- and a vascular accumulation of CD45^+^cells presumed to be microglia. The experimental paradigm was one of passive immunization of transgenic mice at nearly two years of age, old enough to have accumulated substantial Aβ deposits. Consistent with expectations, injection of anti-Aβ antibody diminished deposits in the parenchyma, even those that were mature enough to stain with Congo red. However, vascular deposition of Congo-red staining was elevated by approximately four-fold in the anti-Aβ-treated animals. Pfeiffer *et al*. found similar results in another transgenic line \[[@B19]\]. Further, Wilcock *et al*. now show that the regional accumulation of vascular amyloid was accompanied by an elevated index of hemorrhages and a congregation of CD45^+^cells, presumed to be microglia \[[@B18]\]. Given that stromal microglia show increased signs of activity and contain Aβ after passive Aβ immunization \[[@B20]\], one interpretation is that the immunization-induced shift in amyloid from the parenchyma to the vasculature is mediated by phagocytic microglia attempting to discard the Aβ into the bloodstream. Such a phenomenon is tenuously supported by the analogous transport of pyknotic neuronal nuclei to the vasculature by microglia, observed in 3-D time-lapse videos by Dailey and coworkers \[[@B21]\]. In those images, microglia are occasionally seen to transfer the nuclei to another cell, conceivably a perivascular macrophage or dendritic cell. Thus, it is not clear whether the CD45^+^cells observed by Wilcock and coworkers are microglia or another cell type. It is also unclear whether the accumulation of amyloid and inflammatory cells at the blood vessels represents an arrested state in Aβ clearance or simply a bottleneck in the transport, one that would eventually yield to complete removal of the peptide. However, the appearance of CAA in the human subjects that suffered from acute encephalitis suggests that the vascular accumulation is an untoward event, created or facilitated by inflammation. Another vascular irregularity caused by Aβ has been linked to inflammatory events in both transgenic mice and isolated human blood vessels \[[@B22]\].
The apparent contributions of inflammatory mechanisms to both Aβ clearance and vascular pathology illustrate a somewhat unique example of microglial ambivalence. While many had argued that microglial \"activation\" by Aβ was at least partially responsible for AD-associated degeneration, others had pointed to microglial phagocytosis as a desirable consequence of activation. For the purposes of discussion, the term \"*mal*activation\" will be applied here to microglial activation which produces neurodegeneration. One obvious question is whether there might be a mode of \"activation\" that permits phagocytosis while limiting malactivation. In fact, stimulation of Fc receptors -- the antibody receptors that initiate a good deal of antibody-triggered phagocytosis -- can inhibit cytotoxicity in macrophages \[[@B23]\]. Similarly, phagocytosis of apoptotic cells inhibits macrophage expression of proinflammatory cytokines like IL-1, IL-8, tumor necrosis factor, and several prostanoids through stimulation of a phosphatidylserine receptor \[[@B24]\]. Evidence indicates that malactivation involves the production of reactive oxygen species like superoxide and peroxide, nitric oxide, and excitotoxins (glutamate, quinolinate, and D-serine). If these criteria are germane, malactivation certainly can be suppressed by specific cytokines, such as transforming growth factor β (TGFβ) \[[@B25]\]. Although TGFβ has often been characterized broadly as \"anti-inflammatory,\" it does not inhibit the phagocytic activity of microglia in a setting where another \"anti-inflammatory\" cytokine (IL-4) does \[[@B26]\]. Interestingly, TGFβ1 transgenesis promotes the same apparent shift of Aβ from parenchyma to vessel that is observed after Aβ immunization \[[@B27]\].
Conclusions
===========
While some have argued that CAA is of little consequence in AD \[[@B28]\], the elaboration of the deposition that appears to occur under conditions of \"beneficial inflammation\" is on par with that seen in hereditary cerebral hemorrhage with angiopathy-Dutch type and is certainly a risk factor for devastating levels of hemorrhage. If such a response reflects a broad-acting realignment of cytokine profiles contingent upon immunization, it behooves careful consideration (and extensive animal testing) for any strategy for antibody-mediated reduction of Aβ in the AD brain.
List of abbreviations
=====================
AD: Alzheimer\'s disease
Aβ: amyloid β-peptide
CAA: cerebral amyloid angiopathy
IL-1, -6, -8: interleukin-1, -6, -8
NSAID: nonsteroidal antiinflammatory drug
TGFβ: transforming growth factor β
Competing interests
===================
The author(s) declare that they have no competing interests.
Acknowledgements
================
The author appreciates salary support from NIH funds 1R01 NS046439, 1R01 AG17498, 2P01AG12411-06A10003, and 5R01HD037989
|
PubMed Central
|
2024-06-05T03:55:51.935153
|
2005-1-11
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545943/",
"journal": "J Neuroinflammation. 2005 Jan 11; 2:2",
"authors": [
{
"first": "Steven W",
"last": "Barger"
}
]
}
|
PMC545944
|
Background
==========
Hepatic microcirculatory failure is a major prerequisite for the development of hepatocellular dysfunction in a number of conditions like trauma/hemorrhage, liver transplantation and systemic inflammation. In various inflammatory states, the degree of lethal hepatocyte necrosis can be predicted from the extent of hepatic microcirculatory failure \[[@B1]\], possibly via alterations in the mitochondrial redox state of the liver \[[@B2],[@B3]\]. Previously, our group has shown that the development of systemic inflammation was associated with a disturbance of the hepatic microcirculation, and a subsequent increase in hepatocellular damage \[[@B4],[@B5]\]. The causal mechanisms are not completely understood, but accumulating evidence suggests a dysregulation of stress-inducible vasoactive mediators like endothelins, nitric oxide synthase or heme oxygenase \[[@B6]\]. Moreover, modifications in effector cell function may also alter the response to those mediators \[[@B7]\]. Hepatic microcirculatory failures during various stresses are typically characterized by alterations in the distribution of perfusion, thereby resulting in a disparity between oxygen supply and demand. This impaired nutritive blood flow, together with reduced oxygen availability, decreases cellular high-energy phosphates leading to an early hepatocellular injury and dysfunction. Studies of tissue oxygenation focusing on the relationship between microcirculatory disturbances and oxygen transport dynamics may help to better elucidate the pathophysiological mechanisms involved.
Several methods have been reported in the past couple of years directly quantifying the oxygen distribution in tissues; however, their applicability in tissues, especially in small rodents like mice, is limited due to technical reasons. For instance, microelectrodes measure tissue pO~2~at specific points; but the technique is invasive and consumes oxygen. Electron paramagnetic resonance oximetry techniques or nuclear MRI approaches allow the detection of changes in tissue pO~2~; however, their resolution is too low \[[@B8]\]. A fluorescent membrane, developed by Itoh *et al*. \[[@B9]\] on the basis of an oxygen-quenched fluorescent dye allows the *in vivo*visualization of the tissue pO~2~. This technique allows the visualization of oxygen distribution on tissue surfaces, but this method comprised some technical limitations. The oxygen-sensitive membrane has to be used under gastight and watertight conditions during microscopy and the fluorescent membrane shows a photobleaching effect. Paxian *et al*. \[[@B10]\] recently demonstrated that the intravenous infusion of a special oxygen quenching dye allowed the visualization of the oxygen distribution on the liver surface using intravital videomicroscopy. The fluorescence of the dye was directly dependent on the tissue pO~2~. A disadvantage of this method, especially when used in small rodents like mice, is that it requires changing the continuous intravenous infusion rates of the dye to provide stable plasma concentrations. With mice (increasingly used as laboratory animals) there is a growing need for a method able to reliably detect tissue oxygenation or, at least, hemoglobin oxygen saturation (HbsO~2~) in capillaries of small animals.
The aim of the present study was to investigate whether the utility of a new and simple remission spectroscopy system allows reliable *in vivo*detection of liver sinusoidal HbsO~2~. In a mouse model of early systemic inflammation, we examined whether the detected changes in hepatic HbsO~2~correlated with the established method of NAD(P)H autofluorescence and hepatocellular injury.
Results
=======
Macrohemodynamics
-----------------
Consistent with previous reports \[[@B4],[@B11]\], mean arterial pressure (MAP) was significantly lower in animals after ischemia (I) and reperfusion (R) (3.0 h I/R and 6.0 h I/R) compared to sham animals, but remained normotensive (\> 80 mmHg) throughout the study. MAP did not differ between the I/R groups. Central venous pressure was not different (data not shown).
Blood gas analysis
------------------
The measurement of arterial blood gases carried out after the microscopy procedure showed normal oxygenation, a moderate acidosis, and adequate pCO~2~for all groups (Table [1](#T1){ref-type="table"}).
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Arterial blood gases.
:::
pO~2~(mmHg) pH pCO~2~(mmHg)
------------------------ ------------- ------------- --------------
Sham 128 (46) 7.29 (0.13) 35.8 (11.3)
3.0 h I/R 123 (49) 7.27 (0.15) 36.7 (10.8)
6.0 h I/R 116 (38) 7.26 (0.17) 36.8 (12.4)
6.0 h I/R+endothelin-1 119 (46) 7.26 (0.13) 36.9 (12.9)
Data expressed as Mean (SD); n = 7 for each group
:::
Hepatic sinusoidal hemoglobin oxygen saturation (HbsO~2~)
---------------------------------------------------------
Hepatic sinusoidal HbsO~2~of the different groups are shown in Figure [1](#F1){ref-type="fig"}. Animals treated with 3.0 h I/R have significant lower hepatic HbsO~2~values (56.2 (13.1)) when compared with sham (68.4 (14.1); *p*\< 0.01). No statistically significant differences were observed between 3.0 h I/R and 6.0 h I/R treated animals. However, an obvious shift of hepatic HbsO~2~towards a lower oxygenation was observed when compared with 3.0 h I/R treated animals. Animals treated with 6.0 h I/R and a continuous infusion of endothelin-1 (ET-1) showed significant reduced HbsO~2~values (44.8 (14.7)) when compared with 3.0 h I/R treated animals (56.2 (13.2); *p*\< 0.006). More than half of the measured data from these animals revealed HbsO~2~values lower than 50%. There was no apparent difference in the local tissue hemoglobin (Hb) content detected (data not shown).
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Sinusoidal haemoglobin oxygen saturation (HbsO~2~)**. At least 35 different observation points of the left liver lobe per animal were examined. The frequency distributions of all examined HbsO~2~values per group are shown.
:::
:::
Hepatic tissue redox status
---------------------------
Animals subjected to 3.0 h I/R revealed significantly higher NAD(P)H autofluorescence (141.6 (12.8)); therefore, a significant decline in hepatic tissue oxygenation was observed when compared with sham (100.0 (6.7)) (Figure [2](#F2){ref-type="fig"}). Three hours I/R treated animals failed to show a significant difference in NAD(P)H autofluorescence when compared with the 6.0 h I/R treated animals. Animals treated with 6.0 h I/R and a continuous infusion of ET-1 demonstrated significantly higher NAD(P)H autofluorescence (161.1 (13.8)) when compared to the 3.0 h I/R treated animals (141.6 (12.8)). There was a highly significant correlation found between NAD(P)H autofluorescence and hepatic HbsO~2~detected in the same animal (*p*\< 0.005; r^2^= 0.94), as depicted in Figure [3](#F3){ref-type="fig"}.
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Hepatic tissue redox status**. NAD(P)H autofluorescence, as a marker of the intracellular mitochondrial redox state, was examined using fluorescence intravital videomicroscopy with a filter set consisting of a 365 nm excitation and a 397 nm emission bandpass filter. The complete left liver lobe was systematically scanned and at least 15 different fields of view have been analysed. Fluorescence was densitometrically assessed and expressed as average intensity/liver acinus. \* *p*\< 0.001 vs. sham; \# *p*\< 0.01 vs. 3.0 h I/R; Data expressed as Mean + 2SD; n = 7 for each group.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Correlation between sinusoidal hemoglobin oxygen saturation (HbsO~2~) and tissue redox status**. The mean HbsO~2~values significantly correlated with the corresponding NAD(P)H autofluorescence (*p*\< 0.005; r^2^= 0.94). Data derived from 32 animals.
:::
:::
Hepatic tissue injury
---------------------
Serum alanine aminotransferase (ALT) and serum aspartate aminotransferase (AST) levels are summarized in Table [2](#T2){ref-type="table"}. When compared with sham animals, mice treated with 3.0 h I/R exhibited significantly higher levels of ALT and AST. No significant changes between 3.0 h I/R and 6.0 h I/R animals were detectable. When compared with 3.0 h I/R, mice treated with 6.0 h I/R and a continuous infusion of ET-1 showed significant higher ALT and AST levels. The results of labelling lethally injured hepatocytes with propidium iodide (PI) are shown in Figure [4](#F4){ref-type="fig"}. The 3.0 h I/R treated animals exhibited a significantly increase in lethally injured hepatocytes (120.4 (44.0)) compared with sham (25.7 (17.9)), whereas the 6.0 h I/R group had a significant higher number of dead hepatocytes (260.1 (52.7)) than the 3.0 h I/R treated animals. The treatment of 6.0 h I/R animals with a continuous ET-1 infusion further elevated the degree of lethally injured hepatocytes (361.8 (56.0)) when compared to the 6.0 h I/R treated animals. Regression analysis between lethally injured hepatocytes and hepatic HbsO~2~revealed a significant correlation (*p*\< 0.001; r^2^= 0.86), as shown in Figure [5](#F5){ref-type="fig"}.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Serum levels of alanine aminotransferase (ALT) and aspartate aminotransferase (AST).
:::
Sham 3.0 h I/R 6.0 h I/R 6.0 h I/R+endothelin-1
----------- ------------- ----------------- ----------------- ------------------------
ALT (U/L) 50.2 (16.6) 197.0 (40.4) \* 226.2 (38.5) \* 261.6 (37.8) \*\#\#
AST (U/L) 177 (34) 1825 (410) \*\# 2551 (616) \* 2856 (320) \*\#\#
Data expressed as Mean (SD); n = 7 for each group; \* p \< 0.001 vs. sham; \# p \< 0.02 vs. 6.0 h I/R; \#\# p \< 0.01 vs. 3.0 h I/R.
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Hepatic tissue injury**. Nuclei of lethaly injured hepatocytes were labelled *in vivo*with propidium iodide (PI). PI-labelled nuclei were quantified using fluorescence intravital videomicroscopy with a 510 to 560 nm excitation and an emission barrier filter greater than 590 nm. PI-labelled hepatocytes were expressed as number of cells/10^-1^mm^3^. \* *p*\< 0.001 vs. sham; \# *p*\< 0.001 vs. 3.0 h I/R; \#\# *p*\< 0.01 vs. 6.0 h I/R; Data expressed as Mean + 2SD; n = 7 for each group.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Correlation between sinusoidal hemoglobin oxygen saturation (HbsO~2~) and lethal hepatocyte injury**. There is a significant correlation between the mean HbsO~2~values and the corresponding amount of PI-labelled nuclei (*p*\< 0.001; r^2^= 0.87). Data derived from 32 animals.
:::
:::
Discussion
==========
In the present study, we demonstrate the utility of a remission spectroscopy system for the *in vivo*measurement of murine hepatic sinusoidal HbsO~2~that showed a significant correlation with the established method of NAD(P)H autofluorescence, as well as with the extent of hepatic tissue injury.
Oximetry relies on the detection of the spectral properties of oxygenated and reduced Hb. *In vitro*bench analysis capabilities have spurred the desire to accomplish accurate *in vivo*measurement through various techniques. The 1930\'s and 1940\'s were a particularly active period for oximetry advances culminating in the development of pulse oximeters in the 1970\'s \[[@B12]\]. Remission spectroscopy is based on the same principles of those oximeters, namely because they rely on the emission of white light and measure the total intensity of the backscattered light returned from the tissue. The intensity of the backscattered light is dependant on the amount and absorbance of the Hb in the tissue under observation. Oxygenated Hb has a different absorbance from that of deoxygenated Hb. The analysis of the backscattered light spectrum allows the determination of the HbsO~2~in the tissue. Previously, it has been shown that bilateral hindlimb I/R results in the deterioration of liver microcirculation \[[@B13]\]. Since the hepatic Hb content was not found to be different between groups in this study, the differences in the backscattered light spectra only represent differences in the HbsO~2~.
In the past, we have shown that bilateral hindlimb I/R results in a systemic inflammation with hepatic microcirculatory disturbances, in terms of reduced sinusoidal diameters and sinusoidal volumetric blood flow accompanied by elevated levels of sinusoidal leukocytes \[[@B4],[@B5]\]. These disturbances may result in an imbalance between oxygen supply and oxygen demand. Since the spectra, extinction coefficient, and quantum yield of NADH and NADPH are the same \[[@B14],[@B15]\], they are designated together as NAD(P)H -- this naturally occurring fluorophore transfers electrons to oxygen by means of an electron transport chain located at the inner membrane of mitochondria \[[@B16]\]. Under hypoxic conditions, with no oxygen available to accept electrons from cytochrome a, intracellular NAD(P)H accumulates. Unlike the oxidized form NAD^+^, NAD(P)H is highly fluorescent \[[@B17]\]. Therefore, we compared the changes in NAD(P)H autofluorescence, which reflect the extent of tissue hypoxia, with that of hepatic HbsO~2~obtained by the remission spectroscopy system under pathophysiological conditions. Whether induced by I/R or by the combination of I/R and infusion of ET-1, both analytical methods showed a decrease in hepatic oxygen supply, either as an elevation in NAD(P)H autofluorescence or as a diminution in hepatic HbsO~2~. The significant correlation between remission spectroscopy and NAD(P)H fluorescence indicates that after 3.0 h I/R, 6.0 h I/R and 6.0 h I/R+ET-1, hepatic oxygen supply was compromised. This is further emphasized by the statistical relationship found between hepatic HbsO~2~and the extent of subsequent hepatocyte death.
Both remission spectroscopy and NAD(P)H autofluorescence provide information on the metabolic state of the murine liver. Remission spectroscopy is directly dependent on the HbsO~2~in the sinusoids, whereas NAD(P)H autofluorescence depends upon the mitochondrial redox state and the activity of the mitochondrial electron transport chain. It was previously proposed that during systemic inflammation the NADH/NAD^+^redox potential may increase, and oxygen utilization may be altered \[[@B18]\]. The present study demonstrates a concomitant change in NAD(P)H autofluorescence and hepatic HbsO~2~. Obviously, the observed hypoxia did not occur through altered oxygen utilization, but rather through a reduced oxygen supply induced by sinusoidal microcirculatory disturbances. This corroborates our previous contention that the simultaneous use of remission spectroscopy, and that of NAD(P)H autofluorescence, provides additional information regarding the underlying pathophysiological mechanisms. That technical approach allows the correlation between disturbances in oxygen supply and those of oxygen utilization.
Conclusions
===========
There is a significant reduction in hepatic sinusoidal HbsO~2~during the early stages of systemic inflammation. In parallel, we detected an increasing NAD(P)H autofluorescence representing an intracellular inadequate oxygen supply. Both changes are accompanied by increasing markers of liver cell injury. Future therapeutic interventions should focus on the amelioration of sinusoidal HbsO~2~followed by an improvement in mitochondrial redox state. Remission spectroscopy represents a simple and reliable method for hepatic sinusoidal HbsO~2~determination in small rodents. In combination with NAD(P)H autofluorescence, it provides information on the oxygen distribution, the metabolic state and the mitochondrial redox potential within the hepatic tissue.
Methods
=======
Animals
-------
Male C57/BL6 mice (eight to ten weeks old, weighing 23.7 (11.1) g) were used for all experiments. The experimental protocols were in compliance with the guidelines of the Committee on the Care and Use of Laboratory Animals of the Institute of Laboratory Animal Resources, National Research Council as well as those of Germany. Animals were maintained under controlled conditions (22°C, 55% humidity and 12-hour day/night cycle) with free access to tap water and a standard laboratory chow.
Experimental protocol
---------------------
Mice (n = 7, for each group) were randomly assigned to either a Sham or a hindlimb ischemia/reperfusion (I/R) group. Animals of the I/R groups were treated with 60 minutes bilateral hindlimb ischemia induced by tightening a tourniquet above the greater trochanter of each leg while under anaesthesia. Sham animals were not subjected to ischemia, but remained anaesthetized for the same period of time. Tourniquets were removed just prior to recovery from anaesthesia. The animals were awake during the 3 hours (3.0 h I/R) or the 6 hours (6.0 h I/R) reperfusion periods, and re-anaesthetized for the intravital microscopy procedure.
To further induce liver microcirculatory disturbances and contribute towards a reduction in liver oxygen supply 6.0 h I/R, mice were further randomized to a group treated with a continuous infusion of ET-1 (70 pmol/min., i.v.) starting 15 minutes prior to microscopy. This dose of ET-1 was chosen because it produced alteration in the oxygen distribution, along with derangements in the hepatic tissue perfusion \[[@B19]\].
Surgical procedure
------------------
Animals received anaesthesia, by inhalation, for all procedures. As previously described \[[@B20]\], anaesthesia was performed using isoflurane (Forene, Abbott, Wiesbaden, Germany) in spontaneously breathing animals. The left carotid artery and the left jugular vein were cannulated under sterile conditions. The carotid artery cannula was used for the continuous measurement of systemic arterial blood pressure and heart rate, while central venous pressure was assessed via the jugular vein cannula. Throughout the experiment, normal saline was administered at a rate of 0.4 ml/hr to maintain normal mean arterial pressure. As formerly described \[[@B4]\], and for the realization of the intravital microscopy procedure in anaesthetized animals, a transverse subcostal incision was performed. Briefly, the ligament attachments from the liver to the diaphragm and to the abdominal wall were carefully released. For the evaluation of the hepatic microcirculation by intravital fluorescence microscopy, the animals were positioned on left lateral decubitus and the left liver lobe was exteriorized onto an adjustable stage. The liver surface was covered with a thin transparent film to avoid tissue drying and exposure to ambient oxygen. For equilibrium purposes, a pause of 10 minutes was allowed before data from microscopy and remission spectroscopy was collected. After microscopy, animals were killed by exsanguination, via the insertion of a cannula in the left femoral artery for the collection of arterial blood samples or via cardiac puncture.
Intravital microscopy
---------------------
Details of this technique have been described elsewhere \[[@B4],[@B21]\]. For observations of the liver microcirculation, we used a modified inverted Zeiss microscope (Axiovert 200, Carl Zeiss, Göttingen, Germany) equipped with different lenses (Achroplan × 10 NA 0.25 / × 20 NA 0.4 / × 40 NA 0.6). The image was captured using a 2/3\" charge-coupled device video camera (CV-M 300, Jai Corp., Kanagawa, Japan) and digitally recorded (JVC HM-DR10000EU D-VHS recorder) for off-line analysis. As previously described \[[@B22]\], NAD(P)H autofluorescence, as a marker of the mitochondrial redox state, was assessed using the 10x objective lens. The liver was examined using a filter set consisting of a 365 nm excitation and a 397 nm emission bandpass filter. NAD(P)H autofluorescence was recorded over the complete left liver lobe, allowing at least 15 different fields of view. Non-viable hepatocyte nuclei were labelled *in vivo*with an i.v. bolus of the vital dye PI (0.05 mg/100 g). As previously stated \[[@B21]\], PI-labelled nuclei were used to identify lethally injured hepatocytes. The fluorescent labelling of these nuclei was viewed using the 20x objective lens and a filter set with a 510 to 560 nm excitation and an emission barrier filter greater than 590 nm. Quantification of redox state and cell death was performed off-line by frame-by-frame analysis of the videotaped images using Meta Imaging Series Software (Ver. 6.1; Universal Imaging Corp., Downington, PA, USA). NAD(P)H fluorescence was densitometrically assessed and expressed as \"average intensity/liver acinus\". Gain, black level and enhancement settings were identical in all experiments. PI-labelled hepatocytes were expressed as number of cells/10^-1^mm^3^.
Remission spectroscopy
----------------------
Hepatic sinusoidal HbsO~2~was measured using the remission spectroscopy system Oxygen-to-See (O2C-ATS) supplied with the micro probe VM-3 (Lea Medizintechnik GmbH, Gießen, Germany). White light was continuously emitted via one channel of the micro probe light-guide and was continuously detected via another channel (channel diameter 70 μm). The backscattered light was analyzed in steps of 1 nm (500--650 nm). Each HbsO~2~value was defined by specific Hb spectra. The local tissue light absorbance depends on the total local tissue content of Hb. The local content of Hb was calculated from the local light absorbance and emission. The flexible VM-3 micro probe allowed the detection of oxygen saturation of the left liver lobe placed on the glass slide of the inverted microscope. A special clamping system fixed the micro probe close to the surface of the glass slide and permitted contact-free systematic scanning of the liver lobe (Figure [6](#F6){ref-type="fig"}). At least 35 different observation points per animal were randomly chosen and examined. Before each experiment, the white standard of the micro probe was calibrated according to the technical instructions of the manufacturer.
::: {#F6 .fig}
Figure 6
::: {.caption}
######
**Illustration of the experimental setup**. The flexible probe of the remission spectroscopy system was fixed on a special shaped clamp holder, which allowed the contact free scanning of the left liver lobe from the bottom side of the glass slide. The setup permitted systematic *in vivo*scanning of the liver sinusoidal HbsO~2~, without affecting the organ integrity.
:::
:::
Measurement of serum alanine aminotransferase (ALT) and aspartate aminotransferase (AST) levels
-----------------------------------------------------------------------------------------------
Blood was collected immediately after the microscopy procedure, via cardiac puncture. Blood samples were centrifuged at 6500 g, for 5 min, and the remaining serum analyzed, at 37°C, by means of standard enzymatic techniques.
Blood gas analyses
------------------
Blood samples for blood gas analyses were collected in heparinized syringes, via the insertion of a cannula in the left femoral artery, at the end of the microscopy procedure. The samples were immediately analyzed using the automated blood gas analyzing system Radiometer ABL 700 (Radiometer Medical Aps., Bronshoj, Denmark).
Statistical analysis
--------------------
Data in text and Tables is given as: Mean (SD). Statistical differences between groups and from baseline within each group were determined by ANOVA, followed by the Tukey post-hoc test. The Kolmogorov-Smirnov test was previously used to confirm the normal distribution of data. For checking the nature and extend of the relationship between two variables linear regression analysis was performed. All figures were generated with Sigma Plot (Ver. 8.0) and statistical analyses were performed using Sigma Stat software (Ver. 2.0; SPSS Inc.; München, Germany). Differences were considered significant for *p*\< 0.05.
Authors\' contributions
=======================
CW conceived the design of the study and conducted the laboratory experiments; RB drafted the manuscript and coordinated the study; AK assisted in technical questions. NR participated in design and coordination and OE participated in animal procedures and in drafting the paper. All authors approved and read the final manuscript.
Acknowledgments
===============
This work was supported by the Interdisziplinäre Zenrum für klinische Forschung (IZKF) of the Julius-Maximilians-Universität Würzburg (C. Wunder).
|
PubMed Central
|
2024-06-05T03:55:51.936156
|
2005-1-12
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545944/",
"journal": "Comp Hepatol. 2005 Jan 12; 4:1",
"authors": [
{
"first": "Christian",
"last": "Wunder"
},
{
"first": "Robert W",
"last": "Brock"
},
{
"first": "Alfons",
"last": "Krug"
},
{
"first": "Norbert",
"last": "Roewer"
},
{
"first": "Otto",
"last": "Eichelbrönner"
}
]
}
|
PMC545945
|
Background
==========
Metallothioneins (MTs) were firstly discovered by Margoses and Valle in 1957 \[[@B1]\] as cadmium (Cd) binding proteins. Later, Piscator \[[@B2]\] documented a marked increase of MT in Cd exposed rabbits, as a metal detoxification mechanism. MTs are a family of heavy metal binding proteins with a large degree of sequence homology that have been described in most vertebrate and invertebrate species. They are single-chain proteins, with molecular weight of approximately 6000 Da, characterized by a very high proportion of cysteine residues (30%), resulting in several high affinity Cd and/or zinc (Zn) binding sites \[[@B3]\]. There are two major isoforms, referred to as MT-I and MT-II \[[@B4]\], resolvable through ion exchange chromatography, that have closely related but distinct amino acid sequences and are distributed in most adult mammalian tissues. Recently, a further charge-separable MT isoform (MT-0) \[[@B4]\], and genes for two MT isoforms with restricted tissue distribution MT-III (brain neurons) \[[@B5],[@B6]\] and MT-IV (stratified epithelia) \[[@B7]\] have been described. The potential for wider tissue distribution of MT-III was suggested by recent studies demonstrating the presence of MT-III mRNA and protein in the adult and developing human kidney \[[@B8],[@B9]\].
MT-I and MT-II isoforms are usually expressed in low levels, but are inducible by a variety of metal ions, hormones, inflammatory cytokines and xenobiotics \[[@B10]-[@B12]\]. Induction of MTs is important in detoxification and metal ion homeostasis \[[@B9]\], in protection against reactive oxygen species \[[@B10]\] and in tissue regeneration \[[@B13]-[@B15]\]. MT expression deficiency implicated in carcinogenesis \[[@B16]\] and possible relation of MT over expression and resistance of tumors to anti-cancer therapy \[[@B17]\] has provided evidence of the importance of MT expression in cancer. MT over expression, detected immunohistochemically, has been described in a variety of human tumors, in relation to different stages of tumor development and progression \[[@B18]\]. The involvement of MT and Zn, in processes such as p53 gene activation and protein structure has been referred \[[@B16],[@B19]\]. There is evidence that some human tumors contain high levels of MT, nevertheless, the importance of MT expression in carcinogenic evolution and in patients\' survival is not yet fully understood.
In organs such as kidney, colon and liver, normally implicated in metal ions homeostasis, MT protein is apparently expressed. It would be of great scientific importance to elucidate the pattern of MT expression in tumors developed from these organs, since it could not only delineate the role of MT in carcinogenic transformation but could also provide prognostic information for patients\' outcome. The aim of the present study was to examine the expression of MT in human renal cell carcinoma (RCC) and to correlate the MT positivity, the pattern and extent of MT expression with tumor histological cell type and nuclear grade, pathologic stage and patients\' survival.
Patients and methods
====================
Forty three consecutive patients, 31 men and 12 women, who underwent nephrectomy for RCC comprised the group of our study. Their age ranged from 33 to 85 years (mean age 59.6 ± 11.1). Tumors were histologically classified as clear cell type in 32 cases, papillary type in 2 cases, chromophobe cell type in 4 cases and sarcomatoid type in 5 cases. The 8, 2, 16, 6 and 9 tumors were pathologically staged as T1 or T2N0M0, T2N+M0, T3N0M0, T3N+M0 and T4N+M+, respectively as per the TNM classification \[[@B20]\]. The grade of nuclear atypia according to Fuhrman Grading system \[[@B21]\] was: grade I in 11 cases, grade II in 15 cases, grade III in 6 cases, and grade IV in 11 cases. The patients were followed up from 2 up to 144 months (5 lost in follow up), mean 65.4 months, median 39 months.
Immunohistochemistry
--------------------
Sections of 5 μm thickness were deparaffinized in xylene and rehydrated in graded alcohol series. To remove the endogenous peroxidase activity, sections were treated with freshly prepared 0.3% (v/v) hydrogen peroxide in methanol in dark, for 30 min, at room temperature. Non-specific antibody binding was then blocked using normal rabbit serum (Dakopatts, Glostrup, Denmark) diluted 1:5 in phosphate buffered saline (PBS), for 20 minute. A mouse (IgG1k) monoclonal antibody that reacts with both human MT-I and -II isoforms (Zymed, San Francisco, California, USA) was used in this study. The sections were then incubated for 1 hour, at room temperature, with the primary antibody diluted 1:50 in PBS. After three washes with PBS, sections were incubated for 30 minute at room temperature with rabbit, peroxidase conjugated, anti-mouse serum (Dakopatts) diluted 1:200 in PBS and rinsed three times with PBS. Sections were then incubated with swine, peroxidase conjugated, anti-rabbit serum (Dakopatts) diluted 1:100 in PBS and rinsed three more times with PBS. The resultant immune peroxidase complexes were developed in 0.5% (v/v) 3,3\'-diaminobenzidine hydrochloride (DAB; Sigma, Saint Louis, MO, USA) in PBS containing 0.03% (v/v) hydrogen peroxide. Sections were counterstained with Harris\' hematoxylin and mounted in gelatin (Sigma). Control slides included in MT immunostaining procedure consisted of specific tissues previously shown to express MT (lung cancer) as positive controls, whereas the primary antibody was replaced by PBS in the case of negative controls.
Scoring system
--------------
The stained sections were independently assessed by the pathologists (A.K., S.T., E.M.) without prior knowledge of the clinical data as previously described \[[@B22]\]. Specimens were considered as \"positive\" for MT when more than 5% of tumor cells within the section were positively stained. The intensity of staining was graded as mild (+), moderate (++), and intense (+++). To further evaluate the importance of staining extent, cases were stratified into 3 groups according to the percentage of cells staining positive for MT: group A, 0--25%; group B, 26--50%; and group C, \>50% of MT positive cells. The pattern of MT staining was also characterized as cytoplasmic only, nuclear only, and both cytoplasmic and nuclear.
Statistical analysis
--------------------
Correlation between immunohistochemical data (MT intensity and extent of staining) and clinicopathological data (histologic cell type, nuclear grade, pathologic stage) was assessed using the Chi-square test. The association of intensity and extent of MT immunohistochemical staining with survival was determined by comparing Kaplan-Meier survival curves constructed for different patient groups, and were compared using the log-rank test.
Results
=======
Tubular cells but not glomeruli and interstitial cells of normal autologous renal tissue stained positive (both in the nucleus and the cytoplasm) for MT, although the intensity and extent varied significantly. Positive MT expression was prominent in 21 out of 43 cases (49%), while 22 out of 43 ones (51%) were MT negative. As far as the intensity of staining is concerned, low/moderate intensity was observed in 8 cases, while intense staining was evident in 13 cases. The pattern of positive MT immunostaining observed was either cytoplasmic (7 out of 21 cases, 33.3%) or cytoplasmic and nuclear (14 out of 21 cases, 66.6%). In certain cases of clear cell RCC membranous staining was also observed. Nuclear pattern of staining only, was not observed in any of the RCC cases examined (Figure [1](#F1){ref-type="fig"}). The extent of MT expression in a percentage of up to 25% of tumor cells (negative MT staining included) was observed in 31 out of 43 cases, in a percentage 25 up to 50% of tumor cells in 7 cases, and in a percentage of 50--75% of tumor cells in 5 cases.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Detection of MT expression by immunohistochemistry. Intense membranous, cytoplasmic and nuclear staining in a case of clear cell type RCC (x330).
:::
:::
There was no significant difference between MT intensity of staining and the histological types of RCC cases examined (χ^2^= 5.61, p = 0.46). No statistically significant differences were also observed between MT intensity of staining and stage (χ^2^= 9.24, p = 0.32), or patients\' survival (log rank test: 4.75, p = 0.09) in the cases of RCC examined (Figure [2](#F2){ref-type="fig"}). Statistically significant correlation was found between MT intensity of staining and histological grade, where higher intensity of staining was found in cases presenting high histological grade (χ^2^= 13.63, p = 0.03). However, no such correlation was found in the case of clear cell RCC (χ^2^= 4.75, p = 0.57).
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Cancer-specific survival of patients according to intensity of staining for MT (no staining, green line; mild/moderate staining, blue line; intense staining, red line). No statistically significant difference was detected (p = 0.09).
:::
:::
Non statistical correlation was found among MT extent of staining and histological types (χ^2^= 2.54, p = 0.86), stage (χ^2^= 7.12, p = 0.52) and grade (χ^2^= 6.24, p = 0.39) of RCC cases examined. Statistically significant inverse correlation was found between MT extent of staining and patients\' survival (log rank test: 6.59, p = 0.037) (Figure [3](#F3){ref-type="fig"}). Again, this was not observed in case all other histological types but clear cell RCC were excluded (log rank test 3.36, p = 0.186).
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Survival curves of patients according to MT extent of staining (0--25% of cells, green line; 26--50% of cells, blue line; \>50% of cells, red line). Statistically significant inverse correlation was found between extent of staining and patients\' survival (p = 0.037).
:::
:::
Discussion
==========
MT expression has been observed in different types of human tumors \[[@B18],[@B22]\], including neoplasias of the urogenital tract \[[@B23]-[@B28]\]. However, the biological mechanisms underlying MT over expression in tumors, or the consequences of this over expression are not currently well understood. MT is highly expressed in fetal life \[[@B10]\]; the reappearance of this fetal characteristic in tumors suggests its participation in cellular growth and differentiation \[[@B16],[@B18],[@B19]\].
Developing and adult kidneys consistently express MT-I and MT-II mRNA \[[@B29]\] and the corresponding protein \[[@B28]\], while the MT-III isoform is also expressed in developing renal tissue, adult proximal tubule and renal cell carcinoma cell lines \[[@B8],[@B9]\]. The MT-0 isoform is absent in adult kidneys but it can be found in nonneoplastic tissue from renal and transitional cell carcinoma \[[@B30]\]. The isoform specific expression of MT in RCC has not been so far investigated. Using a monoclonal antibody that reacts to both MT-I and MT-II, we demonstrated positive immunoreaction in 49% of our RCC cases, a percentage close to the 55.7% reported by Tüzel *et al*\[[@B24]\], using the same antibody. Zhang and Takenaka \[[@B24]\] used another commercially available monoclonal antibody with unspecified specificity towards different MT isoforms and found positive immunoreaction for the MT protein in 33% of their cases, while such report is not included in the study of Izawa *et al*\[[@B23]\] who used a polyclonal antibody prepared by their own laboratory.
The cytoplasm and nucleus of normal and malignant cells may both express MT but there is no conclusive data on the functional significance of their subcellular distribution. The majority of the cases in our study expressed MT in both cytoplasm and nucleus, the expression being cytoplasmic only in one third of the cases. Analogous pattern has been reported in other studies \[[@B23]-[@B25]\] while the membranous staining observed in some of our cases as well as in two previous series \[[@B24],[@B25]\] could be explained by the nature of clear cell carcinoma. Differential subcellular expression of MT may be related to either cell proliferation or the induction of apoptosis \[[@B31],[@B32]\]. Recently, Kondo *et al*\[[@B33]\] pointed out the importance of subcellular distribution of MT in drug resistant-prostatic cancer cells, in which the nuclear MT expression rather the cytoplasmic counterpart seems to predominantly confer resistance to cisplatin. On the other hand, nuclear pattern of MT immunostaining has already been correlated with cellular response to stress stimuli \[[@B34]\]. The function of MT to protect the cell from apoptosis could be an explanation for MT over expression observed in high grade cases of RCC \[[@B24]\], as MT is not etiologically correlated with the apoptotic process.
In certain tumors MT over expression has been associated with unfavorable prognostic characteristics such as advanced stage and poor differentiation \[[@B18]\]. In our study, MT intensity of staining did not correlate with stage while it showed an inverse correlation with histological grade. Our results are in accordance to those of other investigators \[[@B23],[@B25]\], who also found an inverse relationship between MT immunoreactivity and tumor grade. In contrast, Zhang and Takenaka \[[@B24]\] reported positive associations between MT expression and tumor grade. The inverse relationship between MT immunohistochemical expression and tumor grade may suggest a role of MT in cellular growth and differentiation, and reflect alterations of intracellular processes leading to a gradual decline of Zn storage and to the subsequent decrease in MT expression. Although no association was found between MT staining intensity and survival, the reduced extent of MT expression significantly correlated with prolonged survival as reported elsewhere \[[@B25]\]. The extent of MT expression may offer an additional, prognostic factor in patients suffering from RCC.
MT concentrations might also prove useful in predicting the efficacy of a particular cancer treatment protocols. Several types of transformed cells enriched for MT have been shown to exhibit greater resistance to chemotherapeutic agents \[[@B17]\]. In this context, the thiolate sulfur of the cysteine residues are thought to act as sacrificial scavengers for radicals and alkylyating agents \[[@B10]\]. Thus, MT might serve as a Cu and Zn reservoir, where their supply may negatively affect the growth of tumor cells. MT could also participate in tumorigenesis by sequestering and then donating these essential bivalent cations to proteins of tumor cells in order to meet metabolic requirements \[[@B35]\].
Conclusions
===========
The current data on the expression of MT in RCC cases examined emphasize the necessity to investigate larger numbers of patients with RCC comparing the staining profile of different MT isoforms with other clinico-pathological parameters and survival status of patients. Currently, it is unknown whether the presence of MT in renal carcinoma cells is related to the induction or inhibition of apoptosis or plays an active role in cell proliferation. Since cytokines may also induce MT expression and immunotherapy is the only, albeit with limited efficacy, available treatment for RCC, the intensity and extent of MT immunostaining should be studied in correlation to the immune status of RCC patients before and after immunotherapy. Thus, it will be possible to elucidate the potential role of MT in renal cell carcinogenesis, as well as its clinical usefulness as a tumor marker and as a tool for selecting patients for adjuvant immunotherapy.
|
PubMed Central
|
2024-06-05T03:55:51.938566
|
2005-1-17
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545945/",
"journal": "World J Surg Oncol. 2005 Jan 17; 3:5",
"authors": [
{
"first": "Dionisios",
"last": "Mitropoulos"
},
{
"first": "Aspasia",
"last": "Kyroudi-Voulgari"
},
{
"first": "Stamatis",
"last": "Theocharis"
},
{
"first": "Efraim",
"last": "Serafetinides"
},
{
"first": "Epaminondas",
"last": "Moraitis"
},
{
"first": "Anastasios",
"last": "Zervas"
},
{
"first": "Christos",
"last": "Kittas"
}
]
}
|
PMC545946
|
Background
==========
The United States Food and Drug Administration (FDA) has defined validation as \"confirmation by examination and provision of objective evidence that computer system specifications conform to user needs and intended uses, and that all requirements can be consistently fulfilled\" \[[@B1]\]. Validation is a process that begins with defining the requirements and ends with ensuring that the needs are being fulfilled consistently. Structured testing is only a part of the validation process. Interaction with researchers emphasized the notion that many people use validation and testing synonymously. To better illustrate the concepts of validation to these same researchers, the following mantra was developed:
Plan what you are going to do,
Do what you planned,
And
Say what you did.
As an introduction to the application of this mantra, consider the following. Clinical trial operations are detailed in numerous manuals and procedural documents with the protocol being the most notable. A clinical trial protocol serves as the master research plan. Great efforts will be employed, both by researchers and by external advisory boards, to ensure that the protocol is sound and is designed to support the clinical hypothesis. Conducting the protocol at clinical sites allows for the execution of the research methodology in a reproducible and orchestrated manner, and in the end, the final report or manuscript summarizes the work performed. If any one of these components is missing from the research plan, the scientific integrity of the research would be questioned or directly refuted. Associating systems validation to this well developed paradigm is at the heart of the \"Plan, Do, Say\" approach to validation. There are many good references available that detail the process of validation. Stokes\' two books are excellent resources for investigators new to validation \[[@B2],[@B3]\]. The FDA provides additional guidance to industry by publishing guidance documents that aid in the interpretation of federal regulations. National organizations such as the Drug Information Association routinely host short sessions addressing validation. Still, for many practicing biostatisticians or other researchers, the concepts of validation may not be clearly understood. Therefore, the intent of this manuscript is to offer guidance on what is a validated computerized system to these individuals and provide a common framework that will enable effective communication.
When revisiting the FDA definition of validation in the context of the scientific method, it becomes clear that \"confirmation by examination and provision of objective evidence that the computer system specifications conform to user needs and intended uses, and that all requirements can be consistently fulfilled\" is essentially applying the scientific method to the life cycle of computerized systems. The definition implies that a validated system is one in which processes are specific and clearly defined, reproducible and consistently applied, and result in a measurable and observable level of quality in the delivered product. Validation of applicable systems is viewed as the best way to ensure that research objectives can be consistently met.
All clinical trials need to collect and deliver results in an efficient and accurate manner while ensuring the integrity of the data. The CONSORT Statement provides guidance to researchers on the format and content of clinical trial reporting \[[@B4]\]. Computerized systems validation is integral to the fulfillment of the CONSORT Statement since a validation process for statistical programs will make the analysis understandable and reproducible by external reviewers. It is the responsibility of the computer system owner to ensure that the system is properly validated. For statistical reporting of biomedical data, the computer system owner is the senior trial biostatistician in consultation with sponsor and principal investigators.
In addition, validation should be considered as a part of the complete life cycle of a computerized system \[[@B2]\]. This life cycle includes the stages of planning, specifying, programming or purchasing, testing, documenting, operating, monitoring and modifying as necessary. The primary focus of regulatory inspection for computerized system validation is to obtain documented evidence to assure that any system is currently operating properly.
Discussion
==========
Just as the \"Plan, Do, Say\" mantra illustrates, there are three main validation deliverables \[[@B5]\]. First, a validation plan must be developed. The validation plan should include the general scope of the system including high-level design features, the personnel responsible for the validation, a timeline, and a summary of other supporting documentation that need to be addressed in order to use the system. The whole deliverable must be approved by management prior to its deployment \[[@B5]\]. Second, there should be documentation that the system has been designed as envisioned in the validation plan \[[@B5]\]. There should be at a minimum complete system specifications / requirements, a traceability matrix and test scripts included in this set of documentation. A test script should be written to ensure that the system requirements function as desired by measuring whether the system produces the expected result, and a traceability matrix cross references the system requirements to individual test script items. Finally, a report should be included with the validation deliverables. The whole deliverable must be approved by management and indicate that the system functions as required \[[@B5]\]. A report also is a place to discuss how deviations to the plan were addressed.
The scope and magnitude of the delieverable will vary from computerized system to computerized system, and it is the responsibility of the management team to decide upon the level. The following section provides an outline to serve as an aid to assembling the validation deliverables for statistical programs.
Statistical program checklist
-----------------------------
1\. Planning activities
\(a) Develop a validation plan
State how the validation process will be conducted, including personnel roles, test parameters, and decision points on what constitutes acceptable test results. Management prior to the development of the detailed specifications should approve the validation plan. The plan may be amended with time.
\(b) Utilize standard operating procedures (SOPs)
SOPs should be available to formalize the procedures used in the validation process as well as establish statistical programming standards. Incorporation of standards will aid in producing reproducible deliverables. The following is a partial list of applicable SOPs:
i\. Validation of statistical programs
ii\. Statistical programming standards
iii\. Statistical program archival: Outlines the necessary steps to archive the analysis program, data sets, and, if necessary, computer hardware so that the results may be reconfirmed at a future date.
\(c) Document training on SOPs
Written SOPs are not useful unless they are incorporated into practice. In order to do this, applicable individuals need to be orientated to the procedures. This orientation session should be documented for regulatory inspection.
\(d) Develop detailed specifications
i\. Data management plan
A. Annotated database structure
B. List of coding conventions for data
C. List of procedures used to process the data
D. Merging criteria (database keys)
E. System environment: analysis package (with version), system hardware, input data structure, output data structure, long-term storage environment
ii\. Analysis objectives
The analysis objectives may vary according to the application; however, for the primary clinical trial report, the protocol or a statistical analysis plan may be sufficient to detail the requirements of the analysis.
\(e) Develop a test plan and/or test script
i\. Mock tables
ii\. Expected results of tests
iii\. Programming standards
iv\. Testing procedures
2\. Execution of the plan
\(a) Retain raw test results
Record individual pass/fail for each step outlined in the test script. A pass occurs when the observed result matches the expected result.
\(b) Note variances and deviations to the test plan
\(c) Document location, time and individuals involved in the testing process
3\. Summary report
\(a) Summarize validation process
\(b) Summarize the variances and deviations
\(c) Summarize test results and provide interpretation when necessary
\(d) Approve by management
Application to statistical programs
-----------------------------------
The definition of a computerized system encompasses both the hardware and software. For analytical applications, the key components of the system also will include the individual programs written to perform the analysis. The use of validated macros aides in reducing the burden introduced by the validation process for an individual system. For example, suppose for the clinical reporting of a trial\'s results, a table is desired that reports the mean and standard deviation or count and percentage for all putative covariates. A SAS macro \[[@B6]\] could be written to compile and export the data from SAS to a word-processing compatible format. This macro would be a candidate for validation since it manipulates raw data, performs calculations, and modifies output from standard SAS output. However, once this macro has been developed, a significant savings in the time required to produce (and verify) publication-ready tables could be possible. The remainder of the discussion highlights some of the key steps in validating this macro.
The validation of the macro begins in the planning stage. For the macro to \"conform to user needs\", the needs need to be clearly identified. The following may describe its user requirements:
1\. Export a word-processing compatible table consisting of three columns;
2\. The table columns must be appropriately labeled;
3\. Column 1 must be the labeled \"Characteristic\";
4\. Column 2 must be the labeled \"Control Group\";
5\. Column 3 must be the labeled \"Intervention Group\";
6\. For categorical covariates, calculate the count and percentage for each level of the variable;
7\. For levels of a categorical variable, the table should indent the formatted description; and
8\. For interval-scaled continuous covariates, calculate the mean and standard deviation.
This set of requirements would then be discussed from a technical perspective. Issues such has how to best calculate the summary statistics could be a topic of discussion. The use of PROC TABULATE and the Output Delivery System (ODS) \[[@B6]\] could be a likely candidate for the summary. Next, discussion pertaining to the specification of covariates and the concatenation of the results is needed. In the process, potential macro parameters would need to be identified. All key decisions are incorporated into the validation plan. This plan is reviewed by the computer system owner and the validation system sponsor.
Once this plan has been approved, a revision log should be kept to ensure traceability of changes that may occur during development.
Once the macro has been developed, the validation system sponsor should work either independently or in conjunction with appropriate staff to review the programming and develop test cases. The test cases are assembled into a test that specifies the system input and expected output. To ensure all user requirements are addressed, a traceability matrix can be utilized. A simple traceability matrix for this validation exercise would list each of the requirements and cross reference the applicable test.
The test script is then implemented under the direction of the validation system sponsor. Documentation of the event should be recorded so that it can be compiled into the validation report. Any discrepancies between the observed and expected results need to be addressed, and, if applicable, documentation on the corrective actions employed to resolve the issue(s) needs to be compiled into the validation deliverables. A revision log of programming changes is desirable; however, in practice, this may prove difficult to document completely. Once the testing has been completed, the computer system owner and validation system sponsor can determine what additional steps are required prior to release. For SAS macros, it is critical to document the parameters, default values of the parameters and system dependencies, then, perhaps include illustrative examples of the macro\'s use. Finally, a written statement by the computer system owner and the validation system sponsor should be issued stating that the macro has been validated and has been deemed acceptable for use.
The above example illustrates a full-scale implementation of a validation process. However, it is acknowledged that individuals and organizations must decide the level of validation that is required. In many cases, more traditional approaches of program verification (code review, double programming, etc.) may be sufficient. The importance of the \"Plan, Do, Say\" mantra is that procedures used for validation are specified and operated on.
Summary
=======
The \"Plan, Do, Say\" approach presented addresses the key deliverables an auditor expects to see in validated systems. By organizing validation activities into a plan, an action or a summary, one begins to gain efficiencies through proper planning. However, individual researchers and support organizations need to evaluate their particular needs for validation and develop a set of procedures to support their needs. The intent of this article is not to provide an ironclad approach to validation that will ultimately meet all needs. However, this design for validation has been conveyed successfully to a variety of audiences. The analogy of the scientific method helps break the technology and nomenclature barrier associated with a more computer science-driven approach to systems validation by associating the importance of validation to the process of designing a research plan.
Competing interests
===================
The author(s) declare that they have no competing interests.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2288/5/3/prepub>
Acknowledgements
================
This work was partially supported by the National Institutes of Health grant DA13727. The author thanks the Systems Development Workgroup of the NIDA Clinical Trials Network for review and discussion of the \"Plan, Do, Say\" approach. In addition, the author thanks Knut Wittkowski for his critical review and helpful comments.
|
PubMed Central
|
2024-06-05T03:55:51.940069
|
2005-1-13
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545946/",
"journal": "BMC Med Res Methodol. 2005 Jan 13; 5:3",
"authors": [
{
"first": "Rickey E",
"last": "Carter"
}
]
}
|
PMC545947
|
Background
==========
Breast cancer is relatively rare in women less than 35 years of age, with this group accounting for less than 4% of the total number of breast cancer cases diagnosed in Western countries \[[@B1],[@B2]\]. Despite the disease being relatively uncommon, it has a severe negative effect on the patients and their families.
It remains controversial whether young age at diagnosis is an adverse prognostic factor in primary breast cancer. While some studies have found that younger patients have worse clinical outcomes than older patients \[[@B3]-[@B7]\], others report younger patients have a more favorable prognosis, or that there is no relationship between outcome and age \[[@B8]-[@B10]\]. Various explanations have been given for these conflicting results, including small numbers of patients comprising the study population, differences in patient selection criteria and differences in the age groupings used in the analyses. Moreover, it has long been debated whether breast cancer diagnosed at a young age is a clinically and etiologically distinct disease from breast cancer diagnosed later in life. Some researchers reported that tumors in younger women were of higher grade, higher proliferation fraction, had more vascular invasion, and expressed fewer estrogen and progesterone receptors compared to tumors in older women \[[@B11]-[@B14]\]. It is important for clinicians to clarify the existing controversy as to whether aggressive treatment for young women with breast cancer is justified.
Breast cancer is the most frequent cancer in Korean women and its incidence is increasing \[[@B15]\]. Breast cancer in young Korean women is a serious problem, with the proportion of young age-onset breast cancer much higher than in western countries. According to the 2002 annual report of the Korean central cancer registry, breast cancers that developed before the age of 35 comprised 9.5% of all breast cancers \[[@B16]\].
The aim of the present study was to retrospectively investigate clinicopathological characteristics and prognosis in a large, ethnically homogeneous group of young breast cancer patients (less than 35 years old) treated with the same strategy at a single institution.
Methods
=======
A retrospective review was performed of the medical records of all consecutive primary invasive breast cancer patients (not including phyllodes tumor) undergoing curative surgery in the Department of Surgery, Seoul National University Hospital between January 1990 and December 1999. Patients with distant metastasis detected at the time of surgery or within 4 months of surgery were excluded. Those patients whose surgical margins were positive for malignancy were also excluded. Patients\' records were reviewed for the following: age of onset, family history of breast cancer in 1st or 2nd degree relatives, histological type of cancer, tumor size in pathology reviews, axillary lymph node status, histological grade (HG: Scarff-Bloom-Richardson classification), nuclear grade (NG: Black\'s nuclear grade), type of surgical procedure and adjuvant therapy administered. Disease was staged according to the American Joint Committee of Cancer (AJCC) system \[[@B17]\]. The \'younger\' group was defined as patients less than 35 years old at the time of breast cancer diagnosis. Expression of immunohistochemical tumor markers such as estrogen receptor (ER), progesterone receptor (PR) and c-erbB2 were determined in over 70% of cases. The expression was determined in assays performed immediately after surgery for each case. The primary antibodies for ER (DAKO, Glostrup, Denmark), PR (DAKO, Glostrup, Denmark) and c-erbB2 (Novocastra, Newcastle, UK) have been previously characterized. A cut-off value of 10% or more positively stained cells out of total cells in ten high-power fields was used in the classification of ER, PR and c-erbB2 expression levels.
Statistical analysis
--------------------
The χ^2^test (Pearson statistic) was used to determine the differences in clinicopathological features between the two groups of patients. The follow-up duration was calculated from the date of diagnosis until the date of death or last contact. The disease-free survival was the time between diagnosis and confirmation of disease recurrence. The overall survival was the time between diagnosis and death as a result of any cause, regardless of recurrence events. Survival estimates were computed using the Kaplan-Meier method \[[@B18]\] and the differences between survival times were assessed by means of the log rank test \[[@B19]\]. Multivariate analyses were carried out using Cox\'s proportional hazards model \[[@B20]\]. All statistical analyses were carried out using the SPSS (version 10.0) software package (Chicago, IL, USA).
Results
=======
A total of 2040 patients were eligible for this study, of which 256 (12.5%) were aged \<35 at the time of diagnosis. The median follow-up was 74 months. Histology showed the incidence of medullary carcinoma was significantly higher than ductal carcinoma in the younger group (p = 0.018). There was a significantly higher incidence of nuclear grade 3 in the younger group than in the older group (p = 0.015). Axillary lymph node status, the most prominent prognostic factor in breast cancer, was not significantly different between the two age groups. Also, neither the family history of breast cancer in 1^st^or 2^nd^degree relatives, T stage, histological grade, c-erbB2 expression, nor ER or PR status were different between the two groups (Table [1](#T1){ref-type="table"}). Frequencies of ER and PR positivity were low, and frequency of c-erbB2 positivity was high, in both age groups compared to frequencies reported in western populations and other Asian studies \[[@B21]-[@B23]\].
The proportion of breast-conserving surgery compared to mastectomy was similar in both groups. Axillary lymph node dissection, at least to the first Berg level \[[@B24]\], was performed in 250 (97.7%) younger patients and 1735 (97.3%) older patients. No sentinel lymph node procedure was performed. Adjuvant radiation therapy was administered to patients who underwent breast-conserving surgery and after mastectomy in patients who had four or more positive lymph nodes or a tumor \>5 cm in diameter. Adjuvant chemotherapy was administered to 68.0% of younger and 58.7% of older patients. The most common chemotherapy regimen was a combination of cyclophosphamide + methotrexate + 5-FU (CMF) for 6 cycles or anthracycline containing regimen (AC). In terms of hormone therapy, tamoxifen was used for as long as 5 years after completion of surgery and adjuvant therapy. We classified a patient as tamoxifen treatment group if she got tamoxifen through more than a year before recurrence. The proportion of tamoxifen treated patients was significantly lower in young age group. Neither the type of surgery nor the postoperative adjuvant choemotherapy was significantly different between the two age groups (Table [2](#T2){ref-type="table"}).
Younger patients had a worse disease free survival (greater probability of recurrence) at all time periods (Fig [1A](#F1){ref-type="fig"}; *p*\< 0.001). At 5 years, the actuarial recurrence rate for patients \<35 years old was 30.4% as compared with 18.7% for older patients. This difference persisted at 10 years, at which time the actuarial recurrence rates were 40.1% and 28.6%, respectively. Overall survival among younger patients was significantly worse than for older patients (Fig [1B](#F1){ref-type="fig"}; *p*= 0.002). The 5-year survival rate was 80.0% for patients aged \<35 years as compared with 88.5% for older patients. Stratified analysis according to axillary lymph node status was performed for disease-free survival. In lymph node-negative patients there was no significant difference in disease-free survival between the two age groups (Fig [2A](#F2){ref-type="fig"}; *p*= 0.223). However, in lymph node-positive patients, disease-free survival was significantly worse in younger patients (Fig [2B](#F2){ref-type="fig"}; *p*\< 0.001).
In multivariate analysis, young age (\<35 years) remained a significant predictor of recurrence when entered into a model containing all potential demographic, pathologic and immunohistochemical variables (Table [3](#T3){ref-type="table"}. Hazard Ratio (HR), 1.7; 95% confidence interval, 1.1--2.6; *p*= 0.010). However, young age was not a significant independent predictor of overall survival in the same Cox model (table not shown. HR, 1.4; *p*= 0.242). Because hormone therapy was done more frequently in older patients than young age group (Table [2](#T2){ref-type="table"}.), we made another multivariate model involving hormone therapy in patients with ER positive and/or PR positive cancer to address the effect of hormone therapy on the prognostic significance of young age. In this analysis, young age was still an independent significant prognostic factor while hormone therapy showed borderline significance (Table [4](#T4){ref-type="table"}.).
Discussion
==========
Our results showed that operable young breast cancer patients (\<35 years old) have a worse prognosis than older patients in terms of both overall survival and recurrence. The difference in disease-free survival was clear in patients with axillary lymph node metastasis, but was not observed in lymph node-negative patients. Even after controlling for differences in distribution of potential prognostic factors, young age remained a significant predictor of recurrence.
The present findings support previous reports showing that women diagnosed with breast cancer at a younger age have a poorer prognosis compared with their older counterparts \[[@B3]-[@B7]\]. However, those reports suffered from limitations including a small younger patient sample size, a study period spanning too many years during which treatments changed, lack of information about pathological and protein markers, and a heterogeneous case population in terms of ethnicity and treatment strategy.
To our knowledge, the present study is the largest to directly compare the prognosis of younger (\<35) breast cancer patients with that of their older counterparts. Moreover, the data in this study were generated from patients of the same ethnicity undergoing treatment at a single institution under the same contemporary strategy of surgery and adjuvant therapy over a relatively short time period (10 years). In addition, this study included a multivariate analysis of the difference in distribution of potential prognostic markers between the two age groups.
The biomarker results in the present study are different to those reported for other populations, including ethnically-related Asian patients. In a study of 1052 Chinese breast cancer patients in Hong Kong, 53% and 61.6% of pre- and postmenopausal women were ER-positive, respectively, and 51.5% and 46.2% were PR-positive, respectively \[[@B21]\]. Those figures are higher than the figures reported in the present study. A recent study of Japanese patients showed 62.2% were ER-positive and only 17.2% were c-erbB2-positive \[[@B22]\]. Merchant et al. found c-erbB2 expression in 30% and 24% of British and Japanese breast cancer patients, respectively \[[@B23]\]. In contrast, Choi et al. reported significant differences in c-erbB2 expression between Korean and white patients (47.5 vs. 15.8%, respectively) using immunohistochemistry and fluorescence in situ hybridization (FISH) techniques. These data suggest c-erbB2 expression may be related to race \[[@B25]\].
Although not a major focus of this study, we found PR and c-erbB2 expression were significant independent predictors of disease recurrence. Currently, the role of PR status as a prognostic factor is not clear, with some evidence to suggest it is useful \[[@B26],[@B27]\] and other evidence to the contrary \[[@B28]\]. As for c-erbB2, its prognostic importance is also controversial. A large number of studies have been published, some reporting positive results and others reporting negative results \[[@B29]-[@B31]\]. The prognostic significance of PR and c-erbB2 in this data set can be investigated further as an independent analysis later.
The St. Gallen Consensus Conferences in 1998 and 2001 concluded that age under 35 was a high risk factor for relapse in node-negative breast cancer patients \[[@B32],[@B33]\]. Kroman et al. \[[@B34]\] reported that young women with low-risk breast carcinoma who did not receive adjuvant treatment had a significantly increased risk of death from the disease. Furthermore, Fowble et al. \[[@B4]\] reported that young women with early stage breast cancer, especially those with lymph node-negative disease, had a relatively worse prognosis than older counterpart. In the present study, although no significant difference was observed between the two age groups in lymph node-negative patients, the pattern of survival curves implied younger patients may have a worse prognosis. It may be that a study with a larger case size and a longer follow-up duration would provide enough statistical power to show a significant difference in prognosis for node-negative patients.
It has been suggested that younger women with breast cancer have a poorer prognosis because they present with later stage disease due to either physician or patient delay in diagnosis. However, in this study, no significant difference was found between the two age groups in terms of tumor size or lymph node status. Moreover, multivariate analysis indicated that young age is an independent negative prognostic factor. This issue of delayed diagnosis is not conclusive now and should be elucidated further in subsequent studies.
One possible limitation of this study is that the control group was heterogeneous and contained a mixture of premenopausal and postmenopausal patients. Adami et al. showed complex pattern of survival as a function of age at diagnosis of breast cancer \[[@B12]\]. However, as shown in Table [4](#T4){ref-type="table"}, young age remained an independent prognostic factor in multivariate analysis even after patients aged over 50 years were excluded from the control group (*p*= 0.008). Although we did not have full information on each patient\'s menopausal status at the time of diagnosis, this result suggests that patients under 35 years have even worse prognosis than relatively \"less young age\" premenopausal patients.
Up to 15--30% of women aged less than 35 years diagnosed with breast cancer are likely to have germ-line BRCA1 or BRCA2 mutations \[[@B35],[@B36]\]. Although we did not investigate BRCA gene mutations in all patients in the present study, we performed BRCA1 and BRCA2 gene mutation scanning in 22 patients who had two or more breast cancer patients in their 1^st^degree relatives. We found four BRCA2 and one BRCA1 mutations that are thought to be disease-causing (data not shown). Only one of these 5 patients was less than 35 years old at the time of cancer development. It is known that young breast cancer patients are more likely to have an inherited form of the disease \[[@B37]\]. However, the current study showed there was no significant difference in the family history of breast cancer between the two age groups. In the recent report by Choi et al. the prevalence of BRCA1 and BRCA2 mutations in Korean women with breast cancer at a young age (\<40) was as high as western population. However, most of the BRCA-associated patients had no family history of breast and/or ovarian cancer. That is, the penetrance appears to be low. They suggested that there may be different genetic and etiologic factors affecting transmission and penetrance of the BRCA genes in Korean patients with breast cancer diagnosed at a young age \[[@B38]\].
Although most current breast cancer investigators agree that young age is an adverse prognostic factor for breast cancer, there have been few studies designed to elucidate the molecular or genetic differences associated with young age breast cancer. Recent CGH (Comparative Genomic Hybridization) analysis suggested that alterations in specific regions on chromosomes might be responsible for the poor outcome of early onset breast cancer \[[@B39]\]. Future research must be focused on this area in order to confirm the characteristics of young age-onset breast cancer at the molecular level.
Conclusions
===========
These results show that operable young breast cancer patients (\<35 years old) have a worse prognosis than older patients in terms of both overall survival and recurrence. Even after controlling for differences in distribution of potential prognostic factors, young age is an independent predictor of recurrence. The underlying biology of young age breast cancer needs to be elucidated and development of tailored treatment for this patient population is crucial.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
WH selected cases, reviewed medical records, analyze data, and drafted the manuscript. SWK participated in the data collection and input. IAP carried out the pathological diagnosis and immunohistochemistry. DK performed the statistical analysis. SWK, YKY, SKO, KJC, and DYN conceived of the study, and participated in its design and coordination. All authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2407/4/82/prepub>
Acknowledgements
================
This work was supported by a grant from the Korea Health 21 R&D Project. Ministry of Health & Welfare, R.O.K (01-PJ3-PG6-01GN07-0004).
We thank former surgical fellows of Seoul National University for their contribution to breast cancer patient database.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
\(A) Disease-free survival curves for women \<35 vs. ≥ 35 years old. Patients younger than 35 had significantly worse outcomes than their older counterparts (*p*\< 0.001). (B) Overall survival curves showing patients younger than 35 had significantly worse outcomes than their older counterparts (*p*= 0.002).
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
\(A) Disease-free survival curves in axillary lymph node-negative patients showing no significant difference between the two age groups (*p*= 0.223). (B) Disease-free survival curves in axillary lymph node-positive patients showing patients younger than 35 had significantly worse outcomes than their older counterparts (*p*\< 0.001).
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Clinicopathological characteristics of younger (\<35) and older age groups
:::
Characteristics Age \<35 (%) (n = 256) Age ≥ 35 (%) (n = 1784) *P*value
----------------------- ------------------------ ------------------------- ----------
Age
20--25 9 (3.5)
26--30 56 (21.9)
31--35 191 (74.6)
36--40 318 (17.8)
41--50 791 (44.3)
51--60 474 (26.6)
60--70 165 (9.2)
71- 36 (2.0)
Family history^a^ 22 (8.7) 135 (7.6) 0.511
Histology
Ductal 238 (93.0) 1624 (91.0)
Lobular 2 (0.8) 30 (1.7)
Medullary 10 (3.9) 29 (1.6) 0.018
others 6 (2.3) 101 (5.7)
T stage
T1 99 (38.7) 770 (43.2)
T2 129 (50.4) 855 (47.9)
T3--4 28 (10.9) 159 (8.9) 0.126
Lymph node metastasis
Negative 138 (53.9) 1063 (59.6)
Positive 118 (46.1) 721 (40.4) 0.084
Histological grade
1--2 78 (58.6) 525 (60.0)
3 55 (41.4) 350 (40.0) 0.767
Nuclear grade
1--2 113 (52.6) 901 (61.3)
3 102 (47.4) 570 (38.7) 0.015
ER
Positive 97 (47.1) 907 (51.8)
Negative 109 (52.9) 843 (48.2) 0.198
PR
Positive 73 (36.7) 708 (43.5)
Negative 126 (63.3) 921 (56.5) 0.068
c-erbB2
Negative 94 (52.8) 647 (46.4)
Positive 84 (47.2) 748 (53.6) 0.106
^a^Patients who have 1st or 2nd degree relatives with breast cancer
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Treatment characteristics
:::
Characteristics Age \<35 (%) Age ≥ 35 (%) *P*value
------------------- ---------------- ----------------- ----------
Surgery
Mastectomy 211 (82.4) 1482 (83.1) 0.796
Conservation 45 (17.6) 302 (16.9)
Chemotherapy
Lymph node (-) 65/138 (47.1) 421/1063 (39.6) 0.281
Lymph node (+) 109/118 (92.4) 627/721 (87.0) 0.674
Radiation therapy 73 (28.5) 462 (25.9) 0.373
Hormone therapy 54 (21.1) 492 (27.6) 0.028
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Multivariate analysis for predictors of recurrence based on the Cox proportional hazards regression model
:::
Variables HR 95% Confidence interval *p*
---------------------- ----- ------------------------- ---------
Age \<35 years 1.7 1.14--2.61 0.010
Tumor size ≥ 2 cm 2.0 1.30--3.07 0.002
Lymph node-positive 3.8 2.64--5.67 \<0.001
Nuclear grade 3 1.4 0.90--2.39 0.124
Histological grade 3 0.9 0.54--1.49 0.675
ER 1.1 0.77--1.63 0.549
PR 2.1 1.41--3.19 0.001
c-erbB~2~ 1.4 1.04--2.05 0.030
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Multivariate analysis for predictors of recurrence involving hormone therapy in ER(+) and/or PR(+) patients
:::
Variables HR 95% Confidence interval *p*
----------------------- ----- ------------------------- ---------
Age \<35 years 2.1 1.14--4.20 0.018
Tumor size ≥ 2 cm 2.2 1.39--3.62 0.001
Lymph node-positive 3.3 1.81--6.14 \<0.001
Nuclear grade 3 1.0 0.45--2.31 0.961
Histological grade 3 1.2 0.55--2.85 0.576
Hormone therapy (yes) 1.6 0.93--2.77 0.086
c-erbB~2~ 1.4 0.81--2.42 0.228
:::
::: {#T5 .table-wrap}
Table 5
::: {.caption}
######
Multivariate analysis for predictors of recurrence based on the Cox proportional hazards regression model after exclusion of patients \>50 years old.
:::
Variables HR 95% Confidence interval *p*
---------------------- ----- ------------------------- ---------
Age \<35 years 1.8 1.17--2.81 0.008
Tumor size ≥ 2 cm 1.7 1.06--2.84 0.028
Lymph node-positive 4.2 2.64--6.82 \<0.001
Nuclear grade 3 1.0 0.60--1.91 0.810
Histological grade 3 1.0 0.58--1.89 0.883
ER 1.1 0.71--1.78 0.625
PR 2.0 1.27--3.37 0.004
c-erbB~2~ 1.4 0.97--2.21 0.069
:::
|
PubMed Central
|
2024-06-05T03:55:51.941731
|
2004-11-17
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545947/",
"journal": "BMC Cancer. 2004 Nov 17; 4:82",
"authors": [
{
"first": "Wonshik",
"last": "Han"
},
{
"first": "Seok Won",
"last": "Kim"
},
{
"first": "In",
"last": "Ae Park"
},
{
"first": "Daehee",
"last": "Kang"
},
{
"first": "Sung-Won",
"last": "Kim"
},
{
"first": "Yeo-Kyu",
"last": "Youn"
},
{
"first": "Seung Keun",
"last": "Oh"
},
{
"first": "Kuk Jin",
"last": "Choe"
},
{
"first": "Dong-Young",
"last": "Noh"
}
]
}
|
PMC545948
|
Background
==========
Biological databases are key computational tools used daily by biologists. Such a large number of biological databases have been developed for biology that the Nucleic Acids Research Journal has published an annual database issue since 1996. From the point of view of the user, these resources are most useful when they are regularly updated and when they provide user-friendly ways to browse, search and view information. These user needs are generally recognized as important requirements by the designers and developers of biological databases. To cope with these requirements, bioinformaticians who develop the biological databases have typically responded by developing increasingly customized software to manage the data and the information (e.g., \[[@B1]-[@B4]\]). In doing so, and to facilitate the software development effort needed to create a biological database, bioinformaticians have used a variety of information technologies. These technologies range from the ones that make it possible to create dynamic web applications (e.g., Common Gateway Interface/CGI, Java Servlets, web application frameworks), to technologies needed to store the data and the information in a persistent manner (i.e., text files \[[@B5],[@B6]\], relational databases \[[@B7]\], frame representation systems \[[@B8]-[@B10]\], object-oriented databases \[[@B11]\]).
In this article, we report on our experience with the Java Data Objects persistence technology and take a critical view at the advantages and drawbacks of this emerging Java persistence standard for the development of advanced biological databases. We have ported the SigPath information management system (see below) to the JDO API and have defined an application-specific benchmark. We used this benchmark to evaluate the performance of two JDO implementations that target either a relational or an object database backend. This article summarizes the performance results that we obtained, announces the availability of the SigPath JDO benchmark (available under the GPL license), and identifies areas where the JDO API could be refined to facilitate portability and scalability of applications.
Data persistence
----------------
Biological databases are built with software that executes on computers. Most biological databases are of a size that could fit entirely in the central memory of modern computers. However, because computers may need to be shutdown for maintenance -- or may crash inadvertently -- data for a given database cannot be kept in computer memory for the life of a biological database. This problem is not specific to biological databases so that a variety of data persistence approaches and technologies are available. The key role of these technologies is to guarantee that data persists safely between the invocations of the programs that may modify the data.
The pros and cons- of persistence technologies for biological databases
-----------------------------------------------------------------------
Data can be stored in **text files with limited structure**and important information can be stored in unstructured text files expressed in English. Unstructured flat-files do not help perform large-scale analyses, structured queries or integrate data across multiple sources, all of which are important requirements for biological databases. Therefore, unstructured files are now widely recognized throughout the field as inadequate for the management of biological information.
**Highly structured data formats**, such as ASN.1 \[[@B12]\] and more recently XML, are a more favored alternative. They can support structured queries, large scale analyses and data federation. Structured file formats, however, do not provide support for concurrent manipulation of the information by several users (e.g., several curators interacting with a submission tool to input new data about one protein in the database). As such, they are adequate for data exchange among systems, but not for concurrent access. Since the file format offers no support for synchronization, locking or complex domain-dependent data validation rules (for XML, XML Schemas are limited to simple validation rules), using structured data formats for biological information storage forces system developers to implement these services explicitly as a layer between the business code and the data storage. For instance, since XML Schemas are not capable of validating XML data with respect to information outside of the scope of the file being validated (such as data in other files or in a database), developers must implement custom validation code. XML Schema focus on syntactic validation, while most applications require semantic validation \[[@B13]\].
**Database management systems**(DBMS) have been historically developed to abstract the services (such as synchronization, business domain constraints) needed by systems that need to support large number of users accessing a shared storage of data. A few types of DBMS exist that differ in the way they represent data. Relational DBMS represent data as tables that contain rows and columns of various types, while Object DBMS support the concept of object classes and object instances directly.
**Relational DBMS**such as Postgres, MySQL or Oracle have been used to store biological information in many laboratories, including ours \[[@B14],[@B15]\]. A short introduction to using RDBMS for biological information storage was recently offered in \[[@B16]\]. Briefly, complex relationships among elements of information are stored in relational databases by expressing relations among records in several tables. The technology is useful for a variety of biological databases, where the mapping between the biological data and the relational data model is simple.
However, the technology has two major drawbacks for advanced biological databases. The first problem is because of a mismatch between the object-oriented programming style and the relational data model. Advanced biological databases often require programs that manipulate tens or hundreds of object classes. Data in the instances of these classes needs to be made persistent, and this requires writing mapping code. The mapping code takes a graph of objects and transfers the data in this graph into records in the various tables of the relational database. Mapping code needs to be developed for the reciprocal operation, from the relational records to the object instance graph. Depending on the complexity of the relationships among objects in the graph, the development of the mapping code may represent a significant part of the code developed for the overall database.
**Object DBMS**such as O2 \[[@B17]\] and FastObjects have been developed to eliminate the need to write mapping code, and to store objects directly in native form in the database. This approach was reported to offer substantial performance improvements and reduced development and maintenance costs for data organized in an object graph with complex relationships.
Java Data Objects Technology
----------------------------
The Java Data Objects Technology (JDO) is a Java application programming interface (API). This API was developed as a Java Specification Request \[[@B18]\] to offer: \"a standard way to store Java objects persistently in transactional data stores\..., a standard way to treat relational database data as Java objects, and a standard way to define transactional semantics associated with those objects.\"
JDO appears as an attractive technology for the development of biological databases for the following main reasons:
1\. It is designed to offer portability across a wide range of transactional stores or database backends, from open-source relational databases to native object oriented databases.
2\. It transparently handles object persistence when relational or object persistence backends are used (the developer only manipulates objects and classes and does not need to write mapping code).
3\. JDO also handles persistence transparently for object oriented databases, where mapping code is not needed.
4\. It is a Java technology that integrates seamlessly with web application servers (e.g., Tomcat, JBoss, etc.) often used to create the web front-ends of a biological database.
A critical evaluation of the JDO technology
-------------------------------------------
Given the stated advantages of the technology we decided to carry out a critical evaluation of JDO to determine if the technology can routinely be used for the development of advanced biological databases. Our evaluation focused on the following questions:
**Portability:**Is JDO a mature API that can guarantee portability of the application across database backends?
**Performance:**If portability is achieved, how do relational and pure object oriented backends compare in term of performance?
**Biological database specific requirements:**Do complex biological databases have specific requirements that JDO 1.0.1 does not address?
To answer these questions, we have ported a biological information management system (the SigPath system, see below) to the JDO 1.0.1 API. (The SigPath system was originally implemented with the ODMG API \[[@B19]\]). In the first step of the port, we compiled the new code with the FastObjects JDO implementation \[[@B20]\] FastObjects JDO is an implementation of the JDO API that connects to the native FastObjects object database. In a second step, we have adapted the existing code to support exchanging the JDO implementation and database backend between the FastObjects implementation and the Solarmetric Kodo implementation of JDO \[[@B21]\]. Kodo is an implementation of JDO 1.0.1 that connects to a variety of relational database backends. The aim of the second development was to modify the code to make it possible to switch from FastObjects JDO to Kodo JDO by changing a configuration property, and then simply recompiling. Our aim was to create a code-base that was fully portable from a relational database backend to an object-oriented database backend to address the portability question.
The SigPath Information Management System
=========================================
SigPath is an open-source project aimed to develop an Information Management System (IMS) to foster modeling and simulation of cell signaling pathways and networks (see the SigPath project \[[@B22]\]) \[[@B23]\]. The SigPath IMS appears to the end-user as a web application that provides search, browsing and visualization capabilities. The project home page provides tutorials that explain how the system is typically used.
Most traditional biological databases focus on one type of database entry (e.g., gene, mRNA, protein, protein motif, protein domain, etc.) and store information in database entries. This approach has been very useful to create detailed catalogs of biological parts and is a critical and essential element of the bioinformatics resources that support modern biological research. However, certain integrative studies, such as systems biology and modeling and simulation of biochemical pathways call for databases that integrate several types of information.
The SigPath IMS is an example of an advanced biological database that encodes information through a number of information types and a set of relationships among them. Figure [1](#F1){ref-type="fig"} illustrates how SigPath encodes information about a biochemical reaction: the reaction is represented as a graph of object instances.
Introduction to the SigPath ontology/database schema
----------------------------------------------------
A fragment of the SigPath ontology is given on Figure [2](#F2){ref-type="fig"} as a UML diagram. The description of the complete set of persistent classes used in SigPath is given on the project web site (\[[@B22]\], see the \"for developers\" tab).
The SigPath system supports several types of biological information, ranging from information to represent small molecules and proteins to the interactions between these molecules. The main information types supported by SigPath are listed on Table [1](#T1){ref-type="table"}. In SigPath, information is represented in an object-oriented manner, with information types often associated with classes. The SigPath object-oriented database schema was adapted from the EcoCyc ontology \[[@B9]\]. Several classes presented on Table [1](#T1){ref-type="table"} have an equivalent in the EcoCyc ontology. In the rest of this article, we will use the terms ontology and JDO database schema indistinctively, as they represent very similar concepts: a class in the object-oriented schema of SigPath is equivalent to a frame in the EcoCyc ontology, and an attribute of an object class is similar to the slot of a frame.
This multiplicity of SigPath information types and the variety of relationships among them makes it important to clearly define what type of information can be represented by the system (i.e., the set of object graphs that could potentially be created and stored in the database). This information is formalized in the SigPath ontology. This ontology is implemented in a JDO database schema. (In the SigPath system, the set of allowed object graphs may be further reduced by adding semantic constraints to the validation mechanism used during information submission). The JDO schema consists of the set of SigPath Java classes that are persistent and of meta-data about these classes. Meta-data is expressed in JDO files and provides information about the classes that cannot be expressed directly in the Java language, for instance, type of the elements for the collection field of the persistent classes. Figure [5](#F5){ref-type="fig"} shows a small JDO file and illustrates the type of information that it provides. A thorough presentation of the structure of JDO files is given in \[[@B24]\], vendor-specific extensions are documented in each JDO implementation.
The SigPath code base has specific characteristic that make it a useful resource for evaluating JDO technology:
• SigPath is an open-source project released under the GPL, so that the benchmark code is freely available for others to study, reproduce our results, or extend the benchmark to other JDO implementations or database backends.
• SigPath is both a web-based application and a batch-oriented application.
• The SigPath code-base includes unit tests \[[@B25]\] that help verify that the application behaves correctly against two different database backends.
• The SigPath system provides varied use cases that exercise different behaviors of the database backend and JDO implementation (see use cases below).
In the next section, we present the methods that we used to evaluate JDO technology for the creation of advanced biological databases.
Results
=======
This section describes the results of the SigPath JDO benchmark and addresses the portability and performance questions described in the introduction.
SigPath: porting from one JDO implementation to another
-------------------------------------------------------
We modified the FastObjects JDO version of SigPath to compile indifferently with the FastObjects and Kodo implementations of JDO. The modifications that we had to make to the project were (i) modifications to the JDO file, (ii) modifications of the code base and (iii) modification of the code base and application data.
### Modifications to the JDO file
The Kodo enhancer tool performs stricter semantic validations on the JDO files than the FastObjects enhancer. Modifications needed to pass the validation tests were:
1\. Added persistence-capable-superclass attribute to classes that have a persistence capable superclass. This attribute is optional for FastObjects, which uses Java reflection by default, but is strictly required by the Kodo implementation (in agreement with the JDO specification).
2\. Removed all interfaces from the JDO file. Enhancing with Kodo failed when interfaces were listed in this file. Since FastObjects requires interfaces to be listed as persistent classes, the SigPath build script conditionally includes such statements in the JDO file when FastObjects is configured. The JDO specification does not mention interfaces, so that the behaviour of JDO implementation is left undefined.
3\. (As a result of 2.) Replaced references to interfaces with references to implementation (e.g., replaced Protein by ProteinImpl) throughout the JDO file.
4\. Added collection element types to all persistent collections. FastObjects requires the type to be specified when the collection is used in a query. Kodo requires the type to be defined for each collection, otherwise Kodo will try to serialize the collection and store it as a binary object. If the persistent class is not serializable, this mechanism will fail. Therefore, for this benchmark, we explicitly defined the collection types for each collection.
5\. Removed field definitions from sub-classes when they refer to fields of a super-class. (e.g., the field \"reactions\" in Model was specified twice in the Model sub-class and in the Pathway super-class). Removing these duplicate declarations is consistent with the JDO specification.
Furthermore, the Kodo enhancer expects classes to be listed in the JDO file in a specific order. The enhancer fails if a class appears in the JDO file before another class that the first class references. Therefore, we reordered the class definitions in the JDO file. (We verified that this is no longer an issue with version 3.0 of Kodo, but keep this description as other JDO implementations may share the same limitation).
Finally, we added Kodo extensions to the JDO file to create indexes on the tables that were used extensively in queries. All changes to the JDO file were consistent with the JDO specification. Index tuning was performed by running the boot and test part of the benchmark and the small molecule import with various indexes choices.
### Modifications to the code base
We modified the code base to work-around a limitation of the Kodo implementation. With Kodo, instances of classes that contain java.lang.Object fields are made persistent with the object field stored as a BLOB in the database. Storing objects as BLOBs puts strong limitations on their use. For instance, it is impractical to query for these objects by their fields (e.g., querying directly for a User instance by the id of the user is not possible if the instance is stored as a BLOB). Storing such fields as BLOBS was therefore not acceptable for certain types of persistence objects, and we implemented the work-around shown on Figure [3](#F3){ref-type="fig"}.
Another code modification was required to work-around a problem with the database backend that did not handle appropriately empty strings (""). The database backend used for this benchmark stored empty strings as null. Reading these strings back from the database resulted in null being obtained from Kodo instead of empty strings. This resulted in several unexpected NullPointerException being thrown during the JUnit tests. Figure [4](#F4){ref-type="fig"} illustrates the approach that we used to work-around this problem.
### Modification to the code base and benchmark/application data
Finally, we had to modify the code base to put a limit on the length of long strings. Using a relational database backends imposes to define the maximum length of each string attribute defined in the persistent classes of the application. For instance, a limit must be set on the name attribute of the SigPathEntityImpl shown on Figure [2](#F2){ref-type="fig"}. We initially used the default maximum length for all fields and found that certain fields could be longer than this limit when running the test and the benchmark. For instance, description fields of ProteinImpl are imported into SigPath from the DE line of SwissProt and TrEMBL entries. Some entries have long descriptions (that can exceed 1,000 characters). To test the impact of this limit on the code of the application, we arbitrarily choose to use a maximum length of 1,000 characters. We excluded from the benchmark input data proteins and small molecules that had aliases or descriptions longer than 1,000 characters, and other entries that would exceed any String field limit. This was done to make sure that the same input data was used for both the FastObjects and the JDO relational benchmarks.
Performance measurements
------------------------
A brief summary of the performance measurements obtained with the SigPath benchmark is given in Table [3](#T3){ref-type="table"}. The table presents time measurements for each use case of the benchmark. The measurements are listed both for the FastObjects JDO implementation (columns marked FO) and for the Kodo implementation. Columns marked %FO/KODO indicate the percentage of the time running the benchmark with FastObjects takes compared to running the benchmark with Kodo. The last column of the table FO/KODO CV indicates the coefficient of variation of the total time across four independent measurements. Small values of CV (1--5%) indicate consistency between the four measurements. However, some use cases showed higher variations (10,11,12,36%), so we report as well the minimum value of the four time measurements for both FO and KODO (in columns marked Min).
The raw data used for the calculation of these performance measures is provided in the supplementary material and on the SigPath JDO benchmark pages. These pages also provide the logs from which the raw data has been collected.
Discussion
==========
Portability
-----------
Our port of SigPath confirms that JDO greatly facilitates the porting of a bioinformatics application from one database backend to another. However, we report here several modifications that we had to make to the SigPath system to achieve this level of portability. This suggests that there is a need to develop JDO compliance tests that could be used to test that a specific implementation of a JDO-aware database is really compliant with the standard. This test suite would validate that JDO enhancers accept correct JDO files and correctly reject JDO files that break the specification. The differences in the interpretation of JDO files that we noticed between FastObjects and Kodo (see Results section) practically limit the portability of JDO applications. This article has presented techniques that can be used to work around these limitations until a JDO compliance test is developed and used. We note that the work arounds that we described may be specific to the two JDO implementations that we tested, and that other work arounds may be needed to achieve portability with other JDO compliant backends.
Surprisingly, we found that an outstanding portability problem is in the way the different JDO back-ends store long strings of characters. While the FastObjects backend put no limitation of the length of long strings, the relational back-end used with Kodo limited the length of long strings to 4,000 characters. This limit had to be chosen and set for each persistent string field used in the application (when the default value was not appropriate). Although 4,000 characters may appear a large limit, it is likely to be reached in bioinformatics application either with textual or with sequence data. When this happens, the application will have to be re-engineered to work around the fixed limit. A work-around could be to use a data type that does not have a length limitation, but these data types also have other limitations (for instance, usually indexes cannot be used on those fields). Whichever solution is chosen, this issue must be considered early during the design of the application. It would be useful if the JDO standard offered a mechanism for the application developers to specify which string length their application requires to function properly with JDO backends. Each enhancer could then check that the application is requesting a maximum string length that is compatible with the database backend and fail early if it does not. (As of now, these types of error will most likely be detected when testing the application.)
Performance
-----------
The SigPath benchmark provides precise measurements of the performance of one biological database application against two JDO compliant database backends. The measurements were performed on use cases that are typical of the activities needed to develop the code of SigPath and to deploy a production SigPath system. (Table [2](#T2){ref-type="table"}. indicates which use cases belong to our software development process and which belong to administrative and curation tasks that we need to carry out to prepare a new release of SigPath).
As shown in Table [3](#T3){ref-type="table"}, performance varies widely with the type of use case, but is overall significantly better with the object database backend. Use cases that perform batch loading of protein information into the database benefited the most from using the native object database FastObjects backend (with loading of data sometimes completed five times faster than with Kodo and a commercial relational backend). An exception to this trend is the SM Import use case, which shows only a 3% performance difference. This use case reads an XML file and loads small molecules into the database. To do so, it checks for each molecule that the accession code of the new molecule does not already exist in the database (this is an error condition that would interrupt the import). Since the database does not contain small molecules, the query used to perform this check returns an empty set for each molecule of the import. It appears that this specific operation is slower with the object-oriented backend that we have used for the benchmark.
The last column of Table [3](#T3){ref-type="table"} indicates the coefficient of variation (CV) of the individual measurements (among four independent executions). The CV values indicate that the performance of certain steps vary significantly from execution to execution. These differences are likely to be caused by the caching behavior of the database server and of the operating system. Caching can occur because we have not restarted the database server between the benchmark runs, or rebooted the machines. These differences may also be caused to a lesser extent by variations in what operating processes were active and the amount of IO wait at the time that the specific use case was executed. We have tried to reduce such causes of variability (see methods) but have not attempted to eliminate them completely (e.g., setup an isolated database server and disable all interactive use of the server). Our rationale is that such variability, including caching, is representative of a typical production system. Given the CV, the average execution time may not be an accurate representation for some use cases, so we report also the minimum execution time across the four independent executions of the benchmark.
The benchmark provides an indication of how well an object-oriented database system performs compared to a relational database backend for the SigPath use cases. A known limitation of benchmarks is that the performance measure that they provide are specific to the application tested, and may not generalize well to other use cases. Also, the SigPath benchmark does not cover multithreaded/multiclient operations. Results may vary depending on the chosen locking strategy and the number of clients/threads running parallel. Given these caveats, however, this benchmark indicates that, for most of the SigPath use cases, the performance of the SigPath system is significantly improved when using a native object database system. Particularities of the SigPath benchmark that may correlate with this result are (i) the complexity of the database schema (75 persistent classes) and (ii) the number of connections that exist among instances of these various classes.
Finally, these results and our distribution of the SigPath benchmark source code can help vendors diagnose performance problems with their implementation of the JDO implementation, and provide users with an objective measure of the performance a given JDO implementation, for similar types of applications.
Biological database specific requirements
-----------------------------------------
During our evaluation of the JDO technology, we have noted that two common requirements of advanced biological databases are currently not being addressed by JDO.
### Support for interfaces
When designing a biological database schema, it is often useful to express that one class shares the properties of two or more classes. In a programming language such as C++ this can be represented as multiple inheritance (one class inherits from two parents) while in a programming language such as Java, this concept is represented with interfaces (one class implements two interfaces). In the context of JDO, consider the class diagram shown on Figure [6](#F6){ref-type="fig"}. The diagram illustrates one way to represent the phosphorylated forms of protein and small molecules. On this diagram, one has represented a \"Phosphorylated\" interface which is implemented by PhosphoProtein and PhosphoSmallMolecule. While this way to represent biological information is useful, JDO does currently not specify the handling of interfaces, so that the design shown on Figure [6](#F6){ref-type="fig"} can not be implemented with JDO in a portable way. (This design would work with FastObjects, but not with Kodo.) The JDO specification should clarify if interfaces must be supported by for a JDO implementation to be compliant with the standard.
### Support for large number of objects
Biological databases often need to manage large number of objects (e.g., large number of proteins, small molecules, etc.). For instance, SigPath stores information about several hundred of thousands of proteins. We found that JDO 1.0.1 lacks some features that would facilitate writing scalable applications.
An example is that the JDO standard does not provide a scalable way to determine the number of persistent instances of a given class. The JDO compliant way to accomplish this operation is to obtain a reference to a collection of instances of this class (using a JDO Extent), and to call the size() method on this collection. Since the collection must first be obtained from the database server before the size() method can be invoked, this procedure takes a time proportional to the number of instances of this class. Most database backends store the number of instances of a certain class in the database and can determine this information in a constant time, so a standard way to obtain this information from a JDO implementation would be very helpful (SigPath can use either a pure JDO extent sizing method, or a vendor-specific method through an extension mechanism implemented in the source code, so that performance can be compared).
A second example is that JDO 1.0.1 does not provide support for queries that return large result sets. Under standard JDO 1.0.1 behavior, traversing a persistent collection (by accessing each element of the collection in turn) brings the entire contents of that collection into memory. This behavior is appropriate for small result sets. However, there are cases where the complete set of instances returned by a query cannot be processed within a single transaction. This occurs for instance when all the results returned by a query do not fit in the fixed memory limit allocated to the Java Virtual Machine. In such cases, it may be necessary to obtain the result of a query in chunks of a certain number of records/instances (for instance 1,000 or 10,000 instances at a time), and process them in independent transactions. Upon transaction commit, memory associated with a chunk is released and can be used to process the next chunk. Implementing this type of scalable processing in an efficient manner usually requires making modifications both in the persistent class of elements in the result set and in the query filter. The class of the elements in the result set can be modified to add an instance identifier that can be used both to sort the instances and to select only those within the current processing chunk. The query filter can be modified to add a clause that selects only instances of the next chunk, based on the identifier introduced in each element. An alternative is to provide an API call to notify the JDO implementation that instances which have been processed can be evicted from memory. Since several vendors already have their own extensions to provide scalability feature, it would be useful for JDO to support such features through a standard API.
Conclusions
===========
Here, we have shown that it is possible to develop a bioinformatics database that can be reconfigured automatically and recompiled to run either against a relational database backend or against an object database backend. The key advantage of this added flexibility is that the bioinformatics database becomes portable with respect to the database backend. This has important implications for the development of open-source bioinformatics databases. In such projects, usually more than one laboratory contributes to developing the software of a specific biological database. Therefore, it is useful if each laboratory can choose a database backend for development and deployment, yet contribute to the project in a shared code base.
The Java Data Objects standard offers the productivity gains of transparent object persistence, and a fine-grained object persistence model useful to represent many biological concepts. We discussed why JDO can appear as an attractive option for the development of advanced biological databases and the type of problems that we encountered when implementing and deploying a biological database against two different JDO implementations. The future JDO standard (JDO 2.0) should address some of the issues that we discussed in this article (e.g., support for interfaces, or for large result sets). When JDO 2 implementations become available, we expect that JDO technology will have a significant impact on the design of high-performance biological databases that need to represent and manage complex biological information types.
Methods
=======
Benchmark use cases
-------------------
To address the performance question, we have developed benchmark use cases. The benchmark use cases were designed to be representative of performance that one would observe when either (i) developing the software of the SigPath system or (ii) preparing a new release of the SigPath IMS (includes loading the database with information from other databases). Our benchmark thus considers both the development and the production stages of the life-cycle of the application.
The use cases, or benchmark steps and a summary of their purpose in the context of the SigPath project are listed on Table [2](#T2){ref-type="table"}.
The boot and test steps make it possible for the SigPath developers to create a sample database and test that important functionalities of the application are working satisfactorily.
**boot**-- The boot step compiles the sources of the project, enhances the JDO persistent classes (a program, called a JDO enhancer, transforms Java class files into persistent classes and allow them to interact with the JDO implementation), creates an empty database and imports information into the database. Importing this information involves parsing an XML file that contains the information, validating this file against the SigPath information exchange XML Schema, validating against additional semantic rules that cannot be expressed with XML Schemas (database lookups are used during this step to connect new instances to instances previously submitted in the database, if needed), and saving new persistent instances to the database. The boot sample data is designed to contain at least one instance of each type of information that can be stored in the SigPath IMS.
**test**-- The test step runs JUnit tests against the data that was imported during the boot step. The JUnit tests assert that information stored in the database corresponds to the information in the boot XML file. For instance, the tests check that the number of persistent instances matches the number found in the boot import file, but also that specific elements of information have been saved accurately. Furthermore, the tests assert various semantic properties of the application and database access code, running queries against the database, navigating through objects, creating new persistent instances or deleting them, etc. The complete set of operations performed in the test is fully described in the source code for the JUnit tests (see edu.mssm.crover.sigpath.test package, and specifically the class MasterTest).
**small molecule import**-- This step creates SmallMoleculeImpl persistent instances (implementation of the SmallMolecule interface shown on Figure [2](#F2){ref-type="fig"}). The data used to create these molecules is a modified form of the NCI open database. Only small molecules that have a name, description, aliases and SMILES representation are imported from NCI Open (the total number of molecules read is 237,771, and the total number of molecules loaded into the database is 45,229). These data are imported and stored in the attributes of SmallMoleculeImpl (most fields: name, description and aliases are inherited from SigPathEntityImpl).
**protein import**-- This series of steps creates ProteinImpl persistent instances (implementation of the Protein interface, shown on Figure [2](#F2){ref-type="fig"}). The data to load these proteins is obtained from a simplified XML format created from SwissProt and TrEMBL data files with SwissKnife \[[@B26]\]. The exact list of files imported is given on Table [2](#T2){ref-type="table"}.
**full text indexer**-- The SigPath system offers users the ability to search entities by keywords. This step builds an inverted full text index \[[@B27]\] that is used by the web application to accelerate keyword-based queries. An inverted full text index links each keyword that occurs in text strings of a SigPath entity (e.g., name, description, aliases) to the SigPathEntity instance that contains the keyword. This step creates 465,679 Keyword instances that link to a total of 345,133 SigPath entities (small molecules or proteins).
**XML import**-- This step is similar to the loading of SigPath information in the boot target, but loads information obtained from the online version of SigPath (this benchmark used the information as of October 14^th^2003). For this benchmark, XML import instantiates 14 SmallMoleculeImpl, 8 ProteinImpl, 121 ComplexImpl, 77 modified chemicals (SmallMoleculeImpl or ProteinImpl), 92 ConcentrationMeasurement, 165 ReactionImpl, 75 EnzymaticReactionImpl, 23 Model, 3 Pathway and 27 PendingReviews.
Benchmark procedure
-------------------
The benchmarks were run as described on the SigPath Project web site (\[??\] see the \"JDO benchmark\" tab). Each benchmark (FastObjects or Solarmetric Kodo with a relational database) was run on a two Xeon 3GHz processor machine with hyper-threading on and 6 Gb of memory. The machine was running Red Hat Advanced Server Linux 2.4.21-4.0.1.ELsmp, and was used both as database server and database client (to minimize the impact of the network on performances). No significant other processes were running on the benchmark machine while the tests were executed. We benchmarked FastObjects t7 server version 9.0.7.185 and Kodo JDO version 2.5.3. Each benchmark was run four times to average the effect of variability in the computational environment that may not have been controlled by our benchmark procedure. The results report the coefficient of variations (mean divided by the standard deviation) of the total running time for each use case and this helps point out cases when the computational environment had an effect on measured times. We believe that these variations are common in a production environment and report the average total running time as well as the minimum total running time for each use case.
Authors\' contributions
=======================
Marko Srdanovic and US implemented significant components of the FastObjects and Kodo JDO ports. Marko Srdanovic collected benchmark data at WMC and US collected similar data at FastObjects. Michael Schwieger and FC designed the study and contributed to the JDO ports. FC drafted the manuscript. All authors read and approved the final manuscript.
Acknowledgements
================
We thank Lucy Skrabanek for assistance with the Kodo implementation port and David Guinther for a technology grant to FC that made this benchmark possible.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Instance graph for representation of a reaction between three chemicals (e.g., *A*↔ *B*+ *C*)**. Some of the relationships that support navigation from reactions to chemicals are shown, as well as links between chemicals. The three chemicals are produced by the \"human\" organism. Quantitative kinetic parameters are also shown for the reaction (backward and forward rate of the reaction). The figure illustrates how a graph of object instances is used to represent biological information corresponding to a biochemical reaction.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Fragment of the UML diagram for the SigPath ontology/database schema**. The Figure shows how instances of certain classes are related and how the relationships among classes can be used to represent information. The link between Reaction and Chemical expresses the information that reactions transforms chemicals, and that chemical can be substrate or products of reactions. The classes and attribute marked in orange are specific to the SigPath ontology and have no equivalent in the EcoCyc ontology.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Work-around for classes that contain an object field.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Work around for classes that have String getter and setters, when empty strings can be made persistent.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Example of JDO file**. This file is used to define the persistent classes that are used in SigPath to represent end-users. Four persistent classes are shown: Address, Affiliation, User and UserRole. The \<field\> element can be used to refer to specific fields of persistent classes (such as the username field of class User on this example). The userRoles field is described to be a collection that contains elements of type UserRole. Elements called \<extension\> make it possible to provide vendor specific directives, such as to define indices on a persistent field.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
**Illustration of the use of interfaces to express is-a relationships among biological concepts**. Circles represent interfaces while boxes represent classes. An arrow from one class to an interface indicate that the class implement the interface, and this relationship can be used to indicate that instances of the class have the properties described by one or several interfaces.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Selected information types supported in SigPath.
:::
**Classes** **Description**
-------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
SmallMolecule A small molecule, such as ATP
Complex A binding complex between two or more molecules
Protein A protein molecule
Chemical A small molecule, protein or complex
Reaction Reaction between molecules: has substrates and products
EnzymaticReaction Associates a reaction to the enzyme that catalyze it
Unit Represents a unit, such as mol/l or /sec
Parameter Associates a value to its unit
ConcentrationMeasurement Associates a concentration to a molecule and the conditions of the measurement
Pathway A set of reactions and enzymatic reactions
Model A quantitative biochemical model, a set of reactions (or enzymatic reactions), initial concentrations for molecules in the model, rates of the reaction, kinetic parameters of the enzymatic reactions.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Overview of the use-cases in the SigPath Benchmark
:::
**Project stage** **Step** **Utility for the SigPath project**
---------------------- --------------------- --------------------------------------------------------------------------------------------------------------------------
development boot Loads a sample set of information that can be used to run JUnit tests and for interactive testing of the web application
development test Performs JUnit tests to verify that key aspects of the system are working correctly
pre-release SM import Import small molecule information (names, aliases SMILES from NCI Open)
pre-release Full Text Indexer 1 Builds an inverted full text index for small molecules imported in previous step
pre-release mam.xml Imports TREMBL mam.dat proteins
pre-release rod.xml Imports TREMBL hum.dat proteins
pre-release hum.xml Imports TREMBL inv.dat proteins
pre-release inv.xml Imports TREMBL inv.dat proteins
pre-release vrt.xml Imports TREMBL vrt.dat proteins
pre-release sprot41\_1.xml import SwissProt 41.dat, part I
pre-release sprot41\_2.xml import SwissProt 41.dat, part II
pre-release Full Text Indexer 2 Builds an inverted full text index for the proteins imported from TrEMBL and SwissProt
pre-release XML Import Imports data from another SigPath database (data is encoded in the SigPath XML exchange format)
traversal simulation TestGetPathways Navigates through the Pathway instances. Used to simulate user navigation on the web site.
traversal simulation TestBenchmark Another benchmark-specific performance test.
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
SigPath benchmark measurement summary.
:::
**[Minimum of four measurements]{.underline}**
------------------------------------------------ -------------- -------------- -------------------- --------------------
**FO Min** **KODO Min** Δ **(KODO -- FO)** **%(FO/KODO) Min**
**Junit Tests** 7809 14962 7153 **52.19%**
**Boot** 3026 3976 950 **76.11%**
**SM Import** 399189 409452 10263 **97.49%**
**FullTextIndexer1** 566858 1804663 1237805 **31.41%**
**mam.xml** 58999 299876 240877 **19.67%**
**rod.xml** 95680 422423 326743 **22.65%**
**hum.xml** 144250 692476 548226 **20.83%**
**inv.xml** 430107 1895850 1465743 **22.69%**
**vrt.xml** 97662 471683 374021 **20.71%**
**sprot41\_1.xml** 332552 1763561 1431009 **18.86%**
**sprot41\_2.xml** 441054 1966868 1525814 **22.42%**
**FullText Indexer2** 2655170 7521535 4866365 **35.30%**
**hum.xml** 144250 692476 548226 **20.83%**
**XML Import** 15859 25441 9582 **62.34%**
**TestGetPathways** 19658 21769 2111 **90.30%**
**TestBenchmark** 6840592 13531171 6690579 **50.55%**
**Totals (ms)** **12229818** **30934447** **18704629** **39.53%**
**[Average of four measurements]{.underline}**
**FO Avg** **KODO Avg** **%(FO/KODO) Avg** **FO/KODO CV**
**Junit Tests** 7875 15414 **51.11%** 2
**Boot** 3064 4319 **71.67%** 12
**SM Import** 402303 413655 **97.26%** 1
**FullText Indexer1** 572934 1818548 **31.51%** 1
**mam.xml** 59745 303455 **19.69%** 3
**rod.xml** 98938 429661 **23.02%** 3
**hum.xml** 152827 698456 **21.89%** 5
**inv.xml** 434253 1915082 **22.68%** 1
**vrt.xml** 100248 476812 **21.03%** 4
**sprot41\_1.xml** 351278 1780467 **19.72%** 9
**sprot41\_2.xml** 473765 1981667 **23.90%** 11
**FullText Indexer2** 2765967 7590177 **36.43%** 4
**XML Import** 16804 39913 **45.44%** 36
**TestGetPathways** 20388 24362 **84.19%** 10
**TestBenchmark** 6886598 13736330 **50.16%** 3
**Totals (ms)** **12346986** **31228318** **39.54%**
:::
|
PubMed Central
|
2024-06-05T03:55:51.944527
|
2005-1-10
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545948/",
"journal": "BMC Bioinformatics. 2005 Jan 10; 6:5",
"authors": [
{
"first": "Marko",
"last": "Srdanovic"
},
{
"first": "Ulf",
"last": "Schenk"
},
{
"first": "Michael",
"last": "Schwieger"
},
{
"first": "Fabien",
"last": "Campagne"
}
]
}
|
PMC545949
|
Background
==========
Accumulation of a huge amount of genome sequence data in recent years and the task of extracting useful information from it, has given rise to many new challenges. One of the biggest challenges is the task of gene prediction and to fulfil this need, several gene prediction programs have been developed (For reviews see \[[@B1]-[@B5]\]). Most of these prediction programs require training based on prior knowledge of sequence features such as codon bias, which in turn are organism specific. In such cases, lack of large enough samples of known genes, as typically seen in a newly sequenced genome, can lead to sub optimal predictions. On the other hand, some gene prediction methods are based on the homology between two or more genomes but these methods are not of much help for gene prediction in case of genomes with no homologues. In addition, most of the gene prediction programs concentrate on the protein-coding regions and RNA genes, that can make up to 5 % of total protein coding genes, are neglected. Hence it is important to design *ab initio*gene prediction programs. One of the important steps towards *ab initio*gene prediction is to develop better promoter and TSS (transcription start site) prediction methods.
Although reasonable progress has been achieved in the prediction of coding region, the promoter prediction methods are still far from being accurate \[[@B6]-[@B9]\] and there are some very obvious reasons for these inaccuracies. One of the major difficulties is that the regulatory sequence elements in promoters are short and not fully conserved in the sequence; hence there is a high probability of finding similar sequence elements elsewhere in genomes, outside the promoter regions. This is the reason why most of the promoter prediction algorithms, which are based on finding these regulatory sequence elements, end up predicting a lot of false positives. Thus it is likely that incorporation of additional characteristics, which are unique to the promoter region, will help in improving the currently available promoter prediction methods.
In our earlier analysis, we observed that in case of bacteria as well as in eukaryotes, various properties of the region immediately upstream of TSS differ from that of downstream region \[[@B10]\]. There are differences in sequence composition as well as in different sequence dependent properties such as stability, bendability and curvature. The upstream region is less stable, more rigid and more curved than downstream region. Some of these observations are supported by other studies carried out independently on genomic sequences \[[@B9],[@B11]-[@B17]\]. Among all types of promoters, the most prominent feature is the difference in DNA duplex stabilities of the upstream and downstream regions. Here, we propose a prokaryotic promoter prediction method, which is based on the stability differences between promoter and non-promoter regions.
Results and discussion
======================
Lower stability of promoter regions in bacterial sequences
----------------------------------------------------------
It is well known that the stability of a DNA fragment is a sequence dependent property and depends primarily on the sum of the interactions between the constituent dinucleotides. The overall stability for an oligonucleotide can thus be predicted from its sequence, if one knows the relative contribution of each nearest neighbour interaction in the DNA \[[@B18]\]. The average stability profiles for three sets of bacterial promoter sequences calculated (using 15 nt moving window) based on this principle is shown in Figure [1](#F1){ref-type="fig"}. It is interesting that the promoters from diverse bacteria, which have quite different genome composition (A+T composition: *E. coli*0.49, *B. subtilis*0.56 and *C. glutamicum*0.46), show strikingly similar features. Promoters from all the three bacteria show low stability peak around the -10 region. The second prominent feature in the free energy profiles of all the three bacteria is the difference in stabilities of the upstream and downstream regions. In all the three groups of promoter sequences, the average stability of upstream region is lower than the average stability of downstream region. But the three sets of promoter sequences differ in their basal energy level, which seems to be dependent on the nucleotide composition of the bacteria.
Detailed analysis of *E. coli*promoter sequences
------------------------------------------------
In order to get a better insight into the stability feature, we carried out a detailed analysis of *E. coli*promoter sequences. Our statistical analysis using \"Wilcoxon signed test for equality of medians\" (see METHODS) shows that the free energy distribution corresponding to a fragment extending from position -148 to 51 in the *E. coli*sequences is appreciably different from the energy distribution calculated in randomly selected windows, at a significance level as high as 0.0001. A comparison of free energy distribution at position -20 (corresponding to the promoter region) with distributions at positions -200 (corresponding to the region upstream of promoter region) and +200 (corresponding to the coding region) is shown in Figure [2](#F2){ref-type="fig"}. It is clearly seen that the region immediately upstream of TSS is much less stable than the other two regions. The average free energy at -20 position is -17.48 kcal/mol while average free energies at the -200 and +200 positions are -19.42 kcal and -20.19 kcal/mol respectively. The Kolmogorov-Smirnov test also confirms that the free energy distribution at position -20 significantly differs from that at -200 and +200 positions at a very high significance level (alpha = 10^-10^).
Details of methodology
----------------------
This difference in free energy and the stability of promoter regions as compared to that of coding and other non-coding regions can be used to search for the promoters. Based on this consideration, a new scoring function D(n) is defined, which will look for differences in free energy of the neighbouring regions of position n:
D(n) = E1(n) - E2(n)
where,
Thus, E1(n) and E2(n) represent the free energy (see METHODS) average in the 50 nt region starting from nucleotide n and neighbouring 100 nt region starting from nucleotide n+99, respectively. The E1 value represents the basal energy level, which is characteristic of the given bacterial genome (e.g. in this case *E. coli*) and the D value represents the free energy difference in the two neighbouring regions. A stretch of DNA is assigned as promoter only if the average free energy of that 50 nt region (E1) and difference in free energy as compared to its neighbouring region (D) is greater than the chosen cut-offs. The protocol followed to calculate the true and false positives and hence sensitivity and precision is presented in the form of a flowchart in Figure [3](#F3){ref-type="fig"}. Identical sensitivity values can be achieved using different combinations of D and E1 cut-off values, which is obvious from the contour plot shown in Figure [4A](#F4){ref-type="fig"}. Similarly, different combinations of D and E1 cut-offs can lead to similar precisions (Figure [4B](#F4){ref-type="fig"}). But we observe that the use of different D and E1 cut-offs, corresponding to a given sensitivity level, results in a wide range of precisions (Figure [5](#F5){ref-type="fig"}). Hence, in order to attain a desired level of sensitivity the D and E1 cut-off values are chosen such that the number of false positives is minimum and the precision is maximum.
Initially, we divided the *E. coli*sequence data into two sets. The E1 and D cut-off values corresponding to different sensitivity levels were obtained for 100 randomly selected sequences (1^st^set). These cut-off values were then applied to a second set consisting of remaining 127 sequences. The sensitivity and precision values calculated for the first and second set match very well. We also found that very similar results can be obtained when we use the whole dataset (Figure [6](#F6){ref-type="fig"}). Hence, we present the results for the whole dataset rather than separately for two sets. The D and E1 cut-offs and the number of false positives corresponding to different levels of sensitivity are given in Table [1](#T1){ref-type="table"}. To confirm the validity of our choice, we used another set of 1000 nt long sequences extracted from the centre of the ORFs, which were more than 2000 nt long. The results corresponding to this set of control fragments are also given in Table [1](#T1){ref-type="table"} and show very few false positives.
In principle, D can also be calculated using equal sized windows, i.e. 50 nt, for both E1 and E2 instead of a 50 nt window for E1 and a 100 nt window for E2. However, our calculations show that use of equal sized windows, for E1 as well as E2 calculations, results in a slightly lesser precision than when 100 nt window is used for E2 calculations (Figure [7](#F7){ref-type="fig"}). Hence, in our promoter predictions, we chose a 100 nt window for E2 calculations.
Comparison with other promoter prediction programs
--------------------------------------------------
A large number of promoter prediction programs have been developed for eukaryotic sequences and are easily accessible, while NNPP \[[@B19],[@B20]\] is the only available prokaryotic promoter prediction program. It is a neural network based method where prediction for each sequence element constituting promoter sequence is combined in time-delay neural networks for a complete promoter site prediction. Some other prokaryotic promoter prediction methods are based on weight matrix pattern searches \[[@B21]-[@B24]\]. One of the representative weight matrix method, proposed by Staden \[[@B21]\], uses three weight matrices corresponding to the -35 sequence, the -10 sequence and the transcription start site. It also takes into account the spacing between the -35 and -10 motifs, as well as the distance between the -10 motif and the transcription start site. A brief comparison of the results obtained by our method and the other two methods (Staden method and NNPP program) is given in Table [2](#T2){ref-type="table"}. It can be clearly seen from Table [2](#T2){ref-type="table"} that for similar sensitivity, our program gives much better accuracy than the other two programs. It is pertinent to mention here that our method differs from the other two methods in one major respect, namely our method tries to find a promoter region while the other two programs try to pinpoint the transcription start site. It may be argued that the lesser number of false positives in our prediction method, as compared to the other two algorithms, may be due to this difference. But even after taking this difference into consideration, the number of false positives predicted by our protocol turns out to be smaller than those predicted by the other two methods. For example, Figure [8](#F8){ref-type="fig"} represents the case of argI and argF genes, where the NNPP program predicts a few extra TSS as compared to our method which correctly picks up a region in the vicinity of TSS. A combination of both the methods can therefore help in reducing the false predictions in the upstream and downstream regions. In principle, by restricting the pattern recognition using NNPP and Staden\'s methods only to the promoter region located initially with the help of our method, one can reduce the number of false positives. This composite approach will also help in pinpointing the TSS, which is not possible by use of our method alone. But at the same time it should be noted that both types of predictions fail to identify some of the promoters (Figure [8](#F8){ref-type="fig"}), e.g. for csiE gene, our program could correctly predict the promoter region but the NNPP program could not locate it. On the other hand, our program failed to find the promoter region for gyrA gene while NNPP could correctly position it. And in case of ilvA gene both the programs did not succeed in identifying the promoter region.
Very recently a study on improvement of NNPP prediction (TLS-NNPP), by combining this method with additional information such as distance between TSS and translation start site (TLS), has been published \[[@B25]\]. With the use of additional information regarding TLS, Burden *et al.*could significantly increase the precision of NNPP. The TLS-NNPP method was tested on 510 *E. coli*sequences of length 500 bp. For comparable sensitivity levels, the precision achieved by TLS-NNPP was 0.188 (sensitivity = 0.452) as compared to 0.109 precision (sensitivity = 0.443) achieved by NNPP. It can be seen that, for similar sensitivity levels, the precision achieved by our method (\~0.7) is higher as compared to both TLS-NNPP and NNPP (Figure-[9](#F9){ref-type="fig"}).
Presence of high densities of promoter like signals in the upstream region of TSS may be one of the reasons why pattern matching programs result in low level of precision. This has been shown recently by a systematic analysis of sigma70 promoters from *E. coli*\[[@B24]\]. In this study a number of weight matrices were generated by analysis of 599 experimentally verified promoters and these were tested on the 250 bp region upstream of gene start site. It was found that each 250 bp region on an average has 38 promoter-like signals. The study also presented a more rigorous patter searching method for locating promoters. With the use of this function the authors reach a sensitivity values of 0.86 but the corresponding precision achieved is only \~0.2. In case of our method, for a sensitivity of 0.9 we obtained a precision of 0.35 (as shown in Figure -[9](#F9){ref-type="fig"}).
Recently Bockhorst *et al.*\[[@B26]\] proposed a very accurate method for predicting operons, promoters and terminators in *E. coli*. This method is based on sequence as well as expression data, but requires prior knowledge of coordinates of every ORF in the genome. We would like to emphasize here that our method is different from other methods in that it is independent of any such prior knowledge about the test gene or the organism and hence holds promise as being useful for promoter prediction in a newly sequenced genome.
The eukaryotic promoter prediction method proposed by Ohler *et al.*\[[@B27]\] is also worth mentioning here. Ohler *et al.*showed that a 30 % reduction of false positives can be achieved by use of physical properties, such as DNA bendability, in addition to other sequence properties of promoters. Interestingly, our method which also uses a physical property gives much smaller number of false positives as compared to Ohler *et al.*\'s method. (For similar sensitivity, number of false predictions in case of Ohler *et al.*\'s method are 1/4740 nt while in case of our method these are 1/8407 nt).
Another vertebrate promoter prediction program, \'Promfind\' \[[@B28]\] identifies differences in hexanucleotide frequencies of promoter and coding region and is algorithmically quite similar to our method. But Promfind differs from our method in two important aspects. First, the Promfind program is developed mainly for vertebrate promoters and second, it assumes that in a given sequence, a promoter is always present and merely predicts its location. This need not necessarily be the case, as some of the sequences may not have any promoter at all. Our program differs from Promfind in that a promoter is predicted only when the sequence satisfies certain criteria and hence is much more appropriate for carrying out genome scale analysis.
Promoter predictions in case of RNA genes
-----------------------------------------
In addition to protein coding genes there are genes present for the non-coding RNAs (ncRNAs), which play structural, regulatory and catalytic roles. It is a difficult task to find out ncRNA genes in a genome because unlike protein coding regions they lack open reading frames and also they are generally smaller in size. In addition, it is also difficult to do a homology sequence search as only the structure of ncRNA is conserved and not the sequence. There are around 156 *E. coli*RNA genes reported on the NCBI site \[[@B29]\] and in addition many more small RNA genes are known to exist. Argaman *et al.*\[[@B30]\] recently identified 14 novel sRNA genes by applying a heuristic approach to search for transcriptional signals. We have checked the performance of our algorithm with respect to the 42 RNA transcription units (TUs) reported in Ecocyc database. Our method could pick up around 57 % RNA TUs, at a cut-off corresponding to 60 % sensitivity. The program works much better in case of rRNA operons than tRNA transcription units. We could correctly pick up promoter regions in 6 out of 7 rRNA transcription units, 17 out of 33 tRNA TUs and 1 out of the 2 remaining RNA types.
Promoter prediction in *Bacillus subtilis*and *Corynebacterium glutamicum*
--------------------------------------------------------------------------
Finally, it is very important to see whether the method works equally well for other organisms which have genome compositions substantially different from that of *Escherichia coli*. Hence, we also tested our method using the promoter sequences from 1) the A+T-rich bacteria, *Bacillus subtilis*and 2) a G+C rich bacteria such as *Corynebacterium glutamicum*. Figure [9](#F9){ref-type="fig"} gives a summary of the predictions in case of bacillus and corynebacterium promoters, along with those of *Escherichia coli*. It can be clearly seen that, at present our method performs optimally for the *Escherichia coli*promoters and also performs quite well in case of *Bacillus subtilis*. The prediction accuracy in case of *Corynebacterium glutamicum*promoters is not as good as that for the other two classes of promoters. However, it should be noted that the number of experimentally determined *Corynebacterium*promoters is much smaller as compared to other two bacteria and a larger dataset is required to arrive at any firm conclusion.
Conclusions
===========
It has often been suggested that use of certain properties of promoters, other than just the sequence motifs, which can distinguish promoters from other genomic regions, could significantly improve the gene prediction methods. Although the lower stability of promoter regions as compared to non-promoter regions has been reported previously, this observation was not incorporated into a promoter prediction program. We have been able to successfully use the differential stability of promoter sequences to predict promoter regions. Our method performs better as compared to currently available prokaryotic prediction methods and is also moderately successful in predicting RNA and bacillus promoter regions. The method certainly needs to be further improved to reduce the number of predicted false positives. This can be achieved by combining the approach presented here, with the earlier reported sequence analysis methods. Such a composite method will also help in pinpointing the TSS within the promoter region identified by our method.
Methods
=======
Promoter sequence sets
----------------------
All the promoter sequences used in this study are 1000 nt long, starting 500 nt upstream (position -500) and extending up to 500 nt downstream (position +500) of the TSS. In order to avoid having multiple TSS in a given 1000 nt sequence, we have excluded all the transcription start sites which are less than 500 nt apart. Our promoter set has 227 *E. coli*promoters, 89 *B. subtilis*promoters and 28 *C. glutamicum*promoters.
### a) *Escherichia coli*promoter sequences
We tested our algorithm using the *Escherichia coli*promoter sequences, which were taken from the PromEC dataset \[[@B31]\]. The PromEC dataset provides a compilation of 471 experimentally identified transcriptional start sites. As mentioned above, after excluding all the transcription start sites which are less than 500 nt apart, the dataset contains 227 promoters. With the help of TSS information, promoter sequences were extracted from *Escherichia coli*genome sequence (NCBI accession no: NC\_000913).
### b) *Bacillus subtilis*promoter sequences
The transcription start sites for *Bacillus subtilis*promoters were obtained from the DBTBS database \[[@B32]\]. The required length sequences around transcription start sites were extracted from the Bacillus genome sequence (NCBI accession no: NC\_000964).
### c) *Corynebacterium glutamicum*promoter sequences
Analysis of *Corynebacterium glutamicum*promoters is carried out on a set of promoters compiled by Pàtek *et al.*\[[@B33]\] based on experimentally determined transcription sites.
### d) RNA promoter sequences
The transcription start positions of RNA transcription units are obtained from the ecocyc dataset. In this set, both computer predicted as well as experimentally determined transcription start sites, are included. In total, we have 7 rRNA TUs, 33 tRNA TUs and 2 TUs of other RNAs.
Free energy calculation
-----------------------
The stability of DNA molecule can be expressed in terms of free energy. The standard free energy change (ΔG^o^~37~) corresponding to the melting transition of an \'n\' nucleotides (or \'n-1\' dinucleotides) long DNA molecule, from double strand to single strand is calculated as follows:
where,
ΔG^o^~ini~is the initiation free energy for dinucleotide of type ij.
ΔG^o^~sym~equals +0.43 kcal/mol and is applicable if the duplex is self-complementary.
ΔG^o^~i,j~is the standard free energy change for the dinucleotide of type ij.
Since our analysis involves long continuous stretches of DNA molecules, in our calculation we did not consider the two terms, ΔG^o^~ini~and ΔG^o^~sym~, which are more relevant for oligonucleotides. In the present calculation, each promoter sequence is divided into overlapping windows of 15 base pairs (or 14 dinucleotide steps). For each window, the free energy is calculated as given in the above equation and the energy value is assigned to the first base pair in the window. The energy values corresponding to the 10 unique dinucleotide sequences are taken from the unified parameters proposed recently \[[@B34],[@B35]\].
Statistical tests
-----------------
### a) Wilcoxon signed test for equality of medians
The free energy distribution at a given position, in the 1000 nt *E. coli*sequences ranging from -500 to +500, was compared to the distribution in a randomly selected set. For this comparison, we followed a similar procedure as adopted by Margalit *et al.*\[[@B11]\]. The random set was chosen such that an energy value per sequence was selected arbitrarily, independent of its position in the sequence. The comparison between the energy distributions was carried out using Wilcoxon signed test for equality of medians. This is a nonparametric test, which is used to test whether the two samples have equal medians or not.
### b) Two-sample Kolmogorov-Smirnov test
We compared the free energy distribution at position -20 (with respect to TSS) with the distributions at the positions -200 and +200 using Kolmogorov-Smirnov two sample test \[[@B36]\].
All the calculations related to the statistical tests were carried out using MATLAB 6.0^®^.
Implementation and scoring of NNPP and Staden\'s method
-------------------------------------------------------
The promoter predictions were also carried out using two other methods *viz.*NNPP and Staden\'s method. NNPP program is available at \[[@B20]\]. All the NNPP predictions were carried out at a score cut-off 0.80.
The implementation of Staden\'s method was carried out as described in \[[@B21],[@B37]\]. The weight matrix search was carried out with the help of PATSER program \[[@B38]\].
In case of NNPP as well as Staden\'s method, the true and false positives were scored as in case of our method (Figure [3](#F3){ref-type="fig"}), with a prediction in -150 to 50 region being considered as a true prediction.
Sensitivity and precision
-------------------------
The sensitivity and precision for the predictions are calculated using the following formulae:
Authors\' contributions
=======================
AK performed the analysis, evaluated the results, and drafted the manuscript. MB suggested the problem, helped with evaluation of the results and the manuscript, also provided mentorship. All authors read and approved the final manuscript.
Acknowledgements
================
During the study, AK was supported by University Grants Commission and Council of Scientific and Industrial Research. We thank Prof. N. V. Joshi for his valuable comments. We also thank Dr MiroslavPátek for the *Corynebacterium*promoter sequences. We are grateful to the two unknown referees for their suggestions.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Overall free energy profile around bacterial TSS**The figure shows the average free energy profiles of A) *Escherichia coli*(227 promoters) and B) *Bacillus subtilis*(89 promoters) C) *Corynebacterium glutamicum*promoters (28 promoters). The profiles extend from 500 nt upstream to 500 nt downstream of transcription start site (positioned at 0, shown as dashed line). The nucleotide sequence position is shown on x-axis. More negative values of free energy indicate greater stability.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Histogram showing the free energy distribution corresponding to upstream region (-200), promoter region (-20) and coding region (+200) in *E. coli*sequences**The free energy distribution corresponding to position -20 (calculated for a 15 nt window extending from -20 to -6) is shown as brown bars. Free energy distribution corresponding to positions -200 (calculated for a 15 nt window from -200 to -186, shown in green bars) and +200 (calculated for 15 nt window from +200 to +214, shown in blue bars) are also shown for comparison. Each bar corresponds to 1 kcal/mol. The average free energies corresponding to -20, -200 and +200 positions are -17.48 kcal/mol, -19.42 kcal/mol and -20.19 kcal/mol respectively.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**A flowchart summarizing our methodology**\* If there are more than one predictions in the 200 nt region (-150 to 50) then only one prediction which is nearest to the TSS is taken as a true prediction. The remaining predictions are counted as false predictions.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Sensitivity and precision contour plots**The E1 value cut-offs are plotted on x-axis while D value cut-offs are plotted on y-axis. The different A) sensitivity and B) precision levels are shown by colours ranging from dark blue to brown, where dark blue corresponds to lowest value and brown corresponds to highest value.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**A plot showing range of precision values obtained for a given sensitivity**The sensitivity (x-axis) and precision (y-axis) corresponding to different E1 and D cut-offs has been plotted.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
**The comparison of sensitivity and precision values from test and \'training\' sets**The sensitivity (x-axis) and precision (y-axis) corresponding to 1) test set (filled circles), 2) training set (open circles) and 3) the whole *E. coli*dataset (red) is shown. The sensitivity and precision values for the test set were calculated using E1 and D cut-offs derived from the training set.
:::
:::
::: {#F7 .fig}
Figure 7
::: {.caption}
######
**Change in precision with the use of different sized windows for E2 calculation**The sensitivity (x-axis) and precision (y-axis) values corresponding to the use of 1) 50 nt window (black) and 2) 100 nt window (red) for E2 calculation.
:::
:::
::: {#F8 .fig}
Figure 8
::: {.caption}
######
**Examples illustrating the predictions with our method as well as NNPP**The promoter predictions for the argF, argI, csiE, gyrA, ilvA genes by our method (red) as well as by NNPP (blue) in the 1000 nt fragments (-500 to 500) with the TSS at the centre. The figure is generated using FEATURE MAP program \[39\].
:::
:::
::: {#F9 .fig}
Figure 9
::: {.caption}
######
**Prediction accuracy of our method in case of promoters from different organisms**The precision (y-axis) of our method in predicting promoter region in different organisms *viz. Escherichia coli*(red), *Bacillus subtilis*(blue) and *Corynebacterium glutamicum*(black) is plotted against various levels of sensitivity (x-axis).
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
The number of false positives obtained for different levels of sensitivity.
:::
Sensitivity Cut-off for D Cut-off for E1 (kcal/mole) Frequency of false positives
------------- --------------- ---------------------------- ------------------------------ ----------
0.13 3.4 -15.99 1/16214 1/261000
0.22 3.4 -16.7 1/11350 1/130500
0.32 3.3 -17.1 1/8407 1/65250
0.40 3.3 -17.55 1/6486 1/29000
0.50 2.76 -17.53 1/3914 1/13737
0.60 2.45 -17.64 1/2467 1/7250
0.70 2.35 -18.07 1/1621 1/2747
0.81 1.9 -18.15 1/1086 1/1878
0.90 0.97 -18.37 1/572 1/967
^a^The false positives in the 1000 nt fragments, with TSS at the centre (-500 to +500).
^b^The false positives in the 1000 nt fragments extracted from the centre of ORFs with length more than 2000 nt.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Comparison of our method with other prokaryotic prediction algorithms vis-à-vis *Escherichia coli*promoters.
:::
TP FP(1/nt)^a^ FP(1/nt)^b^
------------------------- ----- ------------- -------------
Our Program 195 1/780 1/1474
Neural Network \[19\] 195 1/233 1/514
Staden\'s method \[21\] 195 1/65 1/233
^a^The false positives in the 1000 nt fragments with TSS at the centre (-500 to +500).
^b^The false positives in the 1000 nt fragments extracted from the centre of ORFs with length more than 2000 nt.
:::
|
PubMed Central
|
2024-06-05T03:55:51.949417
|
2005-1-5
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545949/",
"journal": "BMC Bioinformatics. 2005 Jan 5; 6:1",
"authors": [
{
"first": "Aditi",
"last": "Kanhere"
},
{
"first": "Manju",
"last": "Bansal"
}
]
}
|
PMC545950
|
Background
==========
Fragile X syndrome, the leading heritable form of mental impairment \[[@B1],[@B2]\], is generally caused by the expansion of a trinucleotide (CGG) repeat element in the fragile X mental retardation 1 (*FMR1*) gene to greater than 200 repeats (full mutation) \[[@B3]\]. Such expansions generally lead to methylation-coupled gene silencing \[[@B4],[@B5]\] and the consequent absence of the *FMR1*protein (FMRP), an RNA-binding protein that is important for neural development and plasticity \[[@B6],[@B7]\].
Although great strides have been made from animal models in our understanding of the neuropathology of fragile X syndrome \[[@B8]-[@B13]\], there is, at present, no adequate animal or cell model to study the detailed molecular biochemistry of the *FMR1*gene. The absence of a suitable model system is a consequence of the inability to clone full mutation alleles, and because no animal system has been found that carries native, full mutation *FMR1*alleles. Thus, all animal or cell model systems for fragile X syndrome are based on *FMR1*(homolog) knock-out constructs. While these models qualitatively recapitulate some of the features of the fragile X syndrome phenotype, they do not address any of the potential consequences of the expanded methylated CGG repeat.
As with most disorders of the human nervous system, it has been impossible to directly study the detailed cellular pathogenic mechanisms that underlie fragile X syndrome, due to the absence of a suitable (human) cell model. However, this barrier may be overcome through the use of neural progenitor cells, which comprise relatively undifferentiated populations of cells in the central nervous system (CNS) that give rise to the broad array of specialized cells, including neurons and glial cells. Long thought to be an exclusive component of the developing CNS, these cells have been shown to exist in the adult CNS \[[@B14]-[@B19]\]. Recent research, demonstrating that these cells can be isolated and cultured \[[@B14],[@B18],[@B20]-[@B22]\] has raised the prospect of using neural progenitor cells as a human cell-appropriate (neuronal and astrocytic) model system for the detailed study of the molecular biology of fragile X syndrome.
Importantly, we have previously shown that neural progenitor cells can be harvested from adult post-mortem brain tissue \[[@B18],[@B20]\], which represents a singular advantage over the use of neural stem cells from fetal sources. In particular, since there is a broad spectrum of clinical involvement in fragile X syndrome \[[@B2]\], the study of neural stem cells from individuals of known phenotype provides us with a closer coupling of the genotype with the clinical phenotype.
In the current report, we describe the successful culturing of neural progenitor cells from an adult male with fragile X syndrome, and some of the characteristics of these cells. We further demonstrate initial efforts to differentiate these progenitor cells into both neuronal and astrocytic lineages. As expected, expression of FMRP is substantially reduced relative to its expression in neural cell culture from an unaffected control.
Methods
=======
Autopsy, brain harvest, and tissue cryopreservation
---------------------------------------------------
Prior to tissue acquisition by the authors, informed consent for the donation of tissues was obtained under the auspices of the protocol for the National Human Neural Stem Cell Resource. This protocol is approved by Institutional Review Board (IRB) of Children\'s Hospital of Orange County, and follows a protocol approved by the UC Davis School of Medicine IRB. All tissues were acquired in compliance with NIH and institutional guidelines. Two patients were used for the present study: a patient with fragile X syndrome and a patient with no neurogenetic disease.
The autopsy followed standard procedures as described \[[@B20]\]. The periventricular zone in the area of the head of the caudate nucleus regions was identified and dissected from the appropriate brain sections. Brain region specimens were then placed in separate Petri dishes and rinsed three times with DGA (see below). Tissues were minced with sterile scalpel blades, triturated in DGF (see below) containing 10% DMSO, taken to -80°C overnight in controlled-rate freezing containers, and then transferred to liquid nitrogen Dewars for long term storage.
Pathology
---------
Portions of fresh brain, heart, and testicular tissue were received, and fixed in 10% formalin for 10 days prior to sampling and processing for paraffin sections in standard fashion. All tissue sections were stained with hematoxylin and eosin, with cardiac valves and aorta additionally stained for elastin and mucin. All sections were examined by standard light microscopy.
Cell culture
------------
The base medium was a high glucose 1:1 DMEM:F12 (Irvine Scientific). The basal medium (DGA) used for all other media was the base medium containing glutamine, penicillin, streptomycin, gentamicin, ciprofloxacin, and amphotericin as previously described \[[@B20]\]. Medium (DGF) used for all washes consisted of DGA containing 20% fetal bovine serum.
All procedures were performed as previously described \[[@B20]\], with modifications. Tissues were quickly thawed, diluted by drop-wise addition and agitated in 10 volumes of DGF, then further dissociated by trituration and three washes in DGF with centrifugation -- no enzymatic digestion was used. Whole tissue homogenates were plated directly on fibronectin-coated tissue culture plates (6-well, tissue culture treated, Falcon) in primary growth medium (PGM) composed of DGF containing 10% BIT 9500 (Stem Cell Technologies), 40 ng/mL basic fibroblast growth factor (FGF-2; InVitrogen), 20 ng/mL epidermal growth factor (EGF; InVitrogen), and 20 ng/mL platelet-derived growth factor-AB (PDGF-AB; Peprotech). Plates had been previously incubated with 200 uL/cm^2^of fibronectin (5 ug/mL; Sigma) overnight at 37°C, the fibronectin solution aspirated, and the plates allowed to air dry before the introduction of tissue homogenates.
Approximately 300 mg fresh tissue was subjected to mincing and trituration, and the resulting crude tissue homogenate plated into six wells of a fibronectin-coated six-well tissue culture plastic plate (≈60 cm^2^total surface area).
After plating, 50% of the medium was replaced, 3 times weekly. Non-adherent cells and debris from the removed supernatant were pelleted by centrifugation and re-introduced into the cultures together with the fresh medium. After 7 days in culture, plates were agitated by sharp rapping with a marking pen and 100% of the culture medium and non-adherent material was removed. Fifty percent of the volume removed was replaced with fresh medium, while the removed medium was centrifuged to pellet cell debris and non-adherent cells and to recover conditioned medium as supernatant. Fifty percent, by volume, of the conditioned medium was then returned to the original plates. The pellet, containing the non-adherent fraction, was resuspended in 50% conditioned medium and 50% fresh medium, by volume, and then transferred to a fresh fibronectin-coated 6-well plate. After one week, the procedure was repeated, except that the non-adherent fraction was discarded. In this way, an additional population of cells was recovered from the non-adherent fraction. All the cells were eventually combined to form a single population of cultured cells.
At near confluence, cultures were passaged by lifting with a solution of Cell Dissociation Buffer (GIBCO) supplemented with trypsin. The cells were washed twice with DGF and plated in 1:1 (conditioned:fresh) medium into a fibronectin-coated T75 flask. Thereafter, and at approximately one week intervals, cells were lifted and similarly plated into twice the surface area from which they were removed. After the cells had reached a confluent surface area of 600 cm^2^, the medium from one T75 flask was exchanged with GM (PGM without serum), and these cells were cultured for two weeks before immunocytochemical analysis, or differentiation and immunocytochemical analysis, as previously described \[[@B20]\].
Immunocytochemistry
-------------------
Immunocytochemistry was performed as previously described \[[@B20]\]. Primary antibodies and dilutions were used as follows: nestin (1:100; mouse; Chemicon), type III β-tubulin (Tuj20; 1:100; mouse; Chemicon), MAP2ab (1:250; mouse; Sigma), GFAP (1:500, guinea pig; Advance Immuno), CD133-APC (1:100; mouse; Miltenyi), NCAM (1:100, rabbit, Chemicon), fusin (1:100, mouse, Chemicon), and FMRP (1:100; mouse; Chemicon). Coverslips were mounted with Prolong^®^Antifade Kit (Molecular Probes, Eugene, OR). Some cells were stained with 4\',6-diamidino-2-phenylindole (DAPI, Sigma) before being rinsed and mounted. Pictures were imaged on an Olympus IX70 Microscope and digitally photographed via a Microfire digital camera (Optronics, Goleta, CA) using Image Pro Plus 4.5 with AFA plugin 4.5 software.
Molecular studies
-----------------
### DNA analysis
Genomic DNA was isolated from approximately 5 × 10^6^neural progenitor cells and from post-mortem sections of about 500 mg of brain tissue using standard methods (Puregene Kit; Gentra). For Southern blot analysis, 10 μg of isolated DNA were digested with *Eco*RI and *Nru*I. The *FMR1*genomic probe StB12.3, labeled with Dig-11-dUTP by PCR (PCR dig synthesis Kit, Roche Diagnostics), was used in the hybridization, as described in Tassone et al. \[[@B31]\]. Genomic DNA was also amplified by PCR using primers c and f \[[@B32]\]; PCR products were detected using a digoxygenin-end-labeled oligonucleotide probe (CGG)~10~. Southern blot and PCR analyses were both carried out using an Alpha Innotech FluorChem 8800 Image Detection System.
### FMR1 mRNA expression levels
Total RNA was isolated from approximately 1 × 10^6^neural progenitor cells and from post-mortem brain tissue using standard methods (Purescript, Gentra Inc. and Trizol). Reverse transcriptase reactions and quantitative fluorescence RT-PCR, using specific primers and probe set for the *FMR1*gene and the control gene (β-glucoronidase; *GUS*), were carried out as described in Tassone et al. \[[@B33]\].
Results
=======
Clinical history
----------------
JS was a 25-year-old man with fragile X syndrome. His history included motor and language delays in childhood, with walking at 20 months, phrases at three years, and sentences at 6 years of age. He was diagnosed with fragile X syndrome at 11 years of age, by cytogenetic testing. His behavior included hyperactivity, anxiety, shyness, poor eye contact, hand flapping, finger biting, and perseverative speech. At age 24, cognitive testing with the WAIS III demonstrated a verbal IQ of 58, performance IQ of 51, and full scale IQ of 51. He did not have autism; childhood autism rating scale (CARS) score was 28.5 (below autism range).
Previous medical history included severe mitral valve prolapse; echocardiography revealed moderate thickening and redundancy of both mitral leaflets, with central mitral regurgitation (grade 2 to 3+ by Doppler), mild tricuspid regurgitation, and moderate left ventricular and mild left atrial enlargement. Upon physical examination at age 24, JS had a long face, prominent ear pinna, high arched palate, and macroorchidism with testicular volume of 60 ml bilaterally. Blood pressure was 142/74, and a grade III/IV systolic and diastolic murmur with click was heard on examination. He died unexpectedly at age 25, presumably due to a cardiac arrhythmia secondary to mitral valve prolapse.
Pathology
---------
A complete autopsy was performed 16 hours after death. Abnormal findings included increased brain weight (1600 gm; normal, 1440 ± 20 g), an enlarged heart, (400 gm; normal, 349 ± 40 g), with features of mitral valve prolapse (myxomatous thickening of both leaflets; \"hooding\" of the posterior leaflet), and increased testicular weight of 73 g (normal average, 25 g). Our laboratory did not receive the intact brain, but portions of it, therefore detailed gross and histological evaluation was not possible. In particular, neither the hippocampus nor the amygdala was available for histopathological examination. Evaluation of available brain sections showed diffuse, acute early ischemic damage corresponding to the manner of death; no intraneuronal inclusions were seen, nor was there spongiosis of white matter in the cerebrum, cerebellum, or middle cerebellar peduncles. Cerebellar folia showed moderate patchy absence of Purkinje cells; because of the manner of death, loss due to ischemic damage cannot be ruled out. However, the majority of Purkinje cells present appeared histologically normal.
Cell culture
------------
Minced brain tissue, derived from the periventricular zone in the area of the head of the caudate nucleus and maintained under proliferation conditions in serum-containing medium with growth factors, yielded viable cells that formed an adherent monolayer on fibronectin-coated plates. When lifted and plated without serum or fibronectin substrate, the cells grew in suspended clusters/spheres (Figure [1](#F1){ref-type="fig"}), similar to previously described neurospheres \[[@B20]\]. The morphology within the adherent population was variable and included small, rounded profiles, medium-sized bipolar and spindle-shaped profiles, and larger cells with polygonal and multipolar morphologies. Once a robust primary culture had been established in the six-well plates (approximately one month after plating of tissue homogenates), the cultures were passaged into T75 flasks approximately once per week until 600 cm^2^of confluent adherent cells had been produced. Cells were then further expanded under serum-free conditions for immunocytochemical analysis, or under growth-factor-free conditions for biochemical analysis.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Phase-contrast photomicrographs of fragile X progenitor cells**The figure shows clusters/spheres during the initial stages (2--3 days after plating) of adherence to a fibronectin substrate. (A) 4×; (B) 10×; (C) 20×. Confluent serum- and growth factor-expanded cultures were serum deprived for one week in the presence of growth factors, then lifted with enzyme-free buffers and transferred to new plates with no fibronectin substrate. After growing the resulting clusters/spheres for two weeks, the clusters/spheres were transferred to new fibronectin-coated plates. Clusters/spheres (black arrows) are abundant and are seen adhering to the substrate. Cells (black arrowheads) can be seen streaming from the spheres and spreading out on the substrate.
:::
:::
Although the methodology used was similar to that previously reported \[[@B20]\], four conditions employed in the current work are noteworthy. (*i*) The cells were grown from cryopreserved, rather than fresh, tissue. This modification allows pathologists with no local access to a stem cell culture laboratory to preserve tissues for later stem cell harvest at a remote collaborating laboratory. (*ii*) No enzymatic digestion, only trituration, was used to generate the crude tissue homogenates. Preliminary studies showed that enzymatic digestion, typically used with fresh tissue, adversely affected our ability to harvest living cells from cryopreserved tissues. (*iii*) Serum (20%) was maintained in the culture until it was expanded to 600 cm^2^of confluent adherent cells, after which the cells were cultured in serum-free medium. Preliminary studies indicated that, unlike cells harvested from fresh tissues, cells harvested from cryopreserved tissues required application of serum for a longer time in culture to sustain a sufficient rate of proliferation. (*iv*) The non-adherent fraction was transferred to new fibronectin-coated plates after one week. Preliminary studies showed that the non-adherent fraction from cryopreserved tissues retained a significant population of viable cells for a longer period of time than that from fresh tissue. These plates were cultured for an additional week before the remaining non-adherent fraction was finally discarded. Cells grown from both sets of plates were combined for expansion.
Immunocytochemistry
-------------------
Immunocytochemical analysis of fragile X neural progenitor cells grown under expansion conditions demonstrated the expression of a range of developmental and mature neural markers (Figure [2](#F2){ref-type="fig"}). Many of the current results are similar to previous findings with control human neural progenitor cell cultures \[[@B20]\]. In particular, the distribution of the multipotential neural progenitor lineage marker, CD133, the neuroepithelial marker nestin, and the neural cell adhesion molecule, NCAM, are found to be widespread in these cultures, consistent with earlier observations of immunocytochemistry and flow cytometry in control human neural progenitor cells \[[@B20],[@B23],[@B24]\]. The CXCL12 (SDF-1) cytokine receptor, CXCR4 (fusin, CD184), and the glial neurofilament marker, glial fibrillary acidic protein (GFAP), are also widely expressed in the fragile X neural progenitor cell cultures, in agreement with a previous report \[[@B20]\]. The expression of β-III-tubulin, a mature neuronal marker, is restricted to subpopulations of cells, again consistent with control human neural progenitor cells \[[@B20]\]. Finally, the primitive neuroepithelial (intermediate filament) markers, nestin and CD133, seen in the proliferating neural progenitor cells (Figure [2](#F2){ref-type="fig"}), disappear under differentiation conditions (data not shown). Importantly, FMRP staining is markedly reduced in the fragile X neural progenitor cells (Figure [3](#F3){ref-type="fig"}) compared to control cultures.
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Staining of fragile X progenitor cells, grown under expansion conditions**In all panels, nuclei are stained with DAPI (blue), while the second color represents antibody staining as follows: (A) multipotential neural progenitor lineage marker, CD133 (red); (B) neural cell adhesion molecule, NCAM (green); (C) CXCL12 (SDF-1) cytokine receptor, CXCR4 (fusin, CD184, red); (D) β-III-tubulin (green); (E) the glial fibrillary acidic protein, GFAP (green); (F) nestin. (100×).
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Markedly reduced staining with anti-FMRP antibody of neural progenitor cells from a fragile X patient**FMRP staining (green) is greatly reduced in the fragile X derived cells (A) relative to progenitor cells from an unaffected control (B). Cell nuclei counterstained with DAPI (blue); both panels, 40×.
:::
:::
Molecular studies
-----------------
Southern Blot analysis of DNA isolated from brain tissue (frontal cortex) showed the presence of a mosaic pattern. Specifically, full mutation alleles were present in 82% of the cells (435, 528, 652, 727, 847 CGG repeats) with the remaining 18% harboring a premutation allele. Sizing of the CGG repeat number by PCR analysis demonstrated the presence of a premutation allele of 90 CGG repeats and an allele with the deletion of the CGG element and the flanking region, the latter of which was not detected by Southern Blot analysis. Sequence analysis of this allele indicated the presence of a deletion extending from nucleotide 13742 to 13915 of the *FMR1*gene (GenBank accession number L29074). Southern blot analysis of DNA isolated from neural progenitor cells revealed the presence only of hypermethylated full mutation alleles (536 and 591 CGG repeats). No premutation alleles were detected by PCR analysis.
The brain *FMR1*mRNA level (frontal cortex) relative to the reference gene (glucuronidase) was low (0.08 ± 0.009, relative value). In agreement with the observed lack of FMRP expression, no *FMR1*message was detected after 40 cycles of PCR in total RNA isolated from the neural progenitor cells.
Discussion
==========
In the current work, we have successfully isolated and cultured neural progenitor cells from post-mortem brain tissue of an adult male with fragile X syndrome, which, to our knowledge, is the first example of the production of adult, human neural progenitor cells for any neurodevelopmental disorder. This result is of particular importance for the study of fragile X syndrome, since the disease-causing CGG repeat expansions have thus far been refractory to cloning into any animal or human cell model. With these cells, we hope to better understand the mechanistic link between the CGG expansion and the disease phenotype, which is known for the donor of the cells.
We have demonstrated the feasibility of generating neural progenitor cell cultures from post-mortem brains of patients with neurogenetic disease. Moreover, as the cultures were generated from cryopreserved tissue, our data suggest that cells can be harvested and processed from the required tissues in locations that are remote from the stem cell culture laboratory. This last point is of particular importance for the study of neurogenetic diseases, which generally affect only 1:20,000 to 1:200,000 live births; a technique that can be implemented at multiple institutions will be necessary to generate sufficient numbers of specimens for statistical analysis.
A major determinant of the proliferative capacity of neural progenitor cells in culture is donor age, with younger donors (particularly fetuses and infants) having greater proliferative capacity than adults \[[@B18]\]. Although progenitor cells can be obtained from fetal or embryonic sources, there are advantages to obtaining cells from post-mortem adult tissue. In using cells derived from adult tissues, one avoids the serious ethical controversies surrounding the use of fetal samples \[[@B25]-[@B27]\]. Moreover, for research aimed at understanding the effects of identified genetic defects on neural development, the phenotypic expression of a particular neurogenetic disease can be ascertained with post-mortem specimens, thus making a correlation possible between *in vitro*and *in vivo*pathophysiology. Due to the broad variability in phenotypic expression in fragile X syndrome (as with many other neurodevelopmental disorders), any such correlations are problematic using tissue obtained at fetopsy.
The method used to isolate the neural progenitor cells in the current study was adapted from the neuro-selective methods developed for culturing CNS stem cells from the brain \[[@B28]\], spinal cord \[[@B29]\], and retina \[[@B30]\] of rodents, as well as the brain of humans \[[@B18],[@B20]\]. Passaged cells expressed a number of immature markers, including the neural stem cell markers CD133 and nestin \[[@B20]\]. Although the currently recognized method for establishing multipotency is clonal derivation followed by differentiation, these cells proliferated poorly when seeded at low density and, therefore, clonal derivation has not been fruitful thus far. Nevertheless, analysis of marker expression provides evidence that these cultures give rise to cells of neuronal lineage (β-III tubulin) and glial lineage (GFAP). Our data are thus most consistent with the interpretation that the immature, highly proliferative neuroepithelial cells in the present study were multipotent neural progenitor cells.
Conclusions
===========
The successful production of neural cells from an individual with fragile X syndrome opens a new avenue for the scientific study of the molecular basis of this disorder, as well as an approach for studying the efficacy of new therapeutic agents.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
PHS conceived of the study, participated in its design and coordination, and drafted the manuscript. FT carried out the molecular genetic studies, participated in the design of the study, and helped to draft the manuscript. CMG performed the neuropathological analyses. HEN carried out the cell culture. BZ performed the immunocytochemical analyses and imaging. RJH performed the clinical evaluation. PJH participated in the design of the study and helped to draft the manuscript. All authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2350/6/2/prepub>
Acknowledgements
================
This research was supported by the CHOC Foundation for Children (PHS), the National Institute of Child Health and Human Development (HD40661, PJH), the Boory and Cooper/Kraff family funds, and by general laboratory support from the UC Davis M.I.N.D. Institute.
|
PubMed Central
|
2024-06-05T03:55:51.952139
|
2005-1-14
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545950/",
"journal": "BMC Med Genet. 2005 Jan 14; 6:2",
"authors": [
{
"first": "Philip H",
"last": "Schwartz"
},
{
"first": "Flora",
"last": "Tassone"
},
{
"first": "Claudia M",
"last": "Greco"
},
{
"first": "Hubert E",
"last": "Nethercott"
},
{
"first": "Boback",
"last": "Ziaeian"
},
{
"first": "Randi J",
"last": "Hagerman"
},
{
"first": "Paul J",
"last": "Hagerman"
}
]
}
|
PMC545951
|
Background
==========
The discovery of cyclic 3\'5\'-adenosine monophosphate (cAMP) by Earl Sutherland in the late 1950s was one of the most significant paradigm shifts in biochemistry \[[@B1]\]. This breakthrough ushered in the concept of second messengers: intracellular molecules which transmit signals in cells and are derived from an extracellular signal. In the past half century, cyclic nucleotides (both cAMP and cGMP) have been implicated in a vast array of biological phenomena in all kingdoms of life \[[@B2]\].
The ubiquitous presence of cyclic nucleotides may be due to several characteristics which make it an ideal second messenger. Cyclic nucleotides are derived in a energetically favourable reaction from common metabolites (ATP and GTP), and can be broken down into non-toxic products (inorganic phosphate and AMP/GMP). The synthesis and degradation of cyclic nucleotides are controlled by enzymes termed adenylate (or guanylate) cyclases and cyclic nucleotide phosphodiesterases, respectively \[[@B3],[@B4]\].
In plants, cyclic nucleotides have endured a checkered research history fraught with complications and setbacks. Despite this, recent work has shown unequivocally that cyclic nucleotides are present in plant cells \[[@B5],[@B6]\], and that they play key roles in the regulation of plant physiology \[[@B7]-[@B9]\]. Furthermore, the recent identification and cloning of adenylate and guanylate cyclases in plants \[[@B7],[@B10]\] may eventually give clues as to what signals the synthesis and degradation of these molecules in plants.
Cyclic nucleotides are able to bind to two distinct protein domains, CNB domains and GAF domains. CNB domains were first identified in the regulatory subunit of mammalian cAMP-dependent protein kinase (RI and RII). Since several CNB domain containing plant proteins have been shown to be directly modulated by cyclic nucleotides, this indicates that the CNB domain in plants is functionally similar to CNB domains in other organisms \[[@B11]-[@B15]\].
GAF domains were initially identified as conserved domains in light sensing molecules but are known as small molecule binding domains in cyclic nucleotide regulated phosphodiesterases, the *Anabaena*cyclic nucleotide stimulated adenylate cyclase and several other proteins \[[@B16]\]. GAF domains have been shown to bind both cAMP \[[@B17]\] and cGMP \[[@B18],[@B19]\]. Recent crystal structures of the GAF domains of human PDE2 \[pdb:1MC0\] and the yeast protein YKG9 \[pdb:1F5M\] have shown that this domain is an alpha/beta two layer sandwich with no structural or sequence homology to the CNB domain \[[@B18],[@B20]\]. Therefore, GAF and CNB domains have evolved independently to bind cyclic nucleotides.
In order to further explore the roles of cyclic nucleotides in plants, we performed a bioinformatics based analysis of the completed *Arabidopsis thaliana*and *Orzya sativa*genomes \[[@B21]-[@B23]\] in order to elucidate the potential targets of cyclic nucleotides in plants.
Results and discussion
======================
GAF domains
-----------
Based on the PDE2 crystal structure 11 residues were proposed to be involved in cyclic nucleotide binding \[[@B18]\], but comparison to the cAMP binding GAF domain of the *Anabaena*adenylate cyclase shows that these residues may only be strictly conserved in mammals. Further complicating our analysis is the fact that GAF domains are known to bind other small molecules such as 2-oxoglutarate, formate and bilins \[[@B24]-[@B26]\]. GAF domains form a structural scaffold which can be utilized to bind several possible small molecules depending on the functional groups on that scaffold. Therefore, their role in cyclic nucleotide signalling must be verified by biochemical means rather than strictly by sequence analysis. Our analysis indicated that in plants there are two types of proteins which contain GAF domains. These are the phytochrome proteins and the ethylene receptor proteins.
Phytochromes
------------
Phytochromes are light sensing signal transduction molecules which function to control several aspects of plant biology. Interestingly phytochromes were found to function upstream of cyclic nucleotides in their signal transduction pathways since their functions can be mimicked by cGMP and calcium in phytochrome knockout cells \[[@B27]-[@B29]\]. The light sensing portion of the phytochrome molecule is a covalently linked bilin molecule which is known to be bound by the GAF domain. Therefore it is unlikely that the GAF domain of the phytochrome is also able to bind cyclic nucleotides directly, although it is clear that cyclic nucleotides are somehow involved in this signalling pathway.
Ethylene receptors
------------------
Ethylene responses have been documented for nearly a century in plants. This gaseous hormone is involved in many aspects of plant physiology, including fruit ripening, organ development, germination, seedling growth, flowering and response to challenges such as pathogens and stress \[[@B30]\]. There are five putative ethylene receptor isoforms in both Arabidopsis and rice as determined by genome sequencing \[[@B31]\]. All known ethylene receptors contain a GAF domain in a cytoplasmic region amino-terminal to the kinase domain. It has been speculated that this domain may be involved in cyclic nucleotide signalling but examination of heterologously expressed, functional ETR1 \[Swiss-Prot: P49333\] showed no detectable cyclic nucleotide binding (G. E. Schaller, personal communication). There are other ethylene receptors which have GAF domains and which to our knowledge have not been tested for cNMP binding, however, to date there is no evidence of cNMP regulation of ethylene receptors. Currently the function and ligands of the GAF domain in ethylene receptors is unknown.
CNB domains
-----------
From the alignment of CNB domains in animals, bacteria and plants, it was apparent that there are some strong similarities, as well as some significant differences (Figure [1A](#F1){ref-type="fig"}). In order to visualize whether plant CNB domains could fold in a similar manner to the other well characterized CNB domains, we generated an *in silico*model. We chose the plant protein which showed highest similarity to known crystal structures (*Arabidopsis thaliana*CNTE1) and based our model on the solved crystal structures of RIα, RIIβ, HCN2, CAP and Epac2 (Figure [1B](#F1){ref-type="fig"} and see [additional file 5](#S5){ref-type="supplementary-material"}).
We then examined our model\'s overall topology as well as its cyclic nucleotide binding site. The basic fold of the domain is two anti-parallel beta sheets consisting of four strands forming a sandwich, ending with an alpha helix (the hinge region). Connecting these sheets are exposed loops, the most important of which is the phosphate binding cassette \[[@B32]\]. It is important to note that our structure models very well against all CNB domains with excellent conservation of all secondary structure and most loops. We calculated the backbone root mean square deviation for our model versus each of the templates as: RIIβ domain 1: 0.76 Å; RIIβ domain 2: 0.83 Å; Epac1 domain 1: 1.08 Å; RIα domain 1: 0.85 Å; RIα domain 2: 0.94 Å; HCN2: 0.82 Å and CAP: 1.12 Å. Our model agrees in general with a previously reported model for the Arabidopsis CNGC2 \[[@B33]\], although a detailed comparison between the two models was not performed. The use of a less distant target (atCNTE1) as well as several templates (seven compared to one) adds to the reliability of our structure. In all mammalian cAMP-binding structures solved, there is a key arginine residue (Arg 209 in RIα) which forms a salt bridge with the cNMP\'s phosphate group. This residue is absent in some CNB domains, despite evidence that at least some of these domains do in fact bind cyclic nucleotides. For example, this residue is absent in the Drosophila ether-a-go-go channel, which is known to be modulated by cyclic nucleotides \[[@B34]\]. Examination of our model shows that in the region near the phosphate, there are two residues which may functionally replace the arginine, Tyr 91 and Ser 92 (Figure [1B](#F1){ref-type="fig"}). In bacterial CAP, hydrogen bonding is provided to the phosphate by the Arg 82 sidechain, Ser 83 amide nitrogen atom and sidechain hydroxyl group, as well as a water molecule. In some mammalian isoforms, the serine residue is changed to an alanine and therefore is only able to provide backbone hydrogen-bonding. In our plant atCNTE1 model, the serine residue is conserved, but the arginine residue is missing. Since there was no good template to model the phosphate binding cassette onto, our model only approximates the position of this loop, and will require verification by other structural studies.
The hydroxyl and amide groups of Ser 92, as well as the hydroxyl group of Tyr 91 are all within proximity of the cNMP phosphate and could play a role in stabilizing the cyclic nucleotide (see blue residues in Figure [1](#F1){ref-type="fig"}). Examination of the region which contacts the base, indicates that our model is most similar to the structure of CAP in this region, so it is likely that the base moeity of a cNMP is bound in a syn orientation as in CAP. This is in agreement with a previously reported atCNGC2 model \[[@B33]\]. Further analysis of the binding site for the nucleotide indicates that it is likely cGMP which binds to atCNTE1. This conclusion is based upon the presence of three residues (Tyr 80, Ser 92 and Ser 109) which could potentially differentiate between cyclic nucleotides, each of which has a preference for cGMP (Figure [1](#F1){ref-type="fig"} and \[[@B35],[@B36]\]). Other conserved structural features of our model are the hydrophobic pocket forming residues Tyr 36, Val 42, Val 43, Tyr 53, Leu 55, Ala 60, Phe 82, Ala 93 and Val 95 and Leu 105 (see green residues in Figure [1](#F1){ref-type="fig"}) as well as several conserved glycine residues which are involved in turns between the beta strands (39, 58 and 83) and the helix capping Asp 109. This residue signals the end of the hinge region alpha helix and is present in most CNB domains examined.
Phylogenetic analysis indicates that the plant CNB domains segregate into three subfamilies (Figures [2](#F2){ref-type="fig"} and [3](#F3){ref-type="fig"}). The phylogenetic distribution of the CNB domain matches their domain context in that CNGC, shaker-type and CNTE proteins form separate groups. Furthermore, for each of the three protein classes, the CNB phylogeny matches the phylogeny of the full-length protein, implying that these proteins obtained the CNB domain prior to isoform duplication (Figure [3](#F3){ref-type="fig"}, \[[Additional file 1](#S1){ref-type="supplementary-material"}\], \[[@B37]\]). Since all three branches have been detected in both Arabidopsis and Oryza, it is likely that the specific plant cyclic nucleotide responses developed prior to monocot-dicot divergence. We did not find CNB domains in any protein kinases, transcription factors or guanine nucleotide exchange factors in our analysis. Each of the three classes of CNB domain containing proteins will be discussed below.
Cyclic nucleotide gated ion channels
------------------------------------
Plant CNGC ion channels were first identified in a screen for calmodulin binding partners in barley \[[@B38]\]. There are now known to be 20 CNGC proteins in Arabidopsis likely indicating a high level of channel redundancy \[[@B37],[@B39]\]. We also detected 16 CNGC proteins in rice by examination of the TIGR Rice Genome Annotaion Resource \[[@B40]\].
Electrophysiological studies have shown the CNGC channels to be permeable to potassium, sodium and calcium \[[@B13]-[@B15],[@B41]-[@B43]\]. Cyclic nucleotides have been shown to activate channel opening in all CNGC proteins examined thus far leading to an influx of cations into the cell \[[@B13]-[@B15],[@B33]\]. Mutagenic screens have shown that mutations in atCNGC2 and atCNGC4 create faulty pathogenic reactions \[[@B13],[@B44]\]. When taken together with data showing that cyclic nucleotides are necessary for pathogen responses and that calcium and potassium influxes are characteristic of early phases of plant pathogen responses \[[@B45]\], this seems to imply that cyclic nucleotides may play a role in controlling plant immune responses.
Finally, work by Maathuis and Sanders \[[@B8]\] has shown that cyclic nucleotides can modulate sodium uptake in Arabidopsis plants, implying that there is a cyclic nucleotide controlled channel which plays a role in salinity tolerance. They showed that cyclic nucleotides are required for limiting sodium uptake in root protoplasts, but the exact molecule (cAMP or cGMP) responsible for this effect has not been pinpointed.
Shaker-type potassium channels
------------------------------
Plant potassium channels fall into two classes, the KCO channels and the shaker-type channels \[[@B37]\]. In addition to the 9 shaker-type channels described in *Arabidopsis thaliana*, we have found 10 channels in *Oryza sativa*. A variety of mutational studies have implicated the shaker-type channels in several key processes involving the movement of potassium including: from the soil (AKT1, KAT3), long distance transport (AKT2), transport into growing pollen tube (AKT6), secretion into xylem sap (SKOR) and transport during guard cell opening either into the cell (KAT1, KAT2) or out of the cell (GORK) \[[@B37]\].
Shaker-type channels are voltage dependent outward (GORK, SKOR) or inward (KAT and AKT) rectifying channels. Analysis of heterologously expressed channels have shown that cyclic nucleotides function to adjust the activation potential of these channels \[[@B11],[@B12]\]. Since cyclic nucleotides have already been implicated in some of the processes controlled by shaker-type channels \[[@B7]-[@B9],[@B46]\], it is reasonable to believe that cyclic nucleotides are physiological regulators of shaker-type potassium channels.
Cyclic nucleotide regulated thioesterases
-----------------------------------------
Initially we detected a short CNB containing protein which was only slightly larger than the domain itself. Sequencing of the EST provided by the Arabidopsis Biological Resource Center \[[@B47]\] showed the protein was actually mis-annotated by the automated gene-finding algorithm. Further analysis indicated that there are two isoforms of this protein in Arabidopsis and one in rice. Each protein contains an amino-terminal CNB domain and a carboxy-terminal acyl-CoA thioesterase domain. Searches of other partially sequenced plant genomes and EST databases indicated that these proteins are present in several plant species, but not in any other division of life and thus represents a novel plant-specific cyclic nucleotide target. Comparison of these protein sequences indicates a high level of conservation, including residues conserved for both catalysis and cyclic nucleotide binding domain structure. Arabidopsis CNTE1 had previously been partially characterized as a thioesterase and shown to have activity versus both 16:0-CoA and 18:1-CoA when over-expressed and partially purified from *E. coli*\[[@B48]\].
Fatty acid synthesis requires the use of acyl-CoA\'s as building blocks for incorporation into lipids. It is therefore possible, that these thioesterases function as scavengers which remove \"irregular\" fatty acids from the pool of available building blocks \[[@B48]\]. Furthermore, the thioesterases could divert fatty acids away from biosynthetic pathways and β-oxidation during germination or during stressful conditions. In most cases when a small molecule binding domain is connected to a catalytic domain on the same polypeptide, the catalytic domain is regulated by the small molecule \[[@B49]\]. The conservation of this protein across planta indicates that the CNB domain likely has a role in controlling the thioesterase activity of this enzyme, but it is unknown at this time exactly what role cyclic nucleotides play in this process. In order to address this we cloned and tried to express the atCNTE1 protein in *E. coli*, however after extensive trials we were unable to express and purify soluble protein.
A cyclic nucleotide dependent protein kinase in plants?
-------------------------------------------------------
We found no PKG or PKA regulatory subunit homologs in the Arabidopsis genome. There has been a long standing controversy in the plant field as to the existence of a plant cyclic nucleotide dependent kinase \[[@B50]-[@B52]\]. As PKA is the major cAMP target in mammalian cells we chose a biochemical approach to further explore the possibility that a PKA-like enzyme may be present in Arabidopsis.
We performed protein kinase assays with extracts of *Arabidopsis thaliana*using the PKA substrate Kemptide. Kemptide is a peptide which has a motif which was confirmed as the optimal substrate for PKA \[[@B53],[@B54]\] and is routinely used in mammalian PKA assays. As Figure [4A](#F4){ref-type="fig"} shows, there was no detectable increase in kinase activity in the plant cell extracts when cAMP or cGMP are added. Fractionation of extracts, as well as testing a range of cyclic nucleotide concentrations also did not allow us to detect any differences in kinase activity with addition of cyclic nucleotides (data not shown). For comparison, adipocyte extracts (a cAMP responsive mammalian tissue) were assayed as well, illustrating the large increase in protein kinase activity in these cells when cAMP is added. Furthermore, blotting of *A. thaliana*extracts with polyclonal antibodies raised against mammalian PKA subunits (both the catalytic and the RIIα subunit) reveals that no structurally similar proteins are present in this extract (Figure [4B](#F4){ref-type="fig"}). Blotting with a monoclonal antibody to the RIIβ subunit gave similar results (not shown). Although there is a weak band present in the Arabidopsis extract which cross-reacted with the catalytic subunit polyclonal antibody, it is likely unrelated to cyclic nucleotide signalling. Protein kinase catalytic domains are very highly conserved \[[@B55],[@B56]\] and therefore a minor amount of cross-reactivity is not unexpected. Further adding to the validity of the western blotting experiment is the observation that several studies have shown that true PKA-like enzymes in non-mammalian eukaryotes do cross react with antibodies raised against the regulatory subunits of mammalian PKA \[[@B57],[@B58]\], whereas this is not detected in our plant extracts.
The lack of evidence for kinase activity could be attributed to substrate specificity, differences in binding affinity or expression levels in plant extracts relative to mammalian extracts, so our experimental approach does not exhaustively rule out the possibility of a cNMP dependent kinase activity. However, we feel that these data in concert with the genomic and blotting data strongly suggest that there is no cNMP dependent kinase in plants. Finally, our data imply that if such a protein exists, it would bear little or no sequence, structural or biochemical similarity to the classically studied mammalian enzyme.
Conclusions
===========
As the understanding of cyclic nucleotide signalling in a variety of systems has progressed, it has been increasingly difficult to describe a general role for cyclic nucleotides in biology. They control \'well-fed\' gene transcription in bacteria, and modulate signal transduction and ion currents in mammals, resulting in a large number of possible physiological responses.
This analysis is potentially limited in that it only analyses cNMP domains which have already been previously identified and characterized in other systems. However, conservation of CNB and GAF domains as the only known cyclic nucleotide binding domains present over a wide cross-section of life indicates that these domains are likely to control most, if not all cyclic nucleotide responses. It is possible however, that plants have evolved entirely novel domains which can be modulated by these second messengers. It will be interesting to compare this *in silico*analysis with future biochemical data regarding the direct effectors of cyclic nucleotide signalling in plants. It is interesting that no homologous proteins in the CNGC, shaker-type or type II acyl-CoA thioesterase families have been found which lack CNB domains. This implies that cyclic nucleotide binding is indispensable to their cellular role.
Although it would have been interesting if this analysis revealed more novel classes of plant cyclic nucleotide binding proteins, the fact that (with the exception of CNTE) all cyclic nucleotide binding proteins had been previously identified indicates that the previously attained biochemical data agrees with our bioinformatic evidence.
The identification of no transcription factors or protein signal transduction molecules with CNB domains implies that cyclic nucleotides may be unable to directly modify the proteome of plant cells. This is in stark contrast to bacterial, yeast and mammalian systems. The only common domain context of CNB domains in animals and plants is the CNGC channels, however, even these channels appear to have evolved independently \[[@B39],[@B59]\]. Therefore it is clear that the roles of cyclic nucleotides in prokaryotic and eukaryotic, as well as plant and animal systems differ and that evolutionarily distant branches of life have evolved different mechanisms by which these molecules are utilized. It is worth pointing out that the ubiquitous presence of cyclic nucleotides in all forms of life may indicate that although the means by which this particular biochemical tool is used differ, it is still an indispensable component of biology\'s toolbox.
Methods
=======
Bioinformatics
--------------
In order to identify the proteins which contain CNB or GAF domains, we initially used the Simple Modular Architecture Research Tool (SMART at smart.embl.heidelberg.de; \[[@B60]-[@B62]\]) to scan all predicted Arabidopsis proteins for CNB and GAF domains in the EMBL, TIGR or NCBI databases. Once redundancies were removed, a list of proteins was generated \[see [additional file 2](#S2){ref-type="supplementary-material"}\]. In order to ensure broad coverage of possible variants, we also examined the Interpro collection of protein sequence analysis algorithms, all of which use slightly different methods \[[@B63]\]. As an additional method, the predicted proteins of the Arabidopsis genome were searched using the BLAST algorithm \[[@B64]\]. As search bait, we used several known cyclic nucleotide binding domains including those from GAF domains (human PDE2A \[Swiss-Prot: O00408\] and Anabaena cyaB1 \[Trembl: P94181\]) as well as CNB domains (human CGK2 \[Swiss-Prot: Q13237\], human RIIβ \[Swiss-Prot: P31323\], human Epac2 \[Swiss-Prot: Q8WZA2\], human rod CNGC \[Swiss-Prot: P29973\] and *E. coli*CAP \[pir: E86000\]). This yielded no new inclusions to our list of proteins, but did confirm each of our previous entries. For examination of the *Oryza sativa spp*. Japonica genome we performed BLAST searches using the aforementioned baits, as well as each of the Arabidopsis proteins. This search was performed using the Blast utility of the TIGR rice database \[[@B40]\]. The criterion for inclusion was that the CNB or GAF domain had to match the consensus motif with an E-value of less than 0.5 over the entire domain as determined by SMART. For newly identified proteins from the *Orzya sativa*, we named them so that they agreed best with the nomenclature of Maser *et al*. \[[@B37]\] \[see additional files [1](#S1){ref-type="supplementary-material"}, [2](#S2){ref-type="supplementary-material"}, [3](#S3){ref-type="supplementary-material"}\]. Sequence alignments were performed using the ClustalX \[[@B65]\] or T-COFFEE algorithms \[[@B66]\]and then inspected visually. Neighbor-joining trees were generated by ClustalX or PHYLIP \[[@B67]\], then were visualized with TreeView \[[@B68]\]. Trees generated using a variety of analysis methods (parsimony, distance and maximum likelihood) yielded similar results to the neighbor-joining trees.
Sequencing of atCNTE1
---------------------
One protein, which appeared to contain only a cyclic nucleotide binding domain and no other motifs was found in the Arabidopsis database. We obtained the clone corresponding to this putative gene from the Arabidopsis Biological Resource Center and sequenced it. Sequencing was performed at the University of Calgary Core Sequencing Facility. We determined that the gene prediction algorithm which scanned the genome improperly predicted the intron/exon structure of this gene. The new gene, which we named cyclic nucleotide regulated thioesterase 1 (atCNTE1) was deposited in the NCBI database \[GenBank: AY874170\]. A subsequent BLAST search using this gene found another isoform of this gene in Arabidopsis (atCNTE2) and one isoform in Rice (osCNTE1) which we also included in our analysis.
Theoretical model of atCNTE1
----------------------------
We generated a model of the atCNTE1 cyclic nucleotide binding domain (residues 28--117) based on an alignment of atCNTE1 with the CNB domains of RIIβ \[pdb: 1CX4\] \[[@B69]\], RIα \[pdb: 1RGS\] \[[@B70]\], CAP \[pdb: 1CGP\] \[[@B71]\], HCN2 \[pdb: 1Q3E\] \[[@B72]\] and Epac1 \[pdb: 1O7F\] \[[@B32]\]. This was submitted to the SWISS-MODEL server via the DeepView program \[[@B73]\]. The alignment was iteratively refined to allow for best agreement of sequence and structural similarity.
Cyclic nucleotide dependent protein kinase assays
-------------------------------------------------
Unless otherwise indicated all chemicals were purchased from Sigma-Aldrich. Assays were performed on extracts of Arabidopsis cells grown in suspension culture \[[@B74]\] or isolated male Wistar rat adipocytes from epididymal fat pads \[[@B75]\]. Both cell types were homogenized in 50 mM Tris pH 7.5, 5% (v/v) glycerol, 0.2 mM phenylmethylsulfonyl fluoride, 1 mM benzamidine and 0.1% (v/v) 2-mercaptoethanol. Adipocytes were lysed by 10 strokes of a dounce homogenizer while plant cells were lysed by two passages through a french press cell at 15 000 psi. The extracts were clarified by centrifugation for 15 min at 4000 RPM in a SS34 rotor at 4°C. These extracts were assayed for kinase activity using ^32^P-γ-ATP (Amersham-Pharmacia), 30 μM Kemptide substrate, 50 mM HEPES pH 7.4, 1 mM dithiothreitol and 10 μM cyclic nucleotide as specified. Reactions were allowed to occur for 10 minutes at 30°C and assays were terminated by spotting onto squares of P81 paper followed by extensive washing with 75 mM phosphoric acid \[[@B76]\]. Assays were performed in duplicate on three separate preparations with error bars indicating the standard error between preparations (n = 3). Protein concentration was determined by the method of Bradford with bovine serum albumin (ICN Biomedicals) as a standard \[[@B77]\].
Western blotting of extracts
----------------------------
Bovine heart PKA catalytic subunit and RII were purified to homogeneity \[[@B78],[@B79]\]. Purified protein was injected into rabbits and serum was obtained according to standard methods \[[@B80]\]. Extracts of adipose and plant cells were prepared as described above. Samples were boiled into SDS-PAGE buffer and separated on a 10% denaturing gel \[[@B81]\]. The proteins were then transferred to nitrocellulose for 2 h at 100V and blocked overnight with 5% (w/v) skim milk powder. Blots were probed with antibodies for 1 h and visualized by enhanced chemiluminesence. The PKA catalytic subunit antibody was affinity purified according to \[[@B82]\] and used at 0.5 μg/mL while RII was used as crude immune serum at a 5000X dilution. For the RII western blots, both a polyclonal and a monoclonal antibody (anti-RIIβ BD Transduction Laboratories) gave identical results.
List of abbreviations
=====================
CAP catabolite activator protein
cAMP 3\'5\'-cyclic adenosine monophosphate
CGK2 cGMP dependent protein kinase 2
cGMP 3\'5\'-cyclic guanosine monophosphate
cNMP 3\'5\'-cyclic nucleotide (cAMP or cGMP)
CNB cyclic nucleotide binding
CNGC cyclic nucleotide gated channel
CNTE cyclic nucleotide dependent thioesterase
Epac exchange protein directly activated by cAMP
GAF domain found in c[G]{.underline}MP-phosphodiesterases, [a]{.underline}denylyl cyclases and [F]{.underline}hlA
GORK guard cell outward rectifying potassium channel
PKA protein kinase A
SKOR stellar potassium outward rectifying channel
SMART simple modular architecture research tool
Authors\' contributions
=======================
DB performed the bioinformatic analysis, modeling and biochemical studies and drafted the manuscript. MEF participated in the modeling. GBG conceived of the study and participated in its design and co-ordination. All authors have read and approved the final manuscript.
Supplementary Material
======================
::: {.caption}
###### Additional File 5
**Co-ordinate file of model generated for Figure**[1b](#F1){ref-type="fig"}.
:::
::: {.caption}
######
Click here for file
:::
::: {.caption}
###### Additional File 1
**Phylogenetic analysis of CNGC and shaker-type channels from *Arabidopsis thaliana*and *Oryza sativa*.**Un-rooted neighbor-joining trees were constructed for (A) shaker-type and (B) CNGC channels using full-length sequences. For the shaker-type channels, numbers at nodes indicate number of trees out of 100 in which the node occurred. These are omitted in (B) for clarity. Scale indicates number of differences per residue. Trees were generated using ClustalX \[[@B65]\] and visualized with TreeView \[[@B68]\].
:::
::: {.caption}
######
Click here for file
:::
::: {.caption}
###### Additional File 4
**Relevant Sequence Alignments**Protein sequence alignments upon which phylogenetic analyses in Figure [2](#F2){ref-type="fig"}, and [additional file 2](#S2){ref-type="supplementary-material"} are based.
:::
::: {.caption}
######
Click here for file
:::
::: {.caption}
###### Additional File 2
**Cyclic nucleotide binding proteins in *Arabidopsis thaliana*and *Oryza sativa*.**Proteins containing CNB domains including species, name, aliases, accession number, other domains present (AR indicates ankyrin repeat, CaM indicates calmodulin binding domain and TE indicates type II acyl CoA thioesterase domain), sequence length, residues encompassing the CNB domain and E-value for the CNB domain as determined by SMART \[[@B62]\]. All accession numbers are from NCBI except osCNGC5b, osCNGC5c and osCNGC18 which were obtained from the MIPS Oryza sativa database <http://mips.gsf.de/proj/plant/jsf/rice/index.jsp>.
:::
::: {.caption}
######
Click here for file
:::
::: {.caption}
###### Additional File 3
**Protein sequences of all sequences analysed in this manuscript.**Sequences are named according to details in the text and are in FASTA format.
:::
::: {.caption}
######
Click here for file
:::
Acknowledgements
================
This work was supported by the Natural Sciences and Engineering Research Council of Canada. DB is supported by a Natural Sciences and Engineering Research Council postgraduate scholarship. MEF is a Biomedical Scholar supported by the Alberta Heritage Foundation for Medical Research.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Analysis of plant CNB domains**. (A) Arabidopsis CNB domains (CNTE1, KAT1 and CNGC2) were aligned against several well studied CNB domains including regulatory subunits of PKA (RIα and RIIβ), Epac1, Epac2, and cyclic GMP dependent kinase 2 (CGK2) from humans, HCN2 from mouse and *E. coli*CAP. Highlighted on the alignment are glycine residues involved in loop structures (dark grey arrows), residues forming the hydrophobic pocket for cNMP binding (green arrows) and residues proposed to contact the phosphate of the cNMP (blue arrows). The highly conserved helix capping acidic residue is shown in red. Secondary structure is denoted by arrows above the alignment, with light blue for alpha helices and pink for beta sheets and is based on the secondary structure of HCN2. (B) A homology model of atCNTE1 was generated from the known structures of CNB domains. Key residues are shown as stick representations and are colored and labeled according to the color scheme described in (A). The cGMP ligand is shown in magenta and is based on the structure of cGMP bound to HCN2 \[pdb: 1Q3E\] superimposed over our model. Figure was generated with Molscript \[83\] and Raster3D \[84\].
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Domain structures of representative CNB domain containing proteins in plants**. Scaled images were generated using SMART \[60, 62\] A) atCNGC2 showing the ion channel, CNB domain and IQ (Calmodulin binding) domain. B) AKT1 showing the ion channel, CNB domain and Ankyrin repeats (ANK). C) atCNTE1 showing the CNB domain and the Acyl-CoA thioesterase domain.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Phylogenetic analysis of plant CNB domains**. Un-rooted neighbor-joining tree of CNB domains in *Arabidopsis thaliana*and *Oryza sativa*. Shading represents the three classes of CNB domains in plants; shaker-type potassium channels are shown in green, acyl-coA thioesterases are shown in red and CNGC channels are shown in blue. Numbers on the nodes indicate the number of possible trees out of 1000 in which that node was present. Scale represents the number of differences per residue. B) This is an expanded view of the unlabelled region in Figure 3A showing the closely related CNGC proteins more clearly. Tree was generated using ClustalX \[65\] and visualized using TreeView \[68\]. See [additional file 4](#S4){ref-type="supplementary-material"} for sequence alignments.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Biochemical evidence for lack of a cyclic nucleotide dependent kinase in *Arabidopsis thaliana***. (A) Protein kinase assays using Kemptide as a substrate. Assays were conducted on identically prepared extracts of Arabidopsis and rat adipose tissue in the presence or absence (control) of 10 μM cyclic nucleotide as indicated. Scale is offset in order to visualize both sets of results. All assays were performed in duplicate from three separate preparations and error bars indicate standard error for three separate preparations. (B) Western blotting of extracts with PKA catalytic (PKAcs) and regulatory (RII) subunit polyclonal antibodies. The PKAcs antibody was affinity purified according to \[82\] and used at 0.5 μg/mL while the RII antibody was used as crude serum at 5000X dilution. Lanes are as follows (A), 10 ng of purified bovine PKAcs or RII, (B) 25 μg clarified crude Arabidopsis extract, (C) 25 μg clarified crude rat adipocyte extract. Positions of mammalian PKA and RII are indicated with arrows.
:::
:::
|
PubMed Central
|
2024-06-05T03:55:51.954701
|
2005-1-11
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545951/",
"journal": "BMC Bioinformatics. 2005 Jan 11; 6:6",
"authors": [
{
"first": "Dave",
"last": "Bridges"
},
{
"first": "Marie E",
"last": "Fraser"
},
{
"first": "Greg BG",
"last": "Moorhead"
}
]
}
|
PMC545952
|
Background
==========
Posttranslational modification of eukaryotic proteins with lipids is a prevalent mechanism for controlling the subcellular localization and activity of these proteins \[[@B1]\]. Most proteins terminating in a CaaX sequence (C, cysteine; \"a\", generally an aliphatic residue; X, the carboxy-terminal residue) are subject to modification by isoprenoid lipids via their ability to serve as substrates for protein farnesyltransferase or protein geranylgeranyltransferase type I \[[@B2]\]. Following covalent attachment of the farnesyl or geranylgeranyl isoprenoid to the Cys thiol of the CaaX sequence, the majority of these proteins are further processed by removal of the carboxyl-terminal aaX residues by an endoprotease termed Rce1 and methylation of the newly-exposed carboxyl group of the isoprenylated cysteine residue by an enzyme termed isoprenylcysteine carboxylmethyltransferase (Icmt) \[[@B3],[@B4]\].
While isoprenoid modification of CaaX proteins is the principal determinant of their membrane targeting, the subsequent steps of proteolysis and methylation are clearly important. Studies performed using small molecule prenylcysteines such as N-acetyl-S-farnesyl-L-cysteine (AFC) and N-acetyl-S-geranylgeranyl-L-cysteine (AGGC) \[[@B5]\] showed that carboxyl methylation is a critical determinant of the hydrophobicity of the farnesylated moiety, whereas in the case of geranylgeranylated counterpart the effect is much less. Similar results were obtained in studies with short prenylated peptides \[[@B6]\]. Ras proteins, which are primarily farnesylated, are largely mislocalized in cells in which the Icmt gene has been disrupted and in cells in which the methylation pathway has been perturbed \[[@B7],[@B8]\], suggesting that this processing step is critical for trafficking and/or stable membrane association. In more biological settings, disruption of either the Rce1 or Icmt genes in mice results in embryonic lethality \[[@B9],[@B10]\], and a very recent study has shown that genetic disruption of Icmt in cells dramatically attenuates their ability to be transformed by the K-Ras oncogene \[[@B11]\]. Altogether, a picture has emerged in which C-terminal methylation of prenylated proteins is thought to contribute substantially to their affinity for cell membranes \[[@B12]-[@B15]\], to influence rates of protein turnover \[[@B11],[@B16]\], and to facilitate functional interactions with other proteins \[[@B17]-[@B19]\].
Although Icmt is a protein methyltransferase, the enzyme can efficiently modify simple prenylcysteines \[[@B6],[@B20]-[@B22]\], a property that has the potential to greatly facilitate kinetic analysis of the enzyme. The only kinetic study of an isoprenylcysteine methyltransferase activity to date was performed prior to the molecular identification of Icmt and involved kinetic analysis of a methyltransferase activity in retinal rod outer segment membranes that was capable of modifying the small molecule substrate N-acetyl-L-farnesylcysteine (AFC) \[[@B20]\]. However, the cloning of mammalian Icmt \[[@B23]\], and the development of an expression system to produce recombinant protein \[[@B8]\], now allows an unambiguous analysis of the properties of this important enzyme. Here we report studies of the kinetic mechanism of recombinant human Icmt produced in Sf9 cells. This analysis was also facilitated by the synthesis of a new small molecule Icmt substrate, biotin-S-farnesyl-L-cysteine (BFC), that allowed development of a facile assay for enzyme activity. Through standard Michaelis-Menten analysis as well as product competition and dead-end inhibitor studies, we demonstrate that catalysis by Icmt proceeds through an ordered sequential mechanism in which the substrate AdoMet binds first and its product AdoHCy is released last.
Results
=======
Characterization of the small molecule Icmt substrate, BFC
----------------------------------------------------------
To overcome the inherent problems associated with use of S-prenylated peptides and proteins as Icmt substrates (expense in production, unwieldy separation techniques, etc), we synthesized a small molecule substrate for the enzyme that contains an appended biotin moiety (Fig. [1A](#F1){ref-type="fig"}) to facilitate the separation required for product analysis. The coupling of biotin moiety to the free amino group of the S-farnesylcysteine, FC, was readily achieved through the use of commercial biotin N-hydroxysuccinimide ester and the product BFC could be readily purified by precipitation or by reverse-phase chromatography (see *Methods*). BFC was characterized in terms of its ability to serve as a substrate for recombinant human Icmt using a standard *in vitro*methylation assay (Fig. [1B](#F1){ref-type="fig"}). The apparent K~m~of BFC was 2.1 ± 0.4 μM, which is essentially identical to the K~m~of 2.1 μM determined for farnesylated, Rce1-proteolyzed K-Ras (data not shown). Hence, BFC is a comparable to K-Ras as an Icmt substrate and has the added advantages of being a stable small molecule that can be readily captured on avidin resins.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**A new small-molecule substrate of Icmt**A. Structure of biotin-S-farnesyl-L-cysteine (BFC) B. Utilization of BFC as a substrate by recombinant human Icmt. Shown are Michaelis-Menten and Lineweaver-Burk plot (inset) of the formation of BFC-\[^3^H\]methylester as a function of BFC concentration. The data shown are the mean of duplicate determinations from a single experiment and are representative of two additional experiments.
:::
:::
Distinguishing ping-pong vs. sequential kinetic mechanisms for human Icmt
-------------------------------------------------------------------------
Detailed knowledge of the kinetic mechanism of an enzyme is very important in providing the foundation for analysis of structure-function studies and in inhibitor design studies. For a two substrate enzyme reaction such as that catalyzed by Icmt, there are three basic mechanisms: i) random sequential, ii) ordered sequential, and iii) ping-pong \[[@B28],[@B29]\]. Initial velocity studies can be used to distinguish between a ping-pong mechanism, in which a product is released between the addition of the two substrates, and the sequential mechanisms, in which both substrates must bind before any product is released. Hence, in the first series of studies, the initial velocity of the Icmt reaction at different substrates concentrations was determined. In a ping-pong mechanism for an enzyme with substrates A and B, the plots of 1/v versus 1/\[A\] at different fixed \[B\], and vice versa, should yield parallel lines \[[@B28]\]. In a sequential mechanism, either ordered or random, the family of reciprocal plots should intersect above the X axis (if the binding of A increase the interaction with B), below the X-axis (if the binding of A decreases the interaction with B), or on the X-axis (if the binding of A has no effect on the interaction of B) \[[@B28]\]. The data obtained in the experiments where AdoMet concentration was varied at a series of fixed BFC concentrations is shown in Fig [2A](#F2){ref-type="fig"}, and Fig. [2B](#F2){ref-type="fig"} shows the data obtained when BFC concentration is varied at fixed AdoMet concentrations. These results show that, in each case, the families of reciprocal plots are not parallel and intersect above the X-axis, ruling out the possibility of a ping-pong mechanism. Moreover, the data for the experiments in which BFC concentration was varied at fixed AdoMet concentrations show an intersection of the reciprocal plots on Y-axis (Fig. [2B](#F2){ref-type="fig"}), indicating that the reaction behaves as rapid-equilibrium bireactant system \[[@B28]\].
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Distinguishing ping-pong vs. sequential kinetic mechanisms for human Icmt**. A. Lineweaver-Burk plot for the methylation of BFC carried out in the presence of (◆) 2 μM, (□) 4 μM and (▲) 8 μM of BFC and varied \[AdoMet\]. B. Lineweaver-Burk plot for the methylation of BFC carried out in the presence of (◇) 5 μM, (■) 10 μM, (○) 20 μM of AdoMet and varied \[BFC\]. For both experiments, the incubation time was 20 min and 0.5 μg of Sf9 membranes containing recombinant human Icmt protein were used for each condition. Data shown are the mean of duplicate determinations from a single experiment, and are representative to two such experiments.
:::
:::
Distinguishing ordered sequential vs. random sequential mechanisms for human Icmt
---------------------------------------------------------------------------------
While the initial velocity studies described above can distinguish between a ping-pong mechanism and a sequential mechanism, they cannot be used to distinguish an ordered from a random mechanism. These two types of sequential mechanisms can be distinguished, however, through the use of so-called \"dead-end\" substrates that cannot go on to form products \[[@B30]\]. Hence, we employed such a dead-end inhibitor of Icmt, farnesylthioacetic acid (FTA) \[[@B31]\], in kinetic studies. To distinguish between the two types of inhibition patterns this compound was examined for its inhibitory properties with respect to both the cognate (BFC) and noncognate (AdoMet) substrates under conditions where the fixed substrate was present at nonsaturating concentrations. The results obtained demonstrate that the FTA is a competitive inhibitor with respect to BFC, with a measured K~i~of 1.2 ± 0.2 μM (Fig, [3A](#F3){ref-type="fig"}, see also Table [I](#T1){ref-type="table"}). In contrast, the experiments performed with varying AdoMet concentrations show a family of parallel lines (Fig. [3B](#F3){ref-type="fig"}), which is a characteristic profile for an uncompetitive inhibition \[[@B28],[@B29]\]. If FTA was able to combine with the free enzyme (i.e. the condition where its cognate substrate BFC binds first), the plots in Fig. [3B](#F3){ref-type="fig"} would have yielded lines that intersect at negative 1/v and 1/\[AdoMet\] values, i.e. FTA should have shown a mixed-type inhibitor with respect to AdoMet \[[@B20]\]. These results indicate that Icmt reaction occurs through an ordered mechanism in which AdoMet binds first to the enzyme followed by BFC binding to the AdoMet-Icmt binary complex.
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Distinguishing ordered sequential vs. random sequential mechanisms for human Icmt**. A. Lineweaver-Burk plot for the dead-end inhibition of BFC methylation by FTA at a fixed concentration of AdoMet. Assays were conducted in the presence of fixed \[AdoMet\] at 5 μM and either (◇) 0 μM, (■) 5 μM or (○) 10 μM of FTA at the indicated concentrations of BFC. B. Lineweaver-Burk plot for the dead-end inhibition of BFC methylation by FTA at a fixed concerntration of BFC. Assays were conducted in the presence of fixed \[BFC\] at 4 μM and either (◆) 0 μM, (□) 2.5 μM or (▲) 5 μM of FTA at the indicated concentrations of AdoMet. Assays were conducted as described in the legend to Fig. 2. Data shown are the mean of duplicate determinations from a single experiment, and are representative to two such experiments.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
K~i~values and type of inhibition for the interaction of a dead-end inhibitor (FTA) and reaction products (AdoHCy, AFCME) with Icmt. The K~m~values obtained for the two substrates of the reaction, BFC and AdoMet, are also shown for comparaison.
:::
**Substrate\\Inhibitor** **AdoHcy** **FTA** **AFCME**
-------------------------- ----------------- ------------------- --------------------
**BFC** **Competitive** **Competitive** **Noncompetitive**
K~m~= 2.1 ± 0.4 μM 3.59 ± 1.03 μM 1.17 ± 0.16 μM 1.91 ± 0.65 μM
**AdoMet** **Competitive** **Uncompetitive** **Mixed-type**
K~m~= 7.8 ± 1.2 μM 3.54 ± 1.12 μM 2.43 ± 0.70 μM
:::
Distinguishing the order of product dissociation from human Icmt
----------------------------------------------------------------
Two products are formed in the reaction catalyzed by Icmt, a methylated prenylcysteine (*in vivo*on proteins, but *in vitro*also on small molecules or peptides) and AdoHcy. To determine the order of the dissociation of these two products, we performed product inhibition using the prenylcysteine methylester, AFCME, and AdoHcy. Examination of the initial rates of product formation at fixed BFC and varying AdoMet concentrations in presence of three concentrations of AdoHcy (Fig. [4A](#F4){ref-type="fig"}), revealed that AdoHcy is a competitive inhibitor with respect to AdoMet with an apparent K~i~of 3.5 ± 1.0 μM (Table [I](#T1){ref-type="table"}). This finding is consistent with both AdoMet and AdoHcy binding to the same form of Icmt, ie. the free enzyme, suggesting that AdoMet binds first in an ordered mechanism and that AdoHcy is the last product released. When the same experiment was performed at varying BFC concentrations, a very similar pattern was observed (Fig. [4B](#F4){ref-type="fig"}), revealing that AdoHcy is also a competitive inhibitor with respect to BFC with a K~i~of 3.6 ± 1.0 μM. While a classic ordered sequential mechanism in which AdoHcy was released last would result in this type of inhibition being non-competitive or mixed-type, this type of behavior in which a single product binds competitively with both substrates can also be observed in ordered bireactant systems that proceed via a rapid equilibrium process \[[@B28]\].
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Distinguishing the order of product dissociation from human Icmt: Inhibition studies with the AdoHcy product**. A. Lineweaver-Burk plot for the product inhibition of BFC methylation by AdoHcy at a fixed concentration of BFC. Assays were conducted in the presence of fixed \[BFC\] at 4 μM and either (◆) 0 μM, (□) 5 μM or (▲) 10 μM of AdoHcy at the indicated concentrations of AdoMet. Assays were conducted as described in the legend to Fig. 2. Data shown are the mean of duplicate determinations from a single experiment, and are representative to three such experiments. B. Lineweaver-Burk plot for the product inhibition of BFC methylation by AdoHcy at a fixed concentration of AdoMet. Assays were conducted in the presence of fixed \[AdoMet\] at 5 μM and either (◇) 0 μM, (■) 2.5 μM, (○) 5 μM, or (●) 10 μM μM of AdoHcy at the indicated concentrations of BFC. Assays were conducted as described in the legend to Fig. 2. Data shown are the mean of duplicate determinations from a single experiment, and are representative to more than \>4 such experiments.
:::
:::
The product inhibition experiments were next repeated with AFCME as the second class of product inhibitor. An ordered sequential mechanism with initial departure of the methylated prenylcysteine product requires that noncompetitive or mixed-type inhibition be observed with respect to both of the substrates of the reaction \[[@B29]\]. The data obtained with L-AFC did indeed demonstrate precisely this \[[@B20]\]. Noncompetitive inhibition by AFCME was observed with respect to the BFC substrate (Fig. [5A](#F5){ref-type="fig"}) with an apparent K~i~of 1.9 ± 0.6 μM, and mixed-type inhibition was observed with respect to AdoMet as a substrate with an apparent K~i~of 2.4 ± 0.7 μM, respectively (Fig. [5B](#F5){ref-type="fig"}). These kinetic results are summarized in Table [I](#T1){ref-type="table"}. In addition, essentially identical results were observed with a second prenylcysteine-type product analog N-acetyl-S-farnesyl-L-cysteine methyl amide (AFCMA) (data not shown). Altogether, the results this kinetic analysis are completely consistent with the reaction catalyzed by Icmt proceeding through an ordered sequential mechanism with the AdoMet substrate binding first and the AdoHcy product being released last.
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Distinguishing the order of product dissociation from human Icmt: Inhibition studies with the AFCME product**. A. Lineweaver-Burk plot for the product inhibition of BFC methylation by AFCME at a fixed concentration of BFC. Assays were conducted in the presence of fixed \[BFC\] at 4 μM and either (◆) 0 μM, (□) 5 μM or (▲) 10 μM of AFCME at the indicated concentrations of AdoMet. Assays were conducted as described in the legend to Fig. 2. Data shown are the mean of duplicate determinations from a single experiment, and are representative to two such experiments. B. Lineweaver-Burk plot for the product inhibition of BFC methylation by AFCME at a fixed concentration of AdoMet. Assays were conducted in the presence of fixed \[AdoMet\] at 5 μM and either (◇) 0 μM, (■) 5 μM or (○)10 μM of AFCME at the indicated concentrations of BFC. Assays were conducted as described in the legend to Fig. 2. Data shown are the mean of duplicate determinations from a single experiment, and are representative to more two such experiments.
:::
:::
Discussion
==========
The studies in this report provide the first detailed kinetic analysis of the molecular entity Icmt, the enzyme responsible for the final step in the maturation of CaaX-type protein in eukaryotic cells. To facility analysis of Icmt activity, we synthesized a new substrate for Icmt, biotin-S-farnesyl-L-cysteine (BFC), that has two properties that improve its utility compared with the only other described small molecule substrate for the enzyme, AFC. First, BFC is utilized by the enzyme with a K~m~(2.1 μM) essentially identical to that of an authentic substrate, farnesylated, Rce1-proteolyzed, K-Ras protein, whereas the apparent Km for AFC is 10-fold higher at 20 μM \[[@B20],[@B31]\]. The second improvement in the BFC substrate is, of course, the appended biotin moiety, which allows facile isolation of product from reaction mixtures using commercially-available avidin resins. Furthermore, this substrate is ideal for use in high-throughput screening for inhibitors of the enzyme since it allows use of scintillation proximity-type assays.
Our first series of kinetic studies was designed to distinguish between the two general kinetic mechanisms that exist for multisubstrate enzymes. The first general mechanism is termed sequential and describes reactions in which all the substrates must bind to the enzyme before the first product is released, whereas reactions in which one or more products are released prior to all substrates being added are termed ping-pong. To distinguish between these reaction types, we performed classic steady state kinetic analyses of Icmt. The results of this analysis revealed that catalysis by Icmt exhibits all the properties of rapid equilibrium bireactant system \[[@B28],[@B29]\] and hence proceeds via a sequential mechanism.
The second series of experiments we undertook was designed to distinguish between a random sequential vs. ordered sequential mechanism; this involved a determination of whether the enzyme could initially form a complex with either substrate or whether must it bind them in a defined order. To verify the assignment of a sequential mechanism for Icmt and to distinguish between an ordered versus a random sequential mechanism, studies with a dead-end inhibitor of Icmt were performed. In a random sequential mechanism, there is no defined order of either substrate binding or product release, whereas in an ordered sequential mechanism the binding of the first substrate is required for the second and the products dissociate in an obligatory order \[[@B28],[@B29]\]. The studies with the dead-end substrate FTA revealed that FTA is a competitive inhibitor with respect to BFC, whereas in a random mechanism FTA would be expected to be a mixed-type inhibitor with respect to AdoMet \[[@B28]\]. Furthermore the double-reciprocal plots from the analysis of FTA as an inhibitor with respect to AdoMet showed the parallel lines indicative of an uncompetitive inhibition, signifying that FTA is unable to bind to the free enzyme. Hence, AdoMet binds first to enzyme and FTA interacts with the AdoMet-Icmt complex since, if BFC combined first with enzyme, the inhibition by FTA should have been competitive with respect to BFC and a mixed-type, noncompetitive or competitive inhibition with respect to AdoMet. Together, these results indicate that the reaction catalyzed by Icmt is proceeds through an ordered sequential mechanism in which the AdoMet substrate binds first.
Following the assignment of a sequential mechanism for Icmt, product inhibition experiments were employed to investigate the order of product dissociation from the enzyme. The finding that AdoHcy is a competitive inhibitor with respect to AdoMet revealed that this compound could interact with free enzyme; ie. it must be the final product released. Also, the finding that the prenylcysteine methylester AFCME exhibited mixed-type inhibition with respect to AdoMet is the result expected if AdoHcy is the last product released. Interestingly, when AdoHcy was examined for its inhibition with respect to BFC a different result was obtained from that of the previous kinetic study of prenylcysteine methyltransferase activity in rod outer segments \[[@B20]\]. With recombinant human Icmt, AdoHcy is a competitive inhibitor with respect to BFC, indicating that the Icmt-AdoHcy complex is inactive and cannot bind to BFC, an observation that further reinforces the conclusion of an ordered sequential mechanism for this enzyme.
The inhibition constants for AFCME determined in our experiments of 1.9 μM and 2.4 μM with respect to BFC and AdoMet, respectively, are substantially different from those in the previous study on the methyltransferase activity in rod outer segments of 41 μM and 73 μM respectively \[[@B20]\]. We believe these results are more likely due to the markedly enriched preparation used in our study rather that to any fundamental difference in the enzyme activities; we used 1000-fold less membrane in our assays and it is likely that the hydrophobic AFCME was substantially \"sopped up\" by the membrane lipid in the previous study, markedly reducing the level available to interact with the enzyme. This phenomenon, which is related to that known as surface dilution kinetics, comes into play when both the concentration of the membrane lipid and of the ligand employed contribute to the kinetic parameter measured \[[@B32]\].
Conclusions
===========
In summary, we have demonstrated that the kinetic mechanism of CaaX protein methylation by Icmt is an ordered sequential mechanism in which the AdoMet substrate associates first with the enzyme and the AdoHCy product dissociates last. This work provides a kinetic framework for the analysis of specific inhibitors of this enzyme that will most assuredly be forthcoming given the recent biological validation of Icmt as a target for blocking oncogenic transformation of cells \[[@B8],[@B11],[@B33]\].
Methods
=======
Materials
---------
Streptavidin-sepharose beads were purchased from Amersham, *trans*, *trans*-farnesyl bromide, biotin N-hydroxysuccinimide ester and S-(5\'-adenosyl)-L-methionine p-toluenesulfonate were purchased from Sigma-Aldrich, L-cysteine was purchased from Novabiochem, \[^3^H-*methyl*\]-S-adenosyl-L-methionine was purchased from Perkin Elmer Life Sciences, farnesylthioacetic acid (FTA) was purchased from Biomol, S-(5\'-adenosyl)-L-homocysteine was purchased from Fluka. The N-acetyl-S-farnesylcysteine methyl ester (AFCME) and N-acetyl-S-farnesylcysteine methyl amide (AFCMA) were synthesized as previously described (\[[@B24]-[@B26]\]). Recombinant human Icmt was prepared by infection of Sf9 cells with a recombinant baculovirus containing the entire open reading frame of the human Icmt cDNA as described \[[@B8]\]. The membrane fraction of the infected Sf9 cells, isolated as described for studies with the Rce1 protease \[[@B27]\], was used as the source of Icmt for all studies described.
Syntheses of biotin-S-farnesyl L-cysteine (BFC)
-----------------------------------------------
S-Farnesyl L-cysteine (FC) was prepared by reaction of farnesyl bromide with L-cysteine in methanol/ammoniac solvent as described \[[@B26]\]. The resultant FC product was separated from L-cysteine by extraction with butanol/H~2~O (1:1); the butanol phase was then evaporated under reduce pressure and the FC product washed several times with hexane to remove residual farnesyl bromide.
Two coupling procedures were used to attach the biotin moiety to the amino group of FC; both involved the use of biotin N-hydroxysuccinimide methylester (biotin-NHS) as the biotinylation agent. In the first method, an excess of biotin-NHS (75 μmol) to FC (7.5 μmol) was used. The two compounds were dissolved in 2.3 ml of DMSO and 200 μl of 1M Hepes, pH 12, was added. Following a 2 h incubation at room temperature, the DMSO was evaporated under reduced pressure and the resulting residue extracted with a solution of butanol/H~2~O (1:1). The butanol phase was dried and the residue dissolved in 100 μl of methanol; this solution was then diluted to 1 ml by addition of 0.1% trifluoroacetic acid (TFA). The addition of the TFA resulted in the appearance of a white precipitate of essentially pure BFC as judged by chromatography and NMR analysis.
The second coupling method used an excess of FC (2.5 mmol) compared to biotin-NHS (0.73 mmol); these two compounds were dissolved in 18 ml of DMSO. To this solution was added 2 ml of 1M Hepes, pH 12, and the coupling allowed to proceed for 2 h at room temperature. As with the first method, the DMSO was evaporated under reduced pressure, the resulting residue extracted with a solution of butanol/H~2~O (1:1), and the butanol phase dried and the residue dissolved in methanol. Because of the excess of FC present with the BFC product, specific precipitation of the BFC product was unsuccessful. Instead, the BFC was purified by preparative HPLC on a C~18~column developed in CH~3~CN/H~2~O (2:3). On this matrix, BFC was well-resolved from FC. The peak fractions were collected and solvent evaporated under reduced pressure; the resulting product was white solid that was \>95% pure as judged by analytical reverse-phase HPLC \[C~18~matrix developed in CH~3~CN/H~2~O/TFA (90:10:0.1). Proton NMR spectra of the products obtained by either method were completely consistent with that expected for authentic BFC.
Icmt assay
----------
The assay developed for Icmt activity involved quantitation of \[^3^H\]methyl incorporation into the small molecule substrate BFC. For the standard assay, reactions were initiated by addition of Sf9 membranes containing Icmt (0.5 μg protein) to an assay mixture containing BFC (4 μM) and \[^3^H\]AdoMet (5 μM, 1.3 Ci/mmol) in100 mM Hepes, pH 7.4 and 5 mM MgCl~2~in a total volume of 45 μl. Reactions were carried out for 20 min at 37°C, whereupon they were terminated by addition of 5 μl of 10% Tween 20. Following termination, streptavidin beads (10 μl of packed beads suspended in 500 μl of 20 mM NaH~2~PO~4~, pH 7.4, contaning 150 mM NaCl) were added, and the mixtures mixed by gentle agitation overnight at 4°C. The beads were harvested by centrifugation in a tabletop microcentrifuge at 10,000 rpm for 5 min and washed 3 times with 500 μl of 20 mM NaH~2~PO~4~, pH 7.4, containing 150 mM NaCl. The beads were then suspended in 100 μl of the same buffer, transferred to scintillation vials, and radioactivity determined. For the kinetic analyses, the concentrations of substrates (AdoMet, BFC) or additional ligands (eg. AdoHcy, AFCME, etc) were varied as detailed in the legends to the appropriate figures.
Abbreviations used
==================
Icmt, Isoprenylcysteine carboxyl methyltransferase; AdoMet, S-adenosyl-L-methionine; AdoHcy, S-adenosyl-L-homocysteine; AFC, N-acetyl-L-farnesylcysteine; AGGC, N-acetyl-S-geranylgeranyl-L-cysteine; TFA, trifluoroacetic acid; BFC, biotin-S-farnesyl-L-cysteine; FTA, S-farnesylthioacetic acid; AFCME, N-acetyl-S-farnesyl-L-cysteine methylester; AFCMA, N-acetyl-S-farnesyl-L-cysteine methyl amide
Authors\' contributions
=======================
RAB carried out all of the studies reported. PJC conceived of the study, and participated in its design and coordination and helped to draft the manuscript. Both authors read and approved the final manuscript.
::: {#F6 .fig}
Figure 6
::: {.caption}
######
**Proposed kinetic mechanism of human Icmt**. See text for details.
:::
:::
Acknowledgements
================
This work was supported by National Institutes of Health Grant GM46372 to P.J.C. and fellowship to R.A.B. from l\'Association pour la recherche contre le cancer. We thank Dr. Rudolph for his helpful assistance in the kinetic experiments and for the use of his kinetic software.
|
PubMed Central
|
2024-06-05T03:55:51.957805
|
2004-12-29
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545952/",
"journal": "BMC Biochem. 2004 Dec 29; 5:19",
"authors": [
{
"first": "Rudi A",
"last": "Baron"
},
{
"first": "Patrick J",
"last": "Casey"
}
]
}
|
PMC545953
|
Background
==========
Until recently multiple gene expression profiling was applied almost exclusively to human and a few model organisms. At present cDNA microarrays are being constructed for new species including teleost fishes \[[@B1]-[@B6]\]. Since EST sequencing projects are carried out with a large number of species, continuous development of new platforms can be expected in the future. We designed a salmonid fish cDNA microarray primarily to characterize responses to stress, toxicity and pathogens. This paper focuses on time-course comparisons of stress responses in rainbow trout and the usage of functional annotation to conduct analyses of gene expression data.
Functional annotation of genes, especially Gene Ontology \[[@B7]\] is increasingly being used for analyses and interpretation of microarray results \[[@B8]-[@B13]\]. We applied Gene Ontology in several modes to facilitate implementation of our research tasks. Furthermore, experimental results generated guidelines for the development of specialized microarrays. Well designed platforms are expected to ensure identification of differentially expressed genes while containing representative coverage from important functional groups. Custom made microarrays include clones from cDNA libraries and/or selected genes, which have advantages and drawbacks. Indiscriminant spotting of EST may result in under representation of many functional classes. On the other hand selection of genes fully relies on annotations and hypotheses, which can be misleading and limit possibilities for nontrivial findings. We used clones from normalized and subtracted cDNA libraries as well as genes selected by the functional categories of Gene Ontology for inclusion onto a microarray targeted at characterizing transcriptome responses to environmental stressors. Designing a new platform requires balancing a large number of genes versus multiple replications of spots, which enhances statistical analyses of data. The rainbow trout microarray was prepared by spotting of relatively small number of genes (1300) in 6 replicates. We show that multiple replications combined with the dye-swap design of hybridization \[[@B14],[@B15]\] allows for accurate detection of relatively small alterations in expression levels, which is important for the functional interpretation of results.
Stress is closely associated with many diverse issues in fish biology and environmental research (reviewed in \[[@B16]\]). Stress is generally defined as the reaction to external forces and abnormal conditions that tend to disturb an organism\'s homeostasis. To illustrate the major trends in the studies of stress in fish, we performed a computer-assisted analysis of Medline abstracts covering this area (Table [1](#T1){ref-type="table"}). Salmonids have been studied more extensively than any other fish species. Research has focused on various biotic and abiotic factors including toxicity, environmental parameters (oxygen, temperature, salinity, acidosis), diseases, social interactions (crowding, aggressiveness) and farming manipulations. Analysis of Medline abstracts indicated physiological processes, cellular structure and selected proteins that have been the major foci of previous fish stress studies. This provided an outline for interpretation of our results. We analyzed the effects of stress on the transcriptome in the brain and kidney, which are considered important target tissues along with muscle, blood cells, liver and epithelia. We report a profound difference of stress response in these tissues and the identification of a diagnostic set of genes.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Thematic associations in studies of fish stress. Computer-assisted analysis of 11129 Medline abstracts was performed as described in Methods. Terms that were over-represented in the abstracts (exact Fisher\'s test, P \< 0.05) are ranked by the numbers of occurrence.
:::
**Category** **Terms (counts/1000 abstracts)**
--------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Species Salmonids (126.4), carp (68.9), eels (67.0), catfish (37.7), tilapia (38.7)
Stressors Toxicity (440.6), temperature (178.3), oxygen (91.5), confinement (52.8), salinity (46.2), hypoxia (54.7), diseases (20.8), crowding (23.6), acidosis (17.9), aggressiveness (11.3)
Messengers Cortisol (208.5), catecholamines (159.4), steroids (92.5)
Tissues Muscle (197.2), blood cells (152.9), pituitary (119.8), liver (123.6), epithelia (96.2), brain (90.6), kidney (89.6), heart (51.9), skin (42.5)
Cellular structures Cytosol (42.5), collagen (17.0), cytoskeleton (15.1), microsome (15.1), microtubule (14.2), lysosomes (13.2), peroxisome (4.7)
Oxidative stress Glutathion (167.9), oxidant (93.4), antioxidant (90.6), peroxide (66.0), radical (55.7), superoxide (40.6), catalase (35.8), redox (18.9)
Other processes Immunity (91.5), secretion (80.2), metabolism (74), transport (56.6), defense (52.8), necrosis (28.3), apoptosis (18.9), phosphorylation (15.1), proteolysis (7.5)
Metabolites Ion (987.7), iron (215.1), glucose (141.5), lactate (67.9), lipid (74.5), zinc (51.9), phospholipid (11.3), triglyceride (11.3), lipopolysaccharide (9.4)
Proteins Enzymes (180.2), heat-shock proteins (84.0), hemoglobin (37.7), metallothionein (37.7), transferase (32.1), phosphatase (26.4), chaperones (21.7), glutathion-S-transferase (17.0), transaminase (17.0), Na/K-ATPase (17.0), aminotransferase (8.5), mitogen-activated kinases (4.7)
:::
Results
=======
1 Design of cDNA microarray
---------------------------
The rainbow trout cDNA microarray was composed of EST and selected genes. The cDNA libraries were prepared from tissues of stressed fish using suppression subtractive hybridization, SSH \[[@B17]\] and a modification of the cap-finder method \[[@B18]\] supplemented with enzymatic normalization \[[@B19]\]. We sequenced 2000 clones and redundancy of the subtracted libraries was markedly greater than that of the normalized (306% and 134% respectively). In addition to EST we selected rainbow trout transcripts from the normalized multi-tissue cDNA library \[[@B20]\] based on their assignment to functional categories of Gene Ontology (stress and defense response, regulation of cell cycle, signal transduction, chaperone activity and apoptosis). The selected genes substantially improved the coverage of many functional classes (Table [2](#T2){ref-type="table"}), though the number of differentially expressed genes in this group was markedly inferior to EST (Figure [1](#F1){ref-type="fig"}). Subtraction cloning enriched genes that showed strong alteration of expression at response to stress (p \< 0.01 or lower, Figure [1A](#F1){ref-type="fig"}), however the SSH clone set did not provide any advantage when microarray was used for the related research tasks (Figure [1B](#F1){ref-type="fig"}).
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Presentation of the Gene Ontology functional categories in the microarray. Table shows the numbers and frequncies of genes in the clone sets that were used for spotting (SSH -- subtracted libraries, EST -- normalized libraries).
:::
**Gene Ontology classes** **N on slide** **SSH** **EST** **Selected**
--------------------------------- ---------------- ----------- ----------- --------------
Response to external stimulus 147 11 (0.07) 48 (0.11) 88 (0.31)
Response to stress 145 7 (0.04) 30 (0.07) 108 (0.38)
Defense response 105 6 (0.04) 34 (0.08) 65 (0.23)
Humoral immune response 42 3 (0.02) 13 (0.03) 26 (0.09)
Apoptosis 79 6 (0.04) 10 (0.02) 63 (0.22)
Cell communication 139 11 (0.07) 45 (0.11) 83 (0.29)
Cell proliferation 82 8 (0.05) 23 (0.05) 51 (0.18)
Cell cycle 62 2 (0.01) 17 (0.04) 43 (0.15)
Signal transduction 114 5 (0.03) 32 (0.07) 77 (0.27)
Receptor activity 49 3 (0.02) 18 (0.04) 28 (0.10)
Intracellular signaling cascade 49 3 (0.02) 15 (0.04) 31 (0.11)
DNA metabolism 47 5 (0.03) 15 (0.04) 27 (0.09)
Transcription 67 9 (0.05) 21 (0.05) 37 (0.13)
Chaperone activity 41 4 (0.02) 12 (0.03) 25 (0.09)
:::
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Performance of the clone sets used for preparation of the microarray.**Figure shows frequencies of genes that were differentially expressed in at least 5 samples at different p-values (Student\'s t-test). **A:**this study (stress response), **B:**related experiments (exposure to aquatic contaminants \[34\], response to stress, cortisol and combination of these treatments, challenge with bacterial antigens, M74 disease). SSH -- subtracted cDNA libraries, EST -- normalized libraries, Select -- genes chosen by the Gene Ontology functional categories.
:::
:::
2 Stress response in the brain and kidney of rainbow trout
----------------------------------------------------------
### 2.1 Differentially expressed genes
Fish were stressed with netting and samples were collected 1, 3 and 5 days after the first exposure. We used plasma cortisol as a stress marker \[[@B21]\]. The hormone levels increased 7.6-fold after 1 day and did not change significantly to the end of experiment (Figure [2](#F2){ref-type="fig"}).
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Plasma cortisol levels.**The data are mean ± SE (n = 4). Difference between the control and stressed fish is significant (Student\'s t-test, p \< 0.05).
:::
:::
Microarray results were submitted to GEO ([GSM22355]{.underline}). Two genes were up-regulated in both tissues at all time-points (Figure [3](#F3){ref-type="fig"}). One is a putative homolog to the mammalian N-myc regulated genes, which are induced with steroid hormones in the brain \[[@B22]\] and kidney \[[@B23]\]. Mitochondrial ADP, ATP carrier can be implicated to both normal functions and cell death \[[@B24]\]. Metallothionein-IL, a classical stress marker was induced to the end of experiment and a similar profile was seen in midkine precursor (growth factor), histone H1.0 and B-cell translocation protein 1. In kidney we observed consistent up-regulation of genes related to energy metabolism, such as mitochondrial proteins (cytochromes b and c, cytochrome oxidases), enzymes (glyceraldehyde 3-phosphate dehydrogenase, fructose-bisphosphate aldolase, serine-pyruvate aminotransferase) and similar profiles were seen in two heat shock proteins and two signal transducers (cytohesin binding protein and GRB2-adaptor). The repressed genes were related to actin binding (coronin and profilin) and immune response (meprin, immunoglobulin epsilon receptor, thymosin and lysozyme).
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Examples of differentially expressed genes**. Pooled RNA from 4 fish was hybridized in dye-swap experiments to two microarrays on which each gene was printed 6 times (total of 12 replicates). Differential expression was analysed with Student\'s t-test (P \< 0.01); the expression ratio is coded with color scale.
:::
:::
Rapid alteration of gene expression was a remarkable feature of stress response in the brain. Only one gene, aquaporin, was up-regulated for the duration of the experiment. Water channel aquaporin plays a key role in water homeostasis being implicated in various physiological processes and pathological conditions \[[@B25]\]. A panel of genes which showed markedly increased expression after 1 day was also suppressed after 5 days. Surprisingly, this group included mainly genes that are predominantly expressed in skeletal or cardiac muscle (myosin light chain 1 and 2, skeletal and cardiac isoforms, myosin heavy chain, troponin I, T and C) or are involved in regulation of muscle contraction (parvalbumin alpha and sarcoplasmic reticulum calcium ATPase). An opposite tendency was shown by a large group of genes however the magnitude of expression changes was smaller. We analysed 5 differentially expressed genes with qPCR and the results were in close concordance with the microarray data (not shown).
### 2.2 Functional classes
The search for enriched Gene Ontology functional categories in the lists of differentially expressed genes found almost no overlap between the tissues (Table [3](#T3){ref-type="table"}). In the brain stress affected binding and transport of metal ions, especially calcium and manganese, chaperones and heat shock proteins, cytoskeleton and microtubules and a number of signaling pathways; whereas, mitochondrion, extracellular structures and peptidases appeared the primary targets in the kidney.
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Enrichment of Gene Ontology categories in the lists of differentially expressed genes. Analysis with exact Fisher\'s test, (p \< 0.05) was made using the composition of microarray as a reference. The numbers of differentially expressed genes and genes on the microarray are in parentheses.
:::
**Brain** **Kidney**
--------------------------------------------- -------------------------------
Intracellular signaling cascade (19/47) Mitochondrion (19/71)
RAS protein signal transduction (6/9) Electron transporters (13/43)
GTPase mediated signal transduction (11/16) Extracellular (19/70)
Chaperones (16/40) Endopeptidases (8/22)
Heat shock proteins (8/16) Metallopeptidases (7/12)
Metal ion binding (31/80) Zinc ion binding (8/24)
Carriers (15/37)
Potential-driven transporters (7/9)
Calcium ion binding (20/41)
Magnesium ion binding (8/14)
Cytoskeleton (27/76)
Myofibril (16/16)
Microtubule-based process (6/6)
:::
Comparison of the differentially expressed genes by the Gene Ontology categories suggested coordinated regulation of various cellular functions in the brain. Early stress response was marked with transient induction of the cytoskeleton proteins and similar profiles were observed in the metal binding proteins and enzymes of carbohydrate metabolism (Figure [4](#F4){ref-type="fig"}). An opposite expression pattern was shown by a large group of genes involved in stress and immune response, regulation of growth and cell cycle, apoptosis, signal transduction and cell to cell signaling. This was in parallel with enhancement of transcription and translation, ubiquitin-dependent protein catabolism and protein folding. In the kidney the temporal alterations were much weaker. Expression of metal binding proteins increased slowly in parallel with peptidases. Strong induction of collagenases coincided with decrease of collagen expression. At the same time a number of metabolic functions were suppressed (oxidative phosphorylation and oxidoreductase activity, amine metabolism and RNA binding).
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Time-course of stress response in the brain and kidney**. Differentially expressed genes were grouped by the Gene Ontology categories and mean log (expression ratios) were analysed with Student\'s t-test. Panel presents examples of categories that showed significant difference between the time points (p \< 0.05). The values are coded with color scale.
:::
:::
3 Stress-responsive genes
-------------------------
Microarray design included genes from functional categories which were expected to be affected by stress (Table [2](#T2){ref-type="table"}). Overall observations of differences in gene expression from this group in response to handling stress were minimal; however, this could be accounted for by its heterogeneity. Therefore we searched for the subgroups of genes with correlated expression profiles within the functional classes using results of 35 microarray experiments conducted by our laboratory. Both factorial and cluster analyses revealed 9 defense response genes that showed tightly coordinated expression being induced with stress. We continued search using the consensus profile of this subgroup and found 47 positively and 1 negatively correlated genes (Pearson r \> \|0.65\|). Of these 29 were identified by the protein products (Figure [5A](#F5){ref-type="fig"}), 19 being from the set of selected clones. Expression of the stress-responsive genes changed significantly in several experiments including this study (Figure [5B](#F5){ref-type="fig"}). They were up-regulated in kidney with stress and injection of cortisol, combination of these treatments showed an additive effect (Figure [5C](#F5){ref-type="fig"}). These genes also responded to the model water contaminants, being induced with low and medium and repressed with high doses (Figure [5D](#F5){ref-type="fig"}).
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Expression of stress-responsive genes**. **A**: Experiments. **1--6**: response to handling stress, this study. Kidney, 1 day (1), 3 days (2) and 5 days (3); brain, 1 day (4), 3 days (5) and 5 days (6). **7--12**: response to handling stress and exogenous cortisol in kidney. Cortisol, 1 day (7) and 3 days (8); stress, 1 day (9) and 3 days (10), combination of stress and injection of cortisol, 1 day (11) and 3 days (12). **13--20**: exposure of yolk sac fry to model contaminants \[34\]. β-naphthoflavone, low (13) and high (14) dose; cadmium, low (15) and high (16) dose; carbon tetrachloride, low (17) and high (18) dose; pyrene, low (19) and high (20) dose. **21--22:**response of yolk sac fry to transportation stress, rainbow trout (21) and Atlantic salmon (22). Ranks are coded with color scale; correlation coefficients (Pearson r) with the mean expression profile are indicated. **B-C:**the mean ranks ± SE of the stress-responsive genes in 3 experiments. **A**: this study; **B**-- response to handling stress and injection of cortisol in kidney; **C**-- exposure of yolk sac fry to β-naphthoflavone, cadmium and pyrene at low, medium and high doses.
:::
:::
Discussion
==========
I Stress response in rainbow trout
----------------------------------
Our study aimed at comparison of time-course of stress response in rainbow trout brain and kidney and finding of a diagnostic set of genes. These tasks were implemented with an aid of Gene Ontology annotation, which was used in several modes. The most straightforward and commonly used approach is counting of Gene Ontology classes in the lists of differentially expressed genes. Statistical inference of enrichment and depletion is made with Z-score of hypergeometric distribution, exact Fisher\'s test or its modifications. Such analyses helped us to interpret differences of stress responses in the brain and kidney (Table [3](#T3){ref-type="table"}). In the brain handling stress mainly affected expression of transcripts for structural proteins (especially cytoskeleton), signal transduction, and binding of metal ions, whereas mitochondria, extracellular structures and peptidases appeared the key targets in the kidney. Computer-assisted analysis of Medline abstracts suggested that most of these themes have not been addressed in the studies of fish stress (Table [1](#T1){ref-type="table"}).
Searches of the enriched Gene Ontology categories associated with differentially expressed transcripts is useful for rapid screens of microarray data; however, it presumes coordinated expression of functionally related genes. This assumption is not valid for many classes, especially large and heterogenous groups, such as stress and defense response. Because the gene composition of microarray is used as a reference, uneven presentation of functional categories can distort the results. Finally, this analysis does not take into account direction and magnitude of differential expression. To overcome these problems, enrichment of Gene Ontology classes is analysed in groups of genes with similar expression profiles revealed with cluster or factorial analyses. In this study we preferred straight comparison of Gene Ontology classes by the mean log expression ratios which helped in interpretation of the time-course of stress response.
In the kidney temporal alterations were relatively weak though significant. Expression of peptidases (especially collagenases) increased steadily, which implied possible degradation of tissue with prolonged stress. We could expect abrupt fluctuations in the rainbow trout brain, since transient induction and up-regulation of gene expression was observed in response to cold in the brain of channel catfish \[[@B4]\]. In our study most differentially expressed genes fell into two groups with distinct temporal profiles which showed remarkable coherence of the functional classes. Early phase was associated with dramatic up-regulation of structural and metal binding proteins, which were repressed in later phases. Expression of genes involved in stress and defense response, apoptosis and signal transduction, cell cycle and growth changed in a reciprocal fashion. Activation of metal binding proteins could be accounted for the role of ions (particularly calcium) in multiple pathways of gene expression regulation in the brain \[[@B26]\]. Motor proteins of cytoskeleton play key roles in the transport of vesicles and the establishment and rearrangement of neuronal networks \[[@B27]-[@B29]\] which also could be implicated to the stress response in fish. However, in mammals these functions are associated with non-muscle isoforms and therefore differential expression of the sarcomeric proteins was unexpected. Additional experiments confirmed induction of these proteins at early phase of stress response. Previously we observed high activity of skeletal α-actin and myosin light chain 2 promoters in the neural tissues of rainbow trout embryos \[[@B30]\]. Sequencing of salmonid fish cDNA libraries provided evidence for the brain expression of sarcomeric proteins, but their role remains fully unknown. At present there is sparse evidence for differential expression of structural muscle proteins in the mammalian brain. For example regulation of troponin I with dextromethorphan (antagonist of excitatory amino acid receptors) was reported in the rat hyppocampus and cortex \[[@B31]\].
Grouping of individual differentially expressed genes by the functional classes reduced noise and enhanced cluster and factorial analyses. This helped to identify stress-responsive genes that showed correlated expression in 35 microarray experiments (22 experiments are shown in Figure [5A](#F5){ref-type="fig"}). Association with stress is well established for most of these proteins and some are used as stress markers. The list of enriched Gene Ontology categories (stress, defense and humoral immune response, signal transduction and response to oxidative stress, p \< 0.05) suggested biological relevance of this group. Computer analysis of Medline abstracts (Table [1](#T1){ref-type="table"}) showed that immunity and metabolism of reactive oxidative species are prioritized in studies of fish stress and these functional categories were enriched in the list of stress-responsive genes. Thus Gene Ontology provided a useful starting point for search of functionally related genes and results of these analyses can be used further for the revision of annotations.
2 Construction of microarrays
-----------------------------
Results of our experiments helped to evaluate the strategy used in construction of the rainbow trout microarray. Researchers developing microarrays for new species are commonly choosing between specially selected genes and clones from normalized and subtracted cDNA libraries. We used SSH, which is at present probably the most popular method of subtraction. Though proven efficient in many studies, this method has a number of drawbacks. Subtraction requires re-association of tester and formation of double-stranded DNA, hence many rare transcripts are not cloned and variations in concentrations of cDNA and hybridization conditions may have strong impacts on library composition. High redundancy is a common feature of the SSH libraries. Apart from these problems, rapid alterations of gene expression observed in this study and many other microarray experiments make the advantages of subtractive cloning ambiguous. Subtraction achieves enrichment of the transcripts, which are over or under represented in the test sample. In many cases one sample will not provide coverage of differentially expressed genes for the whole series, whereas pooling of samples may reduce fluctuations. Furthermore, we observed relatively high ratio of differentially expressed genes among the clones from the unsubtracted cDNA libraries, which are easier for construction and much less redundant. The advantage of subtractive cloning becomes negligible when microarrays are used for different, though related research tasks (Fig. [1B](#F1){ref-type="fig"}).
At present selection of genes for microarrays is facilitated with advances of functional annotation. This helped us to improve presentation of many functional categories (Table [1](#T1){ref-type="table"}) and enhanced interpretation of results. Most of the selected genes did not show differential expression in our studies, however 63% of stress-responsive genes were from this group. In our view, this finding is a strong argument for utilizing Gene Ontology in the development of specialized platforms.
Given the limited number of spots on slides, microarray design requires a careful balance between the number of genes and replication of spots. Apparent advantages of genome-wide platforms are compromised with the problems associated with identifying significantly differentially expressed genes. We preferred combination of multiple spotting and dye-swap normalization, which ensured robust normalization and accurate detection of differential expression at low ratios. Coordinated expression of functionally related genes suggested biological relevance of relatively small alterations in the transcription levels. Selection of differentially expressed genes by the cutoff values would result in loss of valuable information in our experiments. For instance, most of the stress-responsive genes showed small or moderate expression changes, the identification of this group would not be likely without multiple replications.
Conclusions
===========
1\. Combination of EST and selected genes appears a reasonable way for construction of cDNA microarrays. Multiple replications of spots and dye swap design of hybridization ensure robust normalization and high power of statistical analyses. Finding of differential expression at small ratios is essential for the functional interpretation of microarray data.
2\. Stress response in fish brain and kidney is different both by the target functions and time-course. In brain slow progression of adaptive response was preceded with dramatic transient induction of motor and metal ion binding proteins. Prolonged stress was likely to result in slow degradation of extracellular matrix in kidney.
3\. Finding of stress-responsive genes provides possibility for measurement of stress in various conditions and search for the functionally related genes.
Methods
=======
1. Computer-assisted analysis of Medline abstracts
--------------------------------------------------
Search of Medline was made with queries: \"fishes AND stress\" (1060 abstracts) and \"fishes NOT stress\" (10069 abstracts). Abstracts were split into separate words and a list of non-redundant terms was composed. The numbers of abstracts including each term were estimated. The terms were ranked by the Z-scores of hypergeometric distribution and enrichment was analysed with exact Fisher\'s test (p \< 0.05).
2. Experiments with fish, exposure and sampling
-----------------------------------------------
One year old rainbow trout were stressed with netting for 2 min, this treatment was repeated once a day for a duration of 5 days. Fish were killed with over-dose of anaesthetic (MS-222) and blood was taken from the caudal vein. The kidneys and brains were snap-frozen in liquid nitrogen. Plasma cortisol was determined with RIA using Orion Spectra Cortisol kit.
3. Preparation of microarrays
-----------------------------
RNA was extracted with Trizol reagent (Invitrogen) and mRNA was purified with Dynabeads kit (Dynal). SSH cloning was performed as described \[[@B17]\]. For preparation of normalized libraries, synthesis of cDNA with PowerScript reverse transcriptase (Clontech) was primed with oligonucleotides including *EcoRI*and *NotI*sites: 5\'-ACGAGGC[GAATTC]{.underline}ACAGAGAGGAG(T)VN-3\', 5\'-GAGAGAGAGTGGT[GCGGCCGC]{.underline}GGTGTATGGGG-3\'). Double-stranded cDNA was generated using Advantage DNA polymerase mix (Clontech) and PCR primers: 5\'-ACGAGGC[GAATTC]{.underline}ACAGAGAGGAG-3\' and 5\'-GAGAGTGGT[GCGGCCGC]{.underline}GGTGTA-3\'. The PCR products were purified with QIAquick kit (Qiagen), precipitated with ethanol and dissolved to 1 μg/μl in hybridization buffer (1 M NaCl, 50 mM HEPES (pH 8.3), 1 mM EDTA). DNA was denaturated for 5 min at 94°C. Following re-association at 72°C for 16 hours, DNA was ethanol precipitated and digested with 150 U of exonuclease III (MBI Fermentas) for 15 min at 37°C. This treatment eliminates re-associated double-stranded DNA \[[@B19]\]. Single-stranded DNA was PCR amplified, size separated with agarose gel electrophoresis and cloned into pGEM^®^-11Zf (+) (Promega). Normalized and subtracted cDNA libraries were prepared from the stressed fish (whole fry, brain, kidney and spleen of 1-year old fish). A number of clones were from the rainbow trout and Baltic salmon cDNA libraries constructed in University of Turku. The sequences were analysed with stand-alone blastn and blastx \[[@B32]\] Microarray incldued 315 genes selected by the Gene Ontology functional categories. Of these, 282 clones were from the normalized multi-tissue library \[[@B20]\] and the rest were produced with RT PCR. The cDNA inserts were amplified with PCR using universal primers and purified with Millipore Montage PCR96 Cleanup Kit. DNA was spotted onto poly-(L) lysine-coated slides and each clone was printed in 6 replicates.
4. Microarray analyses
----------------------
Total RNA was extracted with Trizol reagent (Invitrogen) and 4 individuals were pooled in each sample. Stressed fish was compared with time-matched control. Labeling with Cy3- and Cy5-dCTP (Amersham Pharmacia) was made using SuperScript III (Invitrogen) and oligo(dT) primer; cDNA was purified with Microcon YM30 (Millipore). We used a dye swap experimental design \[[@B14],[@B15]\] and each sample was hybridized to two microarrays. For the first slide, test and control cDNA were labeled with Cy5 and Cy3 respectively, and for the second array dye assignments were reversed. The slides were pretreated with 1% BSA, fraction V, 5 x SSC, 0.1% SDS (30 min at 50°C) and washed with 2 x SSC (3 min) and 0.2 x SSC (3 min) and hybridized overnight in cocktail containing 1.3 x Denhardt\'s, 3 x SSC 0.3% SDS, 0.67 μg/μl polyadenylate and 1.4 μg/μl yeast tRNA. All chemicals were from Sigma-Aldrich. Scanning was performed with ScanArray 5000 and images were processed with QuantArray (GSI Luminomics). The measurements in spots were filtered by criteria *I/B ≥ 3*and (*I*-*B*)/(*S*~*I*~+*S*~*B*~) ≥ *0.6*, where *I*and *B*are the mean signal and background intensities and *S*~*I*~, *S*~*B*~are the standard deviations. After subtraction of mean background, lowess normalization \[[@B33]\] was performed. Differential expression was analysed with Student\'s t-test (p \< 0.01) and the genes were ranked by the log(p-level).
5. Quantitative RT PCR
----------------------
Primers (Table [4](#T4){ref-type="table"}) were designed to amplify 194--305 b fragments. RNA was processed with Rnase-free Dnase (Promega). Synthesis of cDNA with Superscript III reverse transcriptase (Invitrogen) was primed with oligo(dT). Analyses were carried out using Dynamo SYBR Green kit (Finnzymes) and ABI Prism 7700 (Amersham-Pharmacia).
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Primers used for qPCR.
:::
------------------------------------------------------------------------
**Gene** **Sequence**
-------------------------------- ---------------------------------------
GRB2-related adaptor protein 2 Forward 5\'-GCCAGAGCACCCCAGGAGAT-3\'\
Reverse 5\'-GGCTGAGAGGATGGGGCTGA-3\'
Collagenase type IV Forward 5\'-AACATCAGAAACGCCCTCAT-3\'\
Reverse 5\'-TGGTGGTAGTGGTAGTGGAC-3\'
Troponin T Forward 5\'-TGGGAAGAAGGAAACTGAGA-3\'\
Reverse 5\'-CTCTTACGCAGGGTTGTGAC-3\'
40S ribosomal protein S12 Forward 5\'-AGACCGCACTCATCCACGAC-3\'\
Reverse 5\'-CCACTTTACGGGGTTTTCCT-3\'
EST1 Forward 5\'-CGGAGAAGGAGAACCCACAG-3\'\
Reverse 5\'-CCCTCAAACAAGCAAAGTG-3\'
EST2 Forward 5\'-GCAAATGACAGCCCTCTTAG-3\'\
Reverse 5\'-AGCAGGTTTTCATCAAGGA-3\'
------------------------------------------------------------------------
:::
Author\'s contributions
=======================
AK designed microarray, carried out experiments with fish and data analyses and drafted the manuscript. HK conducted the microarray analyses. PP developed software for annotation of genes and analyses of Medline abstracts. CR constructed the multi-tissue cDNA library and provided the selected genes. SA developed software for management of microarray data and performed the statistical analyses. HM coordinated research. All authors read and approved the final manuscript.
Acknowledgements
================
This study was supported by the National Agency of Technology, Finland. We wish to thank Rolf Sara (CBT, University of Turku) for preparation of cDNA microarrays and Seppo Kukkonen for analyses of plasma cortisol.
|
PubMed Central
|
2024-06-05T03:55:51.960050
|
2005-1-6
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545953/",
"journal": "BMC Genomics. 2005 Jan 6; 6:3",
"authors": [
{
"first": "Aleksei",
"last": "Krasnov"
},
{
"first": "Heikki",
"last": "Koskinen"
},
{
"first": "Petri",
"last": "Pehkonen"
},
{
"first": "Caird E",
"last": "Rexroad"
},
{
"first": "Sergey",
"last": "Afanasyev"
},
{
"first": "Hannu",
"last": "Mölsä"
}
]
}
|
PMC545954
|
Background
==========
Splenic metastases from solid tumors occur in late stage of a disease, so those can hardly be an indication for a surgery. Cancers of the ovaries, lung, breast, stomach, skin and colon are known to metastasize to the spleen \[[@B1]\]. Even though recent reports suggest increasing incidence of splenic metastasis from gynecologic tumors \[[@B2]-[@B8]\], the number of the cases that isolated splenic metastasis is fewer than 25 in the literature worldwide. Among them, solitary parenchymal metastasis would comprise the small portion. Fewer than 15 cases of splenic metastasis occur from the ovaries as a primary site, pathology revealed cystadenocarcinoma.
We present a case of mucinous cystadenocarcinoma that recurred in the splenic parenchyma.
Case presentation
=================
A 29-year-old woman was admitted our hospital on August 1999, had been diagnosed as mucinous tumor of borderline malignancy a year ago. She was followed up at a local clinic. On the follow-up study, the patient\'s CEA level was raised to 43.32 U/L and CT scan showed splenomegaly with cystic lesion. Her past medical history was not significant. She already underwent two surgeries. The first surgery, right salpingo-oophorectomy, was performed at the age of 22 after being diagnosed as dermoid cyst. She was healthy thereafter. Seven years later, left ovarian mass was found on the routine check. At the second surgery, left ovarian mass excision, mucinous tumor of borderline malignancy was diagnosed.
During follow-up after the second surgery, she was referred to our hospital on the suspicion of carcinomatosis peritonei. On the preoperative evaluation (Figure [1](#F1){ref-type="fig"}), splenic lesion, which had been existed for 2 years, was merely noticed as simple cystic lesion unrelated to the ovarian mass.
To exclude peritoneal carcinomatosis, open laparotomy was perfomed. On opening the abdomen, no abnormal gross findings were found except the splenic lesion, which reported as probable metastatic adenocarcinoma on frozen sections. After splenectomy was carried out, peritoneal washings and multiple biopsies on the omentum, peritoneum, mesentery, and left ovary were performed to rule out possible microscopic peritoneal dissemination. Suspecting the transabdominal metastasis, 100 mg of cisplatin was infused into the peritoneal cavity at the end of the operation; however no intraperitoneal recurrence was confirmed after tissue diagnosis.
Final pathologic examination (Figure [2](#F2){ref-type="fig"}) showed metastatic mucinous cystadenocarcinoma. Peritoneal washings and multiple biopsies were all negative. The patient was recovered from the surgery without the evidence of sepsis of severe thrombocytosis. The patient received 5 courses of Taxol and Carboplatin as postoperative chemotherapy.
Two years after the surgery, 3 × 3-sized mass on the left ovary, which assumed to be recurrence, was detected. She has been followed up at outpatient department receiving symptomatic treatment and chemotherapy.
Discussion
==========
The frequency of splenic metastasis has been reported 2.3 to 7.1 per cent from autopsy series of cancer patients \[[@B9]-[@B11]\]. Splenic metastases from the ovaries, uterus, uterine cervix, lung, breast, stomach, skin and colon have been reported, and ovarian cancer comprises the three fourth.
Until now, in the literature, splenic capsular metastasis has been reported to occur in the case of far advanced stage of a disease or in the case of more than one organ was already involved in terminal stage and especially in the case of peritoneal metastasis. When one or more organ in the thoracic cavity and abdominal cavity was involved, the rate of splenic metastasis was 43% \[[@B12]\]. But isolated splenic metastasis is rare; only fewer than 25 cases were reported. Some reported hematogenous metastasis to the parenchyma of the spleen \[[@B13]\]. This mean the spleen is the organ of the privilege \[[@B11],[@B14],[@B15]\]. It could be explained by hypothesis of the role of the splenic capsule as physical barrier; the lack of afferent lymphatics in the splenic parenchyma, the acute angle of the origin and the tortuosity of the splenic artery, the rhythmic contractile properties of the spleen, and the immune competence and possible antineoplastic nature of the splenic tissue itself \[[@B13]\]. In this case, solitary splenic parenchymal metastasis from ovarian epithelial tumor was made hematogenously without intraabdominal dissemination.
Among cases of the isolated splenic metastasis, the ovarian cancers make up the most. After Minazawa et al \[[@B16]\] firstly performed splenectomy in the splenic metastasis of ovarian cancer patient, splenectomy, as a therapeutic modality of splenic metastasis, was supported by similar articles and studies \[[@B17],[@B18]\]. Recently, splenectomy was often included in the cytoreductive surgery of the ovarian cancer \[[@B2],[@B3],[@B19]\].
CT scanning and measurement of serum tumor markers, especially CA 125, are helpful for detecting the recurrence and the infrequent splenic metastasis.
We proposed that splenectomy be a proper therapeutic modality for an isolated splenic metastasis, especially parenchymal metastasis, from an ovarian cancer. When isolated splenic recurrence is suspected on the CT scanning and serum tumor markers, intraabdominal gross findings should be examined meticulously. If only spleen was under suspicion of recurrence, splenectomy would be a proper therapeutic procedure.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
YS carried out the literature search and prepared the manuscript. JC Kim, the corresponding author, was the main operator in charge of the case. CK was involved in the operation and the patient active management. Three authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2407/4/96/prepub>
Acknowledgement
===============
Written consent for publication of the case was obtained from a relative of a patient.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Abdomen CT shows 9 × 8 cm-sized multiseptated cystic lesion with inner calcification in the spleen.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
The section reveals papillary hyperplasia of mucinous lining epithelium with stratification and atypism (H&E stain, ×250).
:::
:::
|
PubMed Central
|
2024-06-05T03:55:51.962918
|
2004-12-22
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545954/",
"journal": "BMC Cancer. 2004 Dec 22; 4:96",
"authors": [
{
"first": "Yang Seok",
"last": "Koh"
},
{
"first": "Jung Chul",
"last": "Kim"
},
{
"first": "Chol Kyoon",
"last": "Cho"
}
]
}
|
PMC545955
|
Background
==========
Razoxane (ICRF-159) (*Raz*), first developed in UK, was the earliest agent against spontaneous metastasis for the murine model (Lewis lung carcinoma) in 1969 \[[@B1]\]. A large volume of papers and projects have been published in the utilities and mechanisms of *Raz*for anticancer actions, like assisting radiotherapy, \[[@B2]\] overcoming multi-drug resistance (MDR) of daunorubicin and doxorubicin \[[@B3]\], inhibiting topoisomerase II \[[@B4]\] and so on. More importantly, *Raz*, as a cardioprotectant of anthrocyclines, has been licensed in 28 countries in 4 continents. Since morpholine groups in some structures were reported to be responsible for cytotoxic or modulative actions on tumors, an anticancer agent, probimane \[1,2-bis (N^4^-morpholine-3, 5-dioxopeprazine-1-yl) propane; AT-2153, Pro\] was synthesized by introducing two morpholine groups into *Raz*in China.\[[@B5]\]. *Raz*and *Pro*belong to *bisdiopiperazines*. Like *Raz*, *Pro*also exhibits anti-tumor activity both *in vivo*and *in vitro*against experimental tumor models in a small scale investigation \[[@B6],[@B7]\] and limited clinical data showed that *Pro*could inhibit human malignant lymphoma even for those resistant to other anticancer drugs \[[@B8]\]. Pro exhibits the same pharmacological effects as *Raz*, like detoxication of *Adriamycin*(*ADR*) induced cardiotoxicities, and synergism with *ADR*against tumors \[[@B9],[@B10]\]. We have found some novel biological effects of *Pro*, like inhibiting the activity of calmodulin (*CaM*), a cell-signal regulator, which can explain anticancer actions and the combined cytotoxic effect of *Pro*and *ADR*\[[@B11]\]. Pro was also shown to inhibit lipoperoxidation (*LPO*) of erythrocytes \[[@B12]\], influence tumor sialic acid synthesis \[[@B13]\] and inhibit the binding of fibrinogen to leukemia cells \[[@B14]\].
Lung cancer is the No 1 killer among all categories of cancers in urban areas in China and many Western countries. The high mortality rate of lung cancer can easily be caused by inducing multi-drug resistance (*MDR*) and by high metastatic occurrence in clinics \[[@B15]\]. Since we assume that *Pro*, like *Raz*may possess useful therapeutic potentialities, we evaluated *in vivo*the chemotherapeutical parameters of *Pro*and *Raz*for lung cancer of both murine and human origins.
Results
=======
Lethal toxicity of Pro and Raz in mice
--------------------------------------
The lethal dosage of *Pro*and *Raz*is tabulated in Table [1](#T1){ref-type="table"}. Since the toxicity of *Pro*and *Raz*seemed to lack sex specificity in mice, we were able to combine their numbers for LD~50~and LD~5~calculations. We used the approximate dosage of LD~5~of Pro (60 mg/kg ip × 7) and Raz (20 mg/kg ip × 7) as equitoxic dosages for further treatment studies.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
The subacute toxicity of Pro and Raz in mice: Mouse survival was observed for 1 month. The numbers of mice in each group were 20 for each of the 5 dosages of a single agent.
:::
Drugs Protocols LD~5~mg/kg LD~50~mg/kg
----------- ----------- ------------ -------------
Probimane ip × 10 66 121
Razoxane ip × 10 23 53
:::
Antitumor and antimetastatic effects of Pro and Raz on LLC
----------------------------------------------------------
Antitumor and antimetastatic effects of *Pro*and *Raz*on *LLC*are tabulated in Table [2](#T2){ref-type="table"} and Table [3](#T3){ref-type="table"}. *Pro*and *Raz*at equitoxic dosages (LD~5~) showed a noticeable anticancer effect on primary tumor growth (inhibitory rates, approximately 30--45 %), and significantly inhibited the formation of tumor metastases (inhibitory rates on pulmonary metastasis \> 90 %, P \< 0.001). Primary tumor growth of *LLC*was inhibited more by *Pro*(48 %) than by *Raz*(40.3%) in a 20 day trial, whereas the inhibition of *Pro*(35.7%) was slightly less than that of *Raz*(40 %) on an 11 day trial. Pro seems to be more persistent than Raz in inhibiting primary tumor growth of *LLC*.
Antitumor effects of bisdioxopiperazines for different schedules and in combination with other anticancer drugs
---------------------------------------------------------------------------------------------------------------
Antitumor effects of *Raz*and *Pro*on *LLC*are included in Table [4](#T4){ref-type="table"}, [5](#T5){ref-type="table"}, [6](#T6){ref-type="table"}. We evaluated 1, 5 and 9 day administration schedules in our study. We found that *Raz*and *Pro*were effective in a statistically significant manner with the 3 injection schedule of the 1, 5 and 9 day administrations on *LLC*. If we administered *Raz*to tumor-bearing mice once on day 1, 5 and 9, there was no difference between treatment and vehicle control. Antitumor effects of *Raz*in combination with *Ble*on *LLC*(73.3 %) were better than those in combination with *Dau*(56.3 %) (Table [5](#T5){ref-type="table"} and Table [6](#T6){ref-type="table"}). *Pro*also showed synergistic effects in combination with *Ble*(Table [7](#T7){ref-type="table"}).
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
The influence of Pro and Raz on primary tumor of LLC (using Student T-test): Route: ip × 7 daily. Experiment term was 11 days. \* P \< 0.05 (treatment vs vehicle control). The numbers of mice were 30 for the control group and 20 for each treatment group. 100 % survival was observed in each group.
:::
Compounds Dosage mg/kg/d Body weight (g) Tumor weight (g) Tumor inhibition%
----------- ---------------- ----------------- ------------------ -------------------
Control \-- 23.3/24.4 2.80 ± 0.04 \--
Razoxane 20 23.3/23.4 1.61 ± 0.03\* 40.0
Probimane 30 23.4/21.6 1.91 ± 0.03\* 32.1
Probimane 60 23.3/23.8 1.80 ± 0.03\* 35.7
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
The influence of Pro and Raz on primary and metastatic tumor of LLC: PTI (%) -- Primary tumor inhibition. MFCPM -- metastatic foci count per mouse. Route: ip × 7 every 2 days. Experiment term was 20 days, \* P \< 0.001(treatment vs vehicle control). The numbers of mice were 30 for both control group and each treatment group. 100 % survival was observed in each group.
:::
Compounds Dosage mg/kg/d Body weigh (g) PTI(%) MFCPM
----------- ---------------- ---------------- -------- -------------
Control \-\-- 22.8/21.4 \-- 30.9 ± 7.3
Razoxane 20 22.7/21.5 40.3 1.2 ± 0.5\*
Probimane 30 23.3/22.5 42.0 1.5 ± 0.5\*
Probimane 60 23.3/20.3 48.0 1.0 ± 0.2\*
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Antitumor effects of bisdioxopiperazines of different schedules on Lewis lung carcinoma: \*Administration every 3 hours, 16 mice were included in each testing group. \*\*p \< 0.05 (treatment vs control), Experimental term was 11 days
:::
Compounds Dosage Schedule Tumor weight Tumor inhibition
----------- -------- ----------------- ----------------- ------------------
Control \-- \-- 2.36 ± 0.05
Razoxane 80 1 time a day 2.49 ± 0.05 -5.5
Razoxane 40 1 time a day 2.32 ± 0.07 1.7
Razoxane 20 1 time a day 2.80 ± 0.06 -18.6
Razoxane 10 3 times a day\* 1.51 ± 0.04\*\* 36.0
Probimane 20 3 time a day\* 1.19 ± 0.05\*\* 49.6
:::
::: {#T5 .table-wrap}
Table 5
::: {.caption}
######
Antitumor effects of Raz on Lewis lung carcinoma in combination with daunorubicin: \*Administration every 3 hours. Experimental term was 11 days
:::
Compounds Dosage Schedule Tumor weight (g) Tumor inhibitions
-------------------- -------- ---------------------- ------------------ -------------------
Control 2.34 ± 0.05
Razoxane (Raz) 10 3 times a day\* 1.57 ± 0.05 32.9
Daunorubicin (Dau) 2 1 time a day 1.10 ± 0.04 53.0
Raz + Dau 10 + 2 3 times/1 time a day 1.02 ± 0.04 56.4
:::
::: {#T6 .table-wrap}
Table 6
::: {.caption}
######
Antitumor effects of Raz on Lewis lung carcinoma in combination with bleomycin: \* Administrate every 3 hours in one day. \*\* p \< 0.01 (treatment vs vehicle control). Experimental term was 11 days
:::
Compounds Dosage Schedule Tumor weight Tumor Inhibition
----------------- --------- ------------------------ ----------------- ------------------
Control \-- \-- 2.46 ± 0.06
Razoxane (Raz) 10 3 times a day\* 1.44 ± 0.07 41.5
Bleomycin (Ble) 15 1 time a day 1.50 ± 0.06 39.0
Raz + Ble 10 + 15 3 times + 1 time a day 0.66 ± 0.05\*\* 73.2\*\*
:::
::: {#T7 .table-wrap}
Table 7
::: {.caption}
######
Antitumor effects of Pro on Lewis lung carcinoma in combination with daunorubicin or bleomycin: \*Administration every 3 hours. Experimental term was 11 days
:::
Compounds Dosage Schedule Body weight Tumor weight (g) Tumor inhibitions
----------- --------- ---------------------- ------------- ------------------ -------------------
Control \-- \-- 20.6/21.6 2.62 ± 0.08
Pro 20 3 times a day 20.6/20.8 1.45 ± 0.07 44.6
Dau 2 1 time a day 20.6/20.0 1.14 ± 0.08 56.5
Ble 15 1 time a day 20.7/21.2 1.36 ± 0.08 48.1
Pro + Dau 20 + 2 3 times/1 time a day 20.6/20.9 1.07 ± 0.05 59.2
Pro + Ble 20 + 15 3 times/1 time a day 20.7/19.8 0.59 ± 0.04 77.5
:::
Antitumor activity of Pro and Raz on LAX-83
-------------------------------------------
The experiments showed that LAX-83 was sensitive to *Raz*(40--60 mgKg^-1^, ip × 5) and *Pro*(80--100 mgKg^-1^ip × 5) with inhibitory rates of 25--32 % and 55--60 % respectively (P \< 0.01 vs control). *CTX*, as a positive anticancer drug (40 mgKg^-1^ip × 5), exhibited antitumor activities against the growth of LAX-83 with an inhibitory rate of 84 %. Obvious necrosis in tumor tissues was observed by histological evaluation of *CTX*and *Pro*treatment groups, but *Pro*showed larger vacuoles than *CTX*. Drug inhibition on tumor volumes were calculated and outlined in Table [8](#T8){ref-type="table"}. We have tested the 5 most commonly used anticancer drugs -- cyclophosphamide (CTX), 5-fluoruoracil (5-Fu), methotrexate (MTX), cisplatin (DDP) and vincristine (VCR) (Table [9](#T9){ref-type="table"}). In the LAX-83 model, CTX has been shown to be the most effective one. The anticancer effect of *Pro*was the same or better than those of MTX, DDP and as well as 5-Fu against LAX-83 tumor growth.
::: {#T8 .table-wrap}
Table 8
::: {.caption}
######
Antitumor activities of Pro and Raz on human tumor LAX-83 using subrenal capsule assay: Route: ip × 5 daily from the day after surgery. \* P \< 0.05, \*\* P \< 0.001 (treatment vs vehicle control). Experiment was completed within 7 days. Tumor volume = 1/2 × width^2^× length (using T-test)
:::
Compounds Dosage mg/kg/d No mice Body weight (g) Tumor volume (mm^3^) Inhibition%
------------------ ---------------- --------- ----------------- ---------------------- -------------
Control \-\-- 16 19.2/21.0 39.8 ± 3.2 \--
Razoxane 40 12 20.8/21.5 29.7 ± 3.0\* 25
Razoxane 60 12 19.8/18.8 27.2 ± 2.8\* 32
Probimane 80 12 20.0/19.6 18.0 ± 2.6\*\* 55
Probimane 100 12 20.0/20.0 15.8 ± 2.6\*\* 60
Cyclophosphamide 40 12 21.0/20.9 6.4 ± 2.0\*\* 84
:::
::: {#T9 .table-wrap}
Table 9
::: {.caption}
######
Antitumor activities of anticancer drugs on human tumor LAX-83 using subrenal capsule assay: Route: ip × 5 daily from the day after surgery. \* P \< 0.05, \*\* P \< 0.001 (treatment vs vehicle control). Experiment was completed within 7 days. Tumor volume = 1/2 × width^2^× length (using T-test)
:::
Compounds Dosage mg/kg/d No mice Body weight (g) Tumor volume (mm^3^) Inhibition%
------------------ ---------------- --------- ----------------- ---------------------- -------------
Control \-\-- 16 20.9/22.5 29.7 ± 3.2 \--
Methotrexate 1.5 12 21.2/21.9 27.4 ± 3.0 7.7
Cis-platin 1.5 12 22.8/21.7 16.6 ± 2.6\*\* 44.1
5-fluoruoracil 37.5 12 21.7/21.4 12.8 ± 2.6\*\* 57.5
Cyclophosphamide 30.0 12 21.0/20.9 5.8 ± 2.3\*\* 80.5
Vincristine 0.3 12 20.8/20.8 7.6 ± 2.2\*\* 74.4
:::
Discussion
==========
Explanations of anticancer and antimetastatic mechanisms of *bisdioxopiperazines*are now inconclusive. The present explanation for the anticancer mechanisms of *Raz*has been attributed to antiangiogenesis and topoisomerase II inhibition.\[[@B16]\] Since the antimetastatic activities of *Raz*and *Pro*were much stronger than those actions against primary tumor growth, this special targeting on metastasis ought to be more useful in clinical cancer treatment. *Raz*and *Pro*show typical characteristics of antiangiogenesis agents, which target small nodule of tumors. Meanwhile, recent reports on drugs targeting *angiogenesis*indicate that most anti-vascular drugs have low or even no effects on most cancers when they are used alone in clinics, but they show synergistic effects in combination with other anticancer drugs. \[[@B17],[@B18]\] Our study shows synergistic anticancer actions of *Raz*and *Pro*with *Ble*or *Dau*basing on this theory. Previous work showed that *Pro*and *Raz*could reduce the cardiotoxicity of *anthrocycline*,\[[@B1],[@B9],[@B10]\] so we may reasonably deduce that they can also reduce the cytotoxicity of *anthrocyclines*. The data in our study suggests that the synergistic effects of *Raz*with *anthrocyclines*are present, but not as potent as those with *Ble*.
Since we have tested the antitumor activity of clinically available anticancer drugs (CTX, 5-Fu, MTX, DDP and VCR) against LAX-83, CTX being the best one, two bisdioxopiperazines studied on this work show overall similar anticancer effective as commonly used drugs. Although the anticancer effects of CTX and VCR are better than those of Pro, for other commonly used drugs, such as DDP, MTX and 5-Fu, the antitumor effects are no better than those of Pro. Since the antitumor effects of MTX and DDP are even less effective than those of *Pro*and *Raz*, we suggest that anticancer effects of *Pro*and *Raz*are within the effective anticancer ranges of commonly available anticancer drugs.
The other useful property of *Pro*is that it is the most water-soluble among the *bisdioxopiperazines*. Most *bisdioxopiperazines*are less water-soluble and given orally in clinics. Although oral administration is easy for patients, bioavailability varies from patient to patient. For some patients who have a poor absorption of *bisdioxopiperazines*in oral administration, *Pro*can be injected *iv*to maintain stable drug levels. Our previous work showed that *Pro*could strongly accumulate in tumor tissue while *Pro*levels in other tissues decrease rapidly \[[@B19]\]. Presently, a stereo-isomer of *Raz*, (dexrazoxane, *ICRF-187*), a water-soluble Raz, is being reinvestigated and has aroused the interests of clinical oncologists. Phase III clinical studies are currently underway in the US. More importantly, *ICRF-187*was licensed in 28 countries in 4 continents. This work shows a noticeable inhibition of *Pro*and *Raz*on lung cancers and suggests possible usage of *Raz*and *Pro*on lung cancer in clinics.
Conclusions
===========
The advantages of *bisdioxopiperazines*in clinical treatment of lung cancers are as follows: (i) *Pro*and *Raz*can inhibit the growth of lung cancers, with and without the help of other anticancer drugs, like *Dau*and *Ble*; (ii) like *Raz*, *Pro*strongly inhibits spontaneous pulmonary metastasis of *LLC*; (iii) since *Pro*can inhibit *CaM*\[[@B11]\], a calcium activated protein that\'s associated with *MDR*and metastatic phenotypes, synergistic anticancer effects of *Pro*and *Raz*can be expected in combination with other anti-cancer drugs, like *Dau*or *Ble*. Now, new concepts of the relationship between tumor metastasis and *MDR*in cancers have been stated,\[[@B20]\] whereas *bisdioxopiperazines*can inhibit both tumor metastasis and *MDR*. As a counterpart of *Raz*, *Pro*might be of interest and have chemotherapeutic potential in clinics.
Methods
=======
Drugs and animals
-----------------
Cyclophosphomide (*CTX*), daunorubicin (*Dau*) and bleomycin (*Ble*), 5-fluororacil (5-Fu), vincristine (VCR), cisplatin (DDP), methotrexate (MTX) were purchased from Shanghai Pharmaceutical Company. Pro and Raz were prepared by Department of Medicinal Chemistry, Shanghai Institute of Materia Medica, Chinese Academy of Sciences. C57BL/6J and Kun-Min strain mice were purchased from Shanghai Center of Laboratory Animal Breeding, Chinese Academy of Sciences. Nude mice (Swiss-DF), taken from Roswell Park Memorial Institute, USA, were bred in Shanghai Institute of Materia Medica, Chinese Academy of Sciences under a specific pathogen free condition. Human pulmonary adenocarcinoma xenograft (*LAX-83*)\[[@B21]\] and Lewis lung carcinoma (*LLC*) were serially transplanted in this laboratory. All animal experiments were conducted in compliance with the Guidelines for the Care and Use of Research Animals, NIH, established by Washington University\'s Animal Studies Committee. Bouin\'s solution consists of water saturated with picric acid: formaldehyde: glacial acetic acid (75: 20: 5, v/v/v).
Lethal dosage determination in mice
-----------------------------------
Mice of Kun-Min strain (equal amount of male and female) were *ip*injected with Pro and Raz daily for 10 successive days. The deaths of mice were counted after 1 month. Lethal dosage of agents was calculated by *Random Probity tests*.
Antitumor and antimetastatic studies of LLC
-------------------------------------------
C57BL/6J mice were implanted *sc*with *LLC*(2 × 10^6^cells) from donor mice. The mice were injected intraperitoneally with drugs daily or every two days for 7 injections. On day 11 or day 20, mice were sacrificed, and locally growing tumors were separated from skin and muscles and weighed, and lungs of host mice were placed into a Bouin\'s solution for 24 h, and then the lung samples were submerged into a solution of 95 % alcohol for 24 h. Finally, the numbers of extruding metastatic foci in lungs were counted.
Antitumor actions of different schedules and in combinations with different drugs
---------------------------------------------------------------------------------
C57BL/6J mice were implanted *sc*with *LLC*(2 × 10^6^cells) from donor mice. Mice were injected intraperitoneally with drugs on day 1, 5, 9. Single injection or 3 injections every 3 hours were used. Tumors were separated and weighed on day 11.
Antitumor activity study of human tumors
----------------------------------------
Nude mice were inoculated with LAX-83 under the renal capsule (SRC method).\[[@B22]\] Nude mice were injected intraperitoneally with drugs daily during next five days after inoculation of *LAX-83*. Then nude mice were sacrificed, and their kidneys were taken out for measurement of tumor sizes using a stereomicroscope a week after transplantation. Tumor volume was calculated as 1/2(ab^2^) where a and b are their major and minor axes of the lump. Kidneys with tumors were paraffin-embedded, sliced and hematoxylin dyed. The tumor tissues were then observed from a light microscope.
Statistical analysis
--------------------
*Student\'s t-test*was used to assess the differences between control and drug treatment groups of above methods.
List of abbreviation used are
=============================
Pro, probimane; Raz, razoxane; CaM, calmodulin; LPO, lipoperoxidation; Dau, daunorubicin; Ble, bleomycin; LLC, Lewis lung carcinoma, LAX-83; a lung adenocarcinoma xenograft; ADR, adriamycin;
Author\'s contribution
======================
The experimental design was made by Bin Xu and Da-Yong Lu.
Experiments were performed by Da-Yong Lu (anticancer activity tests)
The manuscript was written by Da-Yong Lu, and Jian Ding.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Structural formulas of razoxane and probimane
:::
:::
Acknowledgements
================
This work was supported by Science and Technology Foundation of China.
|
PubMed Central
|
2024-06-05T03:55:51.963781
|
2004-12-24
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545955/",
"journal": "BMC Pharmacol. 2004 Dec 24; 4:32",
"authors": [
{
"first": "Da-Yong",
"last": "Lu"
},
{
"first": "Bin",
"last": "Xu"
},
{
"first": "Jian",
"last": "Ding"
}
]
}
|
PMC545956
|
Background
==========
Recombinant adenoviral vectors are highly efficient for in vitro and in vivo gene delivery. They can easily be produced in large numbers, transduce a wide variety of cell types and generate high levels of transgene expression. The AdEasy system is a widely used, simplified system for generating recombinant adenoviral vectors, which are created with a minimum of enzymatic manipulations and by employing homologous recombination in E. coli \[[@B1]\]. The system consists of two adenoviral backbone vectors (pAdEasy-1 with deleted E1/E3 and pAdEasy-2 with deleted E1/E3/E4) and four different shuttle vectors (pShuttle, pShuttle-CMV, pAdTrack, pAdTrack-CMV), into which the desired transgenes are inserted. The polylinkers are surrounded by adenoviral sequences that allow homologous recombination with adenoviral backbone plasmids in E. coli. The shuttle vectors differ by partly carrying a cytomegalievirus (CMV) promoter and GFP as a tracer all of which contain a kanamycin resistance gene. Therefore, the various components can easily be combined depending on the desired purpose. In this paper we describe a simplified and easy method for screening recombinant DNA within the AdEasy system. This Duplex-PCR-method is independent of the transgene or insert and can be used for the complete AdEasy-System. It is characterized by a simple standard protocol and the results can be obtained within a few hours. The PCR is run with two different primer sets. The primers KanaFor and KanaRev hybridizise with the Kanamycin resistence gene and AdFor and AdRev detect the adenoviral backbone. In case of recombinant clones, two diagnostic fragments with a size of 384 bp and 768 bp are generated.
Methods
=======
The presented Duplex-PCR is performed as follows: After Co-transformation of the Pme I-digested shuttle vector with the adenoviral backbone plasmid to E. coli (BJ 5183) and plating on agar (selection on kanamycin), half of the overnight grown colonies are picked and used directly as template for the colony-PCR. The other half of the colony is used for inoculation with LB-kanamycin and then incubation at 37°C. Only the positive, recombinant clones which have been detected by PCR are grown overnight, a minipreparation of DNA is then performed the next morning.
The PCR is run using two different primer sets. The primers KanaFor (5\' CAA GAT GGA TTG CAC GCA GG 3\') and KanaRev (5\'AAG GCG ATA GAA GGC GAT GC 3\') hybridize to the Kanamycin resistence gene and AdFor (5\'GGC TGC TCT GCT CGG AAG AC 3\') and AdRev (5\'GGC ATA CGC GCT ACC CGT AG 3\') detect the adenoviral backbone.
The optimised concentrations of the components for the Duplex-PCR were as follows: 2, 5 mM MgCl~2~, 100 mM dNTPs and 0,2 Units Taq polymerase. The bacteria are denatured at 95°C for 10 minutes. The PCR-products are amplified by 40 cycles of annealing at 58°C (30 sec), extension at 72°C (30 sec) and denaturation at 94°C (30 sec). In our experience, this procedure produced the best results without generating false positive clones (Figure [1](#F1){ref-type="fig"}).
PCR products are analyzed by agarose gel electrophoresis, half of the reaction volume (25 μl) is size fractionated with 80 V for 1 h in 1% agarose in the presence of ethidium bromide and the resulting bands visualized with ultraviolet illumination.
The DNA obtained by small scale alkali lysis from the recombinants is then extracted twice with a phenol-chloroform protocol, precipitated and carefully resuspended in 20 μl RNAse-free water. The construct is linearized with PacI and directly transfected into 911 cells, which are monitored for cytopathic effects, i.e. production of recombinant adenoviruses. The cytopathic effect is usually seen within 5 to 10 days. The expression of the transgene is confirmed by Western blot analysis. The practicability of our procedure was verified with three different transgenes: Cytosin Deaminase (AdCD), p53 (Adp53) and Granulocyte Macrophage Colony Stimulating Factor (AdGM-CSF).
Results and discussion
======================
The conventional way of screening for recombinants after Co-transformation of the linearized shuttle vector with the adenoviral backbone vector in E. coli is by plating on LB/kanamycin, growing the bacteria overnight, then picking the colonies and growing them again for 10--15 hours. Minipreps are then performed and the size is evaluated on agarose gels. A restriction digest with three different restriction enzymes is then done and finally again another agarose gel is run. This is a relatively time-consuming and laborious procedure, which takes about 2 days.
In contrast, the presented alternative protocol allows fast detection of recombinants with a simplified technique by minimizing the amount of necessary steps for generating a recombinant adenovirus. The method is time sparing and cost-effective. In our experience, the above described protocol showed no problems with false negative clones. After optimisation of the PCR protocol, we were able to run the conventional screening method (e.g. by restriction digest) for recombinant clones at the same time as the presented simplified PCR. We found no differences in regard to the final results for the two methods, but it has to be kept in mind that only a limited number of recombinant adenoviruses were actually generated with the new technique. Furthermore, we exclusively then continued our work with the AdCD-virus. Therefore it cannot be ruled out that under other conditions the presented technique may produce other results than the conventional technique. The positive clones were processed and finally transfected into 911 cells. After harvesting recombinant adenovirus, we infected cells with AdCD and checked the expression of the Cytosin deaminase protein by Western blot. The functionality of the gene was proven by FACS analysis. We confirmed expression of the protein as well as its functionality.
Conclusions
===========
The presented protocol allows fast detection of recombinant clones within the AdEasy system with an easy, cost-effective technique. Therefore, this procedure is a potential alternative for screening recombinants within the AdEasy system.
Authors\' contributions
=======================
DA designed the study and was responsible for manuscript preparation, MK, IB optimised the PCR protocol, PK was responsible for manuscript preparation, MB contributed to manuscript preparation and JW was responsible for study design and manuscript preparation. All authors read and approved the final manuscript.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Duplex-PCR Gel electrophoresis for PCR products from Colony Duplex PCR with two primer sets AdFor/AdRev and KanaFor/KanaRev on AdEasy system generating two diagnostic fragments of 384 bp and 768 bp (see arrows) after homologous recombination.(N) negative control; (1) pShuttle/CD; (2) pAdEAsy 1; (3) pAdEasy/CD, e.g. recombinant clone
:::
:::
|
PubMed Central
|
2024-06-05T03:55:51.965960
|
2005-1-11
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545956/",
"journal": "BMC Biotechnol. 2005 Jan 11; 5:1",
"authors": [
{
"first": "Dalibor",
"last": "Antolovic"
},
{
"first": "Moritz",
"last": "Koch"
},
{
"first": "Inga",
"last": "Bohlmann"
},
{
"first": "Peter",
"last": "Kienle"
},
{
"first": "Markus",
"last": "Büchler"
},
{
"first": "Jürgen",
"last": "Weitz"
}
]
}
|
PMC545957
|
Background
==========
Evolution can be measured and studied on a number of different scales, one of which is through the determination and comparison of genetic sequence information. Current-day gene sequences in living organisms have arisen through modifications of an array of ancestral sequences. Duplication with modification is the central paradigm of protein evolution, wherein new proteins and/or new biological functions are fashioned from earlier ones. \[[@B1]\]. To detect these functionally (and structurally) related proteins based upon similarity between their primary nucleic acid or amino acid sequences, a variety of sequence comparison algorithms have been developed. When a new gene is cloned and sequenced, it is now standard practice to use these algorithms to search for similarities between the translated nucleic acid sequence and a protein sequence database such as the NCBI non-redundant protein database (nr) \[[@B2]\]. Sequence similarity that implies similar structure and therefore similar protein or enzymatic function is not definitive proof of such function; however, the results of database sequence similarity searches provide a starting point for researchers attempting to ascertain the function of an unknown gene by supporting the intelligent design of subsequent laboratory experiments.
The basic local alignment search tool (BLAST) \[[@B3],[@B4]\] is by far the most widely used pairwise-based sequence similarity comparison tool. It completes searches more swiftly than other tools, including FASTA, SSEARCH \[[@B5]\], and SCANPS \[[@B6]\]. BLAST uses an efficient, rapid algorithm to look for short segments or words of sequence similarity between two sequences that meet some predefined scoring threshold. After initially locating at least two of these words within a short distance of one another on a common diagonal, the algorithm uses them as \"seeds\" from which to extend the alignment to encompass longer regions of similarity, resulting in high scoring pairs (HSPs). The heuristic algorithm used by BLAST decreases search time dramatically in comparison to that of other search programs. However, the emergence of high-throughput DNA sequencing techniques has increased the size of sequence databases tremendously; thus conventional large-scale BLAST searching against the most commonly used databases has become infeasible on a PC or even a dedicated UNIX server. For this reason, new search strategies are needed.
While BLAST provides a balance between search sensitivity and speed, in many cases a researcher would like to detect more distant sequence similarities by employing search strategies that maximize the sensitivity of the search. A profile-based comparison which, for example, compares a sequence to a hidden Markov model (HMM) representing an empirically derived estimate of all possible evolutionary changes for a protein of a particular function, generally permits identification of a much higher proportion of distantly related sequences \[[@B7]\]. There are two major profile-based comparison tools. PSI-BLAST \[[@B3]\] compares sequences with a profile model constructed dynamically during the initial search phase of a traditional BLAST search, while HMMPFAM in the HMMer package from Sean Eddy at Washington University compares sequences with a well-curated database of HMM profiles as well as models constructed by users. Well-curated profile databases such as Pfam \[[@B8]\] are being developed through the combined efforts of bioinformaticists and molecular biologists. Profile-based comparison has become a reliable way to gather information for predicting structure and function of unknown genes, and tools from the HMMer package are becoming a key centerpiece in many bioinformatics pipelines. The tradeoff of using HMM-based searches for increased sensitivity is the intrinsically slow nature of the Viterbi \[[@B9]\] or forward algorithm used in the application.
Taking advantage of larger databases and more sensitive searching methods necessitates the use of high performance computing (HPC) platforms. Traditionally, HPC has been synonymous with high-priced vector or parallel supercomputers, but rapid advances in microprocessor and network bandwidth technologies are changing the definition \[[@B10]\]. Clusters of connected workstations utilizing commodity microprocessor systems provide enormous benefits in terms of cost and performance. Thus, cluster computing can meet the increased computational needs of resource- and data-intensive bioinformatics applications. HPC environments using workstations connected via high-speed networks are becoming more and more popular in the bioinformatics community \[[@B11]-[@B14]\]. Here we describe the SS-Wrapper package, which provides tools to adapt currently available similarity search applications onto HPC environments implemented through Linux clusters.
Implementation
==============
The objective of this study was to implement a generic wrapper application that could deploy similarity searching applications on a Linux cluster. Design criteria for the wrapper included the ability to deploy applications without the need to alter the original application and the ability to increase the speed of the underlying application in a linear manner dependent on the number of cluster nodes available. To meet these objectives, tools in the SS-Wrapper package were written in C/C++ using the message passing interface (MPI) \[[@B15]\]. Tools in the SS-Wrapper should work with few if any changes on any Linux cluster running MPI utilizing any hardware platform. In addition to the SS-Wrapper executables (which are compiled using the MPI C++ compiler), executables for the underlying similarity search applications (blastall, formatdb, and hmmpfam) appropriate for the hardware platform underlying the cluster will be needed. These can be obtained from NCBI and Washington University in St. Louis. SS-Wrapper is available without charge under the Artistic License described in the Open Source Initiative \[[@B16]\]. The source code can be downloaded via ftp \[[@B17]\].
Using multiple processors is practical only when the computational task is too large to complete on a single processor in a reasonable amount of time. In the case of database searching, the database to be searched may be too large or the set of query sequences used in the search may be too large to accomplish the search on a workstation with a single or even double or quad processor architecture. Since the search must compare every query sequence to every database sequence, parallelizing the process can be accomplished by three different methods. The first method is to split the query sequence file into smaller subsets and apply each subset to one particular node of the cluster in a search against the entire database. The second method calls for the database to be split into a series of smaller files, one of which is distributed to each node of the cluster; then the entire file of query sequences is searched against each of the database segments. Finally, it is possible to use a combination of query sequence splitting and database splitting to accomplish the search.
Using the query splitting approach does not require any modification to the output generated by search programs such as BLAST. When utilizing the database splitting approach, however, a correction must be made to the reported E-value for any particular hit. Most similarity comparison tools provide a statistical measure (e.g., the E-value reported for BLAST and HMMPFAM) that gives an indication of the statistical significance of a match between the query sequence and a particular database hit. This statistical measure is generally influenced by the size of the search space (which includes the total length of the database) and therefore the E-value needs to be recalculated when only a portion of the database is being searched. An added complication is that some search tools require that the database be intact. For example, PSI-BLAST \[[@B3]\] is a variant of BLAST that constructs a sequence profile model based on hits from an initial round of BLAST searching. This profile is then used and refined in subsequent rounds of searching to increase the sensitivity of the overall search. Because PSI-BLAST depends on the integrity of the database to guarantee that the resulting profiles are representative of the entire search space, it is best to perform parallelization using the query splitting approach.
In contrast to database splitting, query splitting offers greater flexibility in that the query file segments can more easily be adaptively distributed to the nodes of the cluster according to the load on each particular node during the search process. As any one node becomes available, another segment of the query file can be distributed to that node during the search process, which improves performance. It is difficult to predict the workload on any one node before the search begins, and the time required to complete a program running on a cluster depends on the processor that finishes last. Adaptive distribution of the workload maximizes resource utilization. Optimizing resource utilization is dependent on finding a balance between having larger numbers of smaller tasks versus increased startup and communication overhead due to distribution of the required query and database files to each node. For this reason, for our query splitting wrapper (QS-search), we adopted a hybrid strategy wherein approximately 90% of the total workload is evenly divided and distributed to each node at the beginning of the search using a modified bucket algorithm (described below). Then the remaining 10% of the workload is divided and distributed to each node as it becomes available after completing its previous task. This strategy is accomplished using a master-slave model, where one node is set aside to act as the master, which is responsible for distributing the workload and supervising the slave nodes, which perform the computations.
In general, the query splitting approach seems to be superior to the database splitting approach due to higher performance and fewer post-search processing tasks. Database splitting does provide a distinct advantage when the database is too large to fit into the physical memory of a single node \[[@B11],[@B14]\]. NCBI BLAST uses memory mapped file I/O for database access. BLAST runs fastest when it can cache the database in memory. When the database size exceeds that of the available memory, however, the database splitting approach can reduce the possibility of swapping the database from physical memory to disk swap space, which could significantly slow the search process. Therefore, we have also developed a wrapper to support database splitting (DS-BLAST). DS-BLAST is specific for BLAST because it is necessary to include code to recalculate E-values following the search.
As indicated above, a modified bucket algorithm is used to split up the query sequences for QS-search, and a similar algorithm is used to split up the database sequences for DS-BLAST. The modified bucket algorithm works as follows: First, the sequences are sorted according to length. In the first cycle, sequences are placed one at a time into each bucket in descending order of length. In the next two cycles, individual sequences continue to be placed into each bucket after first reversing the order of the buckets. Bucket order is reversed every two cycles and the process continues until all sequences have been placed into a bucket. At the end of this process, each bucket contains nearly the same number of sequences, and the total length of all sequences in any one bucket is also approximately the same as the total sequence length of any other bucket. In the end, therefore, the file of query sequences or the file of database sequences is evenly divided in both length and number.
The BLAST E-value is a function of the size of the effective search space, which is dependent on three factors: the number of sequences in the database, the total combined length of all sequences in the database, and the length of the query sequence \[[@B13],[@B18]\]. Figure [1](#F1){ref-type="fig"} shows that when splitting the database into N fragments of equal sequence number and length, the effective search space of each database fragment is approximately 1/N that of the intact entire database. Therefore, the E-value calculated for any particular hit of a query sequence to a database sequence will approximate a linear function dependent on the value of N. For that reason, at the beginning of the search, DS-BLAST lowers the user-provided E-value cut-off to account for the number of nodes used in the search. Following the search, DS-BLAST recalculates the effective search space and each resulting E-value by multiplying by the value of N.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Effective search space vs. database size and query sequence length.**Relationship between effective search space and database size with different query sequence lengths. The GenBank non-redundant protein database (nr) was split evenly according to the modified bucket algorithm in order to construct databases of a size 1/64, 1/32, 1/16, 1/8, 1/4, 1/2, or 1 of the entire nr database. Query sequences of varying lengths were randomly assembled using a Perl script. A BLAST search was then carried out for each query sequence against each database. The effective search space and database size was extracted from the BLAST results and plotted for each query sequence. The length of each query sequence is indicated next to the line which plots the relationship between effective search space and database size for that query.
:::
:::
Results and discussion
======================
Usage
-----
The QS-search executable (qssearch) provides the same interface for all search tools. The command line is as follows:
qssearch -c \<command\> -q \<query\> -d \<database\> -o \<output\> -l \<local scratch\> -x \<database files\>
DS-BLAST uses two executables: dsblast and dsformatdb. dsformatdb is responsible for splitting the database into fragments according to the modified bucket algorithm and then formats these fragments using the NCBI formatdb executable.
The command line for dsformatdb is as follows:
dsformatdb -n \<number\> -c \<command\> -d \<database\> -p \<path\>
The command line for dsblast is as follows:
dsblast -o \<output\> -c \<command\> -l \<local scratch\> -d \<database\> -q \<query\>
The command-line variables are as follows:
• -c command: normal command line used for the underlying application including all desired options
• -q query: query filename in fasta format
• -d database: database filename
• -o output: output filename
• -l local scratch: temporary directory on each node
• -x database files: a space-delimited list of the database file names generated by the search program\'s formatting utility (formatdb for BLAST)
• -n number: desired number of database fragments
• -p path: directory to store database fragments.
Benchmarking
------------
All benchmark experiments were performed on a Linux cluster in the Department of Engineering at the University of Alabama at Birmingham \[[@B19]\]. The cluster consists of one compile node and 64 compute nodes (IBM × 335s), as well as 2 storage servers (IBM × 345s). All machines have 2 × 2.4 GHz Xeon processors, 2 GB of RAM, an 18 GB SCSI hard drive, and are connected via Gigabit Ethernet to a Cisco 4006 switch. The NCBI non-redundant protein database (nr, 733 MB), downloaded from GenBank \[[@B2]\] in August, 2003, was used for testing both DS-BLAST and QS-search; it contained 1,508,485 sequences composed of 492,678,715 amino acids. Release 10.0 of the Pfam \[[@B8]\] database from Washington University (549 MB) was used in benchmarking HMMPFAM under QS-search; it contained 6190 profile models that, when combined, were 1,463,477 residues long. The same set of query sequences was used for all experiments. The query sequences represented all open reading frames of more than 30 amino acids from the genome of monkeypox virus strain WRAIR 7--61 (manuscript in preparation), and totaled 2068 sequences comprising 151,173 amino acids.
Figure [2](#F2){ref-type="fig"} demonstrates that QS-search provided a \>40-fold acceleration for NCBI BLAST when using 64 processors, compared to the speed of 1 processor. Ideally, QS-search should provide an N-fold acceleration when using N processors, but this optimal result is rarely achievable. The major factor limiting the performance gain of QS-search is the time necessary to deliver the database to each node. As more nodes are employed, the portion of time spent searching decreases, but the communication overhead increases. When using QS-search with HMMPFAM, the search resulted in a 58-fold increase in processing speed when using 64 processors compared to a single processor (figure [3](#F3){ref-type="fig"}). QS-search therefore proved to be more efficient when running HMMPFAM in comparison to BLAST. Since the overall time required for the HMMPFAM search is much longer than that for BLAST, the portion of the search time devoted to communication overhead decreases thus increasing overall efficiency.
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**BLAST performance of QS-search.**Performance of QS-search with NCBI BLAST when searching all reading frames (\>30 amino acids) from monkeypox virus against the GenBank non-redundant protein database. Vertical bars represent total time used while the line indicates increase in speed corresponding to the number of processors used.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**HMMPFAM performance of QS-search.**Performance of QS-search for HMMPFAM when searching all reading frames (\>30 amino acids) from monkeypox virus genome against Release 10.0 of the Pfam database. Vertical bars represent total time used while the line indicates increase in speed corresponding to the number of processors used.
:::
:::
Figure [4](#F4){ref-type="fig"} illustrates the performance of DS-BLAST. Since database preprocessing occurs before the search process, the times provided in figure [4](#F4){ref-type="fig"} do not include this preprocessing time. As the same query, database and underlying application (NCBI-BLAST) were used in benchmarking DS-BLAST and QS-search, the results presented in figures [2](#F2){ref-type="fig"} and [4](#F4){ref-type="fig"} are directly comparable and indicate that QS-search appears to be more efficient than DS-BLAST. Two factors cause a reduction in performance of DS-BLAST. The first is the imbalance of the workload between nodes, and the second is the time necessary for the final merge phase of the output results from each node. When using QS-search, the workload is distributed dynamically during execution and therefore is well balanced between nodes. In contrast, when using DS-BLAST, the database is split into segments before the search, so the distribution of the workload between nodes is not as well balanced as for QS-search. The percent load imbalance for DS-BLAST (the time difference between completion of the first and last processors) has been as much as 5% of the total search time; for QS-search, on the other hand, the percent load imbalance is generally much less than 2%, and approached 3% only when 64 processors were used (data not shown). The merge phase of QS-search consists entirely of concatenating the results provided by each node into a single file. In contrast, the merge phase of DS-BLAST must parse the output from each node and combine the results for each single query sequence. As the number of nodes employed increases, the time required for the merge also increases. The advantage of the database splitting approach under limited memory conditions was not apparent in these benchmarks, since the memory available for each node in the cluster used in these experiments was large enough to accommodate the entire 733 MB database.
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Performance of DS-BLAST.**Performance of DS-BLAST when searching all reading frames (\>30 amino acids) from monkeypox virus genome against the GenBank non-redundant protein database. Vertical bars represent total time used while the line indicates increase in speed corresponding to the number of processors used.
:::
:::
We also compared the performance of DS-BLAST to that of mpiBLAST \[[@B11]\] version 1.1.0. We found that DS-BLAST was almost twice as fast as mpiBLAST when both utilized 4 processors and more than twice as fast when both utilized 32 processors (data not shown). Both mpiBLAST and DS-BLAST required a substantial part of the total run time to merge and format the final BLAST output.
Conclusions
===========
To increase the speed and efficiency of sequence similarity search programs, we have developed the SS-Wrapper package, a series of wrapper applications that supports the deployment of sequence similarity searches on high-performance computing clusters. QS-search implements a query sequence splitting approach for the deployment of NCBI BLAST and HMMPFAM. It also will support other similarity search programs, including all variants of NCBI BLAST (blastn, blastp, blastx, tblastn, and tblastx) as well as all options provided by the blastall executable. Because this implementation does not alter the original program, program updates and new programs should be easily accommodated. The output from QS-search is effectively identical to that produced by the underlying program. QS-search is designed to provide optimal load balancing and maximize resource usage when using computer clusters. The performance gain approaches linearity in proportion to the number of processors employed.
When the database is too large to fit into the physical memory of a single node in the cluster, a database splitting approach should outperform the query splitting approach used by QS-search \[[@B11],[@B14]\]. Therefore as a complementary application, the SS-Wrapper package also includes DS-BLAST, which implements a database splitting approach for BLAST searches and provides an effective solution to recalculate the E-value during the post-search phase of processing.
SS-Wrapper provides a suite of tools that makes large sequence similarity searches feasible by deploying the search on a Linux cluster. These tools permit the bioinformatics community to take advantage of the power of high-performance cluster computing. Other tools such as Disperse \[[@B20]\] and TurboBLAST \[[@B21]\] are designed to deploy bioinformatics applications onto loosely connected machines. A more general approach to deployment uses grid computing as an increasingly popular alternative to cluster computing \[[@B22]\]. Grid computing organizes widespread, diverse collections of CPU resources (including desktop workstations, servers, and clusters) into a virtual supercomputer, where these collections of hardware, software, and data resources are organized into a more uniform, manageable, visual whole. In contrast, the CPUs in a Linux cluster are more tightly coupled and specialized. Grid computing has the advantage of utilizing large numbers of CPUs as they become available to the grid. The disadvantage of a grid is in managing the complexity of the disparate architectures of the available CPUs, minimizing overhead, and making maximum usage of network bandwidth. For maximal performance, tools in the SS-Wrapper package have been developed under the assumption of a homogenous Linux cluster in which every CPU is similar. We are currently exploring methods to extend our current work to take advantage of grid computing technologies. To accomplish this, the complexities involved will require significant modification and extension of the applications that are a part of the SS-Wrapper package.
Availability and requirements
=============================
The SS-Wrapper package is freely available under the Artistic License described in the Open Source Initiative. The source code can be downloaded via ftp \[[@B17]\]. Contact <elliotl@uab.edu> for information on obtaining the software. All tools have been tested on an IBM Intel^®^processor-based Linux cluster with LAM/MPI \[[@B23]\] and should be compatible with other implementations of MPI.
Authors\' contributions
=======================
CW was responsible for the conception, design, implementation, and testing of the SS-Wrapper package. EJL contributed to its conception and testing and provided overall project coordination. Both authors have read and approved the final manuscript.
Acknowledgements
================
We would like to acknowledge the Enabling Technology Laboratory \[[@B19]\] in the Department of Engineering at the University of Alabama at Birmingham for giving us the opportunity to use their Linux cluster. We thank Jon Bernard for assistance with the benchmark experiments. We would also like to thank Dr. Mark Buller of St. Louis University for providing us with the genomic sequence of monkeypox virus strain WRAIR 7--61 that was used for the benchmarking experiments. We gratefully acknowledge Dr. Purushotham V. Bangalore, Ms. Catherine B. Galloway, and Dr. Shankar S. Changayil for helpful comments on the manuscript. This work was supported by an NIH/NIAID/DARPA grant (U01 AI48706) to EJL.
|
PubMed Central
|
2024-06-05T03:55:51.966644
|
2004-10-28
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545957/",
"journal": "BMC Bioinformatics. 2004 Oct 28; 5:171",
"authors": [
{
"first": "Chunlin",
"last": "Wang"
},
{
"first": "Elliot J",
"last": "Lefkowitz"
}
]
}
|
PMC545958
|
Background
==========
Patellar tendinopathy
---------------------
Patellar tendinopathy affects athletes in many sports and at all levels of participation, but is of particular concern for elite jumping athletes \[[@B1]\]. Many different types of sport activities have an increased risk for overuse of the patellar tendon including endurance sports (e.g. long-distance running and cross-country skiing) and sports with repetitive demands on strength and technique (e.g. tennis, baseball, volleyball, basketball and ballet) \[[@B2]\]. Athletes who participate in these sports may develop anterior knee pain that presents as tenderness at the inferior pole of the patella. This clinical syndrome is commonly called Jumper\'s knee, or patellar tendinopathy \[[@B3]\]. The term tendinopathy is considered to be the most appropriate clinical description for these chronic painful tendon conditions since there is no evidence of an inflammatory reaction in the chronically degenerated tendon \[[@B4],[@B5]\]. The changes in the tendon are mainly due to chronic collagen fiber degeneration \[[@B6]\], but the cause and source of the pain still remains unclear.
There are few studies on non-surgical treatment of patellar tendinopathy and there is a lack of evidence-based knowledge evaluating the therapy \[[@B7]\]. In-vivo studies in human or animals indicate possible benefits from treatments like heavy pressure \[[@B8]\], therapeutic ultrasound \[[@B9]\] and eccentric strength training \[[@B10]-[@B14]\].
Self-rated inventories for knee function
----------------------------------------
Patient-administrated questionnaires are frequently applied as primary outcome measures in clinical trials and several inventories have been translated from English into Swedish \[[@B15]-[@B17]\]. The WOMAC osteoarthritis index has been tested for reliability and validity in Sweden \[[@B18]\] and compared to quality of life instruments (SF-36 and NHP) \[[@B19]\]. The Knee injury and Osteoarthritis Outcome Score (KOOS) is also a self-administrated instrument measuring outcome after knee injury at impairment, disability, and handicap level with five subscales \[[@B16]\]. Garratt et al determined that KOOS showed good evidence of reliability, validity and responsiveness, and is recommended the score for knee diagnosis like ACL reconstruction, total knee replacements and for arthroplasty patients \[[@B20]\]. The only published clinical scale for patellar tendinopathy problems (VISA-P) was developed in Australia by the Victorian Institute of Sport Assessment in Melbourne \[[@B21]\]. The aim was to assess symptoms, simple tests of function and the ability of subjects to undertake sports. This self-administrated questionnaire has been documented as a reliable instrument for monitoring the progress of rehabilitation \[[@B3],[@B7]\]. It has also been shown to be a valuable tool in the assessment and documentation of recovery from patellar tendinopathy \[[@B22]\]. Even so, the responsiveness and validity of the questionnaire have not yet been fully proven.
The purpose of this study was to translate and cross-culturally adapt the VISA-P score for a Swedish population and to perform a psychometric analysis as well as reliability and initial validity testing of the Swedish VISA-P score.
Methods
=======
Subjects
--------
Fifty-one subjects gave informed consent to participate in this study. The VISA-P score was administered to 17 healthy students \[9 women, 8 men, mean age (± SD) 24 (± 6)\]; a population at risk, the Swedish male national basketball team \[17 men, mean age 26 (± 3)\], and patients with the diagnosis patellar tendinopathy \[17 men, mean age 22 (± 5)\].
The study was approved by the Ethical Committee at the Medical Faculty of the Karolinska Institute, Stockholm (Dnr 00-103).
The VISA-P score
----------------
The VISA-P score consists of eight questions \[[@B21]\], of which six questions concern pain experienced during a range of everyday activities. Two questions deal with the ability to engage in sport activities. All questions are answered on separate scales (0--10), where a higher score indicates a lower level of pain or impairment (Appendix A) \[see [additional file 1](#S1){ref-type="supplementary-material"}\]. The maximal total score is 100 points, which would indicate that the person has no knee pain, good function and can perform fully in sports. The theoretical minimum score is 0 points.
The original VISA-P score lacks information about the selection of items, weighting of each answer and the ranking of the options in the subscales in question 8.
The aim of the present investigation was to get a \"working tool\" for further studies of the usefulness of the instrument for patellar tendinopathy patients in Sweden.
Translation procedure
---------------------
The VISA Tendon Study Group at the University of Melbourne in Australia was informed and gave their consent to a Swedish translation of their original VISA-P score (Karim Khan, personal communication, 2003).
The translation process followed the method described by Beaton et al \[[@B23]\]. This method is currently used by a number of organizations, including the American Association of Orthopaedic Surgeons (AAOS) Outcomes Committee as they coordinate translations of the different components of their outcome batteries \[[@B23]\]. The translation process is divided into five different stages: (I) Translation, (II) Synthesis, (III) Reverse translation, (IV) Expert committee review and (V) Pre-testing.
Initially, two physiotherapists performed two independent translations (I) from English into Swedish. A synthesis (II) of these translations was made, and the consensus of the two translated Swedish versions was documented. Reverse translations (III) were performed independently by three native Anglophones fluent in Swedish. One of the reverse translators was a physiotherapist, one was an economist and the third was a teacher. The three physiotherapists in the expert committee (IV) then made a semantic and idiomatic equivalence analysis between the original source and target Swedish version of the VISA-P questionnaire. The translated questionnaire was pre-tested (V) on 12 individuals, six patients with patellar tendinopathy and six physical education students.
Test-retest reliability
-----------------------
The Swedish VISA-P score (Appendix B) \[see [additional file 2](#S2){ref-type="supplementary-material"}\] was administrated to all 51 participants at Bosön, the Swedish National Sports Confederation Centre (Lidingö, Sweden). The participants completed the questionnaire twice within an interval of one week (range 4--7 days).
The principal investigator administrated the questionnaires at all test occasions, with the exception of six of the tendinopathy patients.
Validity
--------
For validity, the factor structure of the VISA-P score was analyzed with a principal component analysis, Varimax rotation. The number of extracted factors was equal to the number of eigen values above 1.00. Internal consistency of subscales, based on the factor analysis, and the total scale was calculated as a Cronbach α coefficient \[[@B24]\].
For discriminative validity of the VISA-P questionnaires were compared between three groups, each of which were expected to have different levels of scoring.
Statistics
----------
All variables were summarized according to standard descriptive methods \[mean and standard deviation (SD)\] and checked for outliers. No significant deviations from the normal distribution criterion were found. The test-retest reliability was analyzed according to the method described by Bland and Altman, which yields an intra-class correlation (ICC) \[[@B25]\]. Differences between test occasions and groups were analyzed with an ANOVA (analysis of variance for repeated measurements, group \*time). In the post-hoc tests of group differences, Tukey\'s HSD method was applied. A significance level of five percent was applied (two-tailed).
Results
=======
Translation
-----------
The expert committee considered the translation and reverse translation satisfactory.
Test-retest reliability
-----------------------
The test-retest of the Swedish VISA-P score showed high reliability and significance (ICC = 0.97, p \< 0.001). In Figure [1](#F1){ref-type="fig"}, the Bland-Altman plot is showing the difference in total score between occasion one (A) and occasion two (B), plotted against the mean value of both test occasions. There were no significant differences for the total VISA-P score between the first and second test occasions. Each question (Q) was analyzed separately regarding the reliability. Seven out of eight questions has a reliability of more than ICC = 0.8 (range 0.68--0.97).
The score was easy to use and it took about five minutes to complete.
Internal consistency
--------------------
The internal consistency of the total scale was high for the scores both at the first and second occasion, 0.83 and 0.82, respectively.
Factor structure
----------------
The principal component analysis yielded a two-factor solution. The communality, i.e. the degree of explained variance, of one of the questions (Table [1](#T1){ref-type="table"}, \"sit pain-free?\") was below 0.35, and thus not sufficiently explained by this solution. Thus, a three-factor solution was preferred which explained 85% of the total variance, with all communalities above 0.60. The first component comprised of six questions. The second and third components comprised of one question each. This solution showed high stability, being invariant in a second factor analysis of the scores from the second occasion (the amount of explained variance was 83%).
Group differences in the VISA-P score
-------------------------------------
At the first test occasion (A) the mean (± SD) of the VISA-P score in the healthy student group was 83 (± 12), in the basketball players 79 (± 23), and 47 (± 20) in the patient group (Table [2](#T2){ref-type="table"}). In all questions, the patient group had lower scores as compared to the other two groups and statistical significance (p \< 0.05) was observed in all individual questions except the first (\"sit pain-free\"). In Table [1](#T1){ref-type="table"} the post-hoc tests for group differences are presented. The questions concerning pain (\"pain during 10 single leg hops\") had the greatest difference between the groups (F = 12.7, p \< 0.001). Both activity questions (\"currently undertaking sport\" and\" pain during activity\") showed significant (p \< 0.001) differences between the groups.
Discussion
==========
Translation
-----------
The expert committee of the translation process expressed a general agreement of all the questions except one (Q1). During the translation procedure of the VISA-P score, the translation for \"pain\" was debated. Different Swedish words were discussed and compared between the different translators. Translations into the mother tongue, or the first language, more accurately reflected the nuances of the language. Reverse translation into English of the Swedish VISA-P version was without remarks. Thus, the original and translated versions were judged by the expert committee to be congruent.
Test-retest reliability
-----------------------
Over a time interval of one week (range 4--7 days), the Swedish version of the VISA-P score showed high reliability (ICC = 0.97). As compared to other test-retest investigations of this score, this interval is the longest that has been studied \[[@B21]\].
Validation of the VISA-P
------------------------
A factor analysis yielded three factors, of which the first showed the highest correlations with two questions (\"pain during a full weight bearing lunge\" and \"problems squatting\", see Table [1](#T1){ref-type="table"}). The two other factors comprised only one question each, \"currently undertaking sport\" and \"sitting pain-free\", respectively. The separate factor for the question about \"sitting pain-free\" may be an artefact, as this item was the first one where misperceptions of the response dimension were more likely, thereby increasing the risk of higher error or unique variance. Some subjects in the pre-testing group, reported that they had perceived high scores as more pain. Conceptually, this question is equivalent to the questions of the first component. Experiences from the pre-testing resulted in a more detailed instruction for filling out the Swedish questionnaire.
Group difference
----------------
The patellar tendinopathy patients scored lower for all questions in the VISA-P score. The basketball players scored higher than the healthy students in two questions (\"sitting pain-free\" and \"currently undertaking sport\", see Table [2](#T2){ref-type="table"}). The first question was the only question that did not show any statistical significance between the groups and, noteworthy, the lowest score, i.e. highest degree of problem. The reason given above regarding the risk of misperception of the response dimension might be an explanation.
The VISA-P score has not yet been validated for pathological knee conditions other than patellar tendinopathy. Considering the separate questions (Appendix A) \[see [additional file 1](#S1){ref-type="supplementary-material"}\] it would be of interest to test the VISA-P score for patients with anterior knee pain other than patellar tendinopathy.
The significantly higher scores of the basketball players in question 7, \"currently undertaking sport\" (Table [2](#T2){ref-type="table"}) were trivial and obvious, since all of them were active players in Swedish the national team. The standard deviation was nearly twice as high for the patients and basketball players as compared to the healthy students. This reflects the heterogeneity of the first two groups.
Generally, there is a debate concerning scores about the relevance of using the total score or dividing the score in different subgroups. A short clinical scale is often an advantage. The factor analysis as well as the analysis of differences between the groups suggests that the VISA-P score could be abbreviated to two or three items without losing significant clinical information (Table [1](#T1){ref-type="table"}).
An important aspect of a clinical scale is its sensitivity for change or its ability to follow amelioration or exacerbation during treatment. The theoretical range of the VISA score, i.e. the floor and ceiling, is 0--100. The mean total score of the patients was approximately 50 (with a minimum value of 16) and for the control groups 80 (with a maximum value of 100. Thus, there seems to be sufficient scope to follow treatment effects, as well as to follow deteriorations of a risk group. It should be noted, however, that the present study was not designed to study treatment effects or development of a pathological process. The conclusion regarding the sensitivity of the VISA score, thus, awaits empirical support.
Although the mean VISA-P scoring was significantly different between asymptomatic subjects and patients with patellar tendinopathy, the score is not suggested to be a diagnostic test \[[@B21]\]. Therefore the score is considered to be suitable for group and intra-individual comparisons but should be avoided in inter-individual comparison. Another limitation of the score has not been shown to be applicable in a non-athletic population.
Adaptation of a questionnaire for use in a new setting is time consuming and costly. There are specific criteria that investigators should apply when evaluating patient-based outcome measures \[[@B26]\]. That being the case, larger international data collections and better correlations can be made when proper translations are performed and evaluations conducted. Additionally, there is a need for international accepted \'golden standards\' in outcome scores.
In conclusion, the results of the present study suggest that the translated Swedish version (Appendix B) \[see [additional file 2](#S2){ref-type="supplementary-material"}\] of the original Australian VISA-P score (Appendix A) \[see [additional file 1](#S1){ref-type="supplementary-material"}\] had satisfactory test-retest reliability when used to evaluate symptoms, tests of function and ability to undertake sport in patients with patellar tendinopathy.
Authors\' contributions
=======================
AF initiated the study, led the translation process and conducted all test occasions. TS and PR helped with general analysis and writing the article. GE guided and helped the main author with the statistical analyses of the data collected. All four of the authors read and approved the article.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2474/5/49/prepub>
Supplementary Material
======================
::: {.caption}
###### Additional File 1
Appendix A. The original VISA-P score.
:::
::: {.caption}
######
Click here for file
:::
::: {.caption}
###### Additional File 2
Appendix B. The translated and cross-culturally adapted Swedish VISA-P score.
:::
::: {.caption}
######
Click here for file
:::
Acknowledgements
================
The help and encouragement I got from my colleagues at the Sport Medicine Section, Institution of Surgical Sciences, Karolinska Institutet, Stockholm, and at the Elite Sports Centre, Swedish Sports Confederation, Bosön, Lidingö, Sweden is highly appreciated.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Bland-Altman plot of the Swedish VISA-P score in the reliability (test-retest within 1 week) investigation. Each blue dot indicates the difference in the total score, in relation to the mean score, between the two assessments.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Three-factor solution according to a principal component analysis, Varimax rotation. The questions are ordered after their factor loading.
:::
*Component*
-------------------------------------------------------------------------------------- ------------- ------- ------ ------
Q4. Do you have pain when doing a full weight bearing lunge? 0.92 0.18 0.06 0.88
Q5. Do you have problems squatting? 0.90 -0.09 0.25 0.88
Q6. Do you have pain during or immediately after doing 10 single leg hops? 0.88 0.13 0.26 0.86
Q3. Do you have pain at the knee with full active non-weight bearing knee extension? 0.77 0.26 0.12 0.68
Q8. For how long can you manage being train/physically active? 0.75 0.39 0.11 0.73
Q2. Do you have pain walking downstairs with a normal gait cycle? 0.70 0.38 0.24 0.69
Q1. For how many minutes can you sit pain free? 0.23 0.13 0.96 0.99
Q7. Are you currently undertaking sport or other physical activity? 0.17 0.95 0.11 0.94
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Mean score and standard deviation (SD) in three groups of subjects: patients with patellar tendinopathy (P); basket-ball players (B); and healthy students (H), and analysis of variance (ANOVA) of group differences of the Swedish VISA score.
:::
*Patients (n = 17)* *Basketball (n = 17)* *Healthy student (n = 17)* *Total (n = 51)* *ANOVA*
-------------------------------------------------------------------------------------- --- --------------------- ----------------------- ---------------------------- ------------------ --------- ------- ------- ------- ------- ------------ ---------------
Q1. For how many minutes can you sit pain free? A 5.41 2.94 7.53 2.88 7.24 3.52 6.73 3.21 2.29 n.s ns
B 6.06 2.80 7.82 2.65 7.18 3.63 7.02 3.08
Q2. Do you have pain walking down-stairs with a normal gait cycle? A 5.94 2.35 8.47 2.81 8.94 1.34 7.78 2.58 8.70 p \< 0.001 P \< B;P \< H
B 6.94 2.02 8.82 1.98 8.41 2.18 8.06 2.18
Q3. Do you have pain at the knee with full active non-weight bearing knee extension? A 6.59 3.00 8.47 2.53 9.53 1.06 8.20 2.61 6.84 p \< 0.05 P \< H
B 6.29 2.44 8.82 1.97 9.18 1.33 8.10 2.33
Q4. Do you have pain when doing a full weight bearing lunge? A 4.41 2.89 7.65 3.08 8.06 2.19 6.71 3.16 8.97 p \< 0.05 p \< B;p \< H
B 5.24 2.86 7.88 2.83 8.00 2.24 7.04 2.91
Q5. Do you have problems squatting? A 5.76 2.28 7.35 3.26 8.24 1.98 7.12 2.71 4.05 p \< 0.05 p \< H
B 6.65 2.47 7.76 2.88 8.18 2.27 7.53 2.58
Q6. Do you have pain during or immediately after doing 10 single leg hops? A 3.29 2.33 6.65 3.62 8.18 2.55 6.04 3.50 12.66 p \< 0.001 p \< B;p \< H
B 4.29 2.87 6.71 3.88 8.18 2.04 6.39 3.38
Q7. Are you currently undertaking sport or other physical activity? A 5.00 2.48 9.47 1.58 7.35 2.98 7.27 3.00 14.57 p \< 0.001 p \< H\<B
B 4.41 2.62 9.47 1.58 7.00 2.81 6.96 3.14
Q8. For how long can you manage being train/physically active? A 11.35 10.94 23.41 8.27 25.53 5.68 20.10 10.50 13.54 p \< 0.001 p \< B;p \< H
B 12.41 10.25 22.82 8.13 25.53 5.68 20.25 9.86
Total score A 47.76 20.26 79.00 24.18 83.06 12.60 69.94 24.96 16.48 p \< 0.001 p \< B;p \< H
B 52.29 18.98 80.12 22.82 81.65 15.48 71.35 23.33
:::
|
PubMed Central
|
2024-06-05T03:55:51.968854
|
2004-12-18
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545958/",
"journal": "BMC Musculoskelet Disord. 2004 Dec 18; 5:49",
"authors": [
{
"first": "Anna",
"last": "Frohm"
},
{
"first": "Tönu",
"last": "Saartok"
},
{
"first": "Gunnar",
"last": "Edman"
},
{
"first": "Per",
"last": "Renström"
}
]
}
|
PMC545959
|
Background
==========
Amongst the genitourinary cancer, carcinoma of the urinary bladder is one of the leading causes of death in Indian population. Transitional cell carcinoma (TCC) is the commonest histopathological variant where stage and grade are the two important prognostic factors to know the clinical behavior of these tumors. Superficial tumors with different grades behave differently e.g. tumors with high grade recur early and progress to invasive bladder cancer very soon. This behavior of same stage of the tumor but with varied grades is attributed to genetic alterations.
Bladder cancer manifesting from superficial to aggressive muscle invasive tumors undergoes a sequence of genetic alterations. Primary chromosomal aberrations are associated with tumor development while secondary chromosomal abnormalities lead to progression to a more advanced stage. A frequent loss of heterozygosity (LOH) on chromosomes 4, 5, 8, 9, 11 and 17 is considered a major event in the carcinogenesis of bladder cancer \[[@B1],[@B2]\].
Defects in mismatch repair (MMR) genes result in replication errors and genetic instability. Faulty mismatch repair, generally observed as somatic variation in size of microsatellites (short tandem repeat sequences in genome) is referred as microsatellite instability (MSI) \[[@B3]\]. MSI and LOH in bladder cancer have been reported by several investigators \[[@B4],[@B5]\].
A successful treatment of bladder cancer depends on early detection and more specific diagnostic approaches. Preneoplastic changes of the bladder epithelium or superficial tumors as an early event precede invasive bladder carcinomas. Though the higher grade and stage portends a worst prognosis, superficial tumors of same stage and grade have different outcome in different patients. Due to limited utility of these prognosticators in patients with superficial bladder tumor, there is a need to analyze new molecular parameters in predicting the prognosis and risk of recurrence. The following study is based on MSI analysis in tumor tissues to evaluate its utility as a marker for early detection of recurrent bladder carcinomas from lower urinary tract and thus help in deciding treatment modalities.
Methods
=======
Patient selection
-----------------
Total of 44 patients with male & female ratio of (42:2) of TCC with a mean age of 62 years were included for the study after the approval from ethical committee. All the patients selected for the study were not having any familial cancer syndrome or had previous history of cancer to the best of our knowledge.
All the tumors were resected transurethrally from the lower urinary tract. Part of superficial tissue specimen obtained after transurethral resection of bladder tumor (TURBT) was collected immediately in liquid nitrogen. Matched control sample (5 ml of peripheral blood) from all patients was collected in 200 μl of 0.5 M EDTA. The tumor stage and grade was assigned according to the TNM classification by American Joint Committee on Cancer (AJCC-UICC, 1997) \[[@B6]\]. Tumors of superficial nature classified as T1 or Ta while with deep muscular invasion were assigned as T2 or T3. Similarly tumor grading was done as G1 (low grade) and G2 or G3 (high grade). Patients were followed for recurrence (the number of times patient develops the tumor) every three months for 36 months with cytology and cystoscopy. The clinical and pathological characteristics of all the patients are summarized in Table [1](#T1){ref-type="table"}.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Clinical and pathological features of the patients diagnosed with bladder Carcinoma
:::
**Case no.** ^@^**Age/ ^\$^Gender** **Stage** **Grade** **\#Recurrence** **MSI**
-------------- ------------------------ ----------- ----------- ------------------ ---------------------------------------------------
BC 1 45/M T2 High, G3 2 \*High (BAT -- 26, BAT -- 40)
BC 2 59/M T1 High, G2 1 High (BAT -- 26, BAT -- 40, D9S1851)
BC 3 66/M Ta Low, G1 0 High (D9S283, D9S1851)
BC 4 39/M T2 High, G3 0 High (BAT -- 26, BAT -- 40, D9S283)
BC 5 72/M T2 High, G3 0 High (BAT -- 40, D9S1851)
BC 6 59/M T2 High, G2 1 Low (D9S1851)
BC 7 78/M T2 High, G3 0 High (D9S283, D9S1851)
BC 8 52/M Ta High, G2 0 High (BAT -- 40, D9S283, D9S1851)
BC 9 71/M Ta High, G2 0 \*\*Low (BAT -- 40)
BC 10 84/M T1 High, G2 0 \*\*\*MSS
BC 11 55/M T1 High, G1 0 MSS
BC 12 53/M T2 High, G2 0 Low (D18S58)
BC 13 52/M T3 High, G3 0 High (D9S283, D9S1851)
BC 14 40/M T3 High, G3 0 MSS
BC 15 55/M T2 High, G3 0 MSS
BC 16 60/M T2 High, G3 0 High (BAT -- 40, D9S283, D18S58)
BC 17 66/M T1 High, G3 1 High (BAT -- 40, D2S123, D9S283, D9S1851, D18S58)
BC 18 80/M T1 High, G3 1 High (BAT -- 26, D9S283, D18S58)
BC 19 42/M T1 High, G2 0 Low (D9S283)
BC 20 73/M T3a High, G3 0 MSS
BC 21 55/M T2 High, G3 0 Low (D2S123)
BC 22 58/M T2 High, G3 0 High (BAT -- 26, D2S123)
BC 23 70/M T1 High, G2 0 Low (BAT -- 26)
BC 24 53/M Ta High, G3 0 High (D9S1851, D18S58)
BC 25 60/M T2 High, G3 0 High (D9S283, D9S1851)
BC 26 54/M Ta Low, G1 0 MSS
BC 27 72/M Ta High, G2 0 MSS
BC 28 58/M T1 High, G3 0 Low (D9S283)
BC 29 80/M T2 High, G3 0 High (D9S283, D9S1851)
BC 30 64/M Ta High, G2 2 Low (BAT -- 40)
BC 31 74/M Ta High, G3 0 High (D9S283, D9S1851)
BC 32 60/M T1 Low, G1 0 MSS
BC 33 41/M T1 Low, G1 3 Low (BAT -- 26)
BC 34 66/ F T2 High, G3 0 High (D2S123, D9S283)
BC 35 53/M T1 Low, G1 0 Low (D2S123)
BC 36 66/M T1 High, G2 1 High (D2S123, D9S283, D18S58)
BC 37 55/M T2 High, G3 0 High (BAT -- 26, BAT -- 40, D9S1851)
BC 38 65/M T2 High, G3 0 Low (D18S58)
BC 39 69/M T1 Low, G1 0 Low (D9S283)
BC 40 74/M T1 High, G2 1 MSS
BC 41 72/ F T2 Low, G1 0 MSS
BC 42 71/M T1 High, G3 1 Low (D9S283)
BC 43 64/M T2 High, G3 0 High (BAT -- 40, D9S1851)
BC 44 73/M T1 High, G2 1 Low (D9S283)
^@^Age = (years); ^\$^Gender = (M: Male/F: Female); \*MSI -- H = MSI -- High; \*\*MSI -- L = MSI -- Low; \*\*\*MSS = microsatellite stable
\# Recurrence (0, 1, 2, 3) = Number of times the tumor recurred
:::
DNA isolation
-------------
Superficial tumor tissue specimens of histologically confirmed bladder tumors and peripheral blood (frozen) of the same patient were processed for DNA isolation using phenol-chloroform extraction method \[[@B7]\].
MSI analysis
------------
Table [2](#T2){ref-type="table"} demonstrates the characteristic features of the microsatellite markers evaluated in the present study. The primer sequences for mono and dinucleotide microsatellite markers were searched from human genomic database. Polymerase chain reaction (PCR) amplification of DNA was done using primers of concentration of 6 pmol, 200 μM dNTPs, 10 mM Tris -- Cl (pH 8.3), 50 mM KCl, 1.5 mM MgCl~2~, 0.25 units of Taq polymerase (MSI, Fermentas), 100 ng DNA and 2 μCi \[α-^32^P\] dCTP (specific activity: 4000 Ci/mM) (BRIT, India) in a volume of 25 μl. PCR conditions involved an initial denaturation at 95°C for 3 min followed by 30 cycles (95°C for 1 min, 50°C to 60°C for 2 min and 72°C for 3 min) and a final extension at 72°C for 8 min. PCR products were mixed with equal volume of formamide loading dye (95% formamide, 20 mM EDTA, 0.05% bromophenol blue, 0.05% Xylene cyanol), denatured for 5 min at 95°C and loaded onto 8% polyacrylamide gel containing 7 M urea. Gels were run at 55 W for 2 hours, transferred onto a Whatman sheet followed by an exposure to X-ray film (Kodak) for desired time and then developed.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Characteristic features of microsatellite markers examined in urinary bladder tumors
:::
**Microsatellite marker** **Repeat pattern** **Chromosomal location** **\~ PCR product size**
--------------------------- -------------------- ------------------------------------------------ -------------------------
BAT 26 (A)~26~ 5^th^intron of hMSH2, 2p 117 -- 130 bp
BAT 40 (A)~40~ 2^nd^intron of β hydroxy steroid dehydrogenase 94 -- 112 bp
BAX (38 -- 41) (G)~8~ 19q13.3 -- q13.4 94 bp
TGFβ RII (665 -- 737) (A)~10~ 3p22 73 bp
IGFIIR (4030 -- 4140) (G)~8~ 6q26 -- 27 110 bp
HMSH3 (381 -- 383) (A)~8~ 5q 150 bp
D2S123 (CA)~13~TA 2p16 197 -- 227 bp
(CA)~15~(T/GA)~7~
D9S283 (CA)n 9q13 -- q22 178 -- 203 bp
D9S1851 (CA)n 9q22.3 143 -- 159 bp
D18S58 (GC)~5~GA(CA)~17~ 18q22.3 144 -- 160 bp
:::
The tumor was designated unstable if its PCR product had altered band pattern when compared to alleles in corresponding matched blood DNA \[[@B8]\]. Out of he motifs studied, BAT-26 & BAT-40, the mononucleotide poly A repetitive loci have been shown to exhibit polymorphism \[[@B9]\]. The change either borderline or major deletions/ insertions at this loci is compared in tumor tissue of the same patient with the normal tissue in colorectal tumors \[[@B8]\]. Tumors were called MSI-High (MSI-H) when they showed instability at \> 30% of loci and MSI-Low (MSI-L) if they showed at or less than 30% of loci.
Statistical analysis
--------------------
Statistical tests, including, 2 × 2 contingency table, Fisher\'s exact probability test (one or two tailed), Karl Pearson\'s correlation test were applied to assess the relation between the microsatellite instability in tumors and clinicopathological parameters. A student t test was applied to compare the number of genomic alterations between tumors of different grades and stages.
Results
=======
A panel of ten microsatellite markers situated on chromosomes 2, 3, 5, 6, 9, 18 and 19 were screened to look for microsatellite instability in superficial tumor tissues and compared with blood DNA. Alterations were detected in 32 of 44 patients (72.7%).
Out of the six mononucleotide microsatellite markers analyzed, only BAT -- 26 and BAT -- 40 in 17.7% and 24.4% of the cases could demonstrate changes respectively while TGFβ RII, IGFIIR, hMSH3 and BAX were microsatellite stable (MSS). BAT-26 & BAT-40 exhibited borderline changes in tumor DNA as compared to control DNA. We also sequenced the tumor & normal PCR product for these microsatellites to prove the change. The dinucleotide markers- D2S123, D9S283, D9S1851 and D18S58 exhibited altered electrophoretic migration pattern in 15.5%, 40%, 31.1% and 17.8% bladder tumors respectively. The most frequent microsatellite alteration was detected on the markers of chromosome 9 (D9S283 followed by D9S1851) (Fig [1](#F1){ref-type="fig"} and [2](#F2){ref-type="fig"}).
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Changes in allelic pattern indicated by an arrow, observed in superficial tissue (Ts) as compared to blood (germline DNA, N) of patients with bladder carcinoma: (A) Deletion at BAT -- 26; (B) Insertion at BAT -- 40; (C) Insertion at D2S123; (D) Loss of heterozygosity at D9S283; (E) Biallelic alteration at D9S1851 and (F) Deletion at D18S58
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Superficial tumor tissues and blood (control) of bladder tumor patients demonstrated no change at (A) TGFβ RII; (B) BAX; (C) hMSH3 and (D) IGFIIR
:::
:::
A significant association of MSI -- H with T2 or T3 stage tumors (2 × 2 contingency table, p = 0.05) and ≥ G2 grade tumors (Fisher\'s exact probability test, two tailed, p = 0.08) was observed (Table [1](#T1){ref-type="table"}).
Total genomic alterations were analyzed among different stages and grades. Changes were comparable among (24/44) T1-Ta and (20/44) T2-T3 tumors that included 35 (13 insertions, 6 deletions, 10 LOH and 6 BA) and 31 (10 insertions, 8 deletions & LOH and 5 BA) respectively (p \> 0.05). High grade (G2-G3) (37/44) carcinomas encompassed 63 alterations (24 insertions, 14 deletions, 15 LOH and 10 BA) and low grade (G1) (7/44) bladder tumors had only 5 (1 each deletion & BA and 3 LOH) (p = 0.01)
Recurrence of tumors (mean duration of 36 months) was correlated with MSI in 24 superficial (Ta-T1) tumors in patients who turned up for regular follow up (Table [1](#T1){ref-type="table"}). This group comprised 18 high grade and six low grade tumors. In the high grade tumors, 13 were MSI+, and 8 of them (61.5%). showed recurrence while only one (1/5, 20%) MSI- recurred. Amongst low grade tumors (6), 2 recurrences were noted only in MSI+ group and none of the MSI- showed recurrence (Fisher\'s exact probability two tailed test) (p = 0.02 for high grade tumors) and (p = 0.04 for low grade tumors).
Discussion
==========
This present work is a continuation of the previous published work where thirty bladder tumors were analyzed for the presence of MSI at BAT-26, BAT-40, TGFβ RII, IGFIIR, hMSH3 and BAX. The initial results encouraged examining the role of MSI/ LOH in more number of tumors with expanded panel of markers \[[@B7]\]. In this paper, 44 patients of bladder cancer are examined for MSI at BAT-26, BAT-40, TGFbRII, IGFIIR, BAX, & hMSH3 & D2S123, D9S283, D9S1851 & D18S58 dinucleotide repeat motifs. The MSI results are further analysed with clinicopathological features: stage and grade of the tumors & its recurrence in due course of time. The importance of MSI in the diagnosis of recurrence in superficial cases irrespective of grade & thus advocate the bladder cystectomy as a treatment modality in these cases.
Genomic instability measured by changes in tumor tissues as compared to blood of the same patient at repetitive loci was detected in 72.7% cases of bladder carcinoma. It differs from previous observation, which shows infrequent occurrence of MSI in TCC using different microsatellite markers \[[@B10]\]. However, low frequency of MSI with alterations of dinucleotide repeats in TCC of the urinary tract was found as 21% and 16.6% in two independent studies \[[@B11],[@B12]\]. MSI and allelic loss in a series of 26 upper urinary tract tumors using 5 informative microsatellite markers were examined & this study supports the presence of MSI in upper urinary tract which is rare event in bladder cancer \[[@B13]\]. Another study describes MSI and loss of respective MMR protein by immunostaining in a patient with a urothelial carcinoma of the ureter and a strongly positive history of cancer, who was subsequently found to have HNPCC \[[@B14]\].
In the present study a significant association of MSI with tumor stage and grade in sporadic bladder tumors suggested MSI as an early event in tumorigenesis. These results confirm the previous finding where MSI examined in TCC of bladder with low stage and grade using few microsatellite markers mostly confined to chromosome 9 \[[@B15]\]. Another study reports 100% tumor instability as determined by dinucleotide repeat analysis in 14 cases of urinary bladder of different stages and grades \[[@B16]\]. Many studies show relatively high proportion of tumors with mutations in di, tri, and tetra nucleotide repeat motifs, although each tumor exhibits only few such mutations \[[@B4]\]. Recently, a novel form of MSI, termed as EMAST (elevated microsatellite instability at selected tetranucleotide repeats) has been found to be significantly associated with mutations in p53 among the bladder cancer tumors, but no indication of elevated EMAST in tumors with abnormal p53 staining without mutation. EMAST likely reflects a particular pattern of somatic events that are interactive with p53 mutation, particularly common in skin cancer and limited to non-invasive disease in bladder cancer \[[@B17]\]. The difference between these studies and ours may be attributed to the number and identity of microsatellite motifs studied.
Despite clear-cut prognostic differences, genetic alterations were comparable in superficial (Ta-T1) and invasive bladder carcinomas (T2-T3) suggesting the role of MSI in progression of bladder cancer as well. However, strong association of MSI -- H with T2-T3 and G2-G3 was observed. MSI at ≥ 30% of loci has been found in 59.4% (19/32) of TCC bladder, which is not in accordance with reported earlier \[[@B11],[@B12]\]. A good association of MSI -- H with high grade superficial tumors may help in deciding radical surgery to begin with.
Bladder cancer presents as superficial tumor in 75% of the patients, which can easily be removed by transurethral resection (TUR). Around 60--80 % of these treated patients develop recurrence in due course of time. Out of them, 15% progress to higher grade and stage. With so much potential for recurrence, patients need to be followed up with cystoscopy at regular intervals. Although many new tumor markers have been proposed but all have limitations with respect to execution and interpretation in predicting the recurrence of bladder tumors \[[@B18]\]. Among the molecular markers, alterations in p53, p21^WAF/C1P1^, Rb, c-erb B-2 are reported to be associated with tumor recurrence and progression but little is known to address MSI \[[@B19]\]. MSI analysis gives higher sensitivity and easy to execute among other molecular markers, thus making it a valuable marker for detection of recurrence. To the best of our knowledge, this is the first study reporting MSI as a good prognostic marker that correlates with risk of recurrence in superficial (Ta-T1) tumors irrespective of the grade. This may help in deciding radical treatment at an early stage. We could not study the genetic changes during the progression of tumor, which means the extension of superficial tumor confined to the mucosa and submucosa to deep musculature of the bladder. Limitation of this study is a small number of patients but initial trends show a strong correlation of MSI with recurrence irrespective of the grade of the tumor. Further multicentre trial is needed to prove this concept.
Conclusions
===========
MSI has been observed to play important role in evolution, initiation and progression in bladder tumors. Patients with high grade superficial disease are reported to have higher incidence of MSI. Also high frequency of MSI in superficial tumors showing recurrence irrespective of grade may provide an indication for more radical approach to improve the survival.
Competing interest
==================
The author(s) declare that they have no competing interests.
Authors\' contribution
======================
MV carried out the molecular genetic studies, participated in analyzing the data & drafted the manuscript. AM provided the clinical material & information, helped in analyzing the data & designing the manuscript. RDM helped in manuscript drafting. BM participated in its designing of the study & manuscript. All authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2490/5/2/prepub>
Acknowledgement
===============
The author (MV) is thankful to Council of Scientific and Industrial Research (CSIR), Govt. of India, for providing financial assistance.
|
PubMed Central
|
2024-06-05T03:55:51.971579
|
2005-1-12
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545959/",
"journal": "BMC Urol. 2005 Jan 12; 5:2",
"authors": [
{
"first": "Minal",
"last": "Vaish"
},
{
"first": "Anil",
"last": "Mandhani"
},
{
"first": "RD",
"last": "Mittal"
},
{
"first": "Balraj",
"last": "Mittal"
}
]
}
|
PMC545960
|
Background
==========
Recent studies on network science demonstrate that cellular networks are described by universal features, which are also present in non-biological complex systems, as for example social networks or *WWW*. Most networks encountered in real world have scale-free topology, in particular networks of fundamental elements of cells as proteins and chemical substrates \[[@B1]-[@B4]\]. In these networks, the distribution of node degree follows a power-law as *P*(*k*) \~ *k*^-*γ*^(i.e., frequency of the nodes that are connected to *k*other nodes). The degree of a node is the number of other nodes to which it is connected.
One of the most successful models for explaining that scale-free topology was proposed by *Barabási-Albert*\[[@B5]\], which introduced a mean-field method to simulate the growth dynamics of individual nodes in a *continuum theory*framework. However, although that model was a milestone to understand the behavior of real complex networks, it could not reproduce all the observed features in real networks such as clustering dependence. The observed properties of networks with *N*nodes are: scale-free of degree distribution *P*(*k*) \~ *k*^-*γ*^, power-law scaling of clustering coefficient *C*(*k*) \~ *k*^-1^and a high value for the average of the clustering coefficient \<*C*\> and its independence with network size. In particular, the dependence of *C*(*k*) \~ *k*^-1^was one of the results obtained by \[[@B6]\]. In order to bring under a single framework all these observed properties in real networks *Ravasz et al.*(the RSMOB model in what follows) suggested successfully a hierarchical and modular topology \[[@B7],[@B8]\]. In \[[@B8]\], a network with the above mentioned properties was called hierarchical network. We note that this deterministic model is an extension of the original model shown in \[[@B9]\]. It is also worth noticing that this modular topology was also suggested in biological networks by \[[@B10],[@B11]\]. Interestingly, these properties of networks have been found in many non- biological and biological networks. One of them, which is the subject of our study, is the metabolic network.
It is interesting to note that the metabolic network is an example of bipartite networks \[[@B12]\]. In a bipartite network there are two kinds of nodes and edges only connect nodes of different kinds. In the metabolic network these nodes are chemical compounds and reactions. The network generated by the chemical compounds (reactions) is called compound (reaction) projection. A line graph transformation (i.e., each edge between two nodes becomes a node of the transformed network) may relate both projections.
However, although the line graph transformation works fine on bipartite networks, the transformed network (in the particular case of metabolic networks) may not be totally the same as the reaction projection. This issue is discussed in detail later. In addition, we will show by comparing with the experimental data, that this fact does not affect our qualitative results. Furthermore, a detailed analysis of the line graph transformation focused on the degree distribution *P*(*k*) and applied to some real networks can be found in \[[@B13]\]. In that work, similarities and differences between the line graph transformation and the metabolic network are also discussed. There it was found that if the initial network follows a power-law *P*(*k*) \~ *k*^-*γ*^, the transformed network preserves the scale-free topology and in most cases the exponent is increased by one unit as *P*(*k*) \~ *k*^-*γ*+1^.
It is also worth noting that the line graph transformation has recently been applied with success by *Pereira-Leal et al.*\[[@B15]\] on the protein interaction network with the aim to detect functional modules. In that work, the edges (interactions) between two proteins become the nodes of the transformed network (interaction network). By means of the line graph transformation, the interaction network has a higher clustering coefficient than the protein network. By using the TribeMCL algorithm \[[@B16]\] they are able to detect clusters in the more highly clustered interaction network. These clusters are transformed back to the initial protein-protein network to identify which proteins can form functional clusters. At this point, we note that the aim of our study is not to detect functional modules from the metabolic network. In our work the line graph transformation is used successfully to evoke general topological properties related to the clustering degree of the reaction network.
The observed topological properties related to the clustering degree of the metabolic network (in particular, the chemical compound network) have been properly described by means of the RSMOB model. In the present work, our aim is to study the clustering coefficients *C*(*k*) and \<*C*\> of the reaction network by using two approaches: Firstly, we derive mathematical equations of those coefficients in the transformed network. Secondly, we apply the line graph transformation to a hierarchical network. The results from both methods are compared with experimental data of reactions from KEGG database \[[@B14]\] showing a good agreement. Though we started this work motivated by theoretical interest in the line graph transformation, the results provide explanation for the difference of *C*(*k*) between the compound network and the reaction network.
In our work, the hierarchical network is generated by the RSMOB model, where the nodes correspond to chemical compounds and the edges correspond to reactions. While the RSMOB model reproduces successfully the hierarchical properties of the compound network, here we show that this hierarchical model also stores adequate information to reproduce the experimental data of the reaction network. Our study indicates that it is enough to apply the line graph transformation to the hierarchical network to extract that information. While *C*(*k*) follows the power-law *k*^-1.1^for the initial hierarchical network (compound network), *C*(*k*) scales weakly as *k*^0.08^for the transformed network (reaction network). Consequently, we conclude that the reaction network may not be classified as a hierarchical network, as it is defined in \[[@B8]\].
Remark
------
In \[[@B8]\], a network with scale-free topology, scaling law of *C*(*k*) \~ *k*^-1^, and high degree of clustering was called *hierarchical network*. Consequently, the RSMOB model shown in \[[@B7],[@B8]\] was developed to bring these properties under a single roof. Furthermore, in \[[@B7],[@B8],[@B17]\] some networks (in particular metabolic network) were classified as hierarchical network according to the above definition. To be precise, it was argued that the signature of the intrinsic hierarchy (or hierarhical modularity) is the scaling law of *C*(*k*). Moreover, in a more recent work \[[@B18]\], it was claimed that traditional random and scale-free models do not have a hierarchical topology because *C*(*k*) is independent of *k*(i.e., flat plot of *C*(*k*)). In addition, analyses of *C*(*k*) were recently carried out in \[[@B19]\] to uncover the structural organization and hierarchy of non-biological weighted networks. At this point, we must note that we have followed the research done by *Barabási et al*, \[[@B7],[@B8],[@B18]\] and consequently, we have used its definition of hierarchical network in the present work.
However, it is also worth noticing that another way to quantify the hierarchical topology of a network is recently introduced by \[[@B20],[@B21]\]. It is based on the concept of a *hierarchical path*: a path between nodes *i*and *j*is called hierarchical if (1) the node degrees grow monotonously (\"*up path*\"), and it is followed by a path where the node degrees decrease monotonously (\"*down path*\") or (2) the node degrees along this path changes monotonously from one node to the other. The fraction of *shortest paths*in a network, which are also *hierarchical paths*is called *H*. If *H*is very close to 1, the network shows a hierarchical organization. This definition seems interesting, and consequently, as a future work it would be worth to examine some biological networks (in particular, metabolic networks) by using this approach. One remark about this concept is that it focuses on hierarchy and may not contain enough information about modularity or clustering. For a brief discussion of these issues, we refer to some very useful notes written by *Dorogovtsev et al.*\[[@B21]\] (see also \[[@B22]\] for further information about related topics in networks).
Results and discussion
======================
Clustering coefficients *C*(*k*) and \<*C*\>
--------------------------------------------
Recent analyses have demonstrated that the metabolic network has a hierarchical organization, with properties as: scale-free degree distribution *P*(*k*) \~ *k*^-*γ*^, power-law dependence of clustering coefficient *C*(*k*) \~ *k*^-1^and independence with network size of the average clustering coefficient \<*C*\>, where *N*is the total number of nodes in a network \[[@B7]\]. The clustering coefficient can be defined for each node *i*as:
where *n*~*i*~denotes the number of edges connecting the *k*~*i*~nearest neighbors of node *i*to each other, *C*~*i*~is equal to 1 for a node at the center of a fully interlinked cluster, and it is 0 for a node that is a part of a loosely connected cluster \[[@B7]\]. An example can be seen in Fig. [1A](#F1){ref-type="fig"}.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
\(A) Example of clustering in an undirected network. Continuous and dash-dotted lines mean interaction between nodes. In addition, the dash-dotted line defines the only triangle where the node 1 (red) is one of the vertices. The node 1 has 4 neighbors (*k*~*i*~= 4), and among these neighbors only one pair is connected (*n*~1~= 1). The total number of possible triangles that could go through node *i*is 6. Thus, the clustering coefficient has the value *C*~1~= 1/6. High density of triangles means high clustering coefficient. (B) We show an example of the line graph transformation. The initial graph *G*corresponds to one subgraph which belongs to the Lysine Biosynthesis metabolic pathway. This graph is constructed by taking nodes as chemical compounds and edges as reactions. By applying the line graph transformation we find graph *L*(*G*), which is the reaction graph embedded in the graph *G*. The nodes of the graph L(G) are the reactions of the graph *G*\[13\].
:::
:::
Geometrically, *n*~*i*~gives the number of triangles that go through node *i*. The factor *k*~*i*~(*k*~*i*~- 1)/2 gives the total number of triangles that could go through node *i*(i.e., total number of triangles obtained when all the neighbors of node *i*are connected to each other). In the case of Fig. [1A](#F1){ref-type="fig"}, there is one triangle that contains node 1 (dash-dotted lines), and a total of 6 triangles could be generated as the maximum. Hence, the clustering coefficient of node 1 is *C*~1~= 1/6.
On the other hand, the average clustering coefficient \<*C*\> characterizes the overall tendency of nodes to form clusters as a function of the total size of the network *N*. The mathematical expression is:
The structure of the network is given by the function *C*(*k*), which is defined as the average clustering coefficient over nodes with the same node degree *k*. This function is written as:
where *N*~*k*~is the number of nodes with degree *k*, and the sum runs over the *N*~*k*~nodes with degree *k*. A scaling law *k*^-1^for this magnitude is an indication of the hierarchical topology of a network.
Once the theoretical definitions have been introduced, our aim is to analyse how the coefficients \<*C*\> and *C*(*k*) are modified under the line graph transformation.
Line graph transformation to metabolic networks: spurious nodes
---------------------------------------------------------------
Given an undirected graph *G*, defined by a set of nodes *V*(*G*) and a set of edges *E*(*G*), we associate another graph *L*(*G*), called the line graph of *G*, in which *V*(*L*(*G*)) = *E*(*G*), and two nodes are adjacent if and only if they have a common endpoint in *G*(i.e., *E*(*L*(*G*)) = {{(*u*, *v*), (*v*, *w*)}\|(*u*, *v*) ∈ *E*(*G*), (*v*, *w*) ∈ *E*(*G*)}). This construction of graph *L*(*G*) from the initial graph *G*is called line graph transformation \[[@B23]\].
It is worth noting that in a previous work \[[@B13]\] the degree distribution *P*(*k*) was studied by applying line graph transformation to synthetic and real networks. There it is assumed an initial graph *G*with scale-free topology as *P*(*k*) ≈ *k*^-*γ*^. As the degree of each transformed node (i.e., an edge in *G*) will be roughly around *k*, the distribution of the line graph *L*(*G*) should be *k*·*k*^-*γ*^= *k*^-*γ*+1^with degree around *k*. Therefore, it is concluded that if we have a graph *G*with a probability distribution following a power-law as *k*^-*γ*^, then *L*(*G*) will follow a power-law as *k*^-*γ*+1^. The real networks under study were protein-protein interaction, *WWW*, and metabolic networks. In Fig. [1B](#F1){ref-type="fig"}, we can see an example of the line graph transformation applied to a subgraph of the metabolic network.
However, it is important to point out one issue. In metabolic networks, there are cases where spurious nodes appear (see Fig. [2](#F2){ref-type="fig"}). For example, we consider two reactions sharing the same substrate (or product) and at least one of the chemical reaction has more than one product (or substrate). If we apply a line graph transformation to this network, we would obtain more than two nodes in the transformed network, where only two nodes (reflecting two reactions present in the real process) should appear. These spurious nodes appear only when one (or some) reaction(s) in the network has more than one product (or substrate). Therefore, these cases should be computed and transformed by generating only as many nodes in the transformed network as reactions in the real metabolic process. This procedure is called *physical*line graph transformation. In the present work, we have applied this procedure to generate the reaction network by using experimental data from the KEGG database. Experimental data are shown later in Figs. [7B](#F7){ref-type="fig"} and [8](#F8){ref-type="fig"} (blue diamonds). More detailed information about this issue can be found in \[[@B13]\].
::: {#F2 .fig}
Figure 2
::: {.caption}
######
We show two reactions (*R1*, *R2*) sharing a common chemical compound, and both reactions contain more than one product (or substrate). The substrate graph *G*(chemical compounds) is shown dark blue circles. The reaction graph *L*~*real*~(reactions) is shown with light red circles. If we apply a line graph transformation to this network, we would obtain more than two nodes in the transformed network. However, only two nodes (reactions) are present in the real process. These cases are computed and transformed by generating only as many nodes in the transformed network as reactions in the real metabolic process. We call this procedure *physical*line graph transformation.
:::
:::
Equations of *C*(*k*) and \<*C*\> under the line graph transformation
---------------------------------------------------------------------
We assume a graph *G*as it is depicted in Fig. [3A](#F3){ref-type="fig"}. In this graph, edge *a*connects two nodes with degree *k*\' and *k*\". We apply the line graph transformation to this graph *G*and the result of this transformation is the line graph of *G*, *L*(*G*) shown in Fig. [3B](#F3){ref-type="fig"}. We see that, under the line graph transformation, the nodes of *L*(*G*) are the edges of *G*, with two nodes of *L*(*G*) adjacent whenever the corresponding edges of *G*are.
::: {#F3 .fig}
Figure 3
::: {.caption}
######
\(A) Graph *G*with two hubs with degree *k*\' and *k*\" connected by edge *a*. (B) The corresponding line graph *L*(*G*) after the line graph transformation is done. (C) Graph *G*where edges *b*and *b*\' have a common node as endpoint. (D) Line graph of (C). It is worth noticing that (D) has only one more edge than (B). Hence, (D) has one more triangle that go through node *a*than (B).
:::
:::
The clustering coefficient for the node *a*in the transformed network can be written by using Eq. (1) as:
where *k*= *k*\' + *k*\" - 2, because the edge *a*vanishes in the graph *L*(*G*). This equation ignores cases where edges in the graph *G*, *b*and *b*\' for example, have a common node as endpoint (i.e., existence of triangles or *loops*in Fig. [3C](#F3){ref-type="fig"}). However, we can quantify these cases by using a new parameter *l*. As we can see in Fig. [3C--D](#F3){ref-type="fig"}, edges with one common node as endpoint in the graph *G*means one additional edge in the graph *L*(*G*). This additional edge in *L*(*G*) connects two neighbors of node *a*. By following definition of Eq. (1), it means that *n*~*a*~increases its value by one unit. We can consider these cases by increasing one unit the parameter *l*for each common node as endpoint of two edges in the graph *G*(for example, *l*= 1 means one common node). We write Eq. (4) after introducing the parameter *l*as:
where if *l*= 0 means that there are not *loops*and we recover Eq. (4). Though it is more realistic to consider the parameter *l*as a function of *k*\' and *k*\", we have considered *l*as an independent parameter. However, this simplification does not affect the qualitative features of our results. It should be noted that *l*always contributes to increasing the value of *C*~*a*~(*k*) and *C*~*a*~(*k*) ≤ 1 always holds from the definition. In order to study the limits of Eq. (5) we consider the following two cases:
• *a*) *k*\' = *k*\": We analyse the case where both degrees have the same value. We also consider the cases when *l*= 0 and *l*≠ 0 in order to study the effect of triangles. We show the results in Fig. [4](#F4){ref-type="fig"}. For large *k*\', Eq. (5) goes asymptotically to 1/2 for *l*= 0 and *l*≠ 0. We also see that for *k*\' ≥ 25, all lines are very close to 1/2. For low *k*\' and *l*= 0, *C*~*a*~(*k*) takes values from 0.33 (*k*\' = 3) to 0.48 (*k*\' = 20). Hence, we see in Fig. [4](#F4){ref-type="fig"} that higher values of *l*(more triangles) increase the values of *C*~*a*~(*k*).
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Values of *C*~*a*~(*k*) from Eq. (5) calculated by taking *k*\' = *k*\". Number of common nodes as endpoint of two edges (triangles) are indicated by the parameter *l*. The degree of transformed nodes is *k*= *k*\' + *k*\" - 2 because the edge *a*vanishes in the graph *L*(*G*).
:::
:::
• *b*) *k*\" = constant, *k*\' \>\>*k*\": We plot in Fig. [5](#F5){ref-type="fig"} three cases. *k*\" is fixed with constant values as *k*\" = 5 (black), *k*\" = 10 (red), *k*\" = 20 (blue) and *k*\' is a free parameter. We see that *C*~*a*~(*k*) approaches to 1 when *k*\' takes large values. For low *k*\', the case *k*\" = 5 shows a minimum with a few values of *k*\' below 1/2. As we can see with dotted and dash-dotted lines in Fig. [5](#F5){ref-type="fig"}, the presence of triangles (*l*≠ 0) increases the value of *C*~*a*~(*k*). Finally, for *k*\" = 10 and *k*\" = 20, we see that only a few values of *C*~*a*~(*k*) are slightly below 1/2 for low *k*\'. This analysis is complemented by calculating the minimum value of *C*~*a*~(*k*) analytically as: . The value of *k*\', where the function *C*~*a*~(*k*) takes the minimum value, is given by:
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Values of *C*~*a*~(*k*) from Eq. (5) calculated by taking *k*\" with constant values as *k*\" = 5 (black line), *k*\" = 10 (red), *k*\" = 20 (blue) and *k*\' as a free parameter. Dotted and dash-dotted lines show the presence of triangles (*l*≠ 0). Triangles increase the value of *C*~*a*~(*k*).
:::
:::
where positive solution of the square root is written. By substituting this equation into Eq. (5), it is possible to calculate the minimum value of *C*~*a*~(*k*) for each configuration of *l*and *k*\".
From these two cases, we can conclude that for hubs (i.e., those nodes with high degree (*k*\' and *k*\" \>\> 1)) and for highly clustered networks (many triangles *l*\>\> 1), the values of *C*~*a*~(*k*) in the transformed network are between around \[, 1\].
To calculate the distribution of *C*(*k*) in the transformed space (*C*^*T*^(*k*)) we introduce the concept of assortativity. By assortative (disassortative) mixing in networks we understand the preference for nodes with high degree to connect to other high (low) degree nodes \[[@B24]\]. By following *Newman*\[[@B24]\], we define the probability distribution to choose a randomly edge with two nodes at either end with degrees *k*\' and *k*\" as *e*~*k*\'*k\"*~. We also assume that the nodes of the initial network are following a power-law distribution *k*^-*γ*^and have no assortative mixing. Under these assumptions, the probability distribution *e*~*k*\'*k\"*~of edges that link together nodes with degree *k*\' + *k*\" can be written as:
We make a convolution between Eq. (4) and Eq. (7), by summing for all the possible degrees of the two nodes at either end of edges (*k*\', *k*\"), which can generate transformed nodes with degree *k*= *k*\' + *k*\" - 2. Thus, we obtain:
According to the structure of *C*^*T*^(*k*) and the behavior of *C*~*a*~(*k*) exposed above, *C*^*T*^(*k*) will grow smoothly for large *k*, i.e., scaling weakly with the node degree *k*. We have calculated numerically this expression and the results are discussed later in Fig. [7](#F7){ref-type="fig"}.
We have also calculated the analytical expression for \<*C*\>, and we have found that \<*C*\> has a size-independent behavior before and after the line graph transformation is done. We can write the number of nodes with degree *k*as:
and we assume that *C*(*k*) = *A*·*k*^-*α*^, where *A*is a constant. This constant changes when we consider hierarchical networks with different number of nodes in the initial cluster \[[@B7]\]. But it seems natural because in that case the degree distribution *P*(*k*) \~ *k*^-*γ*^of the network also changes. For \<*C*\> before the transformation we can write:
Note that the summation in the denominator begins with *k*= 1 because we renormalize over all the probability distribution.
Furthermore, we can obtain \<*C*\> by using the RSMOB model (explained in next section in detail). This model starts by generating a fully connected cluster of *m*nodes, such that the connectivity of each node is *k*= *m*- 1. In the following iteration, *m*- 1 replicas of the initial cluster are generated, and linked to the central node of the original cluster in such a way that the central node of the original cluster gains (*m*- 1)·(*m*- 1) edges, and its total connectivity being *k*= *m*- 1 + (*m*- 1)^2^. By iterating these procedure, it is easy to see that hubs (i.e., central nodes of each replica) will have connectivities , with *j*= 1, \..., log~*m*~*N*being the iteration number. Therefore, assuming that the degree distribution *P*(*k*) and the clustering coefficient *C*(*k*) are power-laws with exponents *γ*\' and *α*respectively, the expression for \<*C*\> for the hubs reads as:
where *A*\' is a constant adjusted so that \<*C*\>=1 holds for *j*= 1. The upper limit of the summation log~*m*~*N*is obtained by means of the expression *m*^*j*^= *N*, which gives the total number of nodes in the network and  denotes the exponent of the power-law distribution of hubs in the RSMOB model. We must note that in a hierarchical network, the number of nodes with different degree *k*is scarce, therefore the probability distribution of node degree is properly defined as *P*(*k*) = (1/*N*~*tot*~)(*N*~*k*~/Δ*k*), where *N*~*k*~is the number of nodes with degree *k*, *N*~*tot*~is the total number of nodes, and Δ*k*means that nodes with degree *k*are binned into intervals. In addition, we note that for the hierarchical model, Δ*k*changes linearly with *k*. Hence, the exponent of the power-law is given by *γ*= 1 + *γ*\', with  where *m*is the number of nodes in the initial module.
By using Eqs. (10) and (11), we will see later (Tables [1](#T1){ref-type="table"} and [2](#T2){ref-type="table"}) that \<*C*\> converges to a constant. In order to calculate \<*C*\> after the line graph transformation is applied (\<*C*^*T*^\>), we make the substitution *C*(*k*) → *C*^*T*^(*k*) in Eq. (10). As from Eq. (8) we have seen that *C*^*T*^(*k*) is almost constant, we can conclude that \<*C*^*T*^\> also has a constant behavior and it is almost independent with network size. While the scaling law of *C*(*k*) \~ *k*^-1^was proved mathematically in \[[@B6]\], here we have obtained the analytical expressions of *C*^*T*^(*k*), \<*C*\> and \<*C*^*T*^\>.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Results of \<*C*\> evaluated by using Eq. (10) and the needed parameters in that calculation for 3 different setups: *γ*= 1 + *γ*\', where (*P*(*k*) \~ *k*^-*γ*^), *α*(*C*(*k*) \~ *k*^-*α*^), *A*(*C*(*k*) = *A*·*k*^-*α*^). Eq. (10) is a general expression of \<*C*\>.
:::
*m*initial nodes *γ* *α* *A* \<*C*\>
------------------ ------ ----- ------ ---------
3 2.58 1.1 2.34 0.20
4 2.26 1.1 3.68 0.36
5 2.16 1.1 5.18 0.54
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Results of \<*C*\> evaluated by using Eq. (11) for 3 different setups. The exponent of the power-law distribution of hubs is given by . The parameter *α*has same meaning as in Table 1. We also notice that in Eq. (11), *A*\' is adjusted so that \<*C*\>=1 holds for *j*= 1. Eq. (11) is the particular expression of \<*C*\> applied to the RSMOB model.
:::
*m*initial nodes *γ*\' *α* \<*C*\> (Eq. (11)
------------------ ------- ----- -------------------
3 1.58 1.1 0.78
4 1.26 1.1 0.81
5 1.16 1.1 0.83
:::
Line graph transformation to a hierarchical network: numerical results
----------------------------------------------------------------------
The RSMOB model \[[@B8]\] is able to reproduce the main topological features of the metabolic network. We follow the method described in \[[@B8]\] and generate a hierarchical network. Then, we apply the line graph transformation to that network.
Fig. [6](#F6){ref-type="fig"} illustrates the hierarchical network generated by the RSMOB model. The network is made of densely linked 5-node modules (it is worth noticing that the number of nodes in the initial module can be different than 5) that are assembled into larger 25-node modules (iteration n = 1, 5^2^= 25 nodes). In the next step four replicas are created and the peripheral nodes are connected again to produce 125-node modules (iteration n = 2, 5^3^= 125 nodes). This process can be repeated indefinitely \[[@B8]\].
To evaluate *C*(*k*), we have constructed three hierarchical networks with 3, 4, and 5 initial number of nodes. These networks were generated up to 7 (6561), 5 (4096), and 4 (3125) iterations (nodes), repectively. Once we have constructed these three networks, we apply the line graph transformation to them, and we calculate the *C*^*T*^(*k*) clustering coefficient for the transformed networks. In Fig. [7A](#F7){ref-type="fig"} we show the results of the clustering coefficient of the transformed network. Circles, triangles and squares indicate the values of *C*^*T*^(*k*) for the transformed network with 3, 4, and 5 initial nodes, respectively. In Fig. [7A](#F7){ref-type="fig"} we also plot with continuous lines the values of *C*^*T*^(*k*) obtained from Eq. (8). From top to bottom the lines correspond to the networks of 3, 4 and 5 initial nodes, respectively. In Fig. [7A](#F7){ref-type="fig"}, we see that the lines show an acceptable agreement with the overall tendency of data generated by the transformed network. In Fig. [7B](#F7){ref-type="fig"}, we see that the results from theoretical calculation of *C*^*T*^(*k*) via Eq. (8) (lines) are in good agreement with the experimental data (diamonds) from the KEGG database \[[@B14]\]. Moreover, in order to have enough statistics to compare with the analytical expression for the *C*^*T*^(*k*), we have binned into seven intervals the experimental data according to degree *k*(1 \<*k*≤ 8 \< ,\..., 128 \<*k*≤ 256, 256 \<*k*≤ 512), and averaged over the *C*^*T*^(*k*)\'s obtained in that range (red circles). It shows a better agreement between KEGG results and the analytical curves. The only disagreement comes at *k*= 2. This is easy to understand because in the hierarchical model depicted in Fig. [6](#F6){ref-type="fig"}, we can only find *C*(*k*= 2) = 1 for 3 initial nodes by construction of the network. However, in real networks, we could find nodes which have only two neighbors and, in some cases, these neighbors could be connected. In these cases the clustering coefficient takes value one.
In Fig. [8](#F8){ref-type="fig"}, we show the results for *C*^*T*^(*k*) after the line graph transformation is applied to the hierarchical network generated by 4 initial nodes and up to 5 iterations. The results are shown with empty triangles (red) and fitted to the dashed line. We see that *C*(*k*) \~ *k*^-1.1^changed into *C*^*T*^(*k*) \~ *k*^0.08^. We also see that the line graph transformation increases the average of the clustering value of the transformed network. These theoretical results were compared with the experimental data from KEGG \[[@B14]\], finding a good agreement, and supporting the result of a degree-independent clustering coefficient *C*^*T*^(*k*) for the reaction network.
::: {#F6 .fig}
Figure 6
::: {.caption}
######
Hierarchical network generated by using the RSMOB model \[8\]. Starting from a fully connected cluster of 5 nodes, 4 identical replicas are created, obtaining a network of N = 25 nodes in the first iteration n = 1 (5^2^= 25 nodes). We have linked to each other the central hubs of the replicas by following \[7\]. This process can repeated indefinitely. We note that the initial number of nodes can be different than 5.
:::
:::
::: {#F7 .fig}
Figure 7
::: {.caption}
######
\(A) We plot the results of the hierarchical model for *C*^*T*^(*k*) for different configurations. 3 initial nodes and up to 7 iterations (circles), 4 initial nodes and up to 5 iterations (triangles), 5 initial nodes and up to 4 iterations (squares). Prom top to bottom (3 initial nodes (black), 4 initial nodes (red), 5 initial nodes (green)), we show with lines the results of *C*^*T*^(*k*) obtained by means of Eq. (8). (B) The lines have the same meaning as before and the diamonds correspond to the experimental data for reactions from the KEGG database \[14\]. Experimental data involves 163 organisms. Circles (red): Experimental data binned into seven intervals according to degree (1 \<*k*≤ 8 \<, \..., 128 \<*k*≤ 256, 256 \<*k*≤ 512). Figures in log-linear scale.
:::
:::
::: {#F8 .fig}
Figure 8
::: {.caption}
######
Full circles (red) and dot-dashed line (red): *C*(*k*) evaluated with the hierarchical network. Empty triangles (red) and dashed line (red): *C*(*k*) after the line graph transformation is done over the hierarchical network (*C*^*T*^(*k*)). Diamonds (blue): *C*^*T*^(*k*) of reactions data from the KEGG database \[14\]. Empty circles (blue) and continuous line: *C*(*k*) of compounds data from KEGG. Hierarchical model with 4 initial nodes and 5 iterations. Figure in log-log scale.
:::
:::
For \<*C*\> we have evaluated Eq. (10) for 3 different configurations. We have considered 3 initial nodes, 4 and 5 initial nodes nodes up to 7, 5 and 4 iterations, respectively. As it is explained in \[[@B7]\], \<*C*\> approaches asymptotically to a constant value, being independent of the size of the network. The asymptotic value depends on the initial number of nodes. We calculated the values of *γ*corresponding to the degree distribution *P*(*k*) \~ *k*^-*γ*^for each network, and the related constant *A*, which appears in Eq. (10). We show in Table [1](#T1){ref-type="table"} the values of these parameters and the results of \<*C*\> obtained by Eq. (10). These values, as it can be seen in Fig. [9A](#F9){ref-type="fig"}, are below the asymptotic values of \~ 0.66 (circles) and \~ 0.74 (triangles) obtained by using the RSMOB model. However, we have found an explanation for this result. In Fig. [8](#F8){ref-type="fig"}, the full circles at the top of the dash-dotted line correspond to non-hubs nodes. We have checked that these nodes do not follow a power-law, hence the value of *C*(*k*) is being overestimated by the scaling dependence *k*^-1^and it provides a larger value of \<*C*\>. In \[[@B7]\], the values of \<*C*\> from hierarchical model were compared with the experimental values of 43 organisms. The values of \<*C*\> for each organism were around 0.15 -- 0.25. By using the KEGG database we have evaluated the experimental value \<*C*\> for 163 organisms and we obtained an average value of 0.08.
::: {#F9 .fig}
Figure 9
::: {.caption}
######
Dark (black): \<*C*\> is calculated by using the hierarchical network. Light (green): \<*C*^*T*^\> (\<*C*\> after the line graph transformation is applied to the hierarchical network). Circles (3 initial nodes), Triangles (4 initial nodes). Star (red): Experimental \<*C*^*T*^\> for reactions from the KEGG database \[14\]. (B) \<*C*\> is calculated by using Eq. (11). The results show a good agreement and similar tendency to those shown in Fig. 9(a) (dark circles and triangles). Figures in log-linear scale.
:::
:::
We show in Fig. [9A](#F9){ref-type="fig"} the values of \<*C*\> calculated for networks generated by 3 initial nodes (circles) and 4 initial nodes (triangles) by using the RSMOB model. We see that \<*C*\> approaches asymptotically to constant values around \~ 0.66 (circles) and \~ 0.74 (triangles), being independent of the size of the network. Once the line graph transformation is applied, we see that the corresponding values of \<*C*^*T*^\> also approach asymptotically to constant values. Hence, \<*C*^*T*^\> also is size-independent for large *N*(empty circles and triangles). In addition, we have averaged the experimental value of the clustering coefficient for reactions of 163 organisms found in KEGG database and we have obtained the value of \<*C*^*T*^\>= 0.74. We see that the experimental value \<*C*^*T*^\> for reactions is in good agreement with the asymptotic values obtained by the transformed network (empty triangles and circles).
Furthermore, we have also calculated \<*C*\> by using Eq. (11). This equation should reproduce the results of \<*C*\> calculated by using the RSMOB model (dark circles and triangles in Fig. [9A](#F9){ref-type="fig"}). In Fig. [9B](#F9){ref-type="fig"}, we see that the results are qualitatively similar to those shown in Fig. [9A](#F9){ref-type="fig"} (dark circles and triangles).
We remark that the theoretical analysis of \<*C*\> and \<*C*^*T*^\> done here has also been useful to prove that they are independent of network size.
Finally, in Fig. [10](#F10){ref-type="fig"} we plot the hierarchical network (left) and the transformed network (right) by using the graph drawing tool *Pajek*\[[@B25]\]. We see the high degree of compactness of the transformed network. It could be related to the concept of robustness of a network. It means that by removing one node randomly from the reaction network depicted in the Fig. [10](#F10){ref-type="fig"}, the normal behavior of the cell might be preserved by finding an alternative path (reaction) to complete the task. This fact could be a consequence of the high degree of clustering and connectivity between the nodes in the transformed network.
::: {#F10 .fig}
Figure 10
::: {.caption}
######
\(A) Hierarchical network generated by using the model of ref. \[8\] with 4-node modules and up to 2 iterations. (B) Network after the line graph transformation. We see a huge interlinked cluster in the center of figure, which generates the degree-independent clustering coefficient *C*^*T*^(*k*) (it scales weakly as *C*^*T*^(*k*) \~ *k*^0.08^).
:::
:::
Conclusions
===========
We have studied here the clustering coefficients *C*(*k*) and \<*C*\> of the reaction network by applying the line graph transformation to a hierarchical network. This hierarchical network was generated by using the RSMOB model, which reproduces properly the topological features of the metabolic network, in particular the compound network. Our results indicate that by applying the line graph transformation to the hierarchical network, it is possible to extract topological properties of the reaction network, which is embedded in the metabolic network. The RSMOB model stores the adequate information of the reaction network and the line graph transformation is one useful technique to evoke it.
While *C*(*k*) scales as *k*^-1.1^for the initial hierarchical network (compound network), we find *C*(*k*) \~ *k*^0.08^for the transformed network (reaction network). This theoretical prediction was compared with the experimental data from the KEGG database, finding a good agreement. Our results indicate that the reaction network is a degree-independent clustering network. Furthermore, the weak scaling of *C*(*k*) for the reaction network suggests us that this network may not have hierarchical organization. However, further analyses of this network, and in general of all biological networks, by following the concept of hierarchical path are encouraged \[[@B20],[@B21]\].
On the other hand, we have also conducted an analytical derivation for the clustering coefficients *C*(*k*) and \<*C*\>. Expressions for these coefficients were calculated before and after the line graph transformation is applied to the hierarchical network. The agreement obtained by using these expressions was found acceptable, and consequently, they could be useful for further analyses in different networks (biological and non-biological).
The line graph transformation has recently been applied on metabolic networks \[[@B13]\] to study the scale-free topology of the reaction network, and on the protein-protein interaction network to detect functional clusters \[[@B15]\]. The work done here is another important application of this interesting technique.
Authors\' contributions
=======================
JCN conceived of the study, designed and implemented the analyses, and prepared the manuscript. NU carried out computational implementations and experiments. TY participated in the acquisition and processing of data from KEGG database. MK provided conceptual guidance and data from the KEGG database, and conceived the initial idea of the two complementary metabolic networks. TA provided guidance, coordinated and participated in the biological and theoretical analyses, and revised the manuscript. All authors read and approved the final manuscript.
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Definitions of functions and their values before and after the line graph transformation is applied to the hierarchical network. *N*~*k*~: number of nodes of degree *k*. The † symbol means that these dependences were analyzed in the present work, while the \* symbol means that it was studied in our previous work \[13\].
:::
Func. Definition Dependence *before* Dependence *after*(Eq. (11)
--------------- -------------------------------- --------------------- -----------------------------
*P*(*k*) *N*~*k*~/*N* *k*^-*γ*^ *k*^-*γ*+1^\*
*C*~*i*~(*k*) 2*n*/\[*k*~*i*~(*k*~*i*~- 1)\] *k*^-1.1^ *k*^0.08†^
\<*C*\> size-independent^†^ size-independent^†^
:::
Acknowledgements
================
This work was partially supported by Grant-in-Aid for Scientific Research on Priority Areas (C) \"Genome Information Science\" from MEXT of Japan.
|
PubMed Central
|
2024-06-05T03:55:51.974561
|
2004-12-24
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545960/",
"journal": "BMC Bioinformatics. 2004 Dec 24; 5:207",
"authors": [
{
"first": "Jose C",
"last": "Nacher"
},
{
"first": "Nobuhisa",
"last": "Ueda"
},
{
"first": "Takuji",
"last": "Yamada"
},
{
"first": "Minoru",
"last": "Kanehisa"
},
{
"first": "Tatsuya",
"last": "Akutsu"
}
]
}
|
PMC545961
|
Background
==========
Breech presentation
-------------------
Breech presentation occurs when a baby presents with the buttocks or feet rather than head first (cephalic presentation). As breech presentation is related to both fetal size and gestational age, the incidence decreases as pregnancy progresses to 3--4% by full-term\[[@B1],[@B2]\]. Decades of controversy over the safe management of breech birth at term has recently been resolved by an international multicentre randomised controlled trial (the Term Breech Trial, TBT) of planned vaginal breech birth versus planned caesarean section (CS)\[[@B3]\]. This trial was stopped prematurely because of overwhelming benefit favouring planned CS, with a relative risk of 0.33 (95%CI 0.19--0.56) for perinatal/neonatal mortality or serious neonatal morbidity\[[@B3]\]. The TBT results were subsequently added to two small trials in a Cochrane Systematic Review\[[@B4]\]. The reduction in perinatal morbidity and mortality was even more pronounced when the analyses were limited to births in low perinatal mortality countries, such as Australia (RR 0.13; 95%CI 0.05--0.31)\[[@B4]\]. However, planned caesarean section was associated with increased maternal morbidity (RR 1.29; 95% CI 1.03--1.61)\[[@B4]\].
These results have dramatically altered a woman\'s options if she has a breech presentation at term, as CS is now offered as the safest and in many institutions the only, management option. This change has occurred rapidly: The TBT was published in October 2000 and the rate of vaginal breech birth in NSW declined from 17% in 1999 to 14% in 2000 and 4.5% in 2001\[[@B5]\]. However, while safer than vaginal breech birth, planned CS is not without risk\[[@B6],[@B7]\]. Complications include increased risk of pulmonary embolism, infection \[[@B8]-[@B10]\], bleeding\[[@B9],[@B11]\], damage to bladder and bowel\[[@B12]\], slower recovery from the birth\[[@B12],[@B13]\], longer hospitalisation\[[@B11]\], respiratory difficulties for the baby \[[@B14]-[@B16]\], delayed bonding and breastfeeding\[[@B17],[@B18]\], and compromise of future obstetric performance\[[@B17],[@B19]-[@B21]\]. Therefore, the best way to avoid the increased risks associated with term breech presentation is to avoid it altogether, and this is possible via external cephalic version.
External cephalic version (ECV) of the breech-presenting baby
-------------------------------------------------------------
External cephalic version (ECV) is the turning of a breech baby to a cephalic presentation. Systematic review of six well designed randomised controlled trials demonstrates that among women with breech presentation in late pregnancy, ECV reduces both breech presentations in labour (RR = 0.42, 95%CI 0.35--0.50) and caesarean sections (RR = 0.52. 95%CI 0.39--0.71)\[[@B22]\].
Despite clear evidence of effectiveness and potential benefit, many women decline ECV for a variety of reasons. Both breech presentation and ECV success rates are strongly influenced by parity, with success rates reported as low as 25% for women having their first baby\[[@B23]\]. For other women, the inconvenience of extra clinic visits and the need for an IV line for tocolysis may be deterrents\[[@B24],[@B25]\]. Approximately 35% of women undergoing ECV report mild or moderate discomfort during the procedure\[[@B26]\]. Other complications of ECV are either uncommon (e.g. transient fetal bradycardia \[12%\] or dizziness and palpitations from tocolysis \[4%\]) or rare (\<1%) (e.g. profound fetal bradycardia, preterm labour, premature rupture of membranes and bleeding)\[[@B26]\]. The remote possibility of emergency CS (e.g. because of placental abruption following the procedure) is also recognised. For these and other reasons, women may have a preference for planned CS. An Australian study of decision making for CS conducted in 1996 included 62 women with a breech presentation\[[@B27]\]. Of these, 39 women were offered ECV and 12 (31%) \"decided against it\". Further, 37 women were offered vaginal breech birth but 14 (38%) women chose CS\[[@B27]\].
Women\'s views and information needs
------------------------------------
To obtain data on Australian women\'s views and information needs about ECV we undertook a cross-sectional study of women\'s knowledge, attitudes and decision-making preferences for the management of breech presentation\[[@B28]\]. Of 174 pregnant women respondents (97% response rate), almost 90% preferred vaginal delivery but only 66% had heard of ECV. After a brief written explanation of ECV 39% would choose ECV, 22% were unsure and 39% would not choose ECV. The reasons for not choosing ECV included concerns about safety for the baby (13%), that ECV doesn\'t guarantee vaginal delivery (12%) and preference for a caesarean section anyway (8%). Importantly, 95% of pregnant women wanted involvement in decision-making about breech presentation.
Patient participation in clinical decision making
-------------------------------------------------
It is now recognised that many consumers want to participate in clinical decisions about their health \[[@B29]-[@B31]\]. NHMRC states that good medical decision making should take account of patients\' preferences and values\[[@B29]\], thus challenging health professionals to find ways of involving consumers/patients in decisions about their health. Yet little is currently known about how this can be effectively achieved. One method is to provide information to consumers about treatment options and likely outcomes. To assist informed decision making, such information must be unbiased and based on current, high quality, quantitative research evidence. However, patient information materials are often outdated, inaccurate, omit relevant data, fail to give a balanced view and ignore uncertainties and scientific controversies\[[@B31],[@B32]\].
To help patients take a more active role in important clinical decisions, decision aids based on latest research evidence are being developed by several centres (for example the Ottawa Health Decision Center in Canada and the Foundation for Informed Medical Decision Making in the USA). Decision aids are defined by the Cochrane Collaboration\[[@B33]\] as \"interventions designed to help people make specific and deliberative choices among options by providing (at minimum) information on the options and outcomes relevant to the person\'s health status\". Additional strategies may include providing: information on the disease/condition; the probabilities of outcomes tailored to a person\'s health risk factors; an explicit values clarification exercise; examples of others\' decisions; and guidance and coaching in the steps of decision making\[[@B33]\]. Decision aids are *non-directive*in the sense that they do not aim to steer the user towards any one option, but rather to support decision making which is informed, consistent with personal values and acted upon\[[@B34]\]. Decision aids have been found to improve patient knowledge and create more realistic expectations, to reduce decisional conflict (uncertainty about the course of action) and to stimulate patients to be more active in decision making without increasing anxiety\[[@B35]\].
Currently only 38 decision aids worldwide have been developed and carefully evaluated in randomised controlled trials\[[@B33],[@B35]\]. Examples include hormone replacement therapy for postmenopausal women, anticoagulants for atrial fibrillation, PSA testing for prostate cancer and prenatal genetic screening. Until the publication of a series of evidence-based leaflets in the United Kingdom in 2002\[[@B36],[@B37]\], no decision aids have been developed and evaluated in the context of obstetric care, although this is an area in which consumers are known to want to participate actively in decision making\[[@B38]\]. An Australian survey of 790 postpartum women found not having an active say in decisions about pregnancy care was associated with a sixfold increase in dissatisfaction among primiparas and a fifteen fold increase among multiparas\[[@B38]\]. Similarly in the UK, postpartum women rated an explanation of procedures and involvement in decision making as most important to satisfaction with care\[[@B39]\]. Further, neither obstetricians nor midwives appreciated the importance to women of \"being told the major risks for each procedure\"\[[@B39]\].
Decision making and breech presentation
---------------------------------------
The management of breech presentation is a clinical decision that fulfils Eddy\'s criteria for a decision in which patients\' values and preferences should be included\[[@B40]\]. The outcomes for the breech management options (ECV and planned CS), and women\'s preferences for the relative value of benefits compared to risks are variable and could result in decisional conflict. For such a clinical decision, a decision aid would be expected to improve patient knowledge and create more realistic expectations, to reduce decisional conflict and to stimulate patients to be more active in decision making without increasing anxiety\[[@B35]\].
Development and pilot-testing of the decision aid
-------------------------------------------------
In 2002 we developed an evidence-based decision aid for women with a breech presentation in late pregnancy. In developing the decision aid we utilised the NHMRC guideline \"How to prepare and present information for consumers of health services\"\[[@B41]\] (developed in 1999 by a team led by Dr Barratt), and the Ottawa framework established and rigorously tested by the Ottawa Health Decision Center\[[@B34]\]. The decision aid includes a Workbook, Audiotape/CD and Worksheet. The workbook highlights key points (similar to a slide presentation) and the audiotape/CD connects these points in a narrative format, providing more detail than the workbook. The worksheet is a one-page sheet to be completed by the woman to record her decision making steps, to list any questions she needs answered before deciding, and to indicate her preferred role in this decision (she should decide, her health care provider should decide, they should decide together). Most importantly, the DA was designed to be non-directive in that it did not aim to steer the user towards any one option or increase or decrease intervention rates but rather act as an adjunct to care
The decision aid was designed for women to use at home or in the clinical setting, and takes about 30 minutes to complete. The aural component is available on both audiotapes and CDs so participants can choose which they prefer to use. After working through the decision aid, the woman brings her completed worksheet to her next antenatal appointment to discuss her provisional decision with her health care provider before arriving at her final decision. The worksheet is also useful for the practitioner, who can see rapidly from it what evidence the patient has considered, what her values and preferences on this topic are and which way she is leaning in her decision.
The decision aid was developed, pilot tested and revised with extensive consumer involvement, as outlined in the NHMRC guideline on preparing information for consumers\[[@B41]\]. Content was largely driven by consumers\' questions and information needs as determined from the cross-sectional study\[[@B28]\] and from the process of drafting, pilot testing and re-drafting.
A number of draft decision aids (including workbook, audiotape/CD script, and worksheet), were developed and each subjected to pilot testing and revision as we obtained feedback. The process of testing and revising started with the project group. The next phase included a review by a group of national and international content experts, including decision aid experts, obstetricians, midwives, perinatal epidemiologists and psychologists. Once we were convinced that the content was accurate the decision aid was pilot-tested amongst consumers. There were several rounds of consumer review and refinement. We pilot-tested with members of consumer organisation (Maternity Alliance) and in a convenience sample of pregnant and recently pregnant women. The next draft was pilot-tested amongst pregnant women attending the antenatal clinic, who may or may not have had a breech presentation. And finally we formally pilot tested the decision aid with women who had a breech presentation in late pregnancy and were at the point of decision making. Pilot-testing results included: 95% of participants found the decision aid clear and easy to understand and 80% thought there was enough information for them to make a decision. Over 90% found it very helpful and nearly all women would recommend it to others. After reviewing the decision aid, women experienced a significant increase in their knowledge scores, less anxiety, had no difficulty making decisions and were satisfied with their decision.
This study aims to evaluate the ECV Decision Aid for women with a breech-presenting baby in late pregnancy. The decision aid is based on the most recently available evidence and will be evaluated to assess the impact on women\'s satisfaction with decision-making, knowledge, anxiety and pregnancy outcomes. If successful, the results could be applied to a improve consumer information and participation in clinical decisions across a wide spectrum of pregnancy care issues.
Methods
=======
Specific aim
------------
To compare the relative effectiveness of the decision aid with standard care in relation to women\'s knowledge, expectations, satisfaction with and participation in decision making, anxiety and decisional conflict. Secondary outcomes will include service utilisation and perinatal outcomes.
Hypotheses
----------
The primary hypotheses of the study are:
Use of the ECV decision aid by women with a singleton breech-presenting infant in late pregnancy
1\. increases knowledge about breech presentation and the management options
2\. reduces decisional conflict (uncertainty about the course of action)
3\. increases satisfaction with decision making
4\. reduces anxiety
The secondary hypotheses of the study are:
Use of the ECV decision aid by women with a singleton breech-presenting infant in late pregnancy does not influence
1\. uptake of ECV
2\. the proportion of women having a planned caesarean section for breech presentation at term
3\. maternal and infant outcomes
Study design
------------
We will use a randomised trial with the following study groups to assess the impact of the decision aid:
Group 1: Usual care (usual antenatal care provider counselling on the management of breech presentation)
Group 2: Usual care + Decision aid with review by a research midwife
As randomisation will be done at the individual level, there is a risk of contamination of the usual care group if the usual care provider also reviews the decision aid with women in the study group. Therefore the decision aid will be reviewed with a research midwife and the usual antenatal care providers will be blinded to the exact content and format of the decision aid.
Setting
-------
Australian obstetric hospitals that offer external cephalic version.
Participants/eligibility criteria
---------------------------------
Women with a single breech-presenting baby at ≥ 34 weeks gestation and who are clinically eligible for ECV will be invited to participate in the trial.
*Exclusion criteria*are therefore those for ECV and include women presenting with a breech in labour, multiple pregnancies, previous CS, severe fetal anomaly, ruptured membranes and indications for CS anyway. The decision aid will be produced in English and will be designed to be simple and accessible for women with low levels of literacy. The use of audiotapes and graphics will further aid comprehension and ensure that, as far as possible, women with low English literacy need not be excluded.
Procedures, recruitment, randomisation and collection of baseline data
----------------------------------------------------------------------
The study procedure will draw upon the usual schedule of antenatal visits becoming weekly in late pregnancy and the usual management of a breech-presentation in the late pregnancy (Figure [1](#F1){ref-type="fig"}). Women diagnosed with a breech-presenting baby at ≥ 34 weeks gestation will be asked to participate. The research midwife will explain the trial and obtain informed consent, collect baseline data and randomly allocate (using telephone randomisation) women to study or control groups. This is only a minor deviation from current practice. As women of child-bearing age are known to be very mobile, participants will be asked to provide alternate contact details (eg friend or relative) to enhance subsequent follow-up. Private obstetricians will be asked to offer their patients participation in the study. Those interested will be requested to come to the antenatal clinic for randomisation and recruitment. The private obstetrician will provide usual care.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Schema of ECV decision aid trial
:::
:::
Intervention
------------
The study group will receive the decision aid (workbook, tape or CD, and worksheet) and the control group will receive standard information on management options for breech presentation from their usual carer. The study group will be given the opportunity to work through the decision aid while in the antenatal clinic and/or to take home, which ever is most convenient. Many women will also want to discuss the decision with their partner. This pragmatic approach aims to assess the decision aid under the conditions most likely to be applied in real practice. At the next antenatal visit, women in the study group will review their decision aid worksheet and any questions with the research nurse.
Follow-up 1
-----------
All women will be given a follow-up questionnaire to complete prior to their next antenatal clinic visit (see Outcome Measures below for more detail). All women choosing ECV will be given the opportunity to discuss the procedure with the obstetrician providing the service. This discussion is likely to include some of the probabilistic information in the decision aid but will occur after the follow-up questionnaire and will not influence the outcome measures.
Follow-up 2
-----------
At 12--16 weeks post-partum all participants will be mailed a second follow-up questionnaire. This will assess women\'s satisfaction with their decision and the decision making process when the events are past and the outcomes known. (See Outcome Measures below for more detail). Questionnaires will be mailed with reply paid envelopes, with up to two reminder telephone prompts to non-responders.
Blinding and contamination
--------------------------
As in many obstetric interventions double blinding is virtually impossible. The main outcomes of this study are self-reported and the women are clearly not blinded to their treatment allocation. However, we will institute a number of measures aimed at keeping antenatal staff blind to the treatment allocation and preventing contamination of the control group:
• Women will review the decision aid with the research midwife and complete the first questionnaire (primary outcome measures) prior to their next antenatal consultation
• Usual antenatal care providers will be blinded to the content and format of the decision aid
• Regular in-service (educational training) for the antenatal care providers to explain the trial protocol and to make clear the potential effect of unmasking or contamination.
• Monitoring decision aid distribution and keeping them locked up and only accessible by the research midwife
• Asking participants not to reveal their treatment allocation, or share their decision aid material with antenatal staff or other women. If participants do not want to keep their decision aid they will be asked to return it.
• Monitoring the \"usual care\" (control) arm by conducting a run-in period in which women found to have a breech presentation will be asked to complete the 1^st^follow-up questionnaire. Thus we will have a baseline record of knowledge about ECV, anxiety and decisional conflict about the decision and satisfaction with the decision before the DA is in use. Comparison of the data obtained from this run-in period and the control arm will allow us to judge whether, and to what extent, contamination has occurred.
Outcome measures
----------------
### Baseline data collection
Brief baseline data will be collected to assess comparability of the study groups. The baseline assessment will include age, parity, brief socio-demographic data, highest level of education achieved, knowledge and anxiety as assessed by the state component of the short Spielberger anxiety scale\[[@B42]\].
### Primary outcomes
The effectiveness of risk communication to aid patient decision making is best assessed by a combination of cognitive, affective and behavioural outcomes\[[@B43]\]. Thus the primary outcomes of the this study will be
• cognitive: change in knowledge and realistic expectations of the management options and possible benefits and risks of each option
• affective: anxiety, satisfaction with the decision, participation in decision-making and the amount of decisional conflict (uncertainty about which course of action to choose) experienced
• behavioural: actual decision taken and acted upon (see secondary outcomes)
Measures of knowledge and realistic expectations about options for the management of breech presentation and the benefits and risks of ECV will be specific to this project. Thus we will need to develop, and test these measures as part of the project.
Anxiety will be measured by the state component of the short Spielberger anxiety scale which has been extensively used and validated\[[@B42]\]. We do not anticipate the decision aid will increase women\'s anxiety but it is nevertheless important to document any increase or decrease in anxiety attributable to the decision aid.
Satisfaction with the decision will be assessed using the Satisfaction with Decision Scale. Satisfaction with Decision Scale (a very brief six item scale with high reliability) was developed specifically to assess satisfaction with health care decisions\[[@B44]\].
Participation in decision-making will be ascertained using the five-item Degner Control Preferences Scale\[[@B45]\]. This allows respondents to specify the degree of control in decision-making they wish to assume with their doctor.
Decisional conflict will be assessed by the Decisional Conflict Scale which has established reliability, good psychometric properties and is short (16 items)\[46\]. It has been used to evaluate a range of decision aids\[[@B35]\].
Because the decision about ECV must be made within a short timeframe, the outcomes will be measured as soon as practical after the consultation in which the ECV decision was made -- prior to the next antenatal visit. For the primary outcomes this will be within one week of the decision being made (Figure [1](#F1){ref-type="fig"}, 1^st^follow-up).
Satisfaction with the decision and anxiety will be measured again at 12--16 weeks postpartum as the last weeks of pregnancy and the week after birth are associated with a reduction in state anxiety\[[@B47]\]. We are interested to explore whether women\'s views of the decision making process, and the decision they ultimately made, may change with time to reflect on the experience (Figure [1](#F1){ref-type="fig"}, 2^nd^follow-up).
### Secondary outcomes
The aim of the decision aid is to assist patient decision making, and not to influence the direction of the decision taken. Nevertheless, we think it is important to collect service utilisation and pregnancy outcome data so we will record and compare the numbers of ECVs undergone and ECV success rate in both arms of the study, as well as recording and comparing rates of pregnancy complications and perinatal outcomes. Data on ECVs are already prospectively collected for quality assurance, these include fetal lie, parity, success rates and complications. Other perinatal outcomes will be obtained (with informed consent) from the existing computerised obstetric database. These outcomes include mode of delivery (vaginal, emergency or planned CS), enrolment to delivery interval, gestational age, birthweight, Apgar scores, perinatal deaths, Neonatal Intensive Care Unit admission, maternal haemorrhage (antepartum or postpartum) and length of stay.
Statistical issues
------------------
### Sample size
Sample size calculations for the trial (significance 0.05, power 0.8) were determined using the mean difference we would like to detect in women\'s decisional conflict scores and knowledge of options and outcomes. Compared with usual care, decisional conflict was shown in the most current systematic review to be significantly reduced by decision aids; the meta-analysed mean difference was -5.75, 95%CI -8.63, -2.87 (on a scale ranging from 1 lowest to 5 highest decisional conflict; median standard deviation 13.25)\[[@B48]\]. Assuming a mean difference of -5.75 and standard deviation 13.25, we would need approximately 84 women in each arm to demonstrate changes in decisional conflict. The meta-analysis also showed that for nine trials comparing decision aids and usual care, decision aids improved average knowledge scores by 18.75 points (out of 100) (95%CI 13.1 to 24.4, median standard deviation 20)\[[@B48]\]. To show such a difference, assuming mean difference of 18.75 and standard deviation of 20, would require only 18 women in each arm. Because we would like to be able to show differences in decisional conflict if they exist, we have used the larger sample size estimate. Although follow-up will be relatively short term, there will inevitably be some loss to follow-up. To allow for 10% loss to follow-up the sample size calculated above (84) is inflated by 10% to give the effective sample size of 92 women per arm and a total sample of 184 for the trial. This sample size is different from our original application for funding. Originally, we estimated a sample size of 310 women (155 in each arm). This was based on results from a 1999 systematic review of only 2 trials of decision aids versus usual care that had assessed decisional conflict\[[@B35]\]. Subsequent to the submission (January 2001) and funding of this protocol (March 2002) the systematic review was updated (2002) incorporating 6 trials that assessed decisional conflict\[[@B48]\]. At that time we revised the sample size estimate to incorporate the most current research evidence available. Ethics approval was obtained for the protocol amendment.
### Data analysis
Analyses will be by intention to treat, including withdrawals and losses to follow-up. Study groups will be compared in terms of baseline characteristics. As this is a randomised trial, we would anticipate minimal differences in baseline characteristics. If however, important differences are found, these potential confounders will be adjusted for in the analysis of outcomes. For the primary outcomes, the mean score for each measure for each group will be compared using t-tests. If adjustment for confounders is needed a multiple linear regression model will be used. The secondary outcomes will be compared using chi-square tests of significance for categorical data and t-tests for continuous data. If adjustment for confounding is necessary logistic regression and multiple linear regression will be used respectively.
### Interim analysis
An interim analysis will be conducted part way through the study and the results will be reviewed by an independent Data Monitoring Committee. Specifically the incidence of anxiety and decisional conflict in the two randomised groups will be determined after the first 150 women have been enrolled and data have been collected. If there is a significant increase in either of these outcomes at p \< 0.01 (1-tailed) with the decision aid, the trial will be stopped. The trial will also be stopped of it is evident that no clear outcome will be obtained.
Ethical considerations
----------------------
We expect the project to provide ethical benefit. It is possible that some women may experience heightened anxiety as a result of receiving the decision aid during its evaluation by randomised trial. However, a systematic review of decision aids found they improved knowledge without increasing anxiety\[[@B33]\]. Nevertheless we will measure anxiety levels at baseline and follow-up to document any adverse effects. A trained research midwife will interview all women and obtain written consent for the trial. Women will be encouraged to discuss any concerns or anxiety about the project with the research midwife and/or with their usual antenatal care provider. Women will be reassured that they are able to drop out of the study at any time with no adverse effects on the management of their pregnancy. Participation will require women to complete brief self-report questionnaires during and after pregnancy. Working through the decision aid will take \~ 30 minutes and review of the decision and any outstanding questions will be at a routine antenatal visit.
The study has been approved by the Central Sydney Area Health Service Ethics Review Committee (Protocol no. X01-0067) and the University of Sydney Human Ethics Committee (Ref No. 3806).
Confidentiality and data security
---------------------------------
Participants in the trial will be identified by a study number only, with a master code sheet linking names with numbers being held securely and separately from the study data.
To ensure that all information is secure, data records will be kept in a secure location at the University of Sydney and accessible only to research staff. As soon as all follow-up is completed the data records will be de-identified. De-identified data will be used for the statistical analysis and all publications will include only aggregated data.
The electronic version of the data will be maintained on a computer protected by password. All hard copy patient identifiable data and electronic backup files will be kept in locked cabinets, which are held in a locked room accessed only by security code and limited staff.
Data files will be stored for 7 (seven) years after completion of the project as recommended by the NHMRC. Disposal of identifiable information will be done through the use of designated bags and/or a shredding machine.
Outcomes and significance
-------------------------
This project will make an important contribution to a largely neglected aspect of pregnancy care, assisting informed participation by women in clinical decisions that affect their pregnancy. Involvement in decision making is a strong predictor of satisfaction with care in pregnancy and childbirth, yet there are only a few published decision aids for maternity care. A decision aid for the management of breech presentation is both timely and practical as there is new evidence supporting planned CS, dramatically altering the management options. Further, the randomised trial will provide high quality evidence about the effectiveness of the decision aid in supporting shared clinical decision making during pregnancy. If successful, the results of this project could be applied to improve consumer information and participation in clinical decisions across a wide spectrum of pregnancy care. Finally, if the decision aid increases the utilisation of ECV, in addition to reducing breech presentation and CS for breech presentation (and the associated increased hospitalisation and potential morbidities), some women may have more choice of where they give birth as breech presentation precludes birth in birth centres and small rural hospitals.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
CR, AB, CRG, BP, DHS were involved in the conception and design of the study. CR, NN and CRG were responsible for the drafting of the protocol and NN and CR were involved in the development and implementation of the study. All authors have read and given final approval of the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2393/4/26/prepub>
Acknowledgements
================
This study is funded by an Australian National Health and Medical Research Council project grant (211051). Natasha Nassar and Camille Raynes-Greenow are funded by an Australian National Health and Medical Research Council Public Health Postgraduate Research Scholarship. Christine Roberts is funded by an Australian National Health and Medical Research Council Public Health Practitioner Fellowship.
|
PubMed Central
|
2024-06-05T03:55:51.977898
|
2004-12-20
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545961/",
"journal": "BMC Pregnancy Childbirth. 2004 Dec 20; 4:26",
"authors": [
{
"first": "Christine L",
"last": "Roberts"
},
{
"first": "Natasha",
"last": "Nassar"
},
{
"first": "Alexandra",
"last": "Barratt"
},
{
"first": "Camille H",
"last": "Raynes-Greenow"
},
{
"first": "Brian",
"last": "Peat"
},
{
"first": "David",
"last": "Henderson-Smart"
}
]
}
|
PMC545962
|
Background
==========
In recent years, out-of-hours primary care in the Netherlands has been substantially reorganised. Formerly, general practitioners (GPs) used to perform these services in small locum groups (6 to 8 GPs) in which they joined a rota system. Nowadays, out-of-hours care is organised in large-scaled GP cooperatives (45 to 120 GPs) following examples in the UK and Denmark \[[@B1],[@B2]\].
The initiative of reorganising out-of-hours care has come mainly from the profession itself, motivated by increased dissatisfaction with the organisation of former out-of-hours primary care services. This dissatisfaction was mainly due to the high perceived workload (after out-of-hours service a regular day of work followed), and poor separation between work and private life. The main advantage of the reorganisation was the substantial reduction of number of hours a GP has to be on call. Furthermore, the organisation of out-of-hours care became much more professional by installing management, employing doctor\'s assistants, and using chauffeured cars. Studies have indicated that GPs appear to be generally satisfied with out-of-hours care organised in cooperatives\[[@B3]\].
Not only did things change for doctors, but also patients experienced some important changes in out-of-hours primary care. Generally, the reorganisation caused a shift from more personal care to more anonymous care, with increased distance to the GP. Formerly, when patients needed primary care outside office hours, the probability of being seen by their own or a local GP with whom they were familiar, was higher. In addition, when patients contacted the GP during out-of-hours in the past, they were most likely to speak to the GP himself on the phone. Nowadays, the phone is staffed by a doctor\'s assistant who decides what action should follow the patient\'s call. Moreover, out-of-hours care used to be delivered by local GPs, indicating short distances to the GP\'s practice. In large-scale GP cooperatives, the distance to a GP outside office hours will have increased substantially for most patients.
We expected that patient satisfaction would have been reduced after the reorganisation, because factors that guaranteed personal out-of-hours care at a short distance, that may be important to patients, were changed substantially. Furthermore, in Denmark it has been shown that after the out-of-hours primary care reform patient satisfaction dropped significantly\[[@B4],[@B5]\].
Patient satisfaction with out-of-hours primary care has quite often been investigated, especially in the UK \[[@B4]-[@B11]\]. Mostly, comparisons have been made between different types of out-of-hours services. Several of these studies focused on out-of-hours primary care as organised in GP cooperatives. These studies have shown that patients are generally satisfied with out-of-hours primary care organised in GP cooperatives\[[@B5],[@B8],[@B9],[@B11]\]. Nevertheless, patients receiving telephone advice only, appear to be less satisfied compared to those attending the cooperative or those receiving a home visit. In addition, it has been shown that the patient\'s expectation about their contact with the GP cooperative strongly affects the patient\'s overall satisfaction with out-of-hours care\[[@B12]\]. Other variables that appear to be related to overall satisfaction are, access to a car, age, and waiting time\[[@B8]\].
Insight in patient satisfaction with out-of-hours care supplies the health care provider with important information on the patient\'s perception of the quality of that care. During the last years, Dutch GP cooperatives have often received negative publicity in newspapers. The reorganisation has had some important implications for patients, and therefore research on their opinions about current out-of-hours care is warranted. The purpose of this study is to determine patient satisfaction with current out-of-hours care, and to determine how satisfaction is related to different aspects of the patient\'s contact with a GP cooperative.
Methods
=======
Setting
-------
The study was conducted in the province of Limburg in the South of the Netherlands. With respect to out-of-hours primary care, the province is organisationally divided in five regions. Two of these regions each have two GP cooperatives (NL and ML), one region (OZL) has one GP cooperative with two satellite centres, and in the other two regions (WM and MH) only one GP cooperative is operational. All cooperatives but one (MH) are organisationally separate from the emergency department of the local hospital, and are located nearby the hospital. This implies that patients may choose between attending the emergency department and the GP cooperative for medical problems during out-of-hours. The MH cooperative is located at the emergency department of the region\'s only hospital and sees all patients needing out-of-hours care, except for those having a referral for emergency care.
In total, these seven GP cooperatives cover a population of about 1.1 million people (the total Dutch population is over 16 million people), and are fully operational since the 1^st^of September 2001.
Development of the questionnaire
--------------------------------
To determine relevant issues for the questionnaire we interviewed GPs and managers involved with out-of-hours primary care. In addition, we analysed the process for a patient contacting the GP cooperative for all three loci of care (telephone advice, consultation at the cooperative, and home visits) separately to make sure that all facets of the GP cooperative a patient faces would be incorporated in the questionnaire. Moreover, we also analysed unpublished Dutch questionnaires in this field, and the patient satisfaction questionnaire developed by McKinley et al. \[[@B13]\]. Based on these three analyses, we identified a number of relevant elements (initial scales). Next a set of items was developed to enable us to produce multi-item scales. Subsequently, this list was sent to the patient organisation in our province, the two largest health insurance funds, and to the five GP cooperative organisations for commentary. These organisations were asked to critically review the list of items, and to add or remove items if they considered it necessary. After receiving all commentary the questionnaire was adjusted and was submitted to five people not involved in the development but with experience with out-of-hours primary care to check for clarity of the questions.
Finally three questionnaires were constructed for each of the three types of consultations (telephone advice, consultation at the cooperative, and home visit). The three questionnaires differed on items related to the specific type of contact, but general items were the same for all three questionnaires. In this way it was possible to avoid complex skip sections which lengthen the questionnaire and can reduce the response rate. We used a balanced Likert five point scale (strongly agree, agree, neutral, disagree, strongly disagree) to record responses.
The questionnaire related to telephone advice contained six initial scales measuring: accessibility of the cooperative by phone, doctor\'s assistant\'s attitude, questions asked by the assistant, advice given by the assistant, urgency of patient\'s complaint, and overall satisfaction.
The questionnaire related to consultations at the cooperative contained ten initial scales: accessibility of the cooperative by phone, doctor\'s assistant\'s attitude, questions asked by the assistant, urgency of patient\'s complaint, waiting time at the cooperative, waiting room, distance to the cooperative, GP\'s attitude, treatment by GP, and overall satisfaction.
The questionnaire related to home visits contained eight initial scales: accessibility of the cooperative by phone, doctor\'s assistant\'s attitude, questions asked by the assistant, urgency of patient\'s complaint, waiting time until GP arrives, GP\'s attitude, treatment by GP, and overall satisfaction.
In addition, patient characteristics such as, age, gender, level of education, and health insurance (as a measure of social economic status) were recorded. Patients were also asked which type of consultation they expected prior to their contact with the GP cooperative, and whether they thought that the right diagnosis had been made.
Sample
------
From March to June 2003 a sample of 2805 patients -- who had contacted the GP cooperative in their region -- received a questionnaire by mail. Patients received this questionnaire within three weeks after they had contacted the GP cooperative. Sampling was performed per GP cooperative within the four-month period. With respect to patients who received telephone advice only and those who attended the GP cooperative, a computer program randomly selected each fourth patient contact with the GP cooperative backwards from the moment of sampling. Since the number of home visits is limited, all 150 patients, who were visited by a GP from the cooperative, prior to the moment of sampling received a questionnaire. These procedures assured that the time between receiving the questionnaire and the contact with the GP cooperative was not more than three weeks.
Per region 450 questionnaires were sent out; 150 to patients who received only telephone advice, 150 to patients who visited the GP cooperative, and 150 to patients who received a home visit. Because of parallel research, more questionnaires were sent out in one of the regions (WM): 1005 questionnaires equally distributed among the three types of patient contact with the GP cooperative. The study size was chosen based on previous research by McKinley et al\[[@B7],[@B13]\], who presented a study sample of about 1400 patients. We estimated that about half of all questionnaires would be returned, and therefore distributed 2805 questionnaires.
The study was approved by the Institutional Medical Ethics Board.
Reminder and non-respondents interview
--------------------------------------
Three to four weeks after the questionnaire had been distributed, a reminder was sent to patients who had not returned the questionnaire, with the exception of the WM area. Four weeks after the last reminder, a random sample of 100 patients who had not responded, was contacted by phone. They were asked about their reasons not to return the questionnaire, and about their opinion on the contact they had with the GP cooperative. This interview was performed during office hours, during a three-week period.
Statistics
----------
Principal components analysis with varimax rotation was used to test whether the items could be assumed to measure similar aspects or components of patients\' opinions about their contact with the GP cooperative. Next, Cronbach\'s alpha coefficient was calculated to estimate the internal consistency as a measure for reliability for each component. Finally, scale scores were calculated per component by summing the scores per item and expressing the total result as a percentage of the maximum score for each scale\[[@B13],[@B14]\]. Scale scores could range between 0 and 100.
The relationship between individual variables and overall satisfaction was analysed using multiple regression analysis, with subscale satisfaction scores as covariates. Variables that did not significantly contribute to the regression model were excluded from the final model. In case of missing data, listwise deletion of missing cases was applied. All data were analysed using SPSS-pc, version 10.0.5.
Results
=======
Patient characteristics
-----------------------
Seventy-two of the 2805 questionnaires were excluded, either because they could not be delivered (patient had moved or gave a wrong address), the patient had died, or the patient was sent a double questionnaire (multiple contacts). Eventually the response was 42.4% (1160/2733). Generally more women responded to the questionnaire, and about three-quarter of the respondents had public health insurance (table [1](#T1){ref-type="table"}). The age of respondents of those who received telephone advice only was comparable with those who attended the GP cooperative for a consultation. The respondents who received a home visit were generally older; two-third was over sixty years of age.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Patient characteristics.
:::
Telephone advice Consultation at the GP cooperative Home visit
-------------------- ------------------ ------------------------------------ ----------------
n (%) n (%) n (%)
Response 366/908 (40.3) 392/912 (43.0) 402/903 (44.5)
Age
0 -- 20 years 127 (35.5) 146 (39.0) 9 (2.3)
21 -- 40 years 96 (26.8) 81 (21.7) 26 (6.6)
41 -- 60 years 67 (18.7) 82 (21.9) 93 (23.8)
\> 60 years 68 (19.0) 65 (17.4) 263 (67.3)
Total 358 (100) 374 (100) 391 (100)
Gender
Male 148 (42.3) 159 (48.5) 177 (46.0)
Female 202 (57.7) 169 (51.5) 208 (54.0)
*Total* *350 (100)* *328 (100)* *385 (100)*
Level of education
Low 92 (27.2) 91 (25.0) 161 (46.4)
Middle 164 (48.5) 188 (51.6) 131 (37.8)
High 82 (24.3) 85 (23.4) 55 (15.8)
Total 338 (100) 364 (100) 347 (100)
Health insurance
Public 268 (74.4) 283 (73.5) 314 (80.5)
Private 92 (25.6) 102 (26.5) 76 (19.5)
*Total* *360 (100)* *385 (100)* *390 (100)*
:::
Telephone advice
----------------
Forty percent (366/908) of the patients who had received telephone advice only, returned the questionnaire. 67% of these patients responded to be satisfied (44.3%) or very satisfied (22.3%) with their contact with the GP cooperative, and 57% thought that the current out-of-hours care was an improvement compared to the former situation. We identified the same six scales that were initially set to represent patients\' opinions on aspects of primary out-of-hours care (table [2](#T2){ref-type="table"}). All six scales had Cronbach\'s alpha coefficients between 0.64 and 0.93. Detailed information on the scales and items can be found in table [7](#T7){ref-type="table"}.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Description of scales representing patients\' opinion on different aspects of out-of-hours primary care.
:::
Cases Cronbach\'s alpha Scale score
---------------------------------------- ------- ------------------- --------------------------
Scales ^a^ n Mean ± SD (95%CI)
**Telephone advice**
Accessibility by phone 364 0.72 76.5 ± 18.9 (74.6--78.5)
Doctor\'s assistant\'s attitude 363 0.91 72.8 ± 22.1 (70.5--75.1)
Questions asked by assistant 361 0.64 58.6 ± 25.4 (56.0--61.3)
Advice given by assistant 351 0.93 53.7 ± 27.3 (50.8--56.5)
Urgency of complaint 363 0.86 69.1 ± 24.5 (66.6--71.7)
Overall satisfaction 361 0.93 64.2 ± 26.1 (61.5--66.9)
**Consultation at the GP cooperative**
Accessibility by phone 385 0.73 79.3 ± 17.6 (77.5--81.1)
Doctor\'s assistant\'s attitude 386 0.88 79.8 ± 16.3 (78.2--81.4)
Questions asked by assistant 384 0.65 63.5 ± 23.0 (61.2--65.8)
Urgency of complaint 384 0.79 72.0 ± 21.5 (69.8--74.1)
Waiting time at cooperative 387 0.62 61.5 ± 25.8 (58.9--64.1)
Waiting room 381 0.60 65.6 ± 20.3 (63.5--67.6)
Distance to cooperative 388 0.75 66.7 ± 21.2 (64.5--68.8)
Treatment by GP 377 0.93 81.0 ± 18.9 (79.1--82.9)
Overall satisfaction 392 0.88 73.7 ± 19.8 (71.7--75.6)
**Home visit**
Accessibility by phone 391 0.86 80.9 ± 18.4 (79.1--82.7)
Doctor\'s assistant\'s attitude 393 0.90 80.6 ± 18.6 (78.7--82.4)
Questions asked by assistant 383 0.73 59.2 ± 26.6 (56.5--61.9)
Urgency of complaint 383 0.78 86.7 ± 16.0 (85.1--88.3)
Treatment by GP 380 0.96 84.4 ± 19.7 (82.4--86.4)
Waiting time until GP arrives 369 \- 60.0 ± 30.7 (56.8--63.1)
Overall satisfaction 390 0.92 74.6 ± 22.4 (72.4--76.9)
^a^Scale scores range from 0 to 100, where 0 represents very dissatisfied and 100 represents highly satisfied.
:::
::: {#T7 .table-wrap}
Table 7
::: {.caption}
######
Patient satisfaction questionnaire. (Original items are in Dutch)
:::
**Scale 1. Accessibility by phone^t,c,v^**
------------------------------------------------------------------------------------------------------------------------------- -----
It was easy to find the phone number of the GP cooperative^\#^ (+)
It was easy to get through on the telephone (+)
The time until the doctor\'s assistant picked up the phone was short (+)
**Scale 2. Doctor\'s assistant\'s attitude^t,c,v^**
The doctor\'s assistant was friendly on the phone (+)
The doctor\'s assistant had enough time to talk to me on the phone (+)
The doctor\'s assistant seemed to understand the problem (+)
The doctor\'s assistant took my problem seriously (+)
The information given by the doctor\'s assistant was very clear (+)
**Scale 3. Questions asked by the doctor\'s assistant^t,c,v^**
The doctor\'s assistant asked too many questions (-)
I thought it was annoying that the doctor\'s assistant started with noting my personal data before asking about my complaints (-)
**Scale 4. Urgency of complaint^t,c,v^**
I believed my problem was very severe (+)
I thought my problem needed immediate care (+)
**Scale 5. Advice given by doctor\'s assistant^t^**
The doctor\'s assistant\'s information about my problem was good (+)
The advice the doctor\'s assistant gave me was very useful (+)
The telephone advice by the doctor\'s assistant had reassured me (+)
The telephone advice by the doctor\'s assistant was sufficient considering my problem (+)
I thought the doctor\'s assistant was right to give me telephone advice only (+)
**Scale 6. Waiting time at the cooperative^c^**
I thought I had to wait too long at the registration desk (-)
I thought I had to wait too long before the GP came to see me (-)
**Scale 7. Waiting room^c^**
There was enough material (magazines et cetera) in the waiting room to entertain the patients (+)
The waiting room looked very clean (+)
**Scale 8. Distance to the GP cooperative^c^**
I think the travel time from my house to the GP cooperative is too long (-)
The GP cooperative is easy accessible (+)
**Scale 9. Treatment by the GP^c,v^**
The GP took my problem seriously (+)
The GP was friendly (+)
The GP gave me clear information about my problem (+)
The advice the GP gave me was very useful (+)
The GP had enough time for me during the consultation (+)
I was very pleased with the treatment by the GP (+)
**Scale 10. Waiting time until GP arrives^v^**
I thought it took too long for the GP to arrive (-)
**Scale 11. Overall satisfaction^t,c,v^**
I am satisfied about this contact with the GP cooperative (+)
I am satisfied about the time it took to help me (+)
I think the GP cooperative functions very well (+)
*Satisfaction rating on a scale from 1 to 10 regarding the functioning of the GP cooperative^‡^*
*Satisfaction rating on a scale from 1 to 10 regarding the telephone procedure at the GP cooperative^‡^,\**
^t^scale for the patients group who received telephone advice only
^c^scale for the patients group who attended the GP cooperative for a consultation
^v^scale for the patients group who received a home visit
\* this item was excluded from the scale related to patients who attended the GP cooperative
^\#^this item was excluded from the scale related to patients who received a home visit
^‡^these items have been divided by two to reach the same range as the other items.
:::
Overall satisfaction in this group was significantly related to five scales, with a variance explained of 62% (see table [3](#T3){ref-type="table"}.). When patients judged that the right diagnosis had been made overall satisfaction was higher. We found that satisfaction also increased with age. When patients were satisfied with the accessibility of the cooperative by phone, the doctor\'s assistant\'s attitude on the phone, and the doctor\'s assistant\'s advice overall satisfaction was higher.
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Regression analysis with overall satisfaction with out-of-hours primary care as dependent variable of patients who received only telephone advice (adjusted R^2^= 0.615).
:::
Unstandardised coefficients Standardised coefficients
----------------------------------------- ----------------------------- --------------------------- ------- -------- ----------
Constant -2.404 4.302 -0.559
Diagnosis ^(1\ =\ right,\ 0\ =\ wrong)^ 12.345 2.644 0.200 4.668 \< 0.001
Patient\'s age 0.077 0.036 0.076 2.128 0.034
Accessibility by phone ^a^ 0.155 0.054 0.112 2.859 0.005
Doctor\'s assistant\'s attitude ^a^ 0.401 0.067 0.355 5.960 \< 0.001
Doctor\'s assistant\'s advice ^a^ 0.267 0.055 0.282 4.840 \< 0.001
Variables that did not significantly contribute to the regression model: Patient\'s gender, type of health insurance, level of education, expectation about type of consultation, patient\'s perceived urgency of his or her complaint, and opinion on the questions asked by the doctor\'s assistant.
^a^Scale score ranges from 0 to 100, where 0 represents very dissatisfied and 100 represents highly satisfied.
:::
Consultation at the GP cooperative
----------------------------------
Forty-three percent (392/912) of the patients who attended the GP cooperative returned the questionnaire. Approximately 80% of these patients reported to be satisfied (54.6%) or very satisfied (26.3%) with their contact with the GP cooperative, and 61% thought that the current out-of-hours care was an improvement compared to the former situation. We identified nine scales that represent patients\' opinions on aspects of primary out-of-hours care (table [2](#T2){ref-type="table"}), with Cronbach\'s alpha coefficients between 0.62 and 0.93. Two initial scales have been merged into one scale; these were patient\'s opinion on the GP\'s attitude and the treatment by the GP. All other identified scales were the same as the initial scales. Detailed information on the scales and items can be found in table [7](#T7){ref-type="table"}.
Seven variables proved to be predictors of overall satisfaction, with a variance explained of 51% (see table [4](#T4){ref-type="table"}.). Patients, who expected prior to their contact with the cooperative that they were going to be asked to come to the GP cooperative, were generally more satisfied. Those who believed that their medical problem was urgent were less satisfied. Long waiting times and dissatisfaction with the distance to the cooperative also reduced overall satisfaction. When patients were satisfied with the accessibility of the cooperative by phone, the doctor\'s assistant\'s attitude on the phone, and the GP\'s treatment overall satisfaction was higher.
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Regression analysis with overall satisfaction with out-of-hours primary care as dependent variable of patients who went for consultation to the GP cooperative. (adjusted R^2^= 0.501).
:::
Unstandardised coefficients Standardised coefficients
------------------------------------- ----------------------------- --------------------------- -------- -------- ----------
(Constant) -5.249 5.187 -1.012
Expectation about contact \* 4.313 2.113 0.078 2.042 0.042
Accessibility by phone ^a^ 0.095 0.047 0.088 2.022 0.044
Doctor\'s assistant\'s attitude ^a^ 0.165 0.055 0.138 2.981 0.003
Urgency own complaint ^b^ -0.072 0.036 -0.078 -2.008 0.045
Waiting time ^a^ 0.181 0.030 0.241 6.059 \< 0.001
Distance to cooperative ^a^ 0.176 0.035 0.192 4.965 \< 0.001
GP\'s treatment ^a^ 0.454 0.042 0.441 10.756 \< 0.001
Variables that did not significantly contribute to the regression model: Patient\'s age and gender, type of health insurance, level of education, diagnosis (1 = right, 0 = wrong), and opinion on the questions asked by the doctor\'s assistant.
^a^Scale score ranges from 0 to 100, where 0 represents very dissatisfied and 100 represents highly satisfied.
^b^Scale ranges from 0 to 100: 0 represents not urgent and 100 represents very urgent according to the patient.
\* Indicates whether the patient received the type of contact (telephone advice, consultation at the cooperative, or home visit) he or she expected (1 = in accordance with expectation, 0 = not in accordance with expectation)
:::
Home visits
-----------
Almost forty-five percent (402/903) of the patients that received a home visit by a GP from the cooperative returned the questionnaire. About 81% of these patients reported to be satisfied (42.8%) or very satisfied (38.8%) with their contact with the GP cooperative, and 61% thought that the current out-of-hours care was an improvement compared to the former situation. We identified six multi-item scales that represented the patient\'s opinion on different aspects of out-of-hours primary care, with Cronbach\'s alpha coefficients between 0.73 and 0.96. Two initial scales have been merged into one scale; these were patient\'s opinion on the GP\'s attitude and the treatment by the GP. All other identified scales were the same as the initial scales. Detailed information on the scales and items can be found in table [7](#T7){ref-type="table"}.
We found that five variables predicted overall satisfaction, with a variance explained of 51% (see table [5](#T5){ref-type="table"}.). Similar to the group of patients who had received telephone advice only, patients who receive a home visit were generally more satisfied when they believed that the GP of the cooperative had made the right diagnosis. When patients were satisfied with the accessibility of the cooperative by phone, the doctor\'s assistant\'s attitude on the phone, and the GP\'s treatment overall satisfaction was higher. In addition, when patients were satisfied about the waiting time until the GP arrives, overall satisfaction increased.
::: {#T5 .table-wrap}
Table 5
::: {.caption}
######
Regression analysis with overall satisfaction with out-of-hours primary care as dependent variable of patients who received a home visit from a GP from the cooperative. (adjusted R^2^= 0.506).
:::
Unstandardised coefficients Standardised coefficients
----------------------------------------- ----------------------------- --------------------------- ------- -------- ----------
(Constant) -11.650 5.213 -2.235
Diagnosis ^(1\ =\ right,\ 0\ =\ wrong)^ 11.948 2.461 0.207 4.856 \< 0.001
Accessibility by phone ^a^ 0.232 0.059 0.198 3.946 \< 0.001
Doctor\'s assistant\'s attitude ^a^ 0.329 0.061 0.282 5.364 \< 0.001
GP\'s treatment ^a^ 0.260 0.050 0.233 5.155 \< 0.001
Waiting time until GP arrives\*, ^a^ 0.154 0.030 0.218 5.183 \< 0.001
Variables that did not significantly contribute to the regression model: Patient\'s age and gender, type of health insurance, education level, expectation about type of consultation, urgency of own complaint, and opinion on the questions asked by the doctor\'s assistant.
^a^Scale score ranges from 0 to 100, where 0 represents very dissatisfied and 100 represents highly satisfied.
\* Single item scale
:::
Overall satisfaction
--------------------
The means of the three loci of care, adjusted for age, sex, insurance status, and education level, show that there is no difference between overall satisfaction in the group of patients who visited the GP cooperative (75.1 ± 1.31) and those who received a home visit (72.5 ± 1.37) (Table [6](#T6){ref-type="table"}). However, patients who received telephone advice only (66.2 ± 1.30), were significantly less satisfied compared to the other two groups of patients.
::: {#T6 .table-wrap}
Table 6
::: {.caption}
######
Adjusted means for overall satisfaction.
:::
Mean SD 95% CI
-------------------------------- ------ ------ --------------
Telephone advice 66.2 1.30 63.6 -- 68.7
Consultation at GP cooperative 75.1 1.31 72.5 -- 77.6
Home visit 72.5 1.37 69.8 -- 75.2
^a^adjusted for age, sex, insurance status, and level of education.
:::
Non-response
------------
Out of 100 randomly selected patients, who had not returned the questionnaire, we were able to reach 63 by phone. Of these 63 non-respondents 35 (55.6%) were male and 28 (44.4%) were female. Many of them reported that they had forgotten to return the questionnaire (40%). A minority said not to be interested (6.7%) or did not find it needful (6.7%). Most non-respondents (46.7%) gave other reasons like, no time, too difficult, or had lost the questionnaire.
Of these patients, about 71% reported to be satisfied or very satisfied about their contact with the GP cooperative.
Discussion
==========
The results of this study indicate that patients were generally satisfied about their contact with the GP cooperative. Patients who received telephone advice only, however, were less satisfied compared to those who attended the GP cooperative and those who received a home visit. A small majority believes that current out-of-hours care is an improvement compared to the former situation.
The response rate in our study is not as high as presented previously by others who investigated patient satisfaction with out-of-hours primary care\[[@B5],[@B7]-[@B9],[@B11]\]. Reasons for patients not to return the questionnaire in our study were assessed through the non-respondents interview. We found that most patients gave reasons that were not directly related to their contact with the GP cooperative. Therefore, we assume that this reduced response rate may have had little effect on the outcome of our study. In addition, the overall satisfaction in the non-respondents group did not differ much from that of the respondents.
In the process of determining relevant aspects of out-of-hours care to patients, we consulted the province patient organisation and studied discussions on out-of-hours care in newspapers. We have not used patient interviews, although this might have identified other relevant domains of out-of-hours care. However, we think that the current questionnaire captures many relevant domains of out-of-hours care to patients as well as to health professionals.
Based on results of a Danish study \[[@B4],[@B5]\], we expected overall patient satisfaction to be low because our study took place relatively shortly after out-of-hours care had been reorganised. However, we have not assessed patient satisfaction before the reorganisation, and therefore it remains unclear whether satisfaction has changed. Nevertheless, this study showed that more than half of the patients believe that the reorganisation has improved out-of-hours primary care.
We have no reason to believe that the results of this study cannot be generalised to other regions in the Netherlands. Most GP cooperatives in the Netherlands are comparable, with respect to organisation and population size, to those in this study. In addition, the region in our study includes both rural and urban areas. Despite the similarities with out-of-hours primary care in other countries such as Ireland, the UK and Denmark, there are also differences with respect to the way these cooperatives are organised, and therefore care should be taken when generalising these results to other countries.
We identified various factors that are closely related to overall satisfaction. These factors give important insight in aspects of the GP cooperative that really matter in the patient\'s opinion on out-of-hours care. The patient\'s opinion on the doctor\'s assistant\'s attitude on the phone proved to be the strongest predictor of overall satisfaction with respect to those having received telephone advice and those that received a home visit. Also for those attending the GP cooperative, this factor was a relatively strong predictor; in this group the patient\'s satisfaction with the GP\'s treatment was by far the strongest predictor of overall satisfaction. Thus, it appears that the patients\' impression of the first contact they have with the cooperative, which is mostly through telephone, strongly influences overall satisfaction.
In accordance with other studies we found that patients who received telephone advice only, are generally less satisfied with the out-of-hours service, compared to those attending the GP cooperative and those receiving a home visit\[[@B4],[@B5],[@B8],[@B9],[@B11]\]. Patient\'s expectation of care is assumed to be an important factor that influences overall satisfaction\[[@B12]\]. In our study, only 35% of the patients with telephone advice expected that they would receive this type of consultation. In contrast, 85% of the patients that were asked to attend the cooperative or received a home visit found this type of consultation in line with their expectations. This difference in expectation of care may very well explain the difference in overall satisfaction.
It is questionable whether extra information to the public on the process of the telephone triage process will adjust patients\' expectations. Similar to what Salisbury et al\[[@B8]\] suggested, we believe that a shift to an out-of-hours care organisation based predominantly on telephone advice may decrease patient overall satisfaction. Therefore, proper information about the telephone procedure at the GP cooperative is desirable\[[@B15]\]. This information can be supplied by the doctor\'s assistant on the phone, and by written information through folders and posters in GP practices.
Conclusions
===========
This study has shown that patients are generally satisfied with out-of-hours care, but that patients with telephone advice only are less satisfied than those attending the cooperative or receiving a home visit. The patient\'s opinion on several aspects of out-of-hours care can predict overall satisfaction, with different predictors regarding the three types of consultations. However, the accessibility by phone and the doctor\'s assistant\'s attitude on the phone are always significantly related to overall satisfaction, regardless of the type of consultation. This implies that when trying to improve overall satisfaction one should always focus on at least these two factors. The questionnaire used in this study has potential for use as a standardised instrument for assessing satisfaction with out-of-hours care in The Netherlands for either research or service monitoring.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
CU participated in the design of the study, performed the statistical analysis, and drafted this manuscript. AA, SH, PZ, and HC participated in the design of the study, supervised the project, and provided critical edits to this manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1472-6963/5/6/prepub>
Acknowledgements
================
We would like to thank Marloes Elferink, Elles van Cromvoirt, and Dr. Mariette Hubben for their help during the study.
|
PubMed Central
|
2024-06-05T03:55:51.981007
|
2005-1-15
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545962/",
"journal": "BMC Health Serv Res. 2005 Jan 15; 5:6",
"authors": [
{
"first": "CJT",
"last": "van Uden"
},
{
"first": "AJHA",
"last": "Ament"
},
{
"first": "SO",
"last": "Hobma"
},
{
"first": "PJ",
"last": "Zwietering"
},
{
"first": "HFJM",
"last": "Crebolder"
}
]
}
|
PMC545963
|
Background
==========
The complexity of animal body form arises from a single fertilized egg cell in an odyssey of gene expression and regulation that controls the multiplication and differentiation of cells \[[@B1]-[@B3]\]. For over two decades, *Drosophila melanogaster*(the fruit fly) has been a canonical model animal for understanding this developmental process in the laboratory. The raw data from experiments consist of photographs (two dimensional images) of the *Drosophila*embryo showing a particular gene expression pattern revealed by a gene-specific probe in wildtype and mutant backgrounds. Manual, visual comparison of these spatial gene expressions is usually carried out to identify overlaps in gene expression and to infer interactions \[[@B4]-[@B6]\].
Whole fruit fly embryo and other related gene expression patterns have been published in a wide variety of research journals since late 1980\'s. These efforts have now entered a high-throughput phase with the systematic determination of patterns of gene expression \[e.g., \[[@B7]\]\]. As a result, the amount of data currently available has doubled leading to the imminent availability of multiple expression patterns of every gene in the *Drosophila*genome \[[@B7]\]. In addition, the use of micro-array technology to study *Drosophila*development has revealed additional and important insights into changes in gene expression levels over time and under different conditions at a genomic scale \[[@B8],[@B9]\].
With this rapid increase in the amount of available primary gene expression images, searchable textual descriptions of images have become available \[[@B7],[@B10],[@B11]\]. However, a direct comparison of the gene expression patterns depicted in the images is also desirable to find biologically similar expression patterns, because textual descriptions (even using a highly structured and controlled vocabulary) cannot fully capture all aspects of an expression pattern. In fact, there is a need for automated identification of images containing overlapping or similar gene expression patterns \[[@B6],[@B12]\] in order to assist researchers in the evaluation of similarity between a given expression pattern and all other existing (comparable) patterns in the same way that the BLAST \[[@B13]\] technique functions for DNA and protein sequences. Of course, unlike the genomes with four letters and proteomes with 20 letters, all gene expression anatomies cannot be easily reduced to, and thus represented by, a small number of components.
We previously proposed a binary coded bit stream pattern to represent gene expression pattern images \[[@B6]\]. In this digital representation, referred to as the Binary Feature Vector (BFV; BSV in \[[@B6]\]), the unstained pixels in the images (white regions and background) were denoted by a value of 0 and the stained areas (colored and foreground: gene expression) were denoted by a value of 1. Based on the BFV representations of the expression pattern, we proposed a Basic Expression Search Tool for Images (BESTi) \[[@B6]\] with an aim to produce biologically significant gene expression pattern matches using image content alone, without any reference to textual descriptions. We found that the BESTi approach generated biologically meaningful matches to query expression patterns \[[@B6]\].
In this paper, we explore how a more sophisticated Invariant Moment Vectors (IMV, \[[@B14]\]) based digital representation of gene expression patterns performs in generating an ordered list of best-matching images that contain similar/overlapping gene expression patterns to that depicted in a query image. IMV are frequently used in natural image processing (e.g., optical character recognition \[[@B15]\]) and have a number of desirable properties, including the compensation for variations of scale, translation, and rotation. If successful, IMV representations hold the promise of producing significantly shorter computing times for image-to-image matching compared to BFV.
Previously, we had examined the performance of the BFV representation for a limited dataset of early stage images \[[@B6]\]. Here we compare the relative performances of BFV and IMV first using a dataset containing 226 images (from 13 research papers). Then we test for scalability of the BESTi search by using a seven times larger dataset containing 1819 (1593 new + 226 previous) images from 262 additional research papers (list available upon request from the authors). Both datasets contained lateral views of early stage (1--8) embryos.
During these investigations, we also developed another measure of image-to-image similarity for the BFV representation. This measure is aimed at finding images that contain as much of the query image expression pattern as possible, but without penalizing for the presence of any expression outside the overlap region in the target image. In addition, we examined whether partitioning a multi-domain expression pattern into multiple BFV representations, each containing only one domain, yields a better result set.
Recently, Peng and Myers \[[@B16]\] have proposed a different procedure involving the global and local Gaussian Mixture Model (GMM) of the pixel intensities (of expression) to identify images with similar patterns. This GMM method is expected to find images with intensity and spatial similarities. This is different from the BFV and IMV methods examined here, which are intended to find only spatially similar patterns. This focus is important because, as mentioned in \[[@B6]\], the differences in gene expression intensity among images in published literature can arise simply due to use of different techniques, illumination conditions, or biological reasons. However, Peng and Myers method \[[@B16]\] appears to be promising and we plan to examine its effectiveness in a separate paper.
Results and discussion
======================
Data set generation
-------------------
An image database of 226 gene expression pattern images was initially generated using data from the literature \[[@B17]-[@B29]\]. All were lateral images and exhibited early stage (1--8) expression patterns. These images were selected because they had some commonality of gene expression (as seen by the human eye), which allowed us to evaluate the performance of the BESTi in finding correct as well as false matches under controlled conditions. BESTi was also tested for scalability on a larger dataset containing 1819 (1593 plus the 226) lateral views of early stage embryos. These 1593 images were obtained from 262 articles.
In order to present comprehensible result sets in this paper, we have primarily discussed the findings from the dataset of 226 and provided information on how those queries scaled when they were conducted for the larger dataset. In general, our focus was to show the retrieval of biologically significant matches based on both the visual overlap of the spatial gene expression pattern and the genes associated with the pattern retrieved.
Each image was standardized and the binary expression pattern extracted following the procedures described previously \[[@B6]\]. These extracted patterns, their invariant moments (*φ*~*1*~through *φ*~*7*~), and binary feature representations were stored in a database. We also calculated and stored the expression area (the count of the number of 1\'s in the binary feature represented image), the X and Y coordinates of the centroid (, ), and the principal angle (*θ*) for each extracted pattern.
To quantify the similarity of gene expressions in two images, we computed two measures (*S*~S~, *S*~C~) based on the BFV representation (See equations 2 and 3 in **Methods**). *S*~*S*~is designed to find gene expression patterns with overall similarity to the query image, whereas *S*~*C*~is for finding images that contain as much of the query image expression pattern as possible without penalizing for the presence of any expression outside the overlap region in the target image. For a given pair of gene expression patterns (A and B), *S*~*S*~is the same irrespective of which image in the pair is the query image. That is, *S*~*S*~(A,B) = *S*~*S*~(B,A). This is not so for *S*~*C*~, because *S*~*C*~measures how much of the query gene expression pattern is contained in the image. Therefore, *S*~*C*~(A,B) ≠ *S*~*C*~(B,A).
For IMV representation, we computed one dissimilarity measure (*D*~*φ*~, equation 13 in **Methods**). Results from *D*~*φ*~should be compared to that from *S*~S~, as both of these measurements do not depend on the reference image, *i.e.*, *D*~*φ*~(A,B) = *D*~*φ*~(B,A) and, also they capture overall similarity or dissimilarity.
Matches and their biological significance
-----------------------------------------
The effectiveness of the BESTi in finding biologically similar expression patterns was geared towards determining the biological validity of the results obtained from the image matching procedure. All results were based solely on quantitative similarities between images without using any textual descriptions. All images were lateral views from the early stages of fruit fly embryogenesis and were oriented anterior end to the left and dorsal to the top. We refer to the images retrieved as the BESTi-matches.
### Performance of BFV-*S*~*S*~search
Figure [1A](#F1){ref-type="fig"} shows the query image with gene expression restricted to the anterior (left) portion of the embryo, except that the expression is absent at the anterior terminus \[[@B22]\]. The query image depicts the expression of the *sloppy paired*(*slp1*) gene in a wildtype embryo. The BESTi-matches based on the *S*~*S*~measure for the representations are given in Figure [1A1--A8](#F1){ref-type="fig"}. BESTi retrieves images showing similar expression patterns, all of which are from same research article as the query image \[[@B22]\]. These images depict the expression patterns of *sloppy paired*genes (*slp1*and *slp2*) in a variety of genetic backgrounds or in combination with a head gap gene *orthodentical*(*otd*); all of these genes are essential for the pattern formation in *Drosophila*head development \[[@B30]\]. In fact, *slp1*and *slp2*are tightly linked genes found in the *slp*locus of the *Drosophila*genome. They are not only closely related in their primary sequence structure, but also significantly similar in their expression pattern (compare Figure [1A7](#F1){ref-type="fig"} and [1A8](#F1){ref-type="fig"}).
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**BESTi search results with smaller dataset.**Results from the BESTi-search for the same query image \[22\] based on (A) BFV \[*S*~*S*~\], (B) IMV \[*D*~*φ*~\] and (C) BFV \[*S*~*C*~\] representations in the original dataset (226 images); and based on (D) BFV \[*S*~*S*~\] and (E) IMV \[*D*~*φ*~\] representations in the domain database (in which distinct domains of the multi-domain expression patterns were added to the original dataset as additional data points). The search argument and the results retrieved are shown on the left and right of the arrow, respectively. The original data used to generate these expression patterns are shown above this row. BESTi-matches are arranged in descending order starting with the best hit for the given search image. Values of difference in centroids (Δ*C*~*XY*~) and principal angles (Δ*θ*) are also given. Each image is identified by the last name of the first author of the original research article and the figure number with the following abbreviations: Ashe \[19\]; Casares \[20\]; Gaul1 \[28\]; Grossniklaus \[22\]; Hartmann \[24\]; Hulskamp1 \[27\]; Hulskamp3 \[26\].
:::
:::
A search was conducted using the same query image and same distance measure (*S*~*S*~) on the larger dataset. Figure [2](#F2){ref-type="fig"} shows the top-35 matches, which contain all 8 matches shown in Figure [1A](#F1){ref-type="fig"} (images with blue colored legends). This allowed us to directly compare the quality of matches between the two datasets. Analysis of larger database of images yields more matches for the same *S*~*S*~cut-off value, as expected. A visual inspection reveals that these are all relevant images (Figure [2](#F2){ref-type="fig"}), with the larger dataset yielding more images for *otd*(20 images, Figure [2C](#F2){ref-type="fig"}). Images with expression patterns from *slp1*, *slp2*and combined *otd*expression are found in Figure [2A,B](#F2){ref-type="fig"}, and [2D](#F2){ref-type="fig"}. More importantly, searches in the larger dataset provide images containing expression patterns of additional genes: *Kruppel*(Kr), *hunchback*(*hb*), *bicoid*(*bcd*), *nanos*, *snail*, *hu-li tai shao*(*hts*) and *hairy*(Figure [2E--K](#F2){ref-type="fig"}). Since these images did not exist in the smaller dataset, they were not included in the search results in Figure [1A](#F1){ref-type="fig"}. All are biologically useful matches because combinatorial input from gap genes (Kr, *hb*) along with *slp1*establishes the domains of segment polarity genes in the head \[[@B22]\]. As for the *snail*, *hts*and *hairy*genes, there are no known interaction between them and *slp1*(gene in the query image) in the wildtype embryo, but the images show overlap in gene expression due to the genetic backgrounds used \[[@B31]-[@B33]\]. Therefore, they are also biologically relevant matches.
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**BESTi search results for *S*~*S*~with larger dataset.**Comparison of search results from the small (226 images) and large (1819 images) dataset using the *S*~*S*~measure for the same query image (Figure 1A) \[22\]. Panels (A-K) are based on the genes whose expression patterns were retrieved as follows (A) *slp1*, (B) *slp1*and *otd*, (C) *otd*, (D) *slp2*, (E) Kr, (F) *hb*, (G) *hb*and *bcd*, (H) Hb, *bcd*and *nanos*, (I) *snail*, (J) *hts*and (K) *hairy*. Images are referenced with the last name of the first author of the original article and its figure number: Grossniklaus \[22\]; Zhao \[43\]; Gao \[44\]; Wimmer \[45\]; Schulz1 \[46\]; Tsai \[47\]; Janody \[48\]; Stathopoulos \[31\]; Brent \[32\]; Zhang \[33\]. Common search results between the small and large dataset are indicated with dark blue image names.
:::
:::
### Performance of IMV search
We used the same query image for the IMV method applied to the smaller dataset (*D*~*φ*~, results in Figure [1B](#F1){ref-type="fig"}) and compared the results to the BFV-*S*~*S*~search. In this case, we obtain images containing expressions of *hb*, Kr, *tailless*(*tll*), *slp1*, *hairy*and *infra-abdominal*(*iab*) (type I transcript). It is clear that IMV search produces some biologically disconnected matches. For example, Figures [1B2, 1B4--B7](#F1){ref-type="fig"} exhibit no visual overlap in gene expression pattern with the query. Furthermore, even the biologically significant matches were retrieved out of order (Figure [1B1](#F1){ref-type="fig"} before [1B3](#F1){ref-type="fig"}). This happens because *D*~*φ*~retrieves expression patterns that are of similar shape and/or size, regardless of the translation or rotation with respect to the query image.
A comparison of the results from the smaller and larger dataset for the IMV measure is given in Figure [3](#F3){ref-type="fig"}. Twenty-six images were retrieved from the larger dataset when we used the same maximum distance value for the same query image. Of these, only two images were with expression pattern from *slp1*(Figure [3 A1--A2](#F3){ref-type="fig"}). The expression of *bcd*was found in two of the results (Figures [3 B1--B2](#F3){ref-type="fig"}). 13 images containing gap gene expression patterns of Kr, *hb*, *tll*, *giant*(*gt*) and *knirps*(*kni*) (Figures [3 C1--C4, D1--D3, E1--E2, F1--F2, I1](#F3){ref-type="fig"} and [3J](#F3){ref-type="fig"}) were also retrieved. Images with expression patterns of *hairy*, *achaete-scute*complex (AS-C), *iab*(type I transcript), IAB5 enhancer, *ventral nervous system defective*(*vnd*), *short gastrulation*(*sog*) and a combined expression of *bcd*, *nanos*and *cap \'n\' collar*(*cnc*) accounted for the remaining nine (Figures [3 G1--G2, H1--H2, K1, L1, M1, N1](#F3){ref-type="fig"} and [3O1](#F3){ref-type="fig"}). We see that the new results also suffer from the same problems as before. For example, images in Figure [3 C,E,K](#F3){ref-type="fig"} and [3L](#F3){ref-type="fig"} have no common expression pattern with the query image. Hence these are not biologically significant results even though few of them (Figures [3 C1--C4, E1--E2](#F3){ref-type="fig"}) contain expression patterns of developmentally connected genes (Kr and *tll*with *slp1*).
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**BESTi search results for *D*~*φ*~with larger dataset.**Comparison of search results from the small (226 images) and large (1819 images) dataset using the *D*~*φ*~measure for the same query image (Figure 1A) \[22\]. Panels (A-O) are based on the genes whose expression patterns were retrieved as follows (A) *slp1*, (B) *bcd*, (C) Kr, (D) *hb*(D1,D3) and Hb(D2), (E) *tll*, (F) *gt*, (G) *hairy*, (H) AS-C, (I) *hb*and Kr, (J) *kni*, (K) *iab*(type I transcript), (L) IAB5 enhancer, (M) *vnd*, (N) *sog*and (O) *nanos*, *bcd*and *cnc*. Images are referenced with the last name of the first author of the original article and its figure number: Grossniklaus \[22\]; Sauer\[49\]; Tsai\[47\]; Hulskamp1\[27\]; Gaul1\[28\]; Strunk\[50\]; Colas\[51\]; Wu\[52\]; Ghiglione\[53\]; Pankratz\[54\]; Melnick\[55\]; Janody\[48\]; Zhang\[33\]; Parkhurst\[56\]; Zhou\[57\]; Stathopoulos\[31\]. Common search results between the small and large datasets are indicated with dark blue image names.
:::
:::
Since both *S*~*S*~and *D*~*φ*~measures capture the overall similarity or dissimilarity, we can use Figures [2](#F2){ref-type="fig"} and [3](#F3){ref-type="fig"} to compare the relative effectiveness of the BFV and IMV methods on the larger dataset. We clearly see that the BFV method performs much better in retrieving both overlapping and similar expression patterns that are also biologically significant.
In addition to the Hu moments, one could also compute Zernike moments, which are based on the polar coordinate system. Both Hu moments and Zernike moments are susceptible to the same problem namely expression patterns showing a similar shape but translated to different locations in the embryo would be in the same result set. We chose to study the Hu Invariant Moment Vectors mainly because the centroid of the image can be used to distinguish between similarly shaped but translated expression patterns. With Zernike moments, the image must be inherently contained within a unit circle anchored at the centroid \[[@B34]\]. Thus, there is no straightforward method to eliminate the translational problem.
Using the Hu moments, the spatial location problem can be corrected by considering the Euclidean difference in the centroid location expressed in pixels (Δ*C*~*XY*~) of the query and results. In the case of BFV-*S*~*S*~search results in Figure [1 (A1--A8)](#F1){ref-type="fig"}, the maximum Δ*C*~*XY*~is less than or only slightly greater than the minimum Δ*C*~*XY*~for the IMV search results (Figure [1 B1--B8](#F1){ref-type="fig"}). Therefore, in the present case, the IMV-based BESTi search results need to be pared down using the centroid location difference. For example, if we consider results based on a Δ*C*~*XY*~lesser than or equal to 50 pixels, images shown in Figure [1 B2, B4--B7](#F1){ref-type="fig"} would be removed producing a more meaningful result set.
### Performance of BFV-*S*~*C*~search
Figure [1C](#F1){ref-type="fig"} shows the result for the same query image as used in Figure [1A](#F1){ref-type="fig"}, but using the newly devised *S*~*C*~distance for the BFV representation (BFV-*S*~*C*~search). This is expected to retrieve images with gene expression patterns that contain the largest amount of the overlap with the expression pattern in the query image. The top eight hits shown (Figure [1C1--C8](#F1){ref-type="fig"}) all contain over 93% of the query expression pattern: five of the matches are to the expression of *hunchback*(*hb*; C1, C3--C6) and the remaining three are from *slp1*under different genetic backgrounds. As mentioned above, the combinatorial input from gap genes (including *hb*) along with *slp1*establishes the domains of segment polarity genes in the head \[[@B22]\]. Therefore, gene expression patterns found by BFV-*S*~*C*~search are for developmentally connected genes. However, using the same query image, BFV-*S*~*C*~search yielded only two images in common with the BFV-*S*~*S*~results (Figure [1](#F1){ref-type="fig"}; C7 and C8 are the same as A5 and A4, respectively). This difference occurs because *S*~*S*~is designed to find gene expression patterns with overall similarity to the query image (Figure [1A](#F1){ref-type="fig"}), whereas *S*~*C*~is intended for finding images that contain as much of the query image expression pattern as possible and exclusive of the presence of the gene expression in the result image outside the region of overlap with the query image. Therefore, BFV-*S*~*S*~and BFV-*S*~*C*~have the capability of finding gene expression patterns from different biological perspectives.
Using the same minimum similarity value for the BFV-*S*~*C*~in the larger dataset resulted in 55 images, given in Figure [4](#F4){ref-type="fig"}. Gene expression patterns of *slp1*and *otd*accounted for 8 of these images (Figure [4A](#F4){ref-type="fig"} and [4B](#F4){ref-type="fig"}). 22 images contained expression patterns of the various gap genes *hb*, Kr, *kni*and *tll*(Figure [4C, 4E--F, 4I--L](#F4){ref-type="fig"}) that were co-expressed with *bcd*and *nanos*(Figure [4E](#F4){ref-type="fig"} and [4J](#F4){ref-type="fig"}) or with *en*(Figure [4I](#F4){ref-type="fig"}). Five other genes, developmentally connected to the gene, *slp1*, in the query image were also retrieved in this result set (*eve*, *twist*, *dpp*(*decapentaplegic*) \[[@B35]\]; *en*(*engrailed*) \[[@B36]\]; *arm*(*armadillo*) \[[@B37]\]; Figure [4M--Q](#F4){ref-type="fig"}). These images were not found in the top-35 of *S*~*S*~result set, which accentuates the different capabilities of the two BFV similarity measures in retrieving biologically relevant matches. The remaining images had expression patterns of AS-C, *sc*(s*cute*), *snail*, *hairy*, *zen*(*zerknullt*), *run*, Hsp83, *nmo*(*nemo*), Tc\'hb, *iab*, *hts*and *sog*(Figure [4D, 4G--H, 4R--Z](#F4){ref-type="fig"}) which are not known to be directly related to the gene *slp1*. All but seven of these images (Figures [4 D3--D4, H1--H2, R1, X1](#F4){ref-type="fig"} and [4Y1](#F4){ref-type="fig"}) were from a different developmental stage than the query image. Hence, by limiting the results to those from a specific stage, extraneous matches can be removed. The seven images having the same stage as the query image were retrieved because of their significant overlap (more than 94%) with the query gene expression pattern. Thus, we observe that the new distance measure *S*~*C*~has the potential to identify images containing expression patterns of developmentally connected genes, other than those retrieved by *S*~*S*~, thus improving the overall performance of the BFV method and the BESTi tool.
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**BESTi search results for *S*~*C*~with larger dataset.**Comparison of search results from the small (226 images) and large (1819 images) dataset using the *D*~*φ*~measure for the same query image (Figure 1A) \[22\]. Panels (A-Z) are based on the genes whose expression patterns were retrieved as follows (A) *slp1*, (B) *otd*, (C) *hb*, (D) AS-C, (E) *nanos*, *bcd*and Hb, (F) Kr, (G) *sc*, (H) *snail*, (I) *en*and *hb*, (J) *bcd*and *hb*, (K) *kni*and *hb*, (L) *tll*, (M) *eve*, (N) *twist*, (O) *dpp*, (P) *en*, (Q) *arm*, (R) *hairy*, (S) *zen*, (T) *run*, (U) Hsp83, (V) *nmo*, (W) Tc\'hb, (X) *iab*, (Y) *hts*and (Z) *sog*. Images are referenced with the last name of the first author of the original article and its figure number: Grossniklaus \[22\]; Gao \[44\]; Hulskamp1 \[27\]; Hulskamp3 \[26\]; Zhao \[43\]; Gaul1 \[28\]; Tsai \[47\]; Niessing \[58\]; Sauer \[49\]; Parkhurst \[56\]; Janody \[48\]; Schulz2 \[46\]; Yagi \[59\] Cowden \[60\]; Stathopoulos \[31\]; Miskiewicz \[61\]; Schulz1 \[62\]; Goff \[63\]; Sackerson \[64\]; Rusch \[65\]; Steingrimsson \[66\]; Hamada \[67\]; Zhang \[33\]; Klingler \[68\]; Bashirullah \[69\]; Verheyen \[70\]; Wolff \[71\]; Casares \[20\]; Brent \[32\]. Common search results between the small and large dataset are indicated with dark blue image names.
:::
:::
Analysis of multi-domain gene expression patterns
-------------------------------------------------
Due to the presence of multiple areas of expression, some patterns in the database that appeared to contain much better matches (by eye and biologically) to the query image were not found or ranked very high. Hence, we also analyzed multi-domain expression patterns separately for the smaller dataset. Developmental biologists are also interested in finding such patterns as they contain overlaps with the expression domains in the query image. In fact, a large number of the expression patterns available today contain multiple isolated domains of expressions since more than one topologically distinct region of expression may be produced by many genes, transgenic constructs, probes or experimental techniques (multiple staining). In such cases, we need to consider each of these regions individually as well as in the context of the composite pattern. Biologically, it is important to consider them separately because different regions of expression may be under the control of distinct *cis*-regulatory sequences \[e.g., \[[@B28],[@B38]\]\] or may represent the expression of different genes in a multiply-stained embryo.
Separating multi-domain gene expression patterns into individual components was straightforward; we simply generated multiple images from the same initial image and included them in the target dataset. This resulted in 192 additional images (418 total) in the database all of which were components of the initial gene expression patterns. The images were separated into expression regions horizontally and/or vertically depending on the gene expression. For this new set of images, the IMV as well as BFV representations were re-calculated and the BESTi query constructed as above.
Results from BFV-*S*~*S*~and IMV queries for this data set are given in Figures [1D](#F1){ref-type="fig"} and [1E](#F1){ref-type="fig"}, respectively. Now, many images with multiple regions of expression are retrieved in the result set (Figure [1D: D1--D8](#F1){ref-type="fig"}) and many of them show an even better match with the query pattern than those in Figure [1A](#F1){ref-type="fig"} for the BFV-based BESTi search. For instance, gene expression patterns are now retrieved (with more than 55% pattern similarity) from embryos with the expression of *tailless*(*tll*), which is known to interact with *slp1*in defining the embryonic head \[[@B22]\], and with a composite expression of *race (related to angiotensin converting enzyme)*, *sog*(*short gastrulation*) and *eve*(*even-skipped*) due to enhanced *race*expression in the anterior domain caused by a transgenic construct causing ectopic expression of *sog*\[[@B19]\]. Therefore, the strategy of dividing multi-domain expression data into individual domains provides additional flexibility to query individual components or sub-sets of complex expression patterns. Results also improved for IMV (Figure [1E](#F1){ref-type="fig"}), but again the outcome reinforced the need to use the difference in centroid to limit the result set.
Next we examine the performance of *S*~*S*~, *S*~*C*~and *D*~*φ*~in finding BESTi matches for a query pattern with multiple regions of expression (Figure [5A](#F5){ref-type="fig"}). This complex expression pattern consists of anterior and posterior domains caused by enhanced *race*expression resulting from dosage alteration of *dpp*in a *gastrulation defective*(*gd*) mutant background, and a middle stripe due to misexpressed *sog*using an *eve*stripe-2 enhancer \[Figure [2d](#F2){ref-type="fig"} in \[[@B19]\]\]. The results from this query are shown in Figure [5A1--A8](#F5){ref-type="fig"} (only the original image set (226) was used as the target database in this case). We again find that *S*~*S*~finds many images from the same paper as well as some images from other research articles with similar expression patterns. The results correctly include expression pattern of *eve*(Figure [5A4](#F5){ref-type="fig"}), of another pair-rule gene (*ftz*: *fushi tarazu*; Figure [5A6](#F5){ref-type="fig"}), and of two other developmentally related genes \[[@B39],[@B40]\].
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**BESTi search results with multiple domains of expression using smaller database.**Results from BESTi-search for a query image with multiple domains of expression. (A) BFV \[*S*~*S*~\], (B) IMV \[*D*~*φ*~\] and (C) BFV \[*S*~*C*~\] searches for the same expression pattern in the original database (226 images). (D) BFV \[*S*~*S*~\] search using the complete multi-domain expression in the original database and (E) BFV \[*S*~*S*~\] search using only the pattern on the left in the domain database. Search argument and the results retrieved are shown on the left and right of the arrow, respectively. Original data used to generate these expression patterns are shown above this row. BESTi-matches are arranged in descending order starting with the best hit for the given search statistic. Values of difference in centroids (Δ*C*~*XY*~) and principal angles (Δ*θ*) are also given for panels A, B and C. Each image is identified by the last name of the first author of the original research article and the figure number; with the abbreviations as follows: Ashe \[19\]; Arnosti \[17\]; Borggreve \[18\]; Casares \[20\]; Gaul1 \[28\]; Gaul2 \[29\]; Grossniklaus \[22\]; Hartmann \[24\]; Hulskamp1 \[27\]; Hulskamp2 \[25\]; Hulskamp3 \[26\].
:::
:::
When *D*~*φ*~is used as a search criterion, it produces some correct matches in the result set (Figure [5B1--B8](#F5){ref-type="fig"}). However, it generally fails to rank biologically meaningful matches as the best matches. Use of the centroid in this case is also not productive, as most of the matches show very close centroids. The principal angle (*θ*) value calculated does not show a significant difference in the early stage embryos used in this study. The results using the *S*~*C*~based search are given in Figure [5C1--C8](#F5){ref-type="fig"}. They show a number of images in common with the *S*~*S*~results. However, as expected, there are significant differences between the two searches.
The results in Figures [5D](#F5){ref-type="fig"} and [5E](#F5){ref-type="fig"} demonstrate the power of the BESTi-search when the multi-domain expression data are represented in their component patterns (domain database). In this case, all the BESTi searches are based on the use of *S*~*S*~as the search criterion. These searches are based on the complete expression (Figure [5D](#F5){ref-type="fig"}) and on one of its components (bottom-left domain, Figure [5E](#F5){ref-type="fig"}). All, but one, BESTi-matches in Figure [5D](#F5){ref-type="fig"} contain both domains of expression. In contrast, the use of only the left, anterior, domain (Figure [5E](#F5){ref-type="fig"}) in the BESTi search produces many other images in which the gene expression pattern is similar to only the anterior-ventral query pattern. Therefore, the use of individual expression components as search arguments increases the potential of directly identifying different overlapping expression patterns.
Conclusions
===========
We have found that it is possible to identify biologically significant gene expression patterns from a dataset by first extracting numeric signature descriptors and then using those descriptors in a computerized search of the database for expression patterns with similar signatures or maximum pattern similarities. We find that the BFV methodologies provide a longer and more biologically meaningful set of expression pattern matches than IMV. Even though IMV representations will produce much faster retrieval speeds for large collections of embryogenesis images, the lack of biological validity of BESTi-matches retrieved makes IMV undesirable for the present problem. Instead, investigations and strategies aimed at improving the real time performance of the BFV representation will better serve the developmental biological research.
Methods
=======
The wide variety of input methodologies, illumination conditions, equipment, and publication venues involved in the acquisition and presentation of gene expression patterns makes the available gene expression pattern data rather diverse. Extracting a gene expression pattern from its background requires the use of a combination of manual and automatic techniques. Each image is first standardized into a binary image as described in \[[@B6]\]. The standardized images are then represented using the Binary Feature Vector (BFV) \[[@B6]\], and the Invariant Moment Vectors (IMV) \[[@B14]\]. Similarity measures *S*~*S*~and *S*~*C*~are derived from BFV of which, *S*~*S*~is the one\'s complement of the distance metric *D*~*E*~presented in \[[@B6]\] and *S*~*C*~is a new measure introduced in this paper. The third metric *D*~*φ*~is deduced from the invariant moment vectors.
Binary Sequence Vector analysis
-------------------------------
The binary coded bit stream pattern, in which the two possible states indicate staining over or under a threshold value, is called as Binary Feature Vector (BFV). This is referred to as the Binary Sequence Vector (BSV) in \[[@B6]\]. In other words, we represent each image as a sequence of 1\'s and 0\'s, where the black pixels (stained areas) are denoted by a value of 1 and the white pixels (unstained and background) are denoted by a value of 0. This BFV holds the gene expression and localization pattern information of each image.
The expression patterns are ordered by evaluating a set of difference values, *D*~*E*~, between the binary feature vectors of every possible pair of images in the dataset. *D*~*E*~was introduced in \[[@B6]\] and is formally given as,
*D*~*E*~= *Count*(A XOR B)/*Count*(A OR B) (1)
The term *Count*(A XOR B) corresponds to the number of pixels [not]{.underline} spatially common to the two images and the term *Count*(A OR B) provides the normalizing factor, as it refers to the total number of stained pixels (expression area) depicted in either of the two images being compared. For simplicity, we use the one\'s complement of *D*~*E*~, as a measure of similarity of gene expression patterns between two images, *S*~*S*~, is given by the equation
*S*~*S*~= (1 - *D*~*E*~). (2)
*S*~*S*~quantifies the amount of similarity based on the overlap between two expression patterns. *S*~*S*~is equal to 1 when the two expression patterns are identical (*D*~*E*~= 0).
We introduce a new similarity measure in this paper that does not penalize for any non-overlapping region. The measure *S*~*C*~quantifies the amount of similarity based on the containment of one expression pattern in the other given by
*S*~*C*~= *Count*(A AND B)/*Count*(A) (3)
If the entire query image is contained within the result set images found in the database, *i.e.*, there is complete overlap (with respect to the query image) *S*~*C*~is equal to 1. Note that, *S*~*C*~(A,B) ≠ *S*~*C*~(B,A), because the denominator corresponds to the gene expression area of the query image.
Invariant Moment Vector (IMV) analysis
--------------------------------------
Some methodologies of image analysis produce numeric descriptors that compensate for variations of scale, translation and rotation. In the following section, we describe the invariant moment analysis of gene expression data. Invariant moment calculations have been used in optical character recognition and other applications for many years \[[@B15]\].
To calculate these invariant moment descriptors the standardized binary image \[[@B6]\] is converted to a binary representation of the same pattern (BFV). From this binary sequence of the image, the invariant moments and other descriptors are extracted using the following method \[[@B14],[@B41]\]. The continuous scale equation used is
*M*~*pq*~= ∬*x*^*p*^*y*^*q*^*f*(*x*, *y*)*dxdy*, (4)
where *M*~*pq*~is the two-dimensional moment of the function of the gene expression pattern, *f*(*x*, *y*). The order of the moment is defined as (*p*+ *q*), where both *p*and *q*are positive natural numbers. When implemented in a digital or discrete form this equation becomes
We then normalize for image translation using  and  which are the coordinates of the center of gravity, centroid, of the area showing expression. They are calculated as
Discrete representations of the central moments are then defined as follows:
A further normalization for variations in scale can be implemented using the formula,
and  is the normalization factor. From the central moments, the following values are calculated:
where *φ*~*7*~is a skew invariant to distinguish mirror images. In the above, *φ*~*1*~and *φ*~*2*~are second order moments and *φ*~*3*~through *φ*~*7*~are third order moments. *φ*~*1*~(the sum of the second order moments) may be thought of as the \"spread\" of the gene expression pattern; whereas the square root of *φ*~*2*~(the difference of the second order moments) may be interpreted as the \"slenderness\" of the pattern. Moments *φ*~*3*~through *φ*~*7*~do not have any direct physical meaning, but include the spatial frequencies and ranges of the image.
In order to provide a discriminator for image inversion (and rotation), sometimes called the \"6\", \"9\" problem, it has been suggested \[[@B14],[@B42]\] that the principal angle be used to determine \"which way is up\". This is extremely important in embryo images because gene expression at the anterior and posterior regions may simply appear to be mirror images of each other to the invariant moments, but biologically they are completely distinct. The principal axis of the gene expression pattern *f*(*x*, *y*) is the angular displacement of the minimum rotational inertia line that passes through the centroid (, ) and is given as:
The slope of the principal axis is called the principal angle *θ*. It is calculated knowing that the moment of inertia of *f*around the line  is a line through (, ) with slope *θ*. We can find the *θ*value at which the momentum is minimum by differentiating this equation with respect to *θ*and setting the results equal to zero. This produces the following equation:
Using the condition \|*θ*\| \< 45° one can distinguish the \"6\" from the \"9\" and rotationally similar gene expression patterns.
In invariant moment analysis, our initial method of image comparison calculates the Euclidean distance between the images using all moments (*φ*~*1*~through *φ*~*7*~) and combinations of these moments. For example, if the first two invariant moments are used, then
and the distance *D*~*ij*~, between a pair of images *i*and *j*where *i*, *j*= 1, 2,\...n is given by
This can be expanded to use all of the moment variables. Here, the Euclidean distance, *D*~*φ*~, between any two images is calculated as
where *i*and *q*designate images whose distance is being calculated and *j*designates the parameters used in the distance calculation and *j*= 1, 2, \..., 7. This assumes that all moments have the same dimensions or that they are dimensionless.
Using this method, it is possible to rank each of the images in order of their similarity based on, for example, the first two invariant moments that have clear-cut physical meanings. Expansion to include additional moments or parameters can be performed in a number of ways. It is possible to add additional parameters to the distance calculation making sure that each of the parameters has the same dimension. For example, *φ*~1~has the dimension of distance squared, while *φ*~2~has the dimension of the fourth power of distance, thus requiring the square root function to equalize dimensions for comparable distance calculation purposes. In general, the greater number of invariant moments used in the distance calculation, the more selective the ranking. We have also allowed for the use of the centroids and principal angle as a means of list limiting.
Authors\' contributions
=======================
SK originally conceived the project, developed the image distance measures based on the BFV representation, wrote an early version of the manuscript, and edited it until the final version. RG was responsible for writing new and using pre-existing programs to perform the image distance and parameter calculations, helped prepare the figures, searched the literature for gene expression data, maintained the database of gene expression pattern images, and helped in writing the manuscript. BVE provided the IMV method description, managed the day-to-day activities in the project, and did significant editing to produce the manuscript in the desired format for the journal. SP originally proposed the use of invariant moment vectors for biological image analysis, contributed significantly for the image distance and parameter calculations and provided critical feedback during the later stages of revision.
Acknowledgements
================
We thank Dr. Robert Wisotzkey for biological remarks, Dr. Dana Desonie for editorial comments and Dr. Stuart Newfeld for useful suggestions. This research was supported in part by research grants from National Institutes of Health (S.K.) and the Center for Evolutionary Functional Genomics (S.K.) at the Arizona State University.
|
PubMed Central
|
2024-06-05T03:55:51.986118
|
2004-12-16
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545963/",
"journal": "BMC Bioinformatics. 2004 Dec 16; 5:202",
"authors": [
{
"first": "Rajalakshmi",
"last": "Gurunathan"
},
{
"first": "Bernard",
"last": "Van Emden"
},
{
"first": "Sethuraman",
"last": "Panchanathan"
},
{
"first": "Sudhir",
"last": "Kumar"
}
]
}
|
PMC545964
|
Background
==========
Our purpose was to describe both typical and outstanding personal health promotion environments experienced by medical students in U.S. medical schools. Our interest in health promotion among medical students was based on compelling data showing that physicians who have healthy personal habits are more likely to encourage patients to adopt related habits \[[@B1]\]. However, despite the clear possibility that promoting medical student health should therefore be an innovative, efficient, and effective way to improve patient outcomes, no one has examined the extent to which Deans or students believe that this concept is enacted in medical school.
Methods
=======
We collected information through four different modalities: a literature review, a written survey of medical school Deans and students, focus groups of preclinical and clinical medical students and dean, and site visits at and interviews with medical schools with reportedly outstanding student health promotion programs.
Medical student and dean surveys
--------------------------------
There were 17 respondents to the Dean Survey (DS), representing 12 of the 16 schools in the nationally representative Healthy Doc (HD) project \[[@B2]\]. Two deans responded from Mercer, RWJ/UMDNJ, Tulane, UCLA, and University of Pennsylvania, while Colorado, Creighton, Emory, Georgetown, Loma Linda, Medical College of Georgia, and University of Rochester each had one dean respond. It was not always clear whether the Dean of Curriculum or the Dean of Student Affairs was the respondent, therefore we did not differentiate in the analyses by dean type. We also compared Deans\' responses with responses (83% response rate) from the 1336 medical students in the Class of 2003 in these Deans\' schools, as they were about to begin on wards. All medical students in that class were eligible to complete a self-administered questionnaire covering personal and professional health promotion topics. Our sample of schools was designed to be representative of all U.S. medical schools in our geographic distribution, age (our freshman average was 24 vs. 24 nationally), school size (our schools averaged 563 medical students/school vs. 527 nationally), NIH research ranking (our average was 64 vs. 62 nationally), private/public school balance (51% in private schools vs. 41% nationally), under-represented minorities (13% Blacks, Hispanics, and Native Americans, vs. 11% nationally), and gender (45% women vs. 43% nationally)^5--7^Methodology for gathering medical student data in HD has been more fully described elsewhere \[[@B2]\]. DS data were collected between February 2002 and April 2003.
In analyses comparing DS and HD data, DS schools with two respondents were first averaged so that each school is represented by one value (since repeated measures analysis was not available for the desired analyses). Variation between deans representing a school was quite low for all but one pair. By averaging for the five dean pairs (and consequently having a sample size of 12 rather than 17), the tests are conservative. Student opinion scores were also averaged for each of the twelve schools from which we received Dean responses; these averages were then correlated with the Dean\'s scores using Spearman\'s correlation method. For questions with fairly uniform responses by either Deans or students, Wilcoxon\'s Signed Rank Test was used to test if there were consistent differences between student and Dean opinion. The two variables to be correlated were ordinal variables, each with 5 levels. The type of correlation method was therefore limited to a non-parametric method. Additionally, the raw student data was clustered by school, requiring methods suitable for correlated data. Since the non-parametric method needed is not available for correlated data, we determined that the best method was to take the student mean values at each school to correlate with the dean values. While this ignored the student variability within school, this deficit was balanced by the fact that the much smaller n would require much stronger evidence of a relationship to evince a significant result.
Deans were also asked to rate their school relative to other schools. To compare these ratings to students\' opinions, schools were ranked using their mean student scores on each question related to prevention and healthy activities encouraged by the school. All 16 schools in the HD cohort were used in the ranking process (1 = highest, 16 = lowest), not just the 12 schools represented by the responding deans, as the 16 were the intended sample, and are representative of US medical schools \[[@B2]\]. Therefore, the twelve schools for which we have Dean data could have rank values between 1 and 16. For Deans\' survey questions without comparative HD data, only simple descriptive statistics are presented.
Medical student and dean focus groups
-------------------------------------
For our focus groups (conducted in 2002), we identified opportunities where there would be a wide and nationally representative range of medical schools. The first focus group was convened at the AMSA Chapter Officers\' Training Conference (COC) attended by student leaders (primarily rising second years) from every U.S. osteopathic and allopathic medical school. AMSA invited a random sampling of those attending the COC to participate in the focus group. Since the first focus group of students attracted 10 first and second year students, the second focus group was a random sample of 12 clinical students; both student focus groups had an even gender mix. Because Philadelphia has so many medical schools (five), we sampled for the second focus group from those Philadelphia students who were listed in AMSA\'s membership database. Deans of Primary Care were invited to the third focus group convened at the annual conference of the Association of American Medical Colleges. AMSA used the list of Primary Care Deans and invited a random sample of them to attend the focus group; four attended. An outside contractor (Bennett, Petts & Blumenthal) assisted AMSA in developing the focus group guide, conducted all three focus groups, transcribed the conversations and analyzed the notes for trends in responses.
Site visits and interviews
--------------------------
In 2002--2003, we identified medical school campuses with intensive programs in medical student well-being through literature and web searches, recommendations from project advisory panel members, results from the Association of Academic Health Centers\' American Network of Health Promoting Universities assessment, and participants in the HRSA-funded UME-21 project. Site visits and in-depth interviews were conducted using a protocol which sought information and recommendations on the following topics:
• Student well-being programming, including the policies, activities, and evaluation for such efforts as stress reduction, exercise, diet, and mentoring.
• Prevention in the curriculum using the *Healthy People 2010*objectives and how the various topics are integrated, taught, and evaluated.
• Deans\' office support (including financial) for prevention in the curriculum and student wellness activities.
• Student assessments and recommendations regarding their schools\' efforts.
Results
=======
Survey of Deans and medical students
------------------------------------
Most surveyed Deans reported that their schools generally support students\' health, though fewer Deans believe that their school encourages healthy eating (Table [1](#T1){ref-type="table"}). Both Deans and students rate their programs rather positively, and their responses are very highly correlated, though Deans consistently rate their programs even more positively than do students (Table [2](#T2){ref-type="table"}).
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Deans\' beliefs regarding their schools\' student health promotion efforts
:::
**% (n)**
------------------------------------------------------------------------------------------------------ -------------------- ----------- -------------------------------- -------------- -----------------------
**Strongly agree** **Agree** **Neither agree nor disagree** **Disagree** **Strongly disagree**
1.1 Overall, our medical school encourages students to lead healthy lives. 35 (6) 53 (9) 6 (1) 6 (1) 0
1.2 Our medical school curriculum emphasizes preventive medicine in medical practice. 29 (5) 53 (9) 12 (2) 6 (1) 0
1.3 Our medical school encourages extracurricular activities that promote medical students\' health. 35 (6) 35 (6) 18 (3) 6 (1) 6 (1)
1.4 Our medical school tries to minimize student stress. 41 (7) 41 (7) 12 (2) 0 6 (1)
1.5 Our medical school has a good system to help students cope with stress. 29 (5) 47 (8) 18 (3) 6 (1) 0
1.6 Our medical school encourages students\' healthy eating. 12 (2) 47 (8) 35 (6) 6(1) 0
1.7 Our medical school encourages students to exercise. 24 (4) 53 (9) 12 (2) 6 (1) 6 (1)
1.8 Our medical school discourages students from smoking 41 (7) 47 (8) 12 (2) 0 0
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Mean scores\* for and correlation coefficients ^¶^between Deans\' and students\' responses to statements concerning health promotion at medical school.
:::
**Deans\' mean score** **Students\' mean score** **r** **p-value**
-------------------------------------------------------------------------------------------------- ------------------------ --------------------------- ------- -------------
Overall, our medical school encourages students to lead healthy lives. 1.8 2.5 .87 .0002
Our medical school curriculum emphasizes preventive medicine in medical practice. 1.9 2.3 .51 .0912
Our medical school encourages extracurricular activities that promote medical students\' health. 2.1 2.7 .54 .0681
Our medical school tries to minimize student stress. 1.8 3.0 .91 \<.0001
Our medical school has a good system to help students cope with stress. 2.0 2.9 .70 .0110
Our medical school encourages students\' healthy eating. 2.3 3.1 .74 .0064
Our medical school encourages students to exercise. 2.1 2.9 .48 .1139
\*Responses were scored 1 for \"strongly agree\", continuing to 5 for \"strongly disagree\". Therefore higher scores indicate less agreement with the statement.
^¶^Spearman\'s correlation coefficients.
:::
Deans were essentially unanimous in agreeing that faculty members should model healthy behaviors, and that schools should promote health with their students (Table [3](#T3){ref-type="table"}). However, Deans felt less strongly regarding the need for more training in prevention for primary care physicians, or that a physician must have a healthy lifestyle to effectively counsel patients on healthy lifestyles (Table [3](#T3){ref-type="table"}). Students also agreed with these statements, but generally to a lesser extent than Deans (Table [4](#T4){ref-type="table"}).
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Deans\' opinions on the role medical schools and physicians should play in promoting healthy behaviors/prevention
:::
**% (n)**
------------------------------------------------------------------------------------------------------------------------------ -------------------- ----------- -------------------------------- -------------- -----------------------
**Strongly agree** **Agree** **Neither agree nor disagree** **Disagree** **Strongly disagree**
1.9 Medical school faculty members should set a good example for medical students by practicing a healthy lifestyle. 59 (10) 35 (6) 6 (1) 0 0
1.10 Medical schools should encourage students and residents to practice healthy lifestyles. 65 (11) 35 (6) 0 0 0
1.11 Primary Care physicians need more training in prevention. 29 (5) 59 (10) 12 (2) 0 0
1.12 In order to effectively encourage patient adherence to a healthy lifestyle, a physician must adhere to one him/herself. 18 (3) 65 (11) 12 (2) 6 (1) 0
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Dean and student opinions on the need for schools and faculty to promote healthy lifestyles, the need for more prevention training, and the connection between a physician\'s healthy lifestyle and his/her counseling efficacy.
:::
**Deans\' mean score** **Students\' Mean score** **Wilcoxon signed rank test p-value**
------------------------------------------------------------------------------------------------------------------------- ------------------------ --------------------------- ---------------------------------------
Medical school faculty members should set a good example for medical students by practicing a healthy lifestyle. 1.4 2.1 .0015
Medical schools should encourage their students and residents to practice healthy lifestyles. 1.3 1.9 .0015
Doctors need more training in prevention. 1.8 2.1 .0342
In order to effectively encourage patient adherence to a healthy lifestyle, a physician must adhere to one him/herself. 2.1 2.2 .3804
:::
Three-quarters of Deans believed that their medical schools\' attitude toward alcohol was that drinking in moderation was acceptable, though students had more mixed impressions about schools\' alcohol attitudes (Table [5](#T5){ref-type="table"}). Deans believed that their schools did average or better on nearly all health promotion activities (Table [6](#T6){ref-type="table"}), and students\' and Deans\' assessments of their schools are highly correlated (Table [7](#T7){ref-type="table"}).
::: {#T5 .table-wrap}
Table 5
::: {.caption}
######
Deans\' and students\' impressions of their medical schools\' attitudes about alcohol use
:::
**Deans** **Students**
-------------------------------------- ----------- --------------
No obvious attitude 18% 25%
Students shouldn\'t drink at all 6% 13%
Drinking in moderation is acceptable 76% 50%
Drinking is a good release 0% 11%
:::
::: {#T6 .table-wrap}
Table 6
::: {.caption}
######
Deans\' comparisons of their medical school vs. other medical schools
:::
**\"My school does this (circle choice) compared to other schools.\"**
----------------------------------------------------------------------------------- ------------------------------------------------------------------------ -------- --------- -------- -------
3.1 Encourages students to lead healthy lives. 24 (4) 29 (5) 41 (7) 6 (1) 0
3.2 Emphasizes preventive medicine in medical practice. 6 (1) 29 (5) 53 (9) 12 (2) 0
3.3 Encourages extracurricular activities that promote medical students\' health. 24 (4) 18 (3) 41 (7) 18 (3) 0
3.4 Encourages students to exercise. 6 (1) 38 (6) 38 (6) 19 (3) 0
3.5 Helps students minimize/cope with stress. 24 (4) 47 (8) 18 (3) 6 (1) 6 (1)
3.6 Discourages students from smoking. 12 (2) 41 (7) 41 (7) 6 (1) 0
3.7 Discourages drinking as a release for students. 6 (1) 12 (2) 65 (1) 18 (4) 0
3.8 Encourages students\' healthy eating. 6 (1) 24 (4) 65 (11) 6(1) 0
:::
::: {#T7 .table-wrap}
Table 7
::: {.caption}
######
Comparing Dean\'s perceptions of their school\'s health promotion in relation to that of other medical schools\* with school rankings based on students\' opinions.
:::
\"My school **r**^¶^ **p-value**
------------------------------------------------------------------------------------- ------------ -------------
\...encourages students to lead healthy lives.\" .78 .0026
\...emphasizes preventive medicine in medical practice.\" .45 .1433
\...encourages extracurricular activities that promote medical students\' health.\" .70 .0118
\...encourages students to exercise.\" .77 .0051
\...helps students minimize/cope with stress.\" .77/.75^φ^ .0033/.0047
\...encourages students\' healthy eating.\" .68 .0151
\*In which response possibilities were: much less, less, average, more, much more.
^¶^Spearman\'s correlation coefficient.
^φ^The Deans\' survey asked one question that queried both minimizing and coping with stress, while students were asked one about each aspect of stress. Correlations are presented for minimizing stress and coping with stress, respectively.
:::
We also asked a few narrative questions of the Deans only. Deans indicated that Student Affairs and Student Health offices most often had responsibility for handling medical student wellness (responses of 10 and 4 Deans, respectively). Funds for student wellness activities primarily came from student fees and University budgets (9 and 13 Deans, respectively). Activities\' effectiveness was usually unassessed, though some Deans used occasional surveys, data from health programs, student evaluations, and student feedback at meetings/events to help evaluate their programs.
Focus groups
------------
Our three hours of focus groups yielded little information about students\' perceptions of the relationships between their personal and clinical health promotion practices; most students either had not considered this link, or had little to say about it. A few preclinical students reported that their personal wellness is generally linked to their competence as physicians, asserting that \"if we sacrifice our own health from studying too long, staying up too late, stressing out too much about exams, we can\'t take care of other people if we don\'t watch our own health first.\" Several clinical students stated that wellness was difficult to achieve (\"We\'re really stressed, basically\"), and that having access to help/mentorship might help promote wellness for them: \"I \[would like\] having a designated person to whom students can turn at any time. That would be a hotline . . . A counselor.\" Deans generally agreed with the concept of putting a mentoring support system in place. However, both students and deans see few resources in the medical schools directed toward student wellness and what programming that is offered is reactive and small in nature. Both the students and the deans discussed wellness in terms of stress and mental well-being, rather than including physical health factors such as nutrition and exercise. Students felt that the best way to teach prevention would be through skill development and role modeling from faculty who incorporate prevention into their practice. The deans proposed that prevention be integrated throughout the curriculum and not be offered as a separate course; students concurred that more prevention instruction would be optimal and acknowledged that a separate course gives the impression that the content is less important and optional.
Site visits
-----------
We visited three medical schools with especially good and abundant practices around medical student health (Emory, Mercer, and Loma Linda Universities), and several other schools with some activities that seemed also to merit mention. These schools were selected for in-depth interviewing, with the best practices outlined in Table Five being used on medical school campuses.
Conclusions
===========
Prior literature \[ref <http://www.amsa.org/pdf/mswb_bib.pdf>\] has typically examined limited populations of medical students regarding personal health promotion, with few assessments of student well-being or of the success of various interventions, so only limited conclusions can be drawn (a situation that will be improved with this and other publications from HD). However, some trends may be emerging, such as students\' health practices being good in some spheres \[[@B2]\], but not being maintained in medical school \[[@B3]\] and residency \[[@B4],[@B5]\], with an increase in alcohol consumption, and a decrease in socialization and exercise\[[@B6]\]. Poor medical student health habits also include maladaptive behaviors such as students going to school when sick, self-prescribing, and under-using medical care \[[@B7]\]. While medical students\' positive health behaviors may be encouraged by their expanding knowledge and peer and role model support \[[@B2]\], some students may avoid treatment because of concerns that others\' knowledge of their illness may place them in academic jeopardy \[[@B8]\].
Medical student and physician health is of inherent interest, but it is especially of concern because of the well-documented link between physicians\' personal health practices and their patient counseling practices \[[@B1]\]. Despite the clear need in medical school for an emphasis on student wellness, the number of health promotion programs is declining\[[@B9],[@B10]\]: competing demands for faculty time and financial resources are barriers to program implementation, and there is virtually no systematic study of the effects of such programs beyond our HD work with surveying students\' counseling practices and validating these surveys with simulated patients (in review).
We found consistent support from both Deans and students for medical schools\' encouraging healthy student behaviors, though modest follow-through on this support. Though students seemed to have thought little about the relationships between their own personal and clinical health promotion practices, we were especially impressed with the Deans\' unanimity that faculty members should model healthy behaviors. The deans\' support of the relationship between physicians\' personal and clinical health practices, and concern about their institutions\' acting on this relationship bodes well for the role of HD principles in the future of medical education. The correlation between students\' and deans\' responses suggests that deans understand well their students\' health environments. If acted on, this finding (coupled with deans\' beliefs that the environment can and should be improved) could create important positive changes in medical education and in disease prevention.
Competing interest
==================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
EF co-developed the protocol, helped guide analyses, and drafted and revised the manuscript. JH co-developed the protocol, obtained funding, and helped edit the manuscript. LE co-developed the protocol, performed analyses, and helped edit the manuscript. All authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1472-6920/4/29/prepub>
Acknowledgements
================
We would like to thank the Robert Wood Johnson Foundation and the American Cancer Society for their interest in and financial support of this work. We also appreciate the guidance and support provided by our excellent student and faculty advisory panel, and our collaborators, the Association of Teachers of Preventive Medicine, the Association of Academic Health Centers, and the HRSA-funded UME-21 Project. We also acknowledge the contributions of AMSA\'s student leaders, Lauren Oshman, MD, MPH (2003--2004 President) and Jason Block, MD, MPH (2002--2003 AMSA Action Committee Trustee) for help with these data.
|
PubMed Central
|
2024-06-05T03:55:51.989393
|
2004-12-6
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545964/",
"journal": "BMC Med Educ. 2004 Dec 6; 4:29",
"authors": [
{
"first": "Erica",
"last": "Frank"
},
{
"first": "Joan",
"last": "Hedgecock"
},
{
"first": "Lisa K",
"last": "Elon"
}
]
}
|
PMC545965
|
Background
==========
Canadians have become increasingly vocal about the need for improved care for the dying. In response, the Special Senate Committee on Euthanasia and Assisted Suicide\[[@B1]\] declared that Canadian governments should make end-of-life care or palliative care a top priority in the restructuring of the health care system. As the leading cause of adult deaths, it is estimated that 67,400 Canadians died of cancer in 2003\[[@B2]\]. The aging of the Canadian population will result in increasing numbers of individuals presenting to the health care system with advanced chronic illnesses such as cancer\[[@B3]\]. In one Canadian province, Nova Scotia, the number of cancer deaths is projected to increase by 27% from 2,259 in 1999 to 2,870 by 2010\[[@B4]\]. Although cancer is the leading cause of death in those who receive care from comprehensive palliative care programs, access to such programs is hugely variable\[[@B5],[@B6]\].
As these calls for better care for the dying go out, hospital roles in Canadian communities have been redefined. Hospital restructuring has transferred many aspects of inpatient care to community-based care, including the end-of-life or palliative care of those with cancer. End-of-life care is defined broadly as all care provided to dying persons. Multiple providers are involved in such care and include generalist providers such as family physicians, community nurses and other primary care providers, hospital-based providers, specialists, specialist palliative care providers, volunteers and family. Unless otherwise stated \"end-of-life care\" reflects any or all of these kinds of care as received by dying patients. When warranted, we will refer to \"palliative care program\" (PCP) as the comprehensive, organized and specialized program of care for the dying. Such PCPs are often hospital based and include an inpatient palliative care unit, consultant nurses and physicians and may provide consultations and / or care in the community. End-of-life care has been provided in the community in the past, however, this health system restructuring has forced an even greater emphasis on this location for care.
Nova Scotia is a small, east coast Canadian province of just under one million people where the percentage of cancer deaths occurring outside of hospitals rose from 19.8% in 1992--93 to 30.2% in 1997--98, an increase of over 50%\[[@B7]\]. One direct result of this change, we feel, is the need for effective, available, continuous and increasingly complex care of the dying in the community. Options to provide this end-of-life care vary from the coordination and integration of existing community resources including family physicians to specialized palliative care program home support teams providing all the necessary visits. There are no free-standing, community-based hospices available to the dying in Nova Scotia.
While health system restructuring has increased community-based care, restructuring has also affected the context in which family physicians provide this primary medical care. In Nova Scotia, substantial health system restructuring occurred during the study years and initially capped physician incomes, restricted practice locations, downsized hospitals, provided new hospital-in-the-home capacity, initiated redevelopment of a provincial home care program and introduced drug co-payments for seniors receiving government drug benefits.
The purpose of this study was to describe the trends in the provision of family physician visits to those dying of the four cancers with the highest mortality in Nova Scotia and to whom, we believe, would be most frequently seen by a family physician, during the years concurrent with this health care restructuring. We hypothesized that, given our previous research showing the trends of advanced cancer patients spending more time out of hospital, and fewer dying in-hospital, family physician community-based services to them would increase.
Methods
=======
Study subjects included all Nova Scotians who died due to lung, colorectal, breast or prostate cancer from April 1, 1992 to March 31, 1998 as indicated on the Vital Statistics death certificate (International Classification of Diseases. 9^th^revision \[ICD9-CM\]).
This population-based study involved the secondary data analysis of linked administrative health information. Individual level data were obtained from: (1) the Queen Elizabeth II Health Sciences Centre Oncology Patient Information System (OPIS) which encompasses the provincial Nova Scotia Cancer Registry (NSCR) and includes provincial Vital Statistics information, (2) the Nova Scotia Medical Services Insurance Physician Services (MSIPS), and (3) the provincial Hospital Admissions and Separates file (HAS). The MSIPS includes data pertaining to all visits provided by physicians in the province who are remunerated via fee-for-service schedules and for physicians who are paid alternatively that provide \'shadow\' billing information. More than 96% of FPs in Nova Scotia were fee-for-service at the time of this study\[[@B8]\].
Health services were limited to the \'end-of-life (EOL)\', defined in this study as the six months (180 days) prior to the date of death, or from the date of initial cancer diagnosis as recorded in the NSCR to death for persons living less than six months after diagnosis. Service fee codes (for fiscal years 1992--93 to 1995--96) and health service identification numbers (for fiscal years 1996--97 to 1997--98) were used to identify FP visits.
Measures
--------
All visits provided by a family physician were counted for each patient during their end-of-life including those in the FP office, patient\'s home, long-term-care facility (LTC), emergency department (ED), and hospital inpatient settings. Ambulatory visits were defined as all visits, including those to the emergency department, but excluding those made to a patient during a hospital inpatient stay or outpatient procedures. Two \'time of care\' categories were created: regular hours (8:01 am to 5 pm), and after hours care (5:01 pm to 8 am, weekends, holidays). Because a common code was not available to identify the time of visit to hospitalized patients, all hospital inpatient visits were considered as occurring during regular hours.
In addition to the total length of inpatient hospital stay and the number of hospital admissions, we also examined the total number of specialty visits received by these patients. Demographic and clinical variables included sex, age, fiscal year of death, geographic region, time from diagnosis to death, and tumour site (lung, colorectal, breast or prostate).
Analysis
--------
Initial analyses focused on frequency counts and descriptive measures (central tendency, dispersion) of all FP visits, hospital admissions, length of inpatient stay and specialty visits, overall and by fiscal year of death. Each count was expressed as the number of visits provided per 100 \"end-of-life (EOL) person-days\". Linear temporal trends were assessed using negative binomial regression with a logarithmic link function linking the dependent variable (for example, FP visits, hospital admissions, inpatient stay) to fiscal year. Fiscal year was included as a linear predictor adjusting for age and sex, and with log (EOL person-days) as an offset variable. Assessment of a nonlinear trend was made by including fiscal year as both a linear predictor and a qualitative predictor. The Type 3 analysis of year as a qualitative predictor in this model relates to the nonlinear component of trend. Adjusted regression coefficients were exponentiated and reported as rate ratios (RR) with associated 95% confidence intervals (CI). All analyses were conducted using SAS software\[[@B9]\].
Results
=======
In total, 7212 Nova Scotians were identified as having died due to lung, colorectal, breast or prostate cancer over the six-year study period. Males comprised a larger proportion of deaths, as did adults aged 65 years and older, those who died due to lung cancer and survivors of at least 150 days from date of initial cancer diagnosis (Table [1](#T1){ref-type="table"}). The number of deaths across fiscal years remained relatively stable.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Characteristics of adults who died due to lung, colorectal, breast or prostate cancer in Nova Scotia between April 1, 1992 and March 31, 1998
:::
**Characteristic** **Number of deaths (%)**
----------------------------------- --------------------------
**Fiscal year of death**
1992/93 1142 (15.8)
1993/94 1162 (16.1)
1994/95 1243 (17.2)
1995/96 1245 (17.3)
1996/97 1241 (17.2)
1997/98 1179 (16.4)
**Sex**
Female 3126 (43.3)
Male 4086 (56.7)
**Age group (years)**
\< 65 1740 (24.1)
65--74 2151 (29.8)
75--84 2247 (31.2)
85+ 1074 (14.9)
**Cancer cause of death**
Lung 3674 (50.9)
Colorectal 1223 (17.0)
Breast 1243 (17.2)
Prostate 1072 (14.9)
**Survival time (days)**
\<31 723 (10.0)
31--60 470 (6.5)
61--90 340 (4.7)
91--120 277 (3.8)
121--150 222 (3.1)
\>150 5180 (71.8)
**Region of death**
Halifax regional municipality 2293 (31.9)
Cape Breton Island 1153 (16.0)
All other regions of Nova Scotia 3751 (52.1)
:::
In total, 139,641 visits or a median of 13 visits per patient (mean 19.4; standard deviation \[SD\] 20.3), were provided by FPs to patients during their end-of-life with 94% of patients receiving at least one FP visit. Variability across fiscal years was minimal, ranging from 92.9% receiving at least one FP visit in 1993--94 to 95.3% in 1997--98.
The majority of FP visits were provided to hospital inpatients (64%), followed by the office (15%), home (10%), the emergency department (5%) and long-term care (5%). Of visits provided to hospital inpatients, almost 71% were categorized as a \'subsequent hospital visit\', which represents continuing in-hospital care. Other inpatient visits included initial hospital visits (4.7%), visits after four weeks (17.5%), supportive care visits (4.6%) and urgent or emergency care visits (2.5%).
Temporal trends associated with ambulatory and inpatient FP visits per 100 EOL person-days are shown in Figure [1](#F1){ref-type="fig"} along with the average number of days spent as hospital inpatient stay. Ambulatory visits by service location are illustrated in Figure [2](#F2){ref-type="fig"}. After accounting for age, sex and survival time, the total number of FP visits were found to have decreased significantly over the time period (p \< 0.0001 declining from 15.3 visits per 100 EOL person-days in 1992--93 to 11.8 visits per 100 EOL person-days in 1996--97 followed by a small increase to 13.6 visits per 100 EOL person-days in 1997--98. This nonlinear trend is primarily due to the decline and then rise in the number of inpatient visits made by FPs over time. A closer examination of these inpatient visits by category (e.g., initial hospital visits, subsequent visits) did not reveal any major shift in the distribution of inpatient visit types over time. In contrast, total ambulatory visits remained relatively stable over the six-year time period with no evident significant time trends. Stratification of ambulatory visits by location of visit indicate a significant linear trend in emergency department visits over time (p \< 0.01). After accounting for age and sex, patients in 1997--98 made 18% more emergency department visits than patients in 1992--93 (adjusted RR 1.18; 95% CI 1.05, 1.34). No association was evident across time for visits provided in the office, at home or within a long term care facility.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Family physician inpatient and ambulatory visits and length of hospital stay among advanced cancer patients over time
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Ambulatory family physician visits to cancer patients by location over time
:::
:::
Examination of ambulatory FP visits provided during regular hours (8:01 am-5 pm) showed no significant change over time. However, ambulatory visits provided during \'after\' hours (5:01 pm-8 am, weekends) were found to have increased significantly over the study period (p \< 0.03). Compared to patients in 1992--93, patients in 1997--98 received 12% more ambulatory visits after hours (adjusted RR 1.12; 95% CI 1.01, 1.25). Although this increase represented a significant change over time, after hour visits were relatively few compared to the large number of other ambulatory visits and therefore exerted no impact on the overall temporal effect among all ambulatory visits.
Significant temporal trends were evident with respect to both the total number of hospital admissions experienced by the patient and the total number of days they spent as a hospital inpatient. A decline in hospital admissions was seen over time, from 1.2 admissions in 1992--93 to 1.1 admissions per 100 EOL person-days in 1997--98. Compared to 1992--93, patients in all subsequent years experienced fewer hospital admissions after accounting for sex and age. By 1997--98 patients experienced 13% fewer hospital admissions than patients in 1992--93 (adjusted RR 0.87; 95%CI 0.82, 0.93).
Over 85% of patients spent at least one day as a hospital inpatient. Patients spent on average a total of 22.7 days in hospital (SD 27.4; median 14 days; range 0--180 days) or 15.6 days per 100 EOL person-days. Total length of hospital inpatient stays declined from 18.6 days per 100 EOL person-days in 1992--93 to 14.8 in 1997--98. Results from the age and sex adjusted regression analysis indicate total length of hospital inpatient stays in 1997--98 were 21% shorter than experienced in 1992--93 (adjusted RR 0.79; 95%CI 0.71, 0.88).
In total, 82,575 visits were made to a medical specialty during the end of life. The number of visits ranged from one to 169, with a median of 11 visits per patient (mean 11.4; SD 13.5). Age and sex adjusted regression analysis indicate visits to a medical specialty did not change significantly over time.
Discussion
==========
Despite the move to a greater percentage of cancer patients dying out of hospital\[[@B7]\] and despite the findings of this study which show advanced cancer patients are spending more time out of hospital during the end-of-life, we have found no indication of increased family physician involvement in office, home or long-term care settings. There was an increase in FP visits in the emergency department. Many questions follow. By whom and how is the medical component of community-based end-of-life cancer care being provided? We have shown that the number of visits made to a specialist physician has not changed significantly over the same time period. Therefore, is care previously performed by family doctors now being offered by non-physicians? Might these patterns be influenced by changes in the provision of end-of-life care with an increased use of systemic therapies? Are patients receiving adequate and appropriate care at the end-of-life?
When we compare these trends to concurrent trends in the province of Nova Scotia for all types of patients, a number of interesting points emerge. For *all types*of patients in the province from 1992--1999, the total number of office-based, home and long-term-care visits has declined slightly\[[@B8]\]. This is not true for the patients in our study. FPs may be continuing to see cancer patients in the office despite reduced office visits for other types of patients in an attempt to ensure comprehensive care for those with this serious illness.
We expected to see an increase in home visits for those dying of cancer given the longer period people are spending out of hospital and the greater numbers dying at home. This was not the case. Home and long-term care visits among patients remained stable over time. This trend may have been facilitated by a revamped home care system in 1995 providing more nursing and assisted care visits in the home in the latter part of the decade perhaps reducing the need for family physician visiting.
Hospital visits declined quite substantially in the early years of our study and then increased in the final year. Provincial information suggests this trend was true for non-cancer patients as well\[[@B8]\]. Since both length of stay and the number of hospital admissions have declined, it is possible that those admitted in 1997/1998 were those who had been cared for longer in the community but who had reached a critical point where they were sicker than in the previous years and required more visits during these shorter stays. In other words, if patients in hospital had greater severity of illness, greater medical visit intensity may have been required. This may account for the rise in hospital visits in the last study year.
The increase in emergency department visits by family physicians parallels a rise during the study period among the general Nova Scotian population. (personal communication, M. Joyce, Department of Finance, Government of Nova Scotia). This greater emergency department utilization may be a reflection of reduced access to inpatient beds experienced by all patients, including those with advanced cancer.
The lack of increased visits made by family physicians in end-of-life care in the community is concerning when one considers the evidence that their participation is associated with a greater likelihood of home death \[[@B10]-[@B12]\] and less emergency department use\[[@B13]\]. It is not, however, surprising given the overall decline in comprehensiveness of care by individual family physicians in Canada\[[@B14]\].
It is important to note that the advantage of using provincial administrative health databases is that we have information for the entire population regarding cancer mortality and physician and hospital utilization. However, there is no clinical information on severity of disease, which would provide a much better understanding of factors influencing health service utilization. In addition, we do not have concurrent home care utilization data or private long-term care facility data to factor into our modeling.
Conclusions
===========
Despite health care restructuring of the 1990s which resulted in fewer days in-hospital for those dying with cancer in Nova Scotia, there was no concurrent increase in the family physician visits provided to the dying. This may represent a growing unmet need for community-based medical care. However, further research is needed to examine whether end-of-life-care needs are being met by other health providers such as specialized home palliative care nurses, or general home care nursing services.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
FB and BL participated in the conceptualisation and design of the project, the analysis and interpretation of the data, created the first draft of the article, and incorporated co-authors\' comments into the final draft. GJ participated in the design of the project, the interpretation of data, and revising the manuscript. GF participated in the analysis and interpretation of data and the revising of drafts. All authors gave approval to the final version.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2296/6/1/prepub>
Acknowledgements
================
Dr. Burge is supported by a Senior Clinical Research Scholar Career Award from the Faculty of Medicine, Dalhousie University. The authors wish to thank Dr. Ina Cummings for her initial contribution to the design of this project, and Natalie Dawson for manuscript preparation. This project was supported by CIHR grant number MOP-44617.
|
PubMed Central
|
2024-06-05T03:55:51.991913
|
2005-1-4
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545965/",
"journal": "BMC Fam Pract. 2005 Jan 4; 6:1",
"authors": [
{
"first": "Frederick I",
"last": "Burge"
},
{
"first": "Beverley",
"last": "Lawson"
},
{
"first": "Grace",
"last": "Johnston"
},
{
"first": "Gordon",
"last": "Flowerdew"
}
]
}
|
PMC545968
|
Background
==========
Comorbid illness plays an essential, but poorly defined, role in the diagnosis and management of malignant disease. Increasingly, the importance of measuring comorbidity in consistent and quantifiable ways is being recognized. This movement has stemmed in part from a growing consensus that comorbidity confounds the results of clinical trials and limits the generalization of results to older and sicker patients \[[@B1],[@B2]\]. For various reasons, however, the widespread integration of comorbidity into clinical research has yet to be realized. It is our contention that limited accessibility and cumbersome scoring techniques are in part responsible for the limited use of comorbidity indices. We believe that easily accessible tools for calculating comorbidity can increase their use in clinical research.
While multiple comorbidity indices are available, each with unique advantages and disadvantages, no single index has emerged as clearly superior to the others. In fact, we have noted that a distinct trade-off between prognostic utility and ease of use exists. We believe that a scoring system that maximizes ease of use while maintaining prognostic validity represents the optimal balance required for use in clinical research. In addition, a scoring system that can be easily integrated into an electronic medical record will further promote the widespread use of comorbidity data. We have, therefore, chosen to explore the use of the Charlson Comorbidity Index (CCI) as the prototypical comorbidity index in our department. The primary aim of this article, therefore, is to provide an electronic Charlson Comorbidity Index Scoring program and explain its development and use. In an effort to provide the reader with the context in which the electronic application was developed and should be used, the interaction of comorbidity and malignant disease and the validation of the Charlson Index in oncology are discussed. For more detailed reviews of the comorbidity indices outlined in this article, the authors refer readers to three systematic reviews of comorbid illness scoring systems by Extermann and de Groot \[[@B2]-[@B4]\].
Comorbidity and cancer
----------------------
In the 1960\'s, Feinstein initially reported the prognostic importance of patient-related characteristics, such as symptomatology and concurrent illness, in his analyses of differences between actual survival outcomes and those predicted by TNM-based staging among lung cancer patients \[[@B5]\]. In recent years, the direct influence of comorbid illness on treatment decision-making and survival outcomes has been documented for a variety of malignancies including bladder, lung, head and neck, colorectal, breast, and prostate cancers \[[@B6]-[@B11]\].
Hall, et al, for example, evaluated the effect of comorbidity on survival among head and neck cancer patients, concluding that 16% of mortality at 3 years and 18% at 5 years was attributable to comorbid illness alone, with non-cancer causes of death exceeding cancer-related causes of death after 7.5 years \[[@B6]\]. Satariano & Ragland made comparable observations in their study of breast cancer patients. In their analysis, comorbidity increased directly with age (p \< 0.001) and a significant association between comorbidity and the type of treatment received (p \< 0.0001) was observed. After controlling for age, cancer stage, and type of treatment, increasing comorbidity remained significantly predictive of increased all-cause-mortality (1 condition, p = 0.04; and for 2 or 3 conditions, p \< 0.001) \[[@B7]\].
Comorbidity has also demonstrated marked predictive power for survival and treatment allocation among prostate cancer patients. A Netherlands cancer registry study, for example, identified comorbidity as the single most important prognostic factor for 3-year survival, with hazard ratios of 2.0 (95% CI = 1.0--4.3) for a single comorbid illness and 7.2 (95% CI = 3.1--16.6) for 2 or more comorbid conditions, and trends toward fewer radical prostatectomies among men with higher degrees of comorbidity \[[@B8]\]. Total comorbidity counts have been found to be strongly predictive of survival among colon cancer patients as well. In addition to identifying increasing comorbidity with age (p \< 0.0001), Yancik, et al, found raw counts of comorbid conditions to be strongly predictive of survival when used in a model containing age group, disease stage, and gender (p = 0.0007), with risk ratios of 1.11 (95% CI 1.10--1.90) and 1.84 (95% CI 1.39--2.46) for total comorbid illness counts of 5--6 and 7--14 respectively \[[@B9]\]. Additional works by De Marco with colon cancer patients \[[@B10]\], Firat with lung cancer patients \[[@B11]\], and Piccirillo with head and neck cancer patients \[[@B12]\] provide unquestionable support for the importance of comorbidity on survival and treatment-related complications among oncology patients.
Although the preceding examples are not intended to provide a comprehensive review of the influence of comorbidity on survival and treatment-related complications in oncology, they provide a clear demonstration of the effect. In addition, the themes of increasing comorbidity with age and the influence of comorbidity on outcomes and treatment decision-making are illustrated. With these interactions in mind, the investigation of comorbidity has become an area of increasing interest in our department. In particular, we have begun focusing on the use of comorbidity indices and their application in clinical research. For a variety of reasons, which will be explained in forthcoming sections of this work, we have focused on the Charlson Comorbidity Index as the prototypical index on which to base this research.
The Charlson Comorbidity Index
------------------------------
The Charlson Index was developed in 1987 based on 1-year mortality data from internal medicine patients admitted to a single New York Hospital and was initially validated within a cohort of breast cancer patients. The index encompasses 19 medical conditions weighted 1--6 with total scores ranging from 0--37. In the development phase of the index, mortality for each disease was converted to a relative risk of death within 12 months. A weight was then assigned to each condition based on the relative risk (RR); for example, RR \<1.2 = weight 0, RR ≥ 1.2\<1.5 = weight 1, RR ≥ 1.5\<2.5 = weight 2, RR ≥ 2.5\<3.5 = weight 3, and for 2 conditions (metastatic solid tumor and AIDS) = weight 6.
From the weighted conditions, a sum score can be tallied to yield the total comorbidity score. The CCI can be further adapted to account for increasing age. In the validation phase of the CCI, age was also found to be an independent risk factor for death from a comorbid condition. As a result, relative risk was calculated to increase by 2.4 for each additional decade of life. In the same cohort, the relative risk of death for each 1-point increase in CCI score was 2.3. To account for the effects of increasing age, one point can be added to the CCI score for each decade of life over the age of 50 \[[@B13]\].
Reviews of the CCI suggests it has good reliability, excellent correlation with mortality and progression-free survival outcomes, and is easily modifiable, particularly to account for the effect of age. The CCI\'s basic limitations include preservation of data only for the 19 conditions listed in the index, the exclusion of non-malignant hematologic disease, such as anemia, and reduced predictive ability for outcomes \< 6-months. The CCI is praised for its ease of use, short rating time, extractability from other indices, and widespread use \[[@B2],[@B3]\].
Validation of the Charlson Index
--------------------------------
Statistical criteria for the assessment of the validity of a test include content validity, criterion validity, construct validity, and reliability \[[@B4]\]. Although a detailed discussion of statistical tests of validity is beyond the scope of this review, the assessments provide a basis from which to begin an analysis of the validity of the Charlson Index. Statistical criteria of validity, as applied to comorbidity indices, are ultimately dependent upon the comparison of comorbidity indices to each other, as well as subjective assessments of certain criteria, such as content validity and cutoff points for correlation coefficients. The criteria are, therefore, in and of themselves, problematic. Despite these limitations, their application to the common comorbidity indices has been studied extensively.
In a review of validity among comorbidity indices, de Groot, et al, systemically identified articles referring to comorbidity between 1966 and 2000. They compared the Charlson Index with the Cumulative Illness Rating Scale (CIRS), Kaplan-Feinstein Index (KFI), and Index of Coexistent Disease (ICED) and identified correlation coefficients of \> 0.40, \"good\" test-retest reliability and \"moderate to good\" inter-rater reliability for the CCI \[[@B4]\]. In addition, the Charlson Index correlated significantly with mortality, disability, readmission, and length of stay outcomes, suggesting good predictive validity leading de Groot, et al, to conclude that the Charlson Index, as well as the ICED, KFI, and CIRS, is a valid and reliable method for assessing comorbidity in clinical research \[[@B4]\].
A similar review by Extermann suggests the Charlson Index possesses excellent validity and reliability for use in clinical research in oncology. Extermann also reported exceptional predictive validity, correlating the CCI with outcomes involving mortality risk from weeks to years, postoperative complications, length of hospital stay, discharge to nursing home, and progression-free survival among cancer patients. Additionally, inter-rater reliability, by various measures, was reported at 0.74 among a cohort of older general oncology patients and 0.945 within a group of elderly breast cancer patients. Test-retest reliability was also good, ranging from 0.92 among surgical patients and 0.86 among the previously mentioned group of elderly oncology patients. Although Extermann urges some caution based on the tendency of the CCI to result in comorbidity scores that are sometimes lower than those observed with other indices, she concludes that the CCI is easy to use and \"highly suitable for vast cohort studies but may under-detect significant problems resulting in non-lethal endpoints\" \[[@B2]\].
The Charlson Index has demonstrated excellent predictive validity for a variety of clinical outcomes as well as numerous malignancies. As discussed previously, the CCI was developed using a prospective analysis of 1-year mortality rates among internal medicine patients and then validated within a population of 588 breast cancer patients. In the validation phase of Charlson\'s original study, increasing CCI scores were significantly correlated with increased 10-year mortality within a breast cancer cohort (χ^2^= 163, p \< 0.0001), with CCI scores of 0, 1, 2, and 3 predicting 10-year survival rates of 93%, 73%, 52%, and 45%, respectively. In the original manuscript, Charlson, et al, cautioned that their index should be considered preliminary and that it required validation in larger populations \[[@B13]\].
Since the original work by Charlson, et al, the CCI has exhibited substantial prognostic power for both survival and treatment related complications in numerous retrospective studies. Singh, et al, for example, retrospectively analyzed CCI validity within a cohort of head and neck cancer patients. Their analysis revealed reduced median tumor specific survival (12.3 vs. 38.7 months, p = 0.007), and increased risk of cancer death (RR = 2.35) for patients with advanced (≥ 2) CCI scores. The CCI compared similarly to the KFI with respect to frequency of advanced comorbidity (30% for CCI and 32% for KFI) and prognostic power (Spearman correlation coefficient, p \<0.001, r = 0.73). However, the CCI was more applicable to the study population than the KFI, with the KFI successfully applied to only 80% of the study population compared with 100% application of the CCI \[[@B14]\].
Fowler, et al, also examined the validity of the Charlson index in a cohort of men with prostate cancer treated with EBRT or RP. After adjusting for age, a direct relationship between actuarial survival and CCI score (p = 0.00001) was found for all patients. Among individuals with CCI scores of 0, 5 and 10-year survival rates were 86% and 66% compared with 40% and 9% for patients with CCI scores of 3 to 5. Relative mortality risk, based on CCI scores of 0, 1, 2, and 3--5, increased from 1 to 1.7, 2.6, and 5.7, respectively \[[@B15]\].
Additional studies among prostate cancer patients have compared the CCI, KFI, and ICED. Albertsen, et al, for example, found each of the three comorbidity indices had similar power to predict survival (p \< 0.001 for each), with the addition of any of the three indices to Gleason score improving predictive power for survival over Gleason score alone \[[@B16]\]. We also recently reviewed the importance of comorbidity and prognostic utility of the CCI among prostate cancer patients and found that the CCI consistently correlates with reduced survival as well as treatment allocation \[[@B17]\].
The Charlson Index has also been validated as a prognostic indicator for survival in lung cancer cohorts. Firat, et al, recently explored the prognostic importance of comorbidity among patients undergoing surgical resection or definitive EBRT for clinical NSCLC. Within the combined group, both CIRS-G scores ≥ 4 (p \< 0.001) and Charlson score ≥ 2 (p = 0.004) emerged as significant prognostic indicators of reduced overall survival. Examination of the surgical and EBRT groups separately also demonstrated higher CIRS-G and Charlson scores within the EBRT group as compared with the surgical group \[[@B18]\].
The effect of comorbidity on complication rates among lung cancer patients has also been investigated. Brim, et al, for example, identified gender, CCI score 3--4, COPD, and prior tumor within the last 5 years as predictors for major complications (re-thoracotomy, empyema, pleural effusion, bronchopleural fistula, ventilatory support \>72 hours, ventricular arrhythmia, pulmonary embolism, cardiac failure, or myocardial infarction). Charlson scores of 3--4 maintained statistical significance after multivariate regression (OR 9.8, 95% CI 2.1--45.9) \[[@B19]\].
CCI scores have also demonstrated prognostic value, both in terms of postoperative complications and survival among colon cancer patients. Rieker, et al, found raw CCI scores reached 0--2, 3--4, and ≥ 5 in 66%, 25%, and 8% of patients, respectively. With respect to survival, CCI score \>2 emerged as a poor prognostic indicator for overall survival for all stages (p \< 0.001, OR 2.91, 95% CI = 2.00--4.94). Subgroup analysis of stage III and IV patients revealed reduced cancer-specific survival among patients with CCI score \>2 (log rank p \<0.005). CCI scores \> 2 were also correlated with receipt of blood transfusion (p \< 0.021, OR 1.56, 95% CI = 1.07--2.28), postoperative complications (p \< 0.001, OR 2.18, 95% CI = 1.50--3.16), and ICU stay \> 2 days (p \< 0.001, OR 3.28, 95% CI = 1.91--5.64) \[[@B20]\].
Taken together, this series of papers represents a diverse and relatively large experience with the Charlson Index. In each report, CCI scores consistently correlate with disease specific survival, overall survival, or treatment-related complications, confirming its predictive validity.
Implementation
==============
The CCI Calculator provided with this manuscript is based on the original index proposed by Charlson, et al, and is available in the section: supplementary material/table 1/appendix 1 \[see [additional file](#S1){ref-type="supplementary-material"}: CCICalc.xls\]. The calculator was developed using Microsoft Excel/Visual Basic software and can be downloaded from this journal. Simplicity and ease of use were the main design objectives.
Presented as a simple Microsoft Excel tool, it can be easily extended or integrated with other systems that can import Microsoft Excel data, or imported as a flat file. The Calculator functions well with both MS Windows and Macintosh operating systems running any Microsoft Excel version with Macro capabilities and is free to all users of Biomed Central Cancer. There are no restrictions concerning the use of the calculator software. A running CCI score can be calculated by selecting the conditions and age groups within the file. The calculator can be used with or without age modification as proposed by Charlson, et al \[[@B13]\]. It is important to note that the upper limit scores for this calculator are 37 for \"age unadjusted\" and 43 for \"age adjusted.\" Charlson scores \>8--10 have not received extensive evaluation in the comorbidity literature. We intend the calculator to be widely distributed so that use of the CCI can become a routine aspect of clinical research in oncology.
To use the calculator, the user must select \"enable macros\" when prompted to do so as the file opens. To calculate a CCI score, any of the applicable conditions can be selected. All selected conditions will then be displayed in a lighter shade within the table. Corrections can be made by deselecting conditions, which then removes their weighted value from the score. The CCI score can then be totaled, or an age-modified score can be determined by selecting any one of the applicable \"Age by Decade\" groups. Scores totaled without age modification will appear in the \"Age Unadjusted CCI Score\" total and no value will appear in the \"Age Adjusted Score\" total. Scores totaled by selecting an age group without selecting a comorbidity will result in no value for either total and the user will be prompted to \"Reset & Select Condition.\" To reset the program, the \"Reset CCI Calculator\" button can be selected. The calculator can be further modified as needed by changing entries in the \"Data Sheet\" area of the workbook which is hidden in the read-only version of the calculator, but can be unhidden by selecting \"Format,\" then \"Sheet,\" followed by \"Unhide\" from the Excel menu. The \"Data Sheet\" can then be selected and will be viewable. To modify the original Macro, users can contact the authors and the password will be provided on a case-by-case basis.
Results and discussion
======================
The extensive validation of the CCI as a powerful predictor of clinical outcome combined with its simplicity and widespread use in oncology have led to the adoption of the Charlson Index as the prototypical comorbidity index in our department. In addition to validity, our criteria for the use of a comorbidity index focus on simplicity in design, consistency in scoring, and ease of use. It is our contention that many of the commonly used comorbidity indices, such as the ICED, CIRS, and KFI have failed to achieve widespread use because they remain complicated, cumbersome to use, and poorly accessible for use in clinical research. Given the adaptability of the CCI for the inclusion of additional variables, such as age, the CCI also demonstrates marked potential for modification into cancer specific comorbidity indices. We have, therefore, developed a Charlson Comorbidity Calculator based on a Microscoft Excel File to improve the collection of comorbidity data in our department.
Comorbid illness has demonstrated increasing importance as a prognostic factor for survival and treatment-related outcomes in oncology. It confounds the results of clinical trials because the lack of a standardized measurement has resulted in the failure to adjust for comorbidity in statistical analysis of outcomes data \[[@B1],[@B2]\]. It also limits the applicability of clinical research to large segments of the oncology population because protocol designs tend to exclude older and sicker patients \[[@B21],[@B22]\]. Recent reviews consistently identify the CCI, ICED, CRIS and KFI as validated and acceptable measurements of comorbidity and recommend their use in clinical research. Although the ICED, CIRS, and KFI obtain superior prognostic power in some series, the CCI consistently demonstrates statistical validity, particularly in terms of prognostic validity, and remains the most structurally simple, easy to use and well-defined of the comorbidity indices. The ICED and CIRS, for example, both require coding manuals and training courses to be used effectively. The KFI has required extensive modification for use in oncology because it was originally designed to assess comorbidity in diabetic patients. Recent modifications of the KFI for use in oncology, such as those applied by Piccirillo in a head and neck cancer specific modification of the KFI (available in electronic calculator format at <http://oto.wustl.edu/clinepi/calc.html>) also require training courses for effective use \[[@B12]\]. By contrast, the Charlson Index is intuitive, requiring users to select a condition from a defined list, rather than searching for disease value or specific information about disease severity. In our department, the cumbersome requirements for use of the ICED, KFI, and CIRS would reduce compliance with collection of comorbidity data. Furthermore, the increased training requirements and intricacies of these indices may increase variability between scores, as it is unlikely that a single staff member would be responsible for the collection of all data. It is, therefore, our belief that the Charlson Index represents the optimal balance between ease of use and prognostic ability and has, therefore, become the method of choice for the collection of comorbidity data in our department. Accordingly, we developed the CCI calculator to improve compliance with the collection of comorbidity data and as a quality assurance tool to ensure that such data is collected correctly and uniformly.
The use of comorbidity data in clinical research is at an important crossroads, with necessity of its use becoming imperative as electronic capabilities for its assessment become more feasible. As the US population gets older, the use of comorbidity data in clinical trials will only increase in relevance. Current estimates indicate that the elderly will comprise 20% of the population by the year 2030 \[[@B23]\]. Studies of older oncology patients also suggest that the elderly shoulder the majority of cancer burden, with risk rates 11 times greater than those of younger patients, with over 50% of all cancer-related mortality \[[@B24]\]. The rise of comorbidity with increasing age is a theme common to most retrospective studies of comorbidity. In this light, determining the effect of comorbidity on cancer-related survival and treatment-related complications has become increasingly important.
Furthermore, evidence to suggest that comorbidity and performance status represent independent prognostic factors is accumulating. Extermann, et al, for example, examined the relationship between comorbidity and performance status. Both Charlson and CIRS-G were found to have little or no correlation with ECOG performance status, activities of daily living (ADL), or instrumental activities of daily living (IADL). More recently, Repetto, et al, found that among 269 elderly cancer patients with a reported ECOG performance score of \<2, 13% had 2 or more comorbidities, 9.3% had ADL limitations, and 37.7% had IADL limitations. Although a statistical correlation between ECOG performance status, number of comorbidities, and comprehensive geriatric assessment was identified in univariate analysis, only comorbidity, ADL limitation and IADL limitation maintained statistical significance in multivariate analysis. Firat, et al also found CIRS-G and Karnofsy performace status to be independent predictors of outcome in their analysis of prognostic factors in 112 patients enrolled on 4 RTOG trials of stage III lung cancer \[[@B11]\].
Without widespread integration of comorbidity data into clinical research, an increasing number of elderly patients, and their physicians, will be left with treatment recommendations and outcomes data that lack relevance for their age and level of comorbidity.
Concurrently, electronic medical records (EMR) and data collection systems are becoming increasingly common and easy to use, with EMR use among European countries approaching 60% to 90% \[[@B27]\]. The EMR ultimately promises increased physician efficiency and improved clinical outcomes for patients. Contemporary EMR systems have improved outcomes by reducing errors with the use of electronic prescribing systems and improving preventative care with automated reminder systems \[[@B28],[@B29]\]. The MS Excel CCI Calculator provided with this manuscript, for example, could easily be integrated into an EMR for aid in data collection. Such integration would eventually provide an enormous data pool on which to base future research on the prognostic importance of CCI.
To our knowledge, this is the first electronic data collection system offered for the Charlson Comorbidity Index. The simplicity of the index itself, coupled with the simplicity of MS Excel and the Visual Basic programming language, have resulted in a robust electronic CCI calculator that functions well across both Windows and Macintosh platforms. The latest version of the calculator, which is provided with this manuscript, has performed without error consistently on the first (WH), second (RR) and third (SN) authors\' Windows-based PCs.
The major limitations of the CCI calculator lie in the limitations known to comorbidity indices and to the index itself. These include lack of understanding as to the relative importance of various individual conditions on mortality, treatment-related complications and quality of life. Furthermore, failure to include some conditions with particular relevance to cancer patients, such as non-malignant hematopoietic disorders and thromboembolic disorders, as well as uncertainty as to whether a few specific diseases or the overall disease burden is more important for prognosis, remain important considerations limiting use of the CCI \[[@B2],[@B3]\]. Additionally, the CCI has a tendency to underscore comorbidity because it is limited to 19 conditions and because it excludes the primary malignant condition. For example, in a patient with localized prostate cancer, history of COPD and myocardial infarction, the CCI score calculated by a urologist would exclude prostate cancer from the calculation resulting in a score of 2. The same patient might receive a score of 3 by a cardiologist because myocardial infarction, as opposed to prostate cancer, was excluded from the calculation. Another limitation of the CCI lays in the frequent use of grouped CCI scores, or CCI grades, rather than the use of scores as continuous variables. Within an elderly cohort in whom comorbidity is likely to be high, the CCI will have reduced utility if it lacks the ability to distinguish between a score of 2, representing mild to moderate comorbidity, and a score of 8, representing severe comorbidity. With this limitation in mind, we recommend the use of CCI score as a continuous variable.
Despite its limitations, the general oncology literature supports the use of CCI as a prognostic variable in clinical research. It should be emphasized that the CCI is not meant to replace clinical experience and its use in clinical decision-making should be considered investigational. With additional research, CCI methodological limitations can be addressed and the index modified to improve upon its utility. In an effort to improve our understanding of the CCI and identify areas of the index in need of improvement, we are currently investigating the effect of score thresholds on treatment decision-making among prostate cancer experts. We believe that dissemination of the MS Excel CCI Macro will lead to increased use of the CCI for clinical research purposes as well as modification of the CCI to increase its validity and clinical utility. Ultimately, we hope that the comorbidity indices, such as the CCI, will see widespread use in clinical research and eventual integration into EMRs as a result of these efforts.
Conclusions
===========
The Charlson Comorbidity Index has demonstrated excellent predictive validity in numerous cancer-related outcome studies. It has met the criteria for statistical validity as outlined by several authors. In our opinion, the CCI represents the optimal balance between ease of use and prognostic ability. Its simplicity in design also makes its adaptation to include additional variables extremely feasible. We have, therefore, adopted the CCI as an acceptable comorbidity measurement tool in our department and created a Microsoft Excel Macro to facilitate its correct and uniform use in clinical research.
Availability and requirements
=============================
• **Project name**: Charlson Comorbidity Calculator
• **Project home page**: None
• **Operating system(s)**: Windows or Macintosh OS
• **Programming language**: Visual Basic
• **Other requirements**: Microsoft Excel (macro enabled)
• **License**: None
• **Any restrictions to use by non-academics**: None
List of abbreviations
=====================
• **CCI**: Charlson Comorbidity Index
• **ICED**: Index of Co-Existent Disease
• **KFI**: Kaplan-Feinstein Index
• **CIRS**: Cumulative Illness Rating Scale
• **RR**: Relative Risk
• **EMR**: Electronic Medical Record
• **CDSS**: Computer-Based Decision Support Services
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
WH, SN, AJ, and SV carried out the literature review, assembly and editing of the manuscript. RR created the CCI calculator. All authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2407/4/94/prepub>
Supplementary Material
======================
::: {.caption}
###### Additional File 1
A Microsoft Excel (CCICalc.xls) is included with this manuscript and can be found in supplementary material/table 1/appedix 1. A detailed description of the creation of the file and instructions for its use are included in the implementation section of this manuscript. To Calculate a Charlson Comorbidity Index score using the calculator double click on the CCI-Calc.xls icon or open the file from MS Excel. You must select \"enable macros\" when prompted to do so by the MS Excel macro warning pop-up window. A CCI score can then be calculated by selecting the conditions and age groups within the file. Selected conditions will appear in the table as a lighter shade than deselected conditions. As comorbidities are selected a running total of the score will be calculated. Scores totaled without age modification will appear in the \"Age Unadjusted CCI Score\" total and no value will appear in the \"Age Adjusted Score\" total. A selected condition can be deselected by clicking on once on the button for that condition. A score may be calculated without selecting an age category, however Scores totaled by selecting an age group without selecting a comorbidity will result in no value for either total and the user will be prompted to \"Reset & Select Condition.\" Once finished with a calculation, the calculator can be reset by selecting the green \"Reset CCI Calculator\" button. The file is presented in a password protected format so that no changes can be made to the categories and weighting as proposed in the original Charlson Comorbidity Index.
:::
::: {.caption}
######
Click here for file
:::
|
PubMed Central
|
2024-06-05T03:55:51.993925
|
2004-12-20
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545968/",
"journal": "BMC Cancer. 2004 Dec 20; 4:94",
"authors": [
{
"first": "William H",
"last": "Hall"
},
{
"first": "Ramanathan",
"last": "Ramachandran"
},
{
"first": "Samir",
"last": "Narayan"
},
{
"first": "Ashesh B",
"last": "Jani"
},
{
"first": "Srinivasan",
"last": "Vijayakumar"
}
]
}
|
PMC545969
|
Background
==========
The Wnt/β-catenin signal transduction pathway plays a central role in metazoan development, controlling such diverse processes as cell growth, proliferation and organogenesis \[[@B1]\]. Wnt-1 is the prototypic member of this large family of secreted glycoproteins and was originally identified as a gene insertionally activated by mouse mammary tumour virus \[[@B2]\]. Wnt-1 is one of a number of Wnt family members which act to control the cellular level of β-catenin. Wnt proteins bind seven-pass transmembrane receptors of the Frizzled family, and a signal is transduced via Dishevelled to a complex which contains the Adenomatous Polyposis Coli (APC), Axin and Glycogen Synthase Kinase-3β (GSK-3β) proteins \[[@B3],[@B4]\]. This signal antagonizes the phosphorylation of β-catenin by GSK-3β. There are four phosphorylation sites in the N-terminus of β-catenin which, in the absence of Wnt signal, are phosphorylated by Casein Kinase I alpha and GSK-3β \[[@B5],[@B6]\]. This phosphorylation leads to the ubiquitination and subsequent proteasomal degradation of β-catenin \[[@B7]\]. Inhibition of β-catenin phosphorylation by Wnt signalling leads to the accumulation of β-catenin which forms a bipartite complex with members of the TCF/LEF transcription factor family and activates the transcription of target genes, a process which is regulated by multiple interacting factors \[[@B8]\].
Overexpression of Wnt-1 in the mammary glands of transgenic mice leads to extensive hyperplasia and tumorigenesis \[[@B9]\]. *APC*was identified as the tumour suppressor gene mutated in the hereditary colorectal cancer syndrome, Familial Adenomatous Polyposis \[[@B10],[@B11]\]. Mutations in *Axin*and β-*catenin*have also been detected in tumours of the colon and other tissues \[[@B12]\]. Deregulation of this pathway appears to be play a contributory role in a significant proportion of human tumours of epithelial origin and hence, the identification of effector genes of this pathway is an important step towards the elucidation of the mechanisms involved.
Many of the Wnt targets thus far identified are cell-cycle regulators \[[@B13],[@B14]\] and transcription factors \[[@B15]-[@B20]\], and function in a cell-autonomous manner, providing insight into the mechanisms by which tumour cells deregulate proliferation and inhibit apoptosis. Tumours are complex organs composed of tumour cells, stromal fibroblasts, endothelial cells and cells of the immune system; and reciprocal interactions between these cell types in the tumour microenvironment are necessary for tumour growth \[[@B21],[@B22]\]. Here we postulate that proteins secreted by Wnt/β-catenin tumour cells and receptors expressed by these cells may play roles in mediating interactions between neighbouring tumour cells or between tumour cells and their microenvironment. Consequently, in this study we have focussed our attention on identifying novel genes encoding receptors and secreted proteins.
Methods
=======
Cell culture
------------
All reagents were purchased from Sigma unless otherwise noted. HC11 mouse mammary epithelial cells were cultured in 5% CO~2~at 37°C in RPMI 1640, supplemented with 10% Foetal Bovine Serum, 2 mM L-glutamine, 2.5 μg/ml insulin, 5 ng/ml epidermal growth factor and 50 μg/ml gentamycin \[[@B23]\]. HC11-*lacZ*and HC11-Δ*N*β-*catenin*cells were routinely cultured in 2 μg/ml tetracycline to repress transcription of the tetracycline-regulated transgene. HEK293 and MDCK cells were grown in DMEM supplemented with 10% Foetal Bovine Serum.
The HC11-*lacZ*and HC11-Δ*N*β-*catenin*cell lines were generated by infecting the cells with an ecotropic retrovirus (TRE-tTA) in which the *tTA*cDNA is under the control of a tetracycline responsive promoter. Consequently, tTA expression is minimal in the presence of tetracycline and, upon tetracycline withdrawal, tTA activates its own transcription in an autoregulatory manner \[[@B24]\]. HC11 cells expressing tTA were subsequently infected with ecotropic retroviruses derived from RevTRE (Clontech) which directed the expression of either β-galactosidase or ΔNβ-catenin in a tetracycline dependent manner.
Bosc23 cells were used to produce ecotropic retroviruses \[[@B25]\]. Cells were transiently transfected with the appropriate retroviral construct and the supernatant was collected 48 hours post-transfection. Polybrene was added to a final concentration of 5 μg/ml and the supernatant was added to HC11 cells for 24 hours. HC11 cells were then subjected to antibiotic selection using either 250 μg/ml G418 or 200 μg/ml hygromycin B as appropriate.
RNA isolation
-------------
Cell monolayers were washed twice in ice-cold Phosphate Buffered Saline and lysed by addition of Trizol (Invitrogen). Total RNA was isolated according to the manufacturer\'s instructions. PolyA+ RNA was purified from total RNA using Oligotex (Qiagen) according to the manufacturer\'s instructions.
Northern blotting
-----------------
10 μg of total RNA from each cell line was fractionated on a denaturing formaldehyde agarose gel and transferred to a positively charged nylon membrane (Hybond N+, Amersham Pharmacia Biotech) in 10x SSC. Membranes were prehybridised for four hours in 50% (v/v) formamide, 5X SSPE, 2X Denhardt\'s reagent, 0.1% (w/v) SDS and 100 μg/ml denatured herring sperm DNA. Radiolabelled probes were prepared from PCR-amplified cDNA clones using the Rediprime II kit (Amersham Pharmacia Biotech) according to the manufacturer\'s instructions. EST sequences corresponding to the coding sequence of the genes-of-interest were identified by BLAST \[[@B26]\] and obtained from the I.M.A.G.E. consortium through the UK Human Genome Mapping Project Resource Centre (Hinxton, UK). ESTs bearing the following I.M.A.G.E. cloneIDs were used: *Autotaxin*-- 533819; *CD14*-- 2936787; *Ecm1*717050; *Hig2*-- 367488; *Ramp3*-- 615797, *HIG2*4366895). Following overnight hybridisation with the labelled probe, the membranes were washed twice in 1X SSC, 0.1% (w/v) SDS at room temperature for 20 mins, and twice in 0.2X SSC, 0.1% (w/v) SDS at 68°C for 10 mins and exposed to film at -80°C for 48 hours. Bound probe was quantitated using a phosphorimager (Molecular Dynamics).
Western blotting
----------------
Cell monolayers were rinsed twice with ice-cold Phosphate Buffered Saline and total cell lysates were prepared by scraping cells into a minimal volume of 50 mM Tris. HCl pH 7.5, 150 mM NaCl, 0.5% NP40 and Complete protease inhibitor cocktail (Roche). Aliquots containing 80 μg protein from each sample were analysed by SDS-PAGE \[[@B27]\], and transferred electrophoretically to a PVDF membrane. Mouse monoclonal antibodies were used to detect tTA (Clontech), β-catenin (Transduction Laboratories) and EGFP (Santa Cruz Biotechnology). Samples of conditioned medium were concentrated 12-fold using Microcon YM-10 centrifugal filter units (Millipore) prior to analysis.
Construction of plasmids
------------------------
A BgIII fragment containing the *lacZ*cDNA was excised from the CMV-*lacZ*construct (a gift of Trevor Dale) and sub-cloned into BamHI digested RevTRE to make RevTRE-*lacZ*. A plasmid containing a myc-tagged ΔNβ-catenin was obtained from Hans Clevers. The myc-tagged ΔNβ-catenin was excised with KpnI and NotI and the ends were blunted, and subcloned into HpaI digested RevTRE to make RevTRE-Δ*N*β-*catenin*. The mouse *Hig2*open reading frame was amplified by PCR from I.M.A.G.E. cDNA clone 367488 using the primers TTTACTAGTAGGAGCTGGGCACCGTCGCC and TTTTACCGGTGCCTGCACTCCTCGGGATGGATGG. The PCR product was digested with AgeI and SpeI and subcloned into the AgeI and NheI sites in pEGFP-C1 (Clontech) to make the Hig2-EGFP fusion gene. Site directed mutagenesis was carried out by the method of Sawano and Miyawaki (2000) \[[@B28]\]. The primer TGCTGAACCTCGAGGAGCTGGGCATCATG was used to make the Hig2-EGFP(Y8V9/D8D9) mutant.
Transient transfections
-----------------------
Transient transfections were performed using Lipofectamine (Invitrogen) according to the manufacturer\'s instructions. Briefly, 1.5 × 10^5^cells were plated in 3.5 cm wells on the day prior to transfection. Each well was transfected with a total of 0.9 μg DNA under serum-free conditions for six hours, after which the cells were washed and incubated for a further 48 hours before assaying expression.
β-galactosidase activity assay
------------------------------
For the tetracycline dose response curve, 5000 HC11-*lacZ*cells for each condition, were cultured in triplicate in 96 well plates for 72 hours, and beta-galactosidase activity was determined as previously described \[[@B24]\].
Results
=======
Generation of HC11-*lacZ*and HC11-Δ*N*β-*catenin*cell lines
-----------------------------------------------------------
Stable cell lines were generated in which either ΔNβ-catenin or β-galactosidase was expressed in a tetracycline dependent manner. These cell lines were established using a novel autoregulatory system in which the expression level of the tetracycline transactivator (tTA) protein is minimised during routine culture and is induced upon withdrawal of tetracycline with concomitant upregulation of the transgene-of-interest \[[@B24]\]. This strategy helps to minimise deleterious effects due to tTA toxicity.
A dose-response analysis for the HC11-*lacZ*cell line is shown in Figure [1A](#F1){ref-type="fig"}. β-galactosidase expression is effectively repressed at tetracycline concentrations in excess of 20 ng/ml and is strongly induced in the absence of tetracycline. The N-terminal truncation mutant of β-catenin can be detected by western blotting by both its myc-epitope tag and an anti-β-catenin antibody (Figure [1B](#F1){ref-type="fig"}). tTA expression is detectable only in the absence of tetracycline demonstrating the autoregulatory nature of this system.
Microarray analysis
-------------------
Transgene expression was induced in HC11-*lacZ*and HC11-Δ*N*β-*catenin*cells by withdrawal of tetracycline for 72 hours. Total RNA was isolated, from which mRNA was purified. cDNAs were labelled and hybridized to an 8962 element Incyte mouse GEM1 cDNA microarray (Incyte Genomics, Palo Alto, CA). These data are provided as supplementary material (See [Additional file 1](#S1){ref-type="supplementary-material"}). Among those genes upregulated were two genes shown by other workers to be transcriptional targets of this pathway -- *Fibronectin*\[[@B29]\] and *Autotaxin*\[[@B30]\] (data not shown) -- suggesting that our model of Wnt/β-catenin signalling deregulation results in the activation of a set of target genes which overlaps, at least partially, with pathway targets in other cell lines. The microarray experiment described here was performed only once but differential expression was repeatedly validated by northern blotting from independent samples for the genes discussed here.
Validation of targets
---------------------
Five genes were selected for further study -- *Extracellular Matrix Protein 1*(*Ecm1*), *Autotaxin*, *Receptor Activity Modifying Protein 3*(*Ramp3*), *Cd14*and *Hypoxia Inducible Gene 2*(*Hig2*). Each putative target gene was initially subjected to a secondary screen by Northern blotting to confirm the differential expression in response to ΔNβ-catenin (Figure [2](#F2){ref-type="fig"}). RNA samples used for Northern blotting were from independent induction experiments to those used for microarray analysis, thus demonstrating repeatedly by two distinct methods that the transcript levels of these genes are altered in cells overexpressing ΔNβ-catenin. The expression level of each of the transcripts was quantitated using a phosphorimager and normalised to the expression of *Gapdh*mRNA in the samples. The data in Figure [2](#F2){ref-type="fig"} represent film exposure times ranging between 24 and 72 hours. Quantitations were performed using short (one hour or less) exposures to a phosphorimager screen, such that the signal intensity was not saturating.
Molecular cloning of mouse *Hig2*
---------------------------------
*Hypoxia Inducible Gene 2*encodes a 63 amino acid polypeptide and was one of several genes identified in a screen for genes regulated by hypoxia in a human cervical epithelial cell line \[[@B31]\]. HIG2 shares no sequence similarity with other known proteins. In order to facilitate the functional analysis of this gene, ESTs were identified which encoded mouse and rat Hig2, and the sequences of chimpanzee and baboon were inferred from genomic sequence data. A multiple alignment of the inferred amino acid sequences shows that these polypeptides are highly similar (Fig [3A](#F3){ref-type="fig"}). Analysis of these sequences using a Kyte-Doolittle hydrophobicity plot showed that the N-termini of these proteins contain a series of hydrophobic amino acids (Fig [3B](#F3){ref-type="fig"}). This region of hydrophobicity was reminiscent of a signal peptide and sequence analysis using the signal peptide prediction program, SignalP \[[@B32],[@B33]\] supported this possibility.
Hig2 has an N-terminal signal peptide and is secreted
-----------------------------------------------------
To investigate the subcellular localisation of Hig2, a *Hig2*-*EGFP*fusion gene was constructed and expressed in both HC11 and Madin-Darby Canine Kidney (MDCK) cells by transient transfection (Figure [4a](#F4){ref-type="fig"} and [4c](#F4){ref-type="fig"}). In both cell lines, Hig2-EGFP is localised to large round vesicle-like structures in the cytoplasm. Similar observations were made in HEK-293 cells (data not shown). The fluorescence was detected predominantly around the periphery of these structures suggesting that they do not consist of solid masses of aggregated protein. When two aspartate residues were introduced to the putative signal peptide by site-directed mutagenesis, Hig2(Y8V9/D8D9)EGFP, this distinctive subcellular localization was abolished (Fig [4B](#F4){ref-type="fig"} and [4D](#F4){ref-type="fig"}). These large structures did not colocalize with either markers of mitochondria (pDsRed2-mito, Clontech), nor lipid droplets (Nile Red, Molecular Probes) nor with markers of endosomes or lysosomes (pulse-chase analysis with TRITC-dextran); data not shown. However, in live HC11 cells transfected with Hig2-EGFP, observations at high magnification revealed that the cytoplasm of these cells contained many very small solid green vesicles moving along the cytoskeleton. These vesicles were approx 1/100 the size of the large vesicles shown in Fig [4A](#F4){ref-type="fig"} and [4C](#F4){ref-type="fig"}, and were not observed in cells transfected with either Hig2(Y8V9/D8D9)EGFP or EGFP alone. The rapidity of this motion in live cells, even at room temperature, precluded capture of these images but suggested the possibility that measurable amounts of secreted Hig2-EGFP might be found in the culture medium. HEK-293 cells were chosen as they could be transfected at high efficiency (approx 80%), the presence of the green transport vesicles was confirmed, and 48 hours after transfection samples of total cell lysate and conditioned medium were analysed by western blotting. Secreted Hig2-EGFP was detected in the conditioned medium of Hig2-EGFP cells, but not Hig2(Y8V9/D8D9)EGFP cells (Fig [5](#F5){ref-type="fig"}). Multiple bands were detected in cell lysates for both Hig2-EGFP fusion proteins: whether these represent artifactual degradation products or physiologically relevant biological entities is as yet unknown. Such multiple banding has also been observed with other EGFP-fusion proteins targeted to the secretory pathway (Amphiregulin-EGFP, PK unpublished observations). At least one of the bands may result from internal translation initiation at the consensus Kozak initiation sequence of pEGFP-C1 which is located between the Hig2 and EGFP open reading frames.
EGFP was also detected in the conditioned medium. This is consistent with previous reports of GFP secretion via a non-classical Brefeldin A-insensitive pathway \[[@B34]\]. In this study, several cell lines are described (including HEK293) in which wild-type GFP is released from the cell without passing through the golgi apparatus. Thus, it is formally possible that, instead of its secretion being directed by the putative signal peptide, HIG2-EGFP might be released from the cell via this pathway in a manner specifically dependent on the EGFP moiety. The presence of post-translational modifications acquired during endoplasmic reticulum/golgi apparatus mediated secretion would exclude the latter hypothesis. The altered mobility of HIG-2-EGFP in the medium suggested that it might be glycosylated, however the mobility was not changed by treatment with the glycosidase PNGaseF, suggesting that this secreted protein is not glycosylated (data not shown). Previous studies using GFP fused to a signal peptide directing entry into the ER demonstrated that, in this redox environment, the cysteine residues of GFP form intermolecular disulphide bridges which result in oligomerization of GFP molecules \[[@B35]\]. Oligomers of Hig2-EGFP were detected (Fig [5](#F5){ref-type="fig"}, black arrowheads) but no oligomerization of EGFP was observed. Hig2 itself does not contain cysteine residues, thus the oligomerization is mediated by the EGFP domains. These data are consistent with HIG2-EGFP entry into the classical secretory pathway. Collectively, these data demonstrate that Hig2 contains a functional N-terminal signal peptide and is likely a secreted protein.
Expression of *HIG2*in human tumours
------------------------------------
To investigate the relevance of *HIG2*in human tumours, the expression level of this gene was examined in 68 tumour cDNA samples compared to normal adjacent tissue from the same patients using a Matched Tumour/Normal cDNA blot (Clontech) (Figure [6A](#F6){ref-type="fig"}). The levels of *HIG2*were approximately similar in most of the tumour types examined but were strongly and consistently downregulated in most of the cases of kidney and stomach tumours analysed. These data suggest that the downregulation of *HIG2*observed upon deregulated β-catenin signalling *in vitro*may be of clinical relevance in human tumours.
Discussion
==========
cDNA microarray analysis of the transcriptional changes resulting from overexpression of a constitutively active β-catenin revealed a panel of putative target genes of the Wnt/β-catenin pathway in mouse mammary epithelial cells. This differential expression was confirmed by Northern blotting in five cases.
Autotaxin was originally identified as a secreted enzyme with potent motility stimulating activity \[[@B36]\] and has both pyrophosphatase and phosphodiesterase activity \[[@B37]\]. Transplantation experiments in athymic mice showed that ras-transformed NIH-3T3 fibroblasts became significantly more tumorigenic, invasive and metastatic when transfected with *Autotaxin*\[[@B38]\], and purified recombinant Autotaxin has potent angiogenic activity *in vivo*\[[@B39]\]. Autotaxin has been shown to be regulated by both Wnt-1 and retinoic acid \[[@B30]\]. Autotaxin has been shown to have lysophospholipase activity and the effects of Autotaxin on tumour cell motility are mediated by its conversion of lysophosphatidylcholine to lysophosphatidic acid (LPA), a potent signalling molecule \[[@B40],[@B41]\].
Extracellular Matrix Protein 1 was first identified as a novel 85 KDa protein secreted by a mouse osteogenic stromal cell line \[[@B42]\]. *In situ*hybridisation showed that *Ecm1*was strongly expressed in most newly formed blood vessels and experiments using purified recombinant Ecm1 showed that it could increase the proliferation rate of vascular endothelial cells *in vitro*and also stimulate angiogenesis *in vivo*. The ability to induce *de novo*angiogenesis is an absolute requirement for tumours to grow beyond a size which can be readily perfused by oxygen and nutrients from the interstitial fluid. ECM1 is overexpressed in many epithelial tumours including 73% of breast tumours analyzed \[[@B43]\]. Homozygous loss-of-function mutations in the human ECM1 gene were recently identified by linkage analysis as the causative mutations behind Lipoid Proteinosis, a rare autosomal recessive disorder characterized by hyaline deposition in the skin, mucosae and viscera \[[@B44]\]. The identification of *Autotaxin*and *Ecm1*as genes upregulated by activation of this pathway, together with VEGF \[[@B45]\] suggests that deregulation of Wnt/β-catenin signalling during tumour initiation and progression may be one of the factors which promotes tumour angiogenesis.
*CD14*, which can function as both a receptor and a secreted protein, was downregulated upon ΔNβ-catenin expression. CD14 is a glycosyl-phosphatidylinositol-linked cell surface protein, preferentially expressed in monocytes, where it acts as a receptor for Lipopolysaccharide Binding Protein:Lipopolysaccharide complexes \[[@B46]\]. Soluble CD14 (sCD14) is also expressed in mammary epithelial cells *in vitro*and has been detected in human milk where it is postulated to play a role in neonatal immunity \[[@B47]\], and is strongly upregulated in mammary luminal epithelial cells *in vivo*at the onset of involution \[[@B48]\].
*Receptor Activity Modifying Protein 3*(*RAMP3*) was downregulated upon ΔNβ-catenin induction and is one of three members of the RAMP family. These proteins are involved in mediating the cellular response to the neuropeptides calcitonin, calcitonin gene related peptide, amylin and adrenomedullin. The RAMP family members function as chaperones for the seven transmembrane domain G-protein coupled receptors for these neuropeptides, shuttling the receptor to the cell surface and altering receptor glycosylation. The ligand binding phenotype of the receptor is dependent on the RAMP family member with which it is associated \[[@B49]\]. RAMP3-Calcitonin Receptor (CR) heterodimers form a functional receptor for amylin \[[@B50]\], and RAMP3-Calcitonin-Receptor-Like-Receptor (CRLR) heterodimers act as an adrenomedullin receptor \[[@B51]\].
Expression of both CR and CRLR was detected in HC11 by RT-PCR (data not shown) suggesting that functional receptor-RAMP complexes are present in this cell line. Adrenomedullin, the ligand for the CRLR/RAMP3 receptor dimer, functions as a growth factor in several human tumour cell lines \[[@B52]\], in addition to promoting angiogenesis *in vivo*\[[@B53]\] via CRLR/RAMP3 and CRLR/RAMP2 receptor dimers \[[@B54],[@B55]\].
*Hypoxia*-*inducible gene 2*(*Hig2*) was one of several genes identified in a representational difference analysis screen for genes regulated by hypoxia in a human cervical epithelial cell line. The human gene encodes a 63 amino acid polypeptide of unknown function \[[@B31]\]. Expression of mouse *Hig2*was downregulated in HC11 cells overexpressing ΔNβ-catenin. The identification of a group of mammalian orthologues revealed a well conserved hydrophobic region in the N-terminus, reminiscent of a signal peptide. A Hig2-EGFP fusion protein entered the secretory pathway and was detected in conditioned medium of transfected cells. The introduction of a pair of charged amino acids into the hydrophobic region abolished secretion, lending support to the hypothesis that this region contains a functional signal peptide. The nature of the large vesicular structures observed in Hig2-EGFP overexpressing cells is as yet unclear. Mammary epithelial cells are known to contain membrane-enclosed lipid droplets, as well as a variety of vesicular compartments involved in the secretion of casein, citrate, lactose and calcium \[[@B56]\], however the presence of these vesicles in MDCK and HEK293 cells argues that they are not mammary specific. Indeed, co-localization experiments suggest that these structures are neither mitochondria, lysosomes, endosomes nor lipid droplets. Given the demonstration that Hig2 is secreted, these structure most likely correspond to overexpressed Hig2-EGFP in transit through the endoplasmic reticulum and golgi apparatus. As no antibody is available against Hig2, it was not possible to investigate the localisation of the endogenous protein, but these data represent a useful initial step in the functional characterisation of this gene.
Analysis of the expression of *HIG2*using a matched Tumour/Normal tissue cDNA array showed that *HIG2*is widely expressed. In most cases, the levels of *HIG2*in the tumours and the associated normal tissue controls were similar. *HIG2*was, however, strongly and consistently downregulated in the majority of the kidney and stomach tumours analysed. This represents an significant validation of our *in vitro*findings in human tumours, and suggests that HIG2 may exert a tumour suppressive effect *in vivo*. Human HIG2 is located on 7q32.2, a commonly deleted region in several tumour types, most prominently leukaemias and lymphomas \[[@B57]\]. Deletion analysis of 7q in a panel of patients with Splenic Lymphoma with Villous Lymphocytes by Catovsky and colleagues suggests that a critical tumour suppressor is located on 7q32 \[[@B58]\].
Conclusions
===========
The identification of this panel of candidate target genes for this clinically important signal transduction pathway adds to those identified by other workers in a variety of model systems and suggests that, as well as promoting tumour cell proliferation and survival in a cell autonomous manner, this activation of this pathway is likely to have a series of non-cell autonomous effects. Here we have focussed on the identification of Wnt/β-catenin target genes that are either secreted signalling molecules or receptors. It is likely that such targets are involved in mediating autocrine proliferation, promotion of angiogenesis and the mediation of reciprocal communication between Wnt/β-catenin tumours and their microenvironmental milieu.
Competing interests
===================
The authors declare that they have no competing interests.
Authors\' contributions
=======================
PK carried out all of the experimental procedures and drafted the manuscript. PK, TE and AA contributed to the design of the study. All authors read and approved the final version of this manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2407/5/3/prepub>
Supplementary Material
======================
::: {.caption}
###### Additional File 1
cDNA microarray dataset: HC11 ΔNβ-catenin v. HC11 lacZ. This file contains the dataset from the Incyte GEM1 cDNA microarray comparison between HC11 cells overexpressing ΔNβ-catenin (Probe 1) and β-galactosidase (Probe 2). This file may be easily imported into MS Excel, Genespring or other microarray analysis software to facilitate further analysis.
:::
::: {.caption}
######
Click here for file
:::
Acknowledgements
================
The ΔN-β-catenin and β-galactosidase expression vectors, the Bosc23 cell line and the GAPDH probe were gifts of Drs. Hans Clevers, Trevor Dale, Maria Emanuela Cuomo and John Brown respectively. We thank Dr. Derek Radisky for a critical reading of the manuscript. This study was supported, in part, by a Ph.D. studentship from the Institute of Cancer Research.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Generation of stable HC11 cell lines expressing β-galactosidase and Δ Nβ-catenin in a tetracycline dependent manner. **(a)**HC11-*lacZ*cells -- β-galactosidase activity of cell lysates following incubation for three days at the indicated tetracycline concentrations. β-galactosidase expression is effectively repressed at 20 ng/ml tetracycline. The assay was carried out in triplicate and results are presented as the mean β-galactosidase activity (normalised to the protein concentration of the samples) ± standard error. **(b)**Western blot of total cell lysates from HC11-Δ*N*β-*catenin*cells cultured ± tetracycline. The N-terminal deletion mutant of β-catenin is detected by both its myc-epitope tag and the anti-β-catenin antibody. tTA was detected only in the absence of tetracycline, confirming the autoregulatory nature of the induction system.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Confirmation of differential expression of putative target genes in response to ΔNβ-catenin overexpression. The expression level of each gene in the HC11-*lacZ*(L) and HC11-Δ*N*β-*catenin*(ΔN) cell lines was assessed by Northern blotting. The fold-induction was quantitated using a phosphorimager under non-saturating conditions, normalised to the expression of GAPDH mRNA in the samples and is expressed as fold change relative to the expression level of that gene in the HC11-*lacZ*cell line.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Sequence analysis of five mammalian Hig-2 proteins. **(a)**Multiple sequence alignment of Hig2 amino acid sequences of mouse, rat, human, chimpanzee and olive baboon shows that these proteins share strong sequence similarity. **(b)**A representative Kyte-Doolittle hydrophobicity plot using the mouse Hig2 amino acid sequence indicates these proteins have hydrophobic N-termini, reminiscent of a signal peptide.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Subcellular localisation of a Hig2-EGFP fusion protein and mutational analysis of the putative signal peptide in HC11 and MDCK cells. **(a**and **c)**Hig2-EGFP is localised to large spherical vesicles located in the cytoplasm. **(b**and **d)**The introduction of two aspartate residues into the putative hydrophobic signal peptide abolishes the vesicular subcellular localisation of the mutant protein.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Hig2-EGFP fusion protein is secreted from transfected cells. Western blot analysis of total cell lysates and conditioned medium from (1) untransfected HEK293 cells and cells transfected with plasmids encoding (2) EGFP, (3) Hig2-EGFP, (4) Hig2(Y8V9/D8D9)EGFP. Oligomers of HIG2-EGFP formed by intermolecular disulphide bridges are indicated by arrowheads.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
**(a)**Analysis of *HIG2*mRNA expression on a human cDNA array containing matched tumour and normal samples from 68 patients. For each pair, the tumour sample is on top. The bottom row of the array contains cDNA from cell lines (left to right: HeLa, Daudi, K562, HL-60, G361, A549, MOLT-4, SW480 and Raji). **(b)**Quantitative analysis by phosphorimager. Vertical dashed lines indicate ± 2-fold difference. *HIG2*expression is broadly similar in most pairs, but is strongly and consistently downregulated in most of the kidney and stomach tumours analysed.
:::
:::
|
PubMed Central
|
2024-06-05T03:55:51.996954
|
2005-1-10
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545969/",
"journal": "BMC Cancer. 2005 Jan 10; 5:3",
"authors": [
{
"first": "Paraic A",
"last": "Kenny"
},
{
"first": "Tariq",
"last": "Enver"
},
{
"first": "Alan",
"last": "Ashworth"
}
]
}
|
PMC545970
|
Background
==========
Skin ulcers caused by pressure and strains are known by various names: decubitus ulcer, bedsore, ischemic ulcer and pressure ulcer. \"Pressure ulcer\", which indicates the etiology of the lesion, seems the most appropriate name \[[@B1]\]. An estimated 50--80% of individuals suffering from spinal cord injury develop pressure ulcers at least once in their lifetime. Most of these ulcers occur during the first two years after injury, but even after 3--4 years an incidence of 30% has been reported \[[@B2]-[@B4]\].
Although the major challenge is to prevent the occurrence of ulcers \[[@B5],[@B6]\], therapeutic measures merit due attention. Pressure ulcer therapy is among the expensive of medical and surgical interventions \[[@B5]-[@B7]\]. In one study in the United Kingdom, data relating to chronic wound management practice obtained from 15 pressure sore studies showed a cost range of 422--2548 pounds per healed wound for primary dressing, nursing time, wound cleansing and debridements \[[@B8]\]. These figures do not include the much higher costs of hospitalization and plastic surgery. We have tried to find a more effective and cost-efficient method of treatment.
Different methods have been used for preventing and treating pressure ulcers. These include various training programs for patients \[[@B4],[@B9],[@B10]\]; physiotherapy methods employing ultrasound, ultraviolet irradiation and laser treatment \[[@B7]\]; good nutrition emphasizing high protein, high calorie diet and more liquid; electrical stimulation; and application of local ointments and creams such as bacitracin, silver sulfadiazine, neomycin, polymixin, phenytoin and hydrocolloid dressings \[[@B11]-[@B19]\].
The results of the studies conducted so far are incompatible, even contradictory. Most of them considered too few patients and/or lacked a control group. In Iran, 5000 patients suffer from spinal cord injury (SCI): of these, 2000 are lran-lraq war victims and 3000 were handicapped by other causes. In view of the enormous prevalence of pressure ulcers in war victims and other spinal handicap patients, and the importance of these lesions in terms of morbidity, mortality and cost of treatment, we have compared the efficacies of applying hydrocolloid dressing, phenytoin cream and a simple dressing. The aims were to determine: 1. which is the most effective in terms of complete ulcer healing; 2. whether healing rates differ with respect to the ulcer stage (I and II) or location (gluteal, ischial, sacral) using these three different methods.
Methods
=======
The study was a randomized single blind clinical trial involving 2015 Iranian spinal cord injury (SCI) victims of the Iran-Iraq war (1980--1988). The SCI victims were accessed through the mediation and assistance of the Jaonbazan Medical and Engineering Research Center (JMERC) <http://www.jmerc.ac.ir>, the medical and research section of the official governmental body responsible for SCI war victims. The study proposal was reviewed, approved and granted by JMERC.
The medical records of all 2015 subjects were studied to identify cases with pressure ulcers. Where the data were unknown or unreliable, the patients were visited at home or in victims\' long term care centers. Finally, 165 pressure ulcers in 151 patients were identified. All relevant data including patient age and weight, the longevity of the ulcer before our intervention, and the size, stage and location of the ulcer, were collected by the general practitioners in the team. Next, all the patients were examined by one of the authors to confirm their eligibility for the study. The eligibility criteria were: A) Inclusion Criteria: 1. Paraplegia caused by spinal cord injury; 2. Pressure ulcer stage I and II according to Shea classification \[[@B20]\] or National Pressure Ulcer Advisory Panel \[[@B21]\] (Fig. [1](#F1){ref-type="fig"}); 3. Patient\'s informed consent; 4. Smoothness of ulcer area to establish whether adhesive could be used at the site. Exclusion criteria: 1. Addiction; 2. Heavy smoking (more than 20 cigarettes a day or more than 10 packs per year; 3. Concomitant chronic disease (e.g. diabetes mellitus or frank vascular disease such as Buerger\'s disease).
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Two pressure ulcer classifications.
:::
:::
Seventy-four ulcers in 68 patients were excluded because they did not meet these eligibility criteria: 31 ulcers (28 patients) were stage III or higher; 27 ulcers (25 patients) were excluded because of patient\'s smoking/addiction; 5 ulcers (5 patients) had uneven surfaces; 4 ulcers (4 patients) were excluded because of systemic diseases; and 6 patients with 7 ulcers refused to participate (Fig. [2](#F2){ref-type="fig"}). Thus, the study sample comprised 83 patients with 91 pressure ulcers in the ischial, sacral or gluteal areas. These 91 ulcers were allocated to three different groups (30 ulcers each) by stratified randomization. Three therapeutic methods were applied as follow: simple dressing (SD), hydrocolloid dressing (HD), and adhesive and phenytoin cream (PC). Two general practitioners and nine nurses trained in treatment interventions administered the protocols.
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Flow diagram of participants through each stage of the study.
:::
:::
The SD patients were visited twice a day, the PC patients once a day and the HD patients twice a week. All participants were visited and examined in their family homes or nursing homes by general practitioners every two weeks to ensure that the treatments were being properly applied and were consistent among the three groups. There were no differences in the facilities available for patients in family homes versus nursing homes, and all the patients had free access to victims\' long term care centers.
In the SD group, the following steps were taken twice a day. The ulcer was cleaned and washed 3 times with normal saline, then dried with a sterile gauze and, depending on the size of ulcer, covered with wet saline gauze dressing. In the PC group, daily dressing and cleaning of ulcer were similar to the SD group, except that a thin layer of phenytoin cream was applied to the ulcer before the dressing was performed. In the HD group, after the ulcer had been cleaned in a similar manner to the SD group, the hydrocolloid adhesive dressing was applied to the ulcer area. The adhesive dressings were changed twice a week. Any necrotic tissue was debrided before treatment; all debridements preceded ulcer tracing and assignment of the participants to the trial groups. No debridement was allowed after treatment had started. No concomitant topical or systemic antibiotic, glucocorticoid or immunosuppressive agent was allowed during the treatment period. Fortunately, none of our patients needed debridement or the aforementioned concomitant therapies during the study period. There were no differences among the trial groups with respect to other concomitant care measures.
Every two weeks a questionnaire regarding the ulcer\'s status was completed by the general practitioners; and at the end of 8 weeks, the ulcers\' conditions were examined blind by one author and assessed as \"Complete Healing\", \"Partial Healing\", \"Without Improvement\" or \"Worsening\". To measure each ulcer\'s surface area, the ulcer borders were traced on to a paper overlay. This primary schematic representation was then scanned, redrawn and measured by AutoCAD 2000 software.
The primary outcome was whether or not the ulcer was completely healed within 8 weeks. \"Complete ulcer healing\" was defined as: A) For stage I ulcer, intact epidermis, no red area; B) For stage II ulcers, intact dermis and epidermis, no abrasion or ulceration. Other definitions were as follows. \"Partial healing\" = any decrease in ulcer size compared to the baseline ulcer tracing, excluding complete healing. \"Without improvement\" = no change in ulcer size compared to the baseline ulcer tracing. \"Worsening\" = any increase in ulcer size compared to the baseline ulcer tracing. The difference in responses between patients receiving HD and patients receiving the other therapies (PC or SD) were determined \[[@B22]\].
Before the study, we assumed response rates of 30%, 40% and 80% for SD, PC and HD, respectively. Thus, based on the 40% difference, power of 0.85, 95% confidence level and estimated follow-up loss of 10%, 29 patients were required for each study group. The number of ulcers that met the eligibility criteria totaled 91 and all were enrolled in the study.
A random-number table was used to generate the random allocation sequence, and stratified randomization was used to achieve balance between the treatment groups and subgroups (ulcer stages and locations). If a patient had more than one ulcer, all the ulcers were treated by the same method to eliminate the possible complicating factor of treatment interactions. The statistician in the team generated the random allocation sequence. He was informed of the patient list (numbers only) and the ulcer stage and location of each patient.
The treatment category for each patient was determined by the statistician and was delivered in an opaque sealed envelope bearing only the number of the patient. These sealed envelopes were delivered to the general practitioners, along with the list of patients\' numbers and names. After each patient was visited, the appropriately numbered envelope was opened by the general practitioner to determine whether the SD, PC or HD method would be used, then the appropriate intervention commenced.
The authors were blind to the patients\' assignment to trial groups. The general practitioners were also blind to the treatment of each patient up to the start of the study, when they opened the sealed envelopes. After intervention began, both the general practitioners and the nurses knew the trial groups, because significant differences among the three treatment methods precluded blinding. The patients were also aware of the treatment methods, although they initially had equal chances of entering any of the trial groups. Thus, the study was single-blinded and the author who enrolled the patients to the study was blind to treatment assignment. The author who finally assessed the outcomes was also blind to the trial group of each patient.
To maintain the blinded status on assessment of outcomes, the assessor examined the patients, after the ulcer dressings had been removed by the general practitioners, with no knowledge of the trial groups to which they had been assigned. The gross appearance of the ulcers without dressing, whether healed or not, did not indicate the trial group. The assessor was asked during the 8-week outcome assessment to try to identify which treatment has been administered to each patient. Overall, 27.7% of his guesses were correct (25% in the HD group, 32.1% in the PC group and 25.9% in the SD group), so they were were no better than chance; i.e. there were no significant differences among the three trial groups with respect to proportions guessed correctly (P \> 0.2 in all cases).
The study proposal was designed in November 2001, and the recruitment of patients began in March 2002 and lasted about 2 months. Then the patients were allocated to the treatment groups and followed-up for another 2 months. Finally, all the collected data were analyzed within 2 months. Thus the study from proposal to final analysis took about 10 months (November 2001-September 2002). At the end of the study, all the data collected from the patients\' preliminary and complementary questionnaires were analyzed by SPSS software using ANOVA and Chi square tests, and P-values of \<0.05 were assumed significant. The 95% confidence intervals were also calculated and reported \[[@B23]\]. For rare events (more than 20 percent of cross tabulation cells had values less than 5), Fisher\'s exact test was used. Based on stage and location of ulcers, subgroup analyses were performed using the same statistical tests.
Results
=======
Ninety-one ulcers in 83 male patients were treated by one of three methods. The mean age and weight of the patients were 36.64 ± 6.04 years and 61.12 ± 5.08 kg, respectively. Of the 91 ulcers, 33 were stage I and the remaining 58 were stage II. There were no significant differences among the three therapeutic groups in baseline demographic characteristics (table [1](#T1){ref-type="table"}) or in ulcer location (sacral, gluteal, ischial) or stage (I or II) (Fig. [3](#F3){ref-type="fig"} and [4](#F4){ref-type="fig"}).
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Baseline characteristics of study subjects assigned to hydrocolloid, phenytoin and simple dressing groups
:::
**Variables** **Mean Age Of patients (yr ± SD)** **Mean weight Of patients (kg ± SD)** **Mean duration of ulcer before treatment (wk ± SD)** **Mean ulcer size (cm^2^± SD)** **Stage of ulcer (no)**
------------------------------------------------ ------------------------------------ --------------------------------------- ------------------------------------------------------- --------------------------------- ------------------------- ----
**Total n = 83 patients 91 ulcers** 36.64 ± 6.04 61.12 ± 5.08 6.25 ± 6.56 7.54 ± 12.99 33 58
**Hydrocolloid n = 28 patients 31 ulcers** 36.81 ± 6.71 62.26 ± 5.44 7.63 ± 5.59 7.26 ± 15.4 13 18
**Phenytoin n = 28 patients 30 ulcers** 36.5 ± 4.99 60.07 ± 4.39 5.84 ± 8.04 5.12 ± 3.63 9 21
**Simple dressing n = 27 patients 30 ulcers** 36.6 ± 6.17 61 ± 5.03 5.25 ± 5.39 10.27 ± 15.32 11 19
**P-Value of comparing variables of 3 groups** P \> 0.10 P \> 0.10 P \> 0.10 P \> 0.10 P \> 0.62
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Ulcer distribution according to treatment group and location.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Ulcer distribution according to treatment group and stage.
:::
:::
The numbers of ulcers and the degree of improvement in the three therapeutic groups are shown in table [2](#T2){ref-type="table"}. The completion of healing, regardless of location and stage, was better in the HD than in the PC \[23/31(74.19%) vs 12/30(40%); difference 34.19%, 95% CI = 10.85--57.52, (P \< 0.01)\] or the SD \[23/31(74.19%) vs 8/30(26.66%); difference 47.53%, 95% CI = 25.45--69.61, (P \< 0.005)\] groups. Completion of healing of stage I ulcers in the HD group \[11/13(85%)\] was also better than in the SD \[5/11(45%); difference 40%, 95% CI = 4.7--75.22, (P \< 0.05)\] or PC \[2/9 (22%); difference 63%, 95% CI = 29.69--96.3, (P \< 0.005)\] groups. Completion of healing of stage II ulcers was better in the HD group \[12/18(67%)\] than in the SD group \[3/19(16%); difference 51%, 95% CI = 23.73--78.26, (P \< 0.005)\], but there was no significant difference from the PC group \[10/21 (48%); difference 19%, 95 CI = -11.47--49.47, (P \> 0.05)\].
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Healing status of pressure ulcers in 3 treatment groups (hydrocolloid, phenytoin and simple dressing)
:::
**healing status** **Complete** **Partial** **Not improved** **Worsened** **Total**
---------------------------- -------------- ------------- ------------------ -------------- -----------
**Hydrocolloid n = 31** 23 (74.19%) 4 (12.58%) 2 (6.45%) 2 (6.45%) 31 (100%)
**Phenytoin n = 30** 12 (40%) 4 (13.33%) 12 (40%) 2 (6.66%) 30 (100%)
**Simple dressing n = 30** 8 (26.66%) 5 (16.66%) 8 (26.66%) 9 (30%) 30 (100%)
:::
Gluteal ulcers healed more completely in the HD group \[6/6(100%)\] than in the PC \[2/7 (29%); difference 71%, 95% CI = 37.38--100, (P \< 0.005)\] or SD \[1/8(13%); difference 87%, 95% CI = 63.69--100, (P \< 0.001)\] groups. The corresponding figures for ischial ulcers were: HD group 13/18(72%) and SD group 3/14 (21%); difference 51%, 95% CI = 21.2--80.7, (P \< 0.005)\]. The PC group was not significantly different from HD: 8/18(44%); difference 28%, 95% CI = -2.9--58.9, (P \< 0.1)\]. In the case of sacral ulcers, complete healing in HD group did not differ significantly from either of the others. The results were: HD group 4/7 (57%), SD group 4/8(50%); difference 7%, 95% CI = -50--64.15, (P \> 0.35), and PC group 2/5(40%); difference 17%, 95% CI = -39.4--73.4, (P \> 0.20)\].
We performed a second analysis on 83 ulcers in 83 patients. We selected one ulcer per patient using a random number table; 31 of the 83 ulcers were stage I and the remaining 52 were stage II. There were again no significant differences among the trial groups with respect to baseline characteristics (table [3](#T3){ref-type="table"}). This \"per patient\" analysis showed that complete ulcer healing, regardless of location and stage, in the HD group was better than in the PC \[20/28(71.4%, 95% CI = 54.7--88.1) vs 11/28 (39.3%, 95% CI:21.3--57.3); difference 32.1%, 95% CI = 7.4--56.7, (P \< 0.01)\] or SD \[20/28(71.4%, 95% CI = 54.7--88.1) vs 8/27 (29.6%, 95% CI = 12.4--46.8); difference 41.8%, 95% CI = 17.7--65.8, (P \< 0.005)\] groups.
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Baseline characteristics of study subjects assigned to three trial groups considering the patient as unit of analysis(one ulcer per patient).
:::
**Variables** **Mean duration of ulcer before treatment (wk ± SD)** **Mean ulcer size (cm^2^± SD)** **Stage of ulcer (no)**
---------------------------- ------------------------------------------------------- --------------------------------- ------------------------- ----
**Total n = 83** 5.92 ± 6.27 7.78 ± 13.53 31 52
**Hydrocolloid n = 28** 7.12 ± 5.68 7.47 ± 16.4 12 16
**Phenytoin n = 28** 6.11 ± 8.4 5.13 ± 3.67 9 19
**Simple dressing n = 27** 4.47 ± 3.64 10.84 ± 16.32 10 17
**P-Value** P \> 0.20 P \> 0.20 P \> 0.70
:::
All completely healed ulcer patients were followed up by monthly visits from general practitioners for a further 4 months after the end of the trial. They were also examined by the assessor author. No recurrence of ulceration was observed in any of the trial groups during this period.
All patients completed the study and there were no losses to follow up, no treatment withdrawals, no trial group changes and no major adverse events (Fig. [2](#F2){ref-type="fig"}).
Discussion
==========
Diphenyl hydantoin sodium (phenytoin) is an effective anti-epileptic medication. Its capacity to accelerate ulcer healing was reported more than 40 years ago \[[@B24]\]. Since then, it has been used topically for different kinds of wounds and ulcers such as war wounds, sores caused by venous stasis, atrophic ulcers and burns, and positive effects have been reported \[[@B25]-[@B27]\]. Possible mechanisms of action of phenytoin cream on wound healing are as follows: 1. Decrease in serum corticosteroid; 2. Acceleration of assembly and presence of collagen and fibrin in the ulcer area, and stimulation of alkaline phosphatase secretion \[[@B28]\].
The use of HD for healing pressure ulcers dates from about 20 years ago. The benefits of this method in comparison with conventional methods include reduction of bacterial contamination, facilitation of patient movement, improvement in patient\'s psychological condition, more convenience and less pain \[[@B29]-[@B34]\]. Hydrocolloid adhesive dressings absorb water and low molecular weight components from ulcer secretions, so they swell to produce a jelly. This jelly protects the ulcer, and new cells proliferate \[[@B35]\]. Moreover, the jelly stimulates the immune system locally by activating granulocytes, monocytes and the complement system \[[@B36]\], decreasing the effects of bacterial colonization and ensuring autodebridement of the ulcer \[[@B1]\].
Bacterial colonization is likely under the HD layer and is responsible for the unpleasant aroma detected when the dressings are changed, but it should not be misinterpreted as clinical infection. In fact, clinical trials of HD on more than 2000 ulcers have shown a much lower incidence of infection than in other treatment methods \[[@B29],[@B30],[@B33]\]. Thus, the ulcer dry-out method is not considered as useful as it once was, and the current trend is towards a damp method using HD \[[@B37]-[@B41]\].
In this study, the therapeutic effects of HD on gluteal and ischial ulcers were shown to be superior to those of PC and SD. In view of the cost of pressure ulcer management in hospitals and sanitariums and the high expense of plastic surgery \[[@B42]\], and the psychological problems associated with paralysis and pressure management in SCI victims \[[@B35],[@B43],[@B44]\], it seems rational to shift to simpler methods that are more cost efficient and executable by the individual patient \[[@B31],[@B34],[@B35]\]. HD treatment of pressure ulcers is less expensive and more comfortable and will ultimately increase the patients\' self-confidence \[[@B8],[@B45]\]. These adhesives are available in different sizes and brands convenient for use in ulcers of different parts of body. In the most recent products, the appropriate time for changing the adhesive is indicated by a color conversion. In addition, their transparency makes it easy to observe the ulcer\'s status without removing the adhesive and dressing \[[@B46]\].
Although the therapeutic effects of HD on sacral ulcers, in contrast to gluteal and ischial ulcers, did not appear in this survey to be significantly better (p \> 0.05) than phenytoin and simple dressings, nor was it less effective. Whether the lesser healing effect of HD on sacral ulcers corresponds to the pressure effects in this area, or to greater bacterial colonization or other factors \[[@B3],[@B4],[@B6],[@B47]\], needs to be clarified by further studies.
Gross differences among the three treatment modalities precluded double blinding. Blinding the authors to the treatment groups minimized this limitation. The major tasks, i.e. defining the study population, enrolling the participants who met the eligibility criteria and assessing the primary and secondary outcomes, were performed blind by the authors. To reduce differences in baseline demographic characteristics among the treatment groups and subgroups and to minimize losses to follow-up, war-related SCI patients were recruited and all the patients who met the eligibility criteria were enrolled in the study. They were all relatively young males (mean age 36.64 ± 6.04 years) and had good motivation to complete the course of treatment. The results of this trial cannot be extrapolated to stage III or stage IV pressure ulcers or to other types of wounds. Furthermore, the small numbers of gluteal and sacral ulcers preclude definitive statements about differences among the treatment subgroups.
Conclusion
==========
The observed efficacy of HD in the treatment of pressure ulcers suggests that it might be effectively applied to other stage I or stage II pressure ulcers.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors contributions
=====================
MTH designed the study and wrote the proposal, and visited all the patients and examined them for eligibility criteria. HKH designed the study, helped in the recruitment of patients, planned the data analyses, assessed the trial groups for primary and secondary outcomes and wrote the paper. FY reviewed the literature, advised on data analysis and contributed to writing the paper.
Funding
=======
The study was supported by the Jaonbazan Medical and Engineering Research Center, the medical and research section of the official governmental body responsible for SCI war victims.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-5945/4/18/prepub>
Acknowledgements
================
The authors sincerely thank Hassan Rafati, statistics and epidemiology MSc, Dr. Masood Ahmadzad Asl, for their excellent work on randomization and analysis of the data; Dr. Farhad Zargari, who kindly edited the English text of the paper; Dr. Seyed Mortezah Hosseini and Dr. Aghdas Aghaii, who administered the treatment protocols to the trial groups; Seyed Ali Salehi for his computer programming; the nursing staff of the internal medicine and rehabilitation wards of Baqyiatollah Hospital, who administered the interventions to patients; and Elham Mellat, who did the clerical work and typed the paper.
|
PubMed Central
|
2024-06-05T03:55:51.999464
|
2004-12-15
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545970/",
"journal": "BMC Dermatol. 2004 Dec 15; 4:18",
"authors": [
{
"first": "Mohammad Taghi",
"last": "Hollisaz"
},
{
"first": "Hossein",
"last": "Khedmat"
},
{
"first": "Fatemeh",
"last": "Yari"
}
]
}
|
PMC545971
|
Background
==========
While significant strides have been made in medical research over the past several decades, many research results considered important by researchers and expert committees are not being used by health care practitioners. While the value of health services research must be judged by its validity, its utility cannot be taken for granted. There has been an assumption that when research information is available it will be accessed, appraised and then applied \[[@B1]\]. However, knowledge of a research-based recommendation is by itself insufficient to ensure its adoption. While the value of research evidence as a basis for decision making in health care is well established, the incorporation of such evidence into decision-making remains inconsistent \[[@B2]\].
The gap between research evidence and its\' incorporation into practice has led to an increase in research in how to bring new knowledge to bear on everyday health care. Factors influencing the adoption of research evidence have been studied extensively \[[@B3]-[@B5]\]. Personal attributes, time, organizational boundaries, geography and educational background all contribute to decision-makers\' responses to research evidence \[[@B6],[@B7]\]. An area that has received less attention is the incorporation of theory in health services research. The authors of this paper propose a need for a stronger theoretical base in health services research wherein health services research would be more informative and influential, facilitating the adoption of research results into practise. Integrating theory into health services research is an important first step. In this paper we first describe the importance of theory followed by how theory driven research changes the manner researchers interact with decision-makers. We conclude on how theory driven research may influence the training, practice and the funding of health services research.
Discussion
==========
The importance of theory
------------------------
In recent years a number of researchers have advocated a greater role for the use of theory in strengthening the practice of research \[[@B8]-[@B12]\]. However, health services research has continued to focus primarily on evaluating outcomes with less attention to the mechanisms by which these outcomes are produced \[[@B10],[@B13],[@B14]\]. The emphasis on method at the expense of theory has led to several criticisms. Chen and Rossi \[[@B15]\] argue that an atheoretical approach to research is characterized by adherence to a step-by-step cookbook method for doing outcomes studies. In this situation they contend that research is reduced to a set of predetermined steps that are mechanically applied to various interventions without concern for the theoretical implications of intervention content, setting, participants or implementing organizations. The atheoretical approach tends to result in a simple input/output, or black box type of study \[[@B13]\]. Such simple evaluations may provide a gross assessment of whether or not an intervention works under one set of conditions but fail to identify the reasons why. As such, the conclusions are often less than satisfying to consumers of research results and not easily transferable to different settings.
Theory provides a systematic view of a phenomena by specifying the relations among variables and propositions with the purpose to explain or predict phenomena that occurs in the world \[[@B16],[@B17]\]. In health services research theory can provide a framework to understand the relationship between program inputs (resources), program activities (how the program is implemented) and their outputs or outcomes \[[@B11],[@B13]\]. In addition to identifying the mechanisms by which programs are effective, theory may consider program implementation and contextual factors. While it is important to know the extent to which an intervention attains intended outcomes, it is also essential to know what occurred in the implementation of the intervention. Variation in the implementation of the intervention may be due to differences among program providers, target population characteristics, and differences among sites on how the intervention is delivered. Theory also offers the opportunity to specify the contextual conditions that will influence the effectiveness of an intervention. Attitudinal factors at the provider level as well as structural, cultural factors at the organizational level have been under appreciated in exploring variations in health care outcomes \[[@B9],[@B18],[@B19]\]. Understanding the influence that contextual factors have on program implementation and outcomes facilitates successful application of the intervention in alternate settings, therein, addressing the generalizabilty of an intervention.
Theory offers many advantages to the health services researcher. Theory helps to identify the appropriate study question and target group; clarify methods and measurement issues; provide more detailed and informative descriptions on characteristics of the intervention and supportive implementation conditions; uncover unintended effects; assist in analysis and interpretation of results; and, the successful application of an intervention to different settings \[[@B11],[@B12]\].
Theory-driven studies are addressing the challenge of both decision-makers and funding agencies to move beyond simplistic explanations of significance in health services research. Decision-makers are seeking explanations about how an intervention works and whether it will work in a fashion similar to the intervention that was evaluated when applied to a different environment \[[@B10],[@B12],[@B20]\].
Despite these potential benefits, there are a number of reasons offered as to why there has been a failure to integrate theory into research. Ironically, clinical randomized control trials have discouraged the use of theory in health services research. Given the genesis of clinical trial methodology, this may derive, in part, from the very origins of epidemiology, whereby John Snow allegedly ended an epidemic of cholera by removing the handle from the Broad Street water pump, even though he had no concept of what actually caused cholera. By ignoring the need for theory, Snow was able to overcome the fact that the theories he would have needed had not yet been elucidated. Similarly, we know that lung cancer incidence can be reduced by elimination of cigarette smoking, even though we do not know exactly how cigarette smoke causes lung cancer. Experimental trials often determine intervention effects without considering how the component features of an intervention work together to bring about study outcomes \[[@B13],[@B15],[@B21]\]. The more complex the intervention, the more difficult it is to know what the treatment entailed. There is a growing recognition for the need to establish the theoretical bases of interventions. The United Kingdom Medical Research Council recently proposed a framework for the development and evaluation of randomized control trials for complex interventions where theory is viewed as valuable in assisting hypothesis development and steering decisions on strategic design issues \[[@B22]\].
Adopting a theory-driven approach in health services research is not without its challenges. Given the typical training of researchers and the uni-disciplinary nature of the practice the first challenge is the capacity of researchers to engage in theory driven research. Second, a theory driven approach requires organizational conditions that support researchers and decision makers collaborating in the development and testing of theory. Finally, theory development and testing is cumulative in nature, encouraging researchers to pursue a programmatic approach in research. This approach has implications on how funding agencies support health services research. Despite the potential challenges, a theory based approach offers promise for a greater understanding on what happens when interventions work to address social/health problems.
The importance of collaboration with decision-makers
----------------------------------------------------
Collaborative research partnerships between academic researchers and decision-makers describe a relationship and process between individuals from different backgrounds, who together, develop an integrative cooperative approach to resolve a research problem \[[@B23]\]. It has been identified as a significant strategy that holds multiple benefits \[[@B23]-[@B27]\].
Collaborative practice has also been identified as a key strategy in facilitating a theory driven approach. Weiss \[[@B28]\] recommends that the first criterion in selecting a theory to guide the evaluation of a program is to draw the theory out from those associated from the program, including designers of the program, program personnel and relevant clinical staff. The argument is that few programs are theory driven. Rather, they are typically the product of the experience and values of those who are associated with the program. In recent years a number of techniques have been developed for this purpose. Strategies range from unstructured interviews, to highly structured iterative interactions between program personnel and researchers \[[@B29]-[@B31]\]. Perspectives of service providers can be rounded out by a review of the research literature. In fact, a number of researchers suggest a combination of these two approaches \[[@B10],[@B32]\].
Viewing program stakeholders as a key source in developing theory in health services research demands stronger collaboration between researchers and program decision makers \[[@B9]\]. In this fashion, collaborative practice becomes a methodological strategy in health services research. Lomas \[[@B7]\] has stressed that a first step in encouraging meaningful partnerships between researchers and decision-makers is to view linkage and exchange between the two as a process not as a discrete event. Establishing and maintaining ongoing links offers a more comprehensive understanding between the two groups. Researchers uncover the desired program outcomes, the causal change of the program intervention and develop a better understanding of the contextual factors that influence the variation on intervention implementation and outcomes. Similarly, decision-makers will develop a deeper understanding of the research process and thus can influence the development of feasible and sustainable interventions for practice settings.
The role and impact of the researcher and the research process in practice settings have received greater attention in other fields such as program evaluation, nursing, anthropology and community psychology. For example, core principles of community psychology practice include: a) consistency of goals and values between the researcher and the setting, and b) the notion that interventions should have the potential for being \"institutionalized\" or systematically established within the setting in such a way that strengthens the natural resources of the setting \[[@B25]-[@B27]\]. Rather than reinventing the wheel, health services research could benefit from theoretical frameworks developed within these disciplines.
Implications for the practice of health services research
---------------------------------------------------------
Recognizing the importance of theory calls for new expectations in the practice of health services research. There are a number of challenges that must be met in order for these perspectives to gain acceptance in the health services research community.
Evolving perspectives on the practice of health services research require recognition that few disciplines are able to span the breadth of responsibilities associated with the research process. To date there has been a tendency for health services research to be practiced as a uni-discipline where clinical disciplines tend to practice separately from the social science disciplines. A priority is to encourage the formation of research teams that are inter-disciplinary. Pursuing this agenda will promote the formation of research teams that may include: business, anthropology, sociology, psychology, education, engineering, nursing and medicine. Combined disciplinary skills would, in a complementary fashion, address the breadth of skills required in a more complex research environment that includes the development and testing of theory.
A second point concerns broadening the training for those who will practice health services research. By and large, academic training has focused on methodological issues. While a focus on research methods has made an important contribution to the practice of health services research, relying on research methods as a core curriculum has led to limitations in the training of health services researchers such as inadequate attention to the value of theory driven research. As health services research expands its methodological repertoire beyond the classical randomized control trial, researchers face increased ambiguity in attributing the source of intervention impact. It is in this circumstance that theory can guide health services researchers in understanding the causal linkages within an intervention. Further, students are educated in separate departments with little planned, formal activity across disciplines, which discourages co-operative approaches to research and service \[[@B33]-[@B35]\]. Education programs are not generally structured to facilitate the importance of inter-disciplinary strategies. Identifying the processes associated with creating effective linkages between researchers and decision-makers are also not typically part of training. Rethinking the current assumptions and practices regarding the training of health services researchers will enable trainees in health services research to be better prepared for their evolving responsibilities.
Collaboration between researchers and decision-makers are contingent upon supportive organizational conditions for both partners. Researchers have, and most likely will continue to operate from university-based settings where incentives for promotion and tenure can act as barriers to changes in the practice of health services research \[[@B7],[@B36]\]. Most academic institutions award tenure and promote faculty based upon the frequency and quality of publications and on obtaining peer review funding \[[@B36]\]. The time involved in collaborating with decision-makers, joint planning and implementing research often represents activities that are not recognized by tenure promotion committees. As well, these activities may slow the production of research results and the generation of publications. Recognizing these factors requires academic centres to generate new criteria for evaluating contributions to knowledge and practice. Decision-making organizations also play a significant role in ensuring the success of collaborative relationships. The clearest indication of institutional support for research is to provide the time and resources for decision-makers to participate in collaboration activities with researchers.
Funding bodies have the potential to play a significant role in guiding and integrating these considerations into health services research. Research sponsors can develop evaluation criteria that encourage the application of theory. As an example, the Agency for Healthcare Research and Quality (AHRQ) in the USA funded an initiative (Translating Research into Practice) to identify sustainable and reproducible strategies that will: 1) accelerate the impact of health services research on direct patient care; and 2) improve the outcomes, quality, effectiveness, efficiency, and/or cost effectiveness of care through partnerships between health care organizations and researchers \[[@B37]\]. Further, research sponsors are beginning to move away from supporting single shot studies that are conducted in relative independence from one another. This focus on supporting programmatic research should be encouraged. Programmatic research offers a cumulative environment that allows researchers the opportunity to develop and test the application of theory. In a similar fashion, collaborative practice is also best practiced in a programmatic environment. Developing and maintaining linkages with decision-makers is predicated on developing and maintaining long-term relations. Embedded within these linkages are fundamental professional and personal attributes that include; credibility, familiarity, mutual understanding and trust \[[@B38],[@B39]\].
Summary
=======
This paper has examined the importance of theory in health services research. We have argued that by strengthening the role of theory encourages collaborative practice between researchers and decision-makers. It has been noted that a theory driven approach in health services research is not without its challenges. However, given the modest advances towards incorporating research evidence into healthcare decisions, a theory driven approach is well worth the effort. The implication of this approach for health services research is that it has impact on the training and practice of health services research. Institutions and researchers should consider this emerging model of practice if health services research is to fulfill its potential for improving the delivery of care.
Competing interests
===================
The authors declare that there are no competing interests. Disclaimer: The opinions expressed are the authors\' and do not necessarily represent official policy of AHRQ or the Department of Health and Human Services
Authors\' contributions
=======================
KB drafted the manuscript, edited and revised the contents, EO edited and revised the manuscript, KB, EO, MC, RS, DS all contributed to the conceptual development, editing and review of the manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1472-6963/5/1/prepub>
|
PubMed Central
|
2024-06-05T03:55:52.001801
|
2005-1-7
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545971/",
"journal": "BMC Health Serv Res. 2005 Jan 7; 5:1",
"authors": [
{
"first": "Kevin",
"last": "Brazil"
},
{
"first": "Elizabeth",
"last": "Ozer"
},
{
"first": "Michelle M",
"last": "Cloutier"
},
{
"first": "Robert",
"last": "Levine"
},
{
"first": "Daniel",
"last": "Stryer"
}
]
}
|
PMC545995
|
Background
==========
Alopecia areata (AA) is manifested as a sudden loss of hairs without any inflammation or scarring. The hair loss might be seen in a circumscribed area or the whole scalp (alopecia totalis or AT) or whole body (alopecia universalis or AU) \[[@B1]\]. It is a common disease and at any given time, 0.2% of the population has AA and 1.7% of the population will experience an episode of AA during their lifetime \[[@B2],[@B3]\]. The etiology of AA is not known exactly. However factors such as genetic predisposition, autoimmunity, and stress have been suggested \[[@B4],[@B5]\]. The course of disease is not predictable and it is often associated with periods of hair loss and regrowth. The clinical severity of a patient\'s AA may not be a good indicator of subsequent downturn in quality of life or psychological well-being.
The onset of a chronic condition brings with it a range of difficulties that may show considerable variation in their nature and severity as perceived by the patient. In order to make sense of and respond to the difficulties that chronic illness may present, patients construct their own common-sense cognitive model of their condition. Such models are based upon information received from a range of sources including their physician, family, friends, and existing social and cultural notions about health and illness. The resulting system of beliefs can of course be flawed or inaccurate; however there is evidence that it is those beliefs that drive attempts to cope with a condition and issues of compliance with treatment. Patient-held beliefs have important implications for the clinical management of their disease. Studies on patients with psoriasis and acne vulgaris have shown that knowledge of patients about their condition, the course of their disease and current treatments is not appropriate \[[@B6],[@B7]\].
In this study, we examined the system of beliefs held by AA patients and the factors that might influence such beliefs.
Methods
=======
The Illness Perception Questionnaire (IPQ) \[[@B8]\] with a few modifications was given to 80 patients with AA older than 12 years, attending a private skin clinic in Tehran, Iran in 1999. The study was approved by Institutional Review Board of the Center for Research and Training in Skin Diseases and Leprosy. The IPQ was created to provide a theoretically derived measurement instrument suitable for use with any patient population. It has been used in patients with cardiac disease \[[@B9]\], chronic fatigue syndrome \[[@B10]\], diabetes, chronic pain, rheumatoid arthritis \[[@B8]\], and psoriasis \[[@B6]\]. As AA is an asymptomatic disease, we did not use the subscale of \"symptoms\" in our study. Thus the questionnaire that we used consisted of four subscales:
**Cause**subscale (10 items) measures personal ideas about the cause of AA.
**Time line**(3 items) deals with perceptions about how long the disease will last.
**Consequences**(6 items) are concerned with expected effects and outcomes of the illness.
**Cure/Control**(6 items) details beliefs about recovery from or control of the condition.
There were four possible answers for each item in the IPQ to be chosen by patients: I strongly agree, I agree, I do not know, I disagree.
Furthermore, some demographic information such as age, sex, family history of AA, and duration and extent of disease (alopecia areata or AT/AU), and level of education were obtained from the patients to evaluate their influence on patients\' beliefs.
Statistical analysis was conducted by means of SPSS statistical software, version 11.0. Because the data were not normally distributed, nonparametric statistics were used. Correlations were processed by Spearman\'s rank correlation, and differences between means were computed by means of the Mann-Whitney U test. For simplicity of analysis and increasing the power of study, the answers of \"I strongly agree\" and \"I agree\" were grouped together and compared with the answers of \"I don\'t know\" and \"I disagree\" which were grouped together as \"I do not agree\". A p value of less than 0.05 was considered as significant.
Results
=======
A total of 80 patients with AA (38 male, 42 female), with a mean age of 27.5 years (SD 9.3, ranged from 13 to 56 years) were recruited to the study. The mean duration of illness was 7.8 years (SD 7.7, ranged from 1 month to 30 years). In 75% of patients, AA was patchy and it was totalis or universalis in 25%. Fifteen percent of patients had a positive family history of AA in their first degree relatives. Physicians were the main source of patients\' information about their disease in 66.2% of them.
Beliefs about cause
-------------------
Table. [1](#T1){ref-type="table"} shows the percentage of patients \"agreeing\" with each **cause**item. A total of 76.9% of patients believed that stress was a major factor in onset of their illness and older patients were more likely to believe in this (p \< 0.05). Patients who had a belief that their disease was a result of genetic factor were more likely to have a family history of AA and longer duration of the disease (P \< 0.05). Younger patients and those with extensive disease (AT/AU) believed that their illness was because of chance or fate (P \< 0.05).
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Beliefs about causes of alopecia areata (n = 80)
:::
**Causes** **Agree** **Factors influencing beliefs**
------------------- ----------- -------------------------------------------------------------------
Stress 76.9% Older patients (p = 0.012)
My state of mind 59.2% None
My own behavior 47.3% None
Other people 34.2% None
Chance or fate 31.1% Younger patients(p = 0.021), extensive disease (AT/AU)(p = 0.030)
Diet 25.7 % None
Pollution 24.3% None
Germ or virus 21.9% None
Genetic 17.1% Family history of AA(p = 0.006), longer duration(p = 0.017)
Poor medical care 11.8% None
:::
Beliefs about consequences
--------------------------
Majority of the patients (58.2%) believed that their illness had a major consequence on their lives, 53.8% of patients also felt that AA had strongly affected their self-esteem, and 50.6% considered AA as a serious condition. These believes were stronger in younger patients, and in patients who had the disease for a long time (p \< 0.05). Table [2](#T2){ref-type="table"} shows the percentage of patients \"agreeing\" with each **consequence**item.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Beliefs about consequences of having alopecia areata (n = 80)
:::
**Beliefs** **Agree** **Factors influencing beliefs**
-------------------------------------------------------------------- ----------- --------------------------------------------------------------
My disease has had a major consequence on my life. 58.2% Younger patients(p = 0.012), longer duration(p = 0.014)
My disease has strongly affected the way I see myself as a person. 53.8% Younger age at onset(p = 0.003)
My disease has strongly affected the way others see me. 51.3% Younger age at onset(p = 0.022), longer duration(p = 0.012)
My disease is a serious condition. 50.6% Younger age at onset(p = 0.003), younger patients(p = 0.013)
My disease has become easier to live with. 50.6% None
My disease has serious economic and financial consequences. 27.8% Younger age at onset(p = 0.015), longer duration(p = 0.010)
:::
Beliefs about recurrence or chronicity
--------------------------------------
Half of the patients believed whether their disease cleared, it would always come back and forty percent of patients believed that their illness would be likely to be permanent rather than temporary. They were more likely to have a longer duration of disease (p \< 0.05). The minority of patients (25.0%) believed that their illness would last a short time.
Beliefs about cure and control
------------------------------
More than 60% of patients believed that their behavior could determine improvement or worsening of their illness (table [3](#T3){ref-type="table"}). This belief was present in female patients more than male patients (p \< 0.05). 30.4% of patients believed there was very little that could be done to improve their illness. They were more likely to have longer duration of disease (P \< 0.05). Thirty-eight percent of the patients believed that recovery from disease is largely dependent on chance or fate. This belief was stronger in female patients, those with younger age at onset, and patients with extensive disease (p \< 0.05).
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Beliefs about cure and control (n = 80)
:::
**Beliefs** **Agree** **Factors influencing beliefs**
------------------------------------------------------------------ ----------- ---------------------------------------------------------------------------------------------------
What I do can determine whether my disease gets better or worse. 63.3% Female patients(p = 0.010)
My treatment will be effective in curing my disease. 57.5% None
My disease will improve in time. 53.2% Older patients(p = 0.006), older age at onset(p = 0.029)
There is a lot that I can do to control my disease. 52.5% None
Recovery from my disease is largely dependent on chance or fate. 38.0% Female patients(p = 0.014), younger age at onset(p = 0.036), extensive disease (AT/AU)(p = 0.021)
There is very little that can be done to improve my disease. 30.4% Longer duration(p = 0.004)
:::
Discussion
==========
Alopecia areata is a chronic disease which may influence individual or social aspects of patients\' lives. The results of our study confirmed this fact as the majority of patients believed that their illness had strongly affected their lives. It also influenced their self-esteem. The results of studies in other chronic diseases with periods of remission and exacerbation have had different results in this respect. For example in a study on acne patients, the disease had affected patients\' self-image in nearly all of them, but it had no impact on interpersonal relationships, work, or school activities in majority of patients \[[@B7]\]. On the other hand, in a study on patients with psoriasis using the IPQ questionnaire, 68% of patients who suffered from psoriasis, agreed that psoriasis had a major consequence on their lives, and 53.4% agreed that psoriasis had strongly affected the way they saw themselves as a person \[[@B6]\].
The present study also investigated cognitive appraisals held by patients about their illness and showed that such beliefs were not associated in any significant manner with the extent of their condition. Fortune et al also did not find an association between the clinical severity of psoriasis and beliefs held by patients about their condition \[[@B6]\]. Thus the assumption that the objective severity of a condition will be associated in a linear fashion with patient\'s subjective experience in terms of beliefs, coping, or distress is unlikely to be correct. On the other hand, young patients and those with longer duration of disease were more likely to be affected by their disease. This implies that the chronicity of the disease has more influence on patient\'s life than the extent of it.
The results of our study also showed that the beliefs about the consequences of having AA were not influenced by the gender of the patients. Thus men are as vulnerable as women in suffering from the consequences of AA.
Patients with AA, including 77% of patients in this study, often attribute the onset of their disease to a specific stressful life event. In a study on 178 patients, Van der Steen *et al.*showed that emotional stress is not an important factor in the initiation of AA \[[@B11]\]. Brajac *et al.*did not find a significant role of stress in the onset of AA but stressful life events had an important role in triggering of some episodes of disease \[[@B12]\]. On the other hand, Gupta *et al.*found that AA patients who were depressed, were more likely to mention stress as the cause of their disease \[[@B13]\].
In recent studies, psychologic and psychopathologic factors have been analyzed as modulators of neuroendocrinologic, vascular, and immunologic variables; this is far from the initial concept of stress being the causal agent in the illness. In fact, stress may cause its effect by making alterations in immune responses related to neuropeptides, such as the migration of the macrophages, vasodilator or vasoconstrictor responses, phagocytosis, lymphocytic cellular immunity, and expression of some factors of leukocytic adhesion to the microvascular endothelium \[[@B14]\]. In addition, the adaptation to the illness is regarded as an important factor with regard to prognosis. However the exact cause of AA is not known and such events are very common, making it difficult for the investigator to prove that they are in fact involved in causing or precipitating the disease. In this study, one-third of the patients believed in chance or fate as the cause of AA, and this belief was stronger in younger patients and those with extensive disease (AT/AU).
This study also showed that as AA lasts, the patients feel more hopeless about time line and treatment modalities of their disease. The majority of patients had no hope to get rid of their disease. Almost half of patients expected their disease to relapse after it disappeared. Such perpetual stresses are hardly endurable. The psychiatrists\' intervention may alleviate patients\' stress and improve their quality of life.
This study was performed in a private dermatology clinic. The possibility of socioeconomic homogeneity among recruited patients may be biasing the results. So it should be considered that these results may be different in patients with AA in different socioeconomic and cultural backgrounds.
Conclusion
==========
There is a need for accessible, accurate, community-based education on the natural history of AA, the effectiveness and expected duration of treatment. The inadequacy of information provided by current sources is evident in ongoing misconceptions on causality and the perceptions of respondents. Incorporating information on this disease may facilitate patient into therapeutic selection, enhance understanding of treatment options and improve patient compliance.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
AF participated in the design and conduct of the study and preparation of the manuscript.
MRF participated in the conduct of the study and statistical analysis.
BG participated in the conduct of the study and statistical analysis.
YD participated in the design of the study and preparation of the manuscript.
All authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-5945/5/1/prepub>
|
PubMed Central
|
2024-06-05T03:55:52.004104
|
2005-1-12
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545995/",
"journal": "BMC Dermatol. 2005 Jan 12; 5:1",
"authors": [
{
"first": "Alireza",
"last": "Firooz"
},
{
"first": "Mehdi Rashighi",
"last": "Firoozabadi"
},
{
"first": "Behnaz",
"last": "Ghazisaidi"
},
{
"first": "Yahya",
"last": "Dowlati"
}
]
}
|
PMC545996
|
Background
==========
Coronary artery disease is the leading cause of death among Americans \[[@B1]\]. Hospitalization for acute coronary syndromes (ACS), which includes both acute myocardial infarction (AMI) and unstable angina, is common and costly. Many patients admitted with ACS to primary hospitals (i.e. those without on-site cardiology subspecialty services, including cardiac catheterization facilities) are transferred to tertiary hospitals for cardiac catheterization and consideration of coronary revascularization.
The coordination and integration between primary and tertiary hospitals has important implications for integrated health care delivery systems. The Veterans Health Administration (VHA) is one of the largest vertically integrated health care delivery systems in the United States \[[@B2]\]. The VHA is organized in 21 regional networks. Regionalization has been adopted by many integrated health care delivery systems, both to improve quality and to increase efficiency \[[@B3]-[@B6]\]. In most VHA regions, a single tertiary hospital is associated with one or more primary hospitals. A particular challenge in the VHA is providing access to sub-specialty cardiology services for patients hospitalized with acute coronary syndromes because primary hospitals are often geographically distant from tertiary hospitals \[[@B5]\].
Treatment guidelines for acute coronary syndromes \[[@B7]-[@B10]\] suggest that some diagnostic tests and therapies can be performed at most primary VHA hospitals, while others, such as cardiac catheterization and coronary revascularization, require transfer to a tertiary hospital. Well-functioning transfer processes are critical to making a policy of regionalization work. In addition, there are strong financial and organizational incentives to provide care within an integrated health care system like VHA rather than referring to non-VHA hospitals, even when this requires transfer to distant tertiary hospitals \[[@B11]\]. In the VHA, transfers within the system represent cost savings, while transfers out, by and large, represent cost increases. In addition to cost issues, there are also coordination of care concerns that are addressed through within-system transfer, particularly in a system with a common electronic medical record. However, the constraint on within-system transfer is that patients requiring urgent or emergent transfer to receive definitive care should be transferred to the nearest facility with capacity to provide care, even if this requires a transfer out of the system. Issues related to cost differences due to transfer within and outside integrated health care systems are most applicable in the United States, where the multiplicity of payers is a major financial concern; in other countries with integrated national health care, or single payer, systems, these issues are less relevant, although issues of care coordination may still be important.
The objective of this study was to evaluate the association between structural components of clinical integration and patient transfer rates from VHA primary hospitals to tertiary hospitals, both within and outside the VHA system for patients with ACS. We hypothesized that primary VHA hospitals with structural components of clinical integration present would have a higher rate of within-system transfer of ACS patients than primary VHA hospitals lacking these components.
Methods
=======
The VHA Access to Cardiology study was a prospective cohort study of 2,733 patients with a primary discharge diagnosis of either acute myocardial infarction (ICD9-CM 410.xx) or unstable angina (ICD9-CM 411.xx) discharged over a one year period (March 1, 1998 through February 28, 1999) from 24 VHA hospitals in five regions, including Minnesota and the Dakotas, the Southwest, the Rocky Mountains, the Pacific Northwest, and Southern California. Patient demographics, clinical characteristics, and specific processes of care including hospital transfer were obtained as part of the Access to Cardiology Study.
All patients admitted to one of the 12 primary VHA hospitals in the study were eligible for this analysis (n = 862 out of the 2,733 in the larger Access to Cardiology study). The remaining 12 VHA Medical Centers were tertiary hospitals with cardiology services and cardiac catheterization laboratories on site. These were not the focus of the analysis reported in this paper. We excluded 107 patients because they were initially admitted to a private hospital and transferred into a primary VHA hospital. In addition, we excluded 3 patients who were transferred from one primary VHA hospital to another. Finally, 27 patients had missing data in the variable indicating prior history of congestive heart failure, which was included in the final analysis. As a result, a total of 725 patients from 12 primary VHA hospitals were included in these analyses. The study protocol was approved by the Human Subjects Committee at the University of Washington, and by Institutional Review Boards and Research and Development Committees at each participating VHA hospital.
Transfer rates
--------------
Patient transfer from a primary VHA hospital to a tertiary hospital (either VHA or private) was the primary outcome for this study. Secondary outcomes included both transfer from a primary VHA hospital to a tertiary VHA hospital, and transfer from a primary VHA hospital to a private (non-VHA) tertiary hospital.
Transfers to a tertiary VHA hospital were considered transfers within the system, while transfers to a private hospital were considered transfers outside the system. Transfer data were available for all 725 patients in the study cohort. We constructed two binary variables for the analyses: transfer to any tertiary care hospital (yes/no), and transfer to a tertiary VHA hospital versus transfer to a private (non-VHA) hospital.
Clinical integration
--------------------
The key independent variable for this study was clinical integration of cardiac services. We defined clinical integration \[[@B12],[@B13]\] as the extent to which patient care services, in this case cardiology consultation services, are coordinated across the units and hospitals in the VHA providing care to cardiology patients. We measured clinical integration of cardiac services using three binary variables to indicate the presence or absence of these structural elements of clinical integration: a) a VHA staff cardiologist on-site at least episodically at the primary VHA hospital (either through a full or part time VHA staff cardiologist on site, or through periodic visits by a VHA staff cardiologist from the affiliated tertiary VHA hospital); b) a referral coordinator at the tertiary referral VHA hospital; and c) a referral coordinator at the primary VHA hospital. Referral coordinators at primary VHA hospitals are generalists, in that they facilitate referrals, transfers, and sometimes consultations for patients with many different kinds of diseases or health problems. In contrast, at tertiary VHA hospitals, referral coordinators are often associated with particularly sub-specialties, and work closely with these specialty services to provide assistance to referring hospitals and providers in determining whether transfer, referral, or consultation is advisable, and expediting the processes. These were all hospital level variables. We combined the two groups, VHA staff cardiologist on site and periodic visits by a VHA staff cardiologist, for two reasons. First, only one of the 12 primary hospitals in the sample had an on site VHA cardiologist, and the sample size in that group was too small to analyze independently. Second, in our interviews with Chiefs of Cardiology at the tertiary VHA hospitals, there was unanimity in their beliefs that either type of VHA cardiologist being available in a primary hospital produced more appropriate referrals, and improved interactions between providers at the primary hospital and the VHA tertiary cardiology service.
The data used to construct these measures came from on-site interviews conducted with Chiefs of Cardiology at each of the tertiary VHA hospitals associated with the primary VHA hospitals included in this study. During on-site interviews, Chiefs of Cardiology were asked to describe all of the primary VHA hospitals that refer ACS patients to them on a regular basis, and to identify the presence or absence of each of the structural elements of clinical integration. Interviews followed a structured protocol, ensuring uniform data collection. In all cases, the Chiefs of Cardiology were able to provide detailed information about the services available at both the tertiary and primary VHA hospitals. We also asked the Chief of Cardiology about the degree of competitiveness for cardiac services in the local markets for each of the primary VHA hospitals. This was an ordinal variable, with three levels: non-competitive; moderately competitive; or highly competitive market. In all cases, the Chief of Cardiology was able to answer the questions about market competition in the primary hospital market without difficulty, indicating considerable awareness of market conditions and the impact these had on their referral base.
In addition, we constructed two separate variables to control for patient distance from the primary VHA hospital to which they were initially admitted, and to control for the distance between primary and tertiary VHA hospitals. The patient distance variable was measured as the distance from the patient\'s home zip code centroid to the primary VHA hospital. The distance between the primary and tertiary referral VHA hospitals was measured in miles using VHA national databases. We tested different specifications of the distance variables, concluding that it was best to enter the distance between primary and tertiary VHA hospital as a continuous variable, whereas it made no difference in the results of the estimation what form we used for patient distance to primary VHA hospital. In the final analyses, it was dichotomized at greater than or equal to 100 miles -- approximately two hours driving time. The patient distance variable is measured at the patient level, while the hospital distance variable is measured at the hospital level.
We included several measures of patient clinical characteristics, including age 65 or over; prior history of chronic obstructive pulmonary disease, bleeding disorder (such as hemophilia or anticoagulation therapy), smoking, prior percutaneous coronary intervention (PCI), or chronic heart failure; having a \"Do Not Resuscitate\" order, and several measures of seriousness or urgency of condition during the index admission in the primary VHA hospital: ST segment elevation on electrocardiogram or elevated cardiac enzymes at presentation; and a composite variable indicating the presence of a serious event during admission. Presence of a serious event during admission was a binary variable taking the value \"1\" if at least one of the following conditions was present: angina persisting more than 24 hours after admission; hypotensive episode; heart failure during admission; cardiac arrest; or positive stress test during admission. All of these variables were abstracted from the medical record.
Analyses
--------
We explored the bivariate association between clinical integration variables, distance variables, patient characteristic variables, and patient transfer using one-way analysis of variance with Scheffe correction for multiple comparisons.
To construct the most parsimonious models using the full set of candidate independent variables (clinical integration variables and patient characteristics), we used backward stepwise logistic regression, beginning with all available patient clinical characteristics that have been shown to be significant in predicting mortality outcomes for ACS patients in prior studies. We eliminated variables from the model if the p-value for the variable was greater than 0.1. A number of the candidate variables, including many of the history and co-morbidity variables, were found to be insignificant, and we created a summary variable described above which included many of the highly significant variables from the index hospital admission (details available from authors). C-statistics for each of the final models ranged from 0.77 to 0.85. We used Stata SE version 8.2 for all analyses.
We then investigated the relationship between clinical integration of cardiac services and transfer rates using random effects logistic regression \[[@B14]\], correcting for cluster sampling by hospital and region and controlling for distance and patient characteristics that reflect cardiac disease severity and therefore may affect the likelihood of transfer. Two models were estimated, one for transfer to any tertiary care hospital, and the second to estimate the conditional probability that the patient was transferred to a VHA tertiary hospital versus transfer to a non-VHA tertiary hospital, given that they were transferred. Random effects logistic regression allowed us to control for the effects of clustering on both the hospital and regional (Veterans Integrated Service Network, or VISN) level. The intra-class correlation of overall transfer with hospital and VISN jointly was 0.12 (p = 0.006), suggesting the need to control clustering at both levels.
Results
=======
Among the 12 primary VHA hospitals included in the sample, the mean rate of transfer was 42% (319 of 725). Mean rate of transfer to a tertiary VHA hospital was 31% (237 of 725), and to a private hospital was 11% (82 of 725). Most patients were transferred in order to receive cardiac catheterization or coronary revascularization. In addition, 37% of patients were treated in primary VHA hospitals that were over 250 miles from their tertiary referral VHA hospital, and 18% of patients lived over 100 miles from the primary VHA hospital to which they were admitted.
Three of the 12 primary care VHA hospitals had a VHA cardiologist available at least episodically on site; six had a referral coordinator at the associated tertiary center; and four had a referral coordinator at the primary VHA hospital. The distribution of these components is shown in Figure [1](#F1){ref-type="fig"}. Five of the twelve hospitals had none of the three components of integration.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Distribution of integration components across the 12 primary VHA hospitals
:::
:::
Unadjusted associations
-----------------------
The bivariate associations between the patient characteristic variables, clinical integration variables, and type of transfer are shown in Table [1](#T1){ref-type="table"}. All of the patient characteristics except history of chronic obstructive pulmonary disease were strongly and positively associated with transfer to a tertiary hospital. Distance between primary and tertiary VHA hospital was significantly different between the three groups, with overall transfer being associated with increased distance between the primary and tertiary VHA hospital. The degree of market competition was also significantly associated with transfer, principally to tertiary private hospitals. Each of the three individual components of integration were significantly associated with transfer from primary VHA.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Patient and facility characteristics by transfer type
:::
**Variable** **Overall for study sample N = 755** **Not transferred N = 436** **Transferred to tertiary VHA hospital N = 237** **Transferred to tertiary private hospital N = 82** **p-value\***
-------------------------------------------------------------------------------- -------------------------------------- ----------------------------- -------------------------------------------------- ----------------------------------------------------- ---------------
**Patient age 65 and over** 58.0% 63.1% 52.3% 48.8% 0.005
**Prior medical history**
**Chronic obstructive pulmonary disease** 37.1% 40.6% 31.9% 33.7% 0.067
**Bleeding disorder** 3.6% 2.1% 5.5% 6.2% 0.035
**Smoker** 31.6% 26.8% 41.8% 27.2% \<0.001
**Prior percutaneous coronary intervention** 15.2% 11.7% 21.5% 15.8% 0.003
**Chronic heart failure** 23.0% 28.9% 13.1% 18.8% \<0.001
**Course of index hospital admission**
**ST segment elevation on EKG** 17.8% 12.8% 19.0% 39.0% \<0.001
**Cardiac enzymes abnormal on presentation** 52.5% 52.0% 46.4% 71.3% \<0.001
**Do not resuscitate during hospitalization** 5.3% 6.7% 2.1% 5.2% 0.039
**In-hospital event\*\*** 47.3% 37.8% 62.9% 52.4% \<0.001
**Distance, market and integration variables**
**Distance from patient home zip code centroid to hospital \>100 miles** 18.1% 15.6% 21.1% 22.0% 0.128
**Distance from primary VHA to tertiary VHA hospital in miles** 281 270 285 326 0.045
**Degree of market competition (1 = not competitive; 3 = highly competitive)** 1.74 1.82 1.57 1.79 \<0.001
**VHA cardiologist on site** 30.6% 29.8% 36.3% 19.5% 0.015
**Tertiary VHA hospital has referral coordinator** 54.7% 56.4% 60.8% 30.5% \<0.001
**Primary VHA hospital has referral coordinator** 33.0% 28.9% 43.9% 24.4% \<0.001
\* p-value obtained from ANOVA testing difference between means for patients not transferred, transferred to VHA tertiary hospital, or transferred to non-VHA tertiary hospital for continuous variables, chi-square test of inference for categorical variables
\*\* Presence of at least one of the following adverse events during admission: angina persisting more than 24 hours after admission; a hypotensive episode; an episode of heart failure; cardiac arrest; or positive stress test during admission
:::
Risk-adjusted association: transfer to any tertiary care hospital
-----------------------------------------------------------------
Results of the random effects logistic regressions for transfer to any tertiary care hospital are shown in Table [2](#T2){ref-type="table"}. Patient factors increasing the likelihood of transfer to a tertiary hospital included being a smoker; history of chronic heart failure; ST-segment elevation on presenting electrocardiogram; in-hospital events (presence of at least one of the following events during admission: angina persisting more than 24 hours after admission; a hypotensive episode; an episode of heart failure; cardiac arrest; or positive stress test during admission); and distance from patient home to hospital more than 100 miles.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Results of random effects logistic regression of transfer to any tertiary care hospital
:::
**Variable** **Odds ratio** **p-value** **Lower limit 95% CI** **Upper limit 95% CI**
-------------------------------------------------------------------------------- ---------------- ------------- ------------------------ ------------------------
**Patient age 65 and over** 0.69 0.06 0.48 1.01
**Chronic obstructive pulmonary disease** 0.48 \<0.001 0.31 0.74
**Bleeding disorder** 0.68 0.04 0.47 0.98
**Smoker** 3.28 0.01 1.32 8.12
**Prior percutaneous coronary intervention** 1.30 0.18 0.89 1.91
**Chronic heart failure** 2.10 \<0.001 1.33 3.32
**ST segment elevation on presenting electrocardiogram** 2.07 \<0.001 1.32 3.26
**Cardiac enzymes abnormal on presentation** 0.92 0.65 0.64 1.31
**Do not resuscitate during hospitalization** 0.29 \<0.001 0.12 0.65
**In-hospital event\*** 3.14 \<0.001 2.21 4.46
**Distance from patient home zip code centroid to hospital \>100 miles** 1.71 0.02 1.08 2.70
**Distance from primary VHA to tertiary VHA hospital in miles** 0.998 0.03 0.997 0.999
**Degree of market competition (1 = not competitive; 3 = highly competitive)** 0.55 \<0.001 0.41 0.73
**VHA cardiologist on site** 0.48 \<0.001 0.29 0.79
**Tertiary VHA hospital has referral coordinator** 0.39 \<0.001 0.23 0.69
**Primary VHA hospital has referral coordinator** 6.53 \<0.001 3.29 12.98
\* Presence of at least one of the following adverse events during admission: angina persisting more than 24 hours after admission; a hypotensive episode; an episode of heart failure; cardiac arrest; or positive stress test during admission
:::
Patient factors that decreased the likelihood of transfer to any tertiary hospital included history of chronic obstructive pulmonary disease, or bleeding disorder; and having a do not resuscitate (DNR) order during the hospital admission. In addition, the further the distance between primary and tertiary VHA, the less likely patients were to be transferred at all, and the more competitive the market for cardiac care, the less likely that the patient was transferred to a tertiary care hospital.
All three components of integration were significantly associated with transfer to tertiary care, although in different directions. After adjustment for patient and other characteristics, the presence of a VHA staff cardiologist and having a referral coordinator at the tertiary VHA hospital decreased the likelihood of transfer to any tertiary care hospital. In contrast, the presence of a referral coordinator at the primary VHA hospital increased the probability of transfer to a tertiary hospital.
Risk-adjusted association: transfer to tertiary VHA hospital vs. tertiary non-VHA hospital
------------------------------------------------------------------------------------------
The results of this analysis are shown in Table [3](#T3){ref-type="table"}. Patient factors associated with transfer to VHA rather than private tertiary hospital included prior history of percutaneous coronary intervention, and history of chronic heart failure. Patient factors associated with transfer to private rather than VHA tertiary hospital included elevated ST-segment on presenting electrocardiogram, abnormal cardiac enzymes on presentation, and presence of a do not resuscitate order during the hospitalization.
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Results of conditional random effects logistic regression of transfer toVHA tertiary care compared to private tertiary care hospital
:::
**Variable** **Odds ratio** **p-value** **Lower limit 95% CI** **Upper limit 95% CI**
-------------------------------------------------------------------------------- ---------------- ------------- ------------------------ ------------------------
**Patient age 65 and over** 1.42 0.29 0.75 2.71
**Chronic obstructive pulmonary disease** 0.56 0.15 0.25 1.23
**Bleeding disorder** 1.10 0.75 0.60 2.03
**Smoker** 1.14 0.84 0.34 3.77
**Prior percutaneous coronary intervention** 3.67 \<0.001 1.91 7.04
**Chronic heart failure** 2.05 \<0.001 1.43 2.95
**ST segment elevation on presenting electrocardiogram** 0.27 \<0.001 0.14 0.51
**Cardiac enzymes abnormal on presentation** 0.30 0.02 0.11 0.81
**Do not resuscitate during hospitalization** 0.14 \<0.001 0.04 0.54
**In-hospital event\*** 1.47 0.31 0.70 3.08
**Distance from patient home zip code centroid to hospital \>100 miles** 2.10 0.10 0.86 5.10
**Distance from primary VHA to tertiary VHA hospital in miles** 1.00 0.35 0.99 1.00
**Degree of market competition (1 = not competitive; 3 = highly competitive)** 0.19 0.06 0.03 1.05
**VHA cardiologist on site** 1.17 0.85 0.23 6.06
**Tertiary VHA hospital has referral coordinator** 20.62 \<0.001 4.50 94.47
**Primary VHA hospital has referral coordinator** 1.38 0.69 0.27 6.99
\* Presence of at least one of the following adverse events during admission: angina persisting more than 24 hours after admission; a hypotensive episode; an episode of heart failure; cardiac arrest; or positive stress test during admission
:::
The degree of market competition was not significantly associated with transfer to VHA versus private tertiary hospital. Neither of the distance variables were associated with transfer either to VHA or non-VHA tertiary hospitals. Furthermore, only one of the individual integration variables entered separately were significantly associated with likelihood of transfer to tertiary VHA versus private hospital, and although the parameter estimate for the variable indicating presence of a referral coordinator at the tertiary hospital was large and significant, it was very imprecise (i.e. large standard error). This is probably due to the relatively small number of patients included in the estimation (N = 319) and uneven splits among hospitals, clustered by VISN.
Discussion
==========
The goal of this study was to investigate the association between measures of clinical integration of care and transfer of patients with acute coronary syndromes in the VHA. In particular, we evaluated whether structural components of clinical integration, such as the presence of referral coordinators and on-site cardiologists, were associated with patient transfer within and/or outside of the VHA healthcare system. In multivariate analysis, the presence of referral coordinators located at primary care VHA hospitals increased the overall likelihood of transfer of ACS patients. In contrast, having a VHA staff cardiologist available or a referral coordinator at a tertiary VHA hospital significantly decreased the likelihood of any transfer to a tertiary care hospital. Finally, we found that only one of the three integration components, presence of a referral coordinator at the tertiary VHA hospital, was significantly associated with transfer to a tertiary VHA hospital compared to a non-VHA tertiary hospital.
Our finding that referral coordinators at primary care hospitals increase the likelihood of transfer to tertiary care hospitals is consistent with prior studies demonstrating that referral coordinators increase the ease of referral and frequency of transfer \[[@B5],[@B15]-[@B19]\]. Presence of a referral coordinator at the primary hospital means that a knowledgeable staff person, not a physician but usually a clinician such as a nurse, is available to coordinate and facilitate what can otherwise be a very cumbersome process of referral and transfer. This individual usually locates and communicates with tertiary care providers and facilitates paperwork and other processes required for patient transfer.
However, our finding that the presence of a referral coordinator at a tertiary VHA hospital was negatively associated with transfer appears contradictory. It is possible that referral coordinators at the tertiary centers may facilitate consultation, which may, at least for lower risk patients, appropriately reduce the need for transfer. However, it is of some concern that these referral coordinators may be serving in a gatekeeper role with regard to transfer decisions. Future research should focus on the role and decision-making associated with these referral coordinators. Of note, when transfer did occur, the presence of a referral coordinator at the tertiary VHA hospital was positively associated with transfer to VHA facilities rather than non-VHA facilities. This suggests that referral coordinators may function differently with different kinds of patients, decreasing overall transfer rates but facilitating within-system transfer when transfer occurred.
In general, we found that transfers to tertiary care were largely associated with patient characteristics appropriate to transfer: sicker and more urgent patients, except for those for whom more intensive care may not be indicated (e.g. DNR status), were significantly more likely to be transferred. In particular, patients with ST-segment elevation on their presenting electrocardiogram and abnormal cardiac enzymes were significantly more likely to be transferred, most likely for coronary revascularization. These patients are most likely to benefit from revascularization \[[@B5],[@B9]\], and their higher probability of transfer suggests that appropriate triage and risk stratification took place in the primary VHA hospitals providing their care. In addition, we found that these patients were more likely to be transferred to non-VHA tertiary hospitals, presumably because these hospitals were closer to the primary VHA hospital than the affiliated tertiary VHA hospital, indicating appropriate out-of-system transfer for the most urgent patients who could benefit from rapid access to tertiary care. The finding that DNR status appears to be associated with transfer to private tertiary rather than VHA tertiary hospital may be due to small cell size, combined with other characteristics of the small number of patients with that status among those who were transferred at all (9 of 319).
Distance between the patient\'s home and primary VHA hospital was significantly associated with increased likelihood of subsequent transfer to a tertiary care hospital. This may indicate that patients who live further from the hospital take longer to present and are therefore sicker on arrival, leading to the requirement for higher levels of care. Also of interest, distance between primary and tertiary VHA hospitals was significantly associated with a decreased likelihood of transfer, indicating that in situations where primary and tertiary VHA hospitals are further apart, primary VHA hospitals may elect to keep more ACS patients rather than transfer them at all. Future research is needed on the appropriateness of transfer of ACS patients, as it is not clear that variation in transfer based on distance between hospitals represents appropriate variation in care.
The finding that cardiologist availability at the primary VHA hospitals was associated with less transfer to tertiary care hospitals may reflect that local or distant cardiology consultation was sufficient in some cases (e.g. lower risk patients) to avoid transfer. Similarly, the availability of a transfer coordinator at the tertiary VHA hospital may have provided an avenue for consultation and avoidance of transfer in some cases. Future studies are needed to define the mechanisms of association between reduced transfers and both on-site cardiology availability and tertiary hospital transfer coordinators.
The findings of this study, that referral coordination is associated with transfer from primary to tertiary hospitals, but may operate differently for different types of patients, and may have one mechanism of operation within a health care system and another outside that system, have potential application outside VHA. Previous studies \[[@B20]\] have found that patients\' access to needed services, such as revascularization after acute myocardial infarction, has a significant effect on mortality outcomes. Services such as referral coordination, which increase the likelihood that a patient will be transferred, can reduce the negative impact of receiving initial care in a hospital without specialized tertiary services, such as cardiac catheterization. These findings are potentially relevant in all health care systems where hospitals have different levels of service. Even though they are based on a relatively small patient sample size, the implications of the findings -- that referral coordinators at primary hospitals increase the probability of transfer, with the link to better outcomes at tertiary centers \[[@B21]\] with a full range of treatment options -- should spark discussion in a health care system such as VHA about recommending use of referral coordinators in primary hospitals.
Limitations
-----------
First, we were not able to conduct full-scale validation and reliability testing of the clinical integration measures, which would have required a larger sample of hospitals participating in the study to conduct split-sample validation. Second, we used structural, rather than process, elements of integration in this analysis. We focus on structural elements both because they are relatively easier to measure (present or not), and because in Donabedian\'s widely accepted model of quality in health care, structure precedes process and outcome \[[@B22],[@B23]\]. Third, clinical integration is a complex multi-faceted construct which we captured in a relatively simplistic way. However, we wanted to see if measures that would be straightforward to implement in a health care system like the VHA, such as referral coordinators, had an impact on this key process of care. We measured other components of integration, including communication methods, provider satisfaction with communication methods, and overall perception of how well referral and consultation worked in providing care to ACS patients. Individually, these factors were not as strongly linked to the transfer process as the three structural components we present in this analysis.
Fourth, because transfer is closely related to patient outcomes, especially for ACS patients \[[@B21]\], careful modeling of the relationship between transfer and mortality and morbidity outcomes is essential. We plan to conduct future analyses on the relationships between patient characteristics, transfer, and mortality and morbidity outcomes. In addition, it is important to note that most veterans over the age of 65 are dually eligible for Medicare as well as VHA benefits, and previous analyses have shown that a majority of veterans with acute myocardial infarction, even among those who use VHA hospitals, receive care for AMI in private hospitals \[[@B24],[@B25]\]. This study was designed only to assess transfer of veterans who went to primary VHA hospitals for their ACS care.
Conclusions
===========
We found that referral coordinators located at primary care VHA hospitals increase the overall likelihood of transfer of ACS patients. Referral coordinators at tertiary VHA hospitals and the presence of on-site cardiologists appeared to decrease the likelihood of transfer. Only one component of integration, presence of a referral coordinator at the tertiary hospital, was associated with within-system compared to out-of-system transfer. These findings have significant potential implications for the VHA. One of the goals of an integrated health care system is to maintain optimal coordination between its component parts \[[@B12]\]. This study demonstrates that simple structural components of care, such as a referral coordinator at either a primary or tertiary care hospital, can have an impact on a key process of care above and beyond patient characteristics.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
AES participated in the design and conduct of the study, conducted the analyses and wrote the manuscript. SLP participated in conducting the project, and assisted in writing the manuscript. DJM participated in writing the manuscript. NRE participated in the design and conduct of the study. NDS participated in writing the manuscript. JSR participated in the statistical analyses and co-wrote the manuscript. All authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1472-6963/5/2/prepub>
Acknowledgements
================
This study was funded by the VA Health Services Research and Development Service, ACC 97-079. The views expressed in this paper do not reflect the views of the Department of Veterans Affairs.
|
PubMed Central
|
2024-06-05T03:55:52.005517
|
2005-1-13
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545996/",
"journal": "BMC Health Serv Res. 2005 Jan 13; 5:2",
"authors": [
{
"first": "Anne E",
"last": "Sales"
},
{
"first": "Sandra L",
"last": "Pineros"
},
{
"first": "David J",
"last": "Magid"
},
{
"first": "Nathan R",
"last": "Every"
},
{
"first": "Nancy D",
"last": "Sharp"
},
{
"first": "John S",
"last": "Rumsfeld"
}
]
}
|
PMC545997
|
Background
==========
Programs that promote condom use for HIV prevention typically monitor their progress through survey-based indicators, such as the percentage of the population who ever used a condom or the percentage who used a condom in their last sex act with a casual or regular partner \[[@B1],[@B2]\]. Such information is routinely collected in national surveys, such as the Demographic and Health Surveys (DHS) and the CDC Reproductive Health Surveys \[[@B3],[@B4]\]. In addition, HIV prevention programs often monitor the number of condoms sold and/or the number distributed free of charge.
The purpose of this study is to explain inconsistencies between information on reported levels of condom use and data on the number of condoms sold and distributed. Understanding the apparent inconsistencies between sales and survey data will help clarify to what extent the concerns about condom wastage, misreporting, and other related problems are founded. It will also provide guidance for improving the monitoring of condom sales and distribution, and for improving survey questionnaires. To achieve these objectives, we use survey data from six Demographic and Health Surveys to estimate the total annual number of sex acts in a country, and the total number of condoms used in those sex acts, and compare the totals with reported data on condom sales and distribution.
At least in some instances, survey information on condom use and condom sales records appear to be inconsistent \[[@B5],[@B6]\]. For example, in some countries we observe steady increases in reported condom sales while survey indicators suggest that there has been no significant increase in the percentage of condom use in last sex across survey rounds. In Zimbabwe, sales of socially marketed *Protector Plus*condoms increased from 1.9 million in 1997, to 4.8 million in 1998, to 8.9 million in 1999. Data on public sector condom distribution, which we discuss later in this paper, suggest that public sector sales also increased substantially. Yet, nationally representative surveys indicate that condom use in last sex stayed constant between 1996 and 1999 at roughly 34% for males and 17% for females \[[@B6],[@B7]\]. Similarly, in Tanzania, sales of socially marketed *Salama*condoms increased steadily between 1995 and 2000, as did condom distribution by the Ministry of Health. However, survey data indicate that condom use at last intercourse remained roughly constant between 1996 and 1999, for both men and women \[[@B5]\]. These discrepancies suggest that either the data on reported levels of condom use or the data on condom sales and distribution are inaccurate, or possibly that both are inaccurate.
Inaccuracies in the number of condoms sold or distributed are likely because sales figures typically represent sales to the trade (i.e., sales to wholesalers and distributors) rather than sales to consumers. Consequently, the recorded sales numbers will include condoms that are being stocked at various levels of the distribution chain. In addition, some of the condoms that are sold and/or distributed may be wasted or smuggled to other countries.
In addition to these potential problems with condom sales data, there are concerns that reported condom use in surveys may be inaccurate. For example, there are concerns that respondents may overreport condom use because they do not want to admit to the interviewer that they are engaging in risky sexual behavior. There are also concerns that condom use may be underreported because condoms are frequently used with sex workers, which stigmatizes condom use. Women may also underreport condom use because it is a male method. Some questionnaires try to overcome this by asking \"The last time you had intercourse, was a condom used?\" rather than \"The last time you had intercourse, did you use a condom?\" \[[@B3]\].
Methods
=======
Sources of data
---------------
This study uses two types of data: data on condom sales and distribution, and survey data on self-reported condom use. We restrict our analysis to data from four countries in sub-Saharan Africa (Kenya, Tanzania, Nigeria, and Zimbabwe), largely because these countries have strong condom social marketing programs and therefore relatively good data on condom sales and distribution. In addition, Tanzania and Zimbabwe are two of the countries where discrepancies between condom distribution and condom use have been noted.
Data on sales of socially marketed condoms were obtained from DKT International\'s Social Marketing Statistics \[[@B8]-[@B14]\], while data on donor-supplied public sector condoms were obtained from UNFPA and USAID \[[@B15],[@B16]\]. Data on commercial condom sales are not readily available, but for recent years very rough estimates were obtained from Population Services International\'s MIS database \[[@B17]\]. As commercial sales tend to be negligible in the countries under consideration, the lack of accurate data on commercial sales is unlikely to have a significant effect on our findings.
The survey data used in this study include the following Demographic and Health Surveys (DHS): Kenya (1998), Nigeria (1999), Tanzania (1996, 1999), and Zimbabwe (1994, 1999). Each of the six surveys comprises a representative sample of females aged 15--49 and of males 15--54 (note that the upper age limit varies for men, see Table [1](#T1){ref-type="table"}). For more detailed information on the sampling methods and the data collection, we refer the reader to the DHS reports for these surveys \[[@B18]-[@B23]\].
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Data available in selected DHS surveys on frequency of intercourse and probability of condom use
:::
Country Year Sex Age range Time since last intercourse Frequency of intercourse Condom use during last intercourse Frequency of condom use
---------- ------ ------- ----------- ----------------------------- -------------------------- ------------------------------------ -------------------------
Kenya 1998 Men 15--54
Women 15--49
Nigeria 1999 Men 15--64
Women 15--49^1^
Tanzania 1996 Men 15--54
Women 15--49
1999 Men 15--59
Women 15--49
Zimbabwe 1994 Men 15--54
Women 15--49
1999 Men 15--54
Women 15--49
Note: ^1^The age range for women in the 1999 NDHS is 10 to 49. To enhance comparability, we restricted our analysis to women aged 15 to 49.
:::
Determining the total annual number of condoms used in a population requires information on the frequency of intercourse. Unfortunately, recent sexual behavior surveys typically do not allow the quantification of the number of sex acts \[[@B24]\].
While some of the DHS surveys from the late 1980s and early 1990s did ask respondents about the frequency of intercourse in a fixed time interval (e.g., frequency of intercourse in the past month), such a question has not been included in recent surveys \[[@B25]\]. For example, the standard questionnaire for DHS surveys implemented since 1997 does not include a question on the frequency of intercourse. In the surveys included in our study, the 1994 Zimbabwe survey was the only one that included a question on the self-reported frequency of intercourse (see Table [1](#T1){ref-type="table"}). However, the DHS surveys do ask respondents about the time since they last had intercourse \[[@B3],[@B26]\]. Hence, our analysis estimates the total annual number of sex acts on the basis of reported data on time since last intercourse \[[@B5],[@B27],[@B28]\]. Depending on the survey, it may or may not be possible to differentiate the frequency of intercourse by partner type. Differentiation by partner type may be important, as it is believed that men who admit having a nonmarital partner are unlikely to underreport the frequency of intercourse \[[@B24]\].
All DHS surveys asked whether respondents used a condom in their last sex act. We use this information to estimate the probability of condom use, and, subsequently, to estimate the total annual number of condoms used in the country.
General estimation procedure
----------------------------
In theory, estimating the total number of condoms used in a population is straightforward. The estimated mean number of condoms used per sexually active person (*C*) equals the product of the frequency of intercourse, or the number of sex acts (*F*), and of the probability of condom use (*p*):
*C*= *F*× *p* (1)
The total number of condoms used (*C*^*T*^) then can be calculated by multiplying *C*with the proportion of individuals who are sexually active (*s*) in the population at risk and with size of the population at risk (*N*):
*C*^*T*^= *N*× *s*× *C* (2)
Since the frequency of intercourse and the probability of condom use are known to vary by age and marital status \[[@B25],[@B27],[@B29]-[@B32]\], it is advisable to estimate these coefficients separately for various subpopulations and subsequently to calculate a weighted average for the entire population. In this paper, we stratified our estimates by the respondents\' age and marital status. The formula to calculate the mean annual number of condoms used per sexually active respondent is:
where *w*~*a*~is the weight for age group *a*, *m*~*a*~and (1 - *m*~*a*~) are the proportion of married and unmarried respondents in age category *a*, and *s*~*am*~and *s*~*au*~are the proportion of sexually actives where the subscripts *am*and *au*refer to the rates for married and unmarried respondents in age category *a*, respectively.
We used five-year age categories, and based the age weights on the age distribution within the household file of the DHS, as no other reliable data on the age structure of the population in these countries were available (preliminary analyses with one-year age groups produced similar results). Marital status and the marital status weights were derived from the individual respondent files of the DHS. Following the DHS definition, we define marriage as formal marriage or living together. Information on current sexual activity, defined as having had sex at least once in the past year, was also obtained from the individual respondent files of the DHS. Data on the countries\' population size were obtained from the 2003 World Bank World Development Indicators and are summarized in Appendix \[see [Additional File 1](#S1){ref-type="supplementary-material"}\].
Although the above procedure is simple, data on the two main components, *F*and *p*are not readily available and need to be estimated. The following sections describe the procedures for estimating them.
Methods for estimating frequency of intercourse
-----------------------------------------------
This section describes methods to estimate frequency of intercourse. Three types of estimation methods are presented: 1) estimation based on the reported frequency of intercourse during a four-week period, 2) methods based on the proportion of respondents reporting intercourse the day before the interview, and 3) survival analyses based on the time since last intercourse.
All methods follow a similar strategy: 1) Estimate the mean likelihood or frequency of intercourse for a specific time unit (e.g., for a day, one week, or four weeks) for each of the subpopulations, and 2) estimate the mean frequency of intercourse per year for the entire population by calculating a weighted average of the subpopulation results. The general formula is:
where *F*~*i*~stands for the annual frequency of intercourse estimated by method *i*, *f*~*iam*~and *f*~*iau*~for the estimated mean likelihood or frequency of intercourse per time unit using method *i*for married and unmarried persons in age category *a*, respectively, and *n*~*i*~the number of time units for this method in a year.
Some surveys asked married respondents separate questions about the time since last intercourse with the respondents\' spouse and with the respondents\' other partners. Such questions were included in the 1998 Kenya and 1996 Tanzania DHS surveys. For these surveys, the formula becomes:
where the *b*subscript in *F*~*ib*~indicates that for married respondents marital and extramarital sex were included separately.
### Method F~1~
When self-reported data on the frequency of intercourse during the past four weeks are available, such as in the 1994 Zimbabwe DHS survey, the annual number of sex acts can be estimated by extrapolation. Assuming the past four weeks are representative of the respondents\' behavior, the mean annual number of sex acts can be estimated by multiplying this four-week frequency with 13 (*n*~1~= 13). However, because few recent surveys contain this type of information, it is generally necessary to use other estimation methods.
### Method F~2~
The frequency of intercourse can be estimated on the basis of the proportion of respondents reporting intercourse the day before the interview \[[@B5]\]. Assume each of a group of individuals has 104 sex acts per calendar year (i.e., two sex acts per week). Assuming one sex act per day that intercourse occurs, the probability of intercourse on any given day during the calendar year would equal 104/365, or 0.285. Hence, it is expected that, on average, 28.5% of the population will have intercourse on any given day.
In other words, the proportion of the population reporting intercourse on any given day equals the daily probability of intercourse. Therefore, the annual number of sex acts can be estimated by multiplying the proportion of respondents who had intercourse the day before the interview by 365.
The advantage of this method is that it is simple to calculate, and that use of data that refer to the day before the interview minimizes recall problems. The disadvantage is that the method does not take into account that some people may have more than one sex act in a day (i.e., only one of those sex acts will be counted), so that the frequency of intercourse may be slightly underestimated. In turn, the impact of this more frequent intercourse on condom use may be somewhat greater than results would indicate, as the uncounted numbers may represent commercial sex workers with a relatively high condom use.
Another problem with this method is that for some surveys the percentage of respondents reporting last having intercourse the day before the survey does not appear to be reliable. For example, in the 1998 Kenya survey the percentage of respondents reporting last having sex one day before the survey was smaller than the percentage last having sex two days before the survey (4.1% vs. 8.9%). Similarly, in the 1999 Nigeria survey 0.7% reported last having intercourse one day before the survey, compared to 10.0% who reported having sex two days before the survey. In the other surveys, the percentage reporting last having sex the day before the survey is slightly higher than the percentage last having sex two days before the survey. While it is unclear why so few respondents in the Kenya and Nigeria surveys reported last having intercourse the day before the survey, the implication is that the *F*~2~estimates for these surveys appear to be unrealistically low.
### Method F~3~
A third alternative is to estimate frequency of intercourse based on data on the duration since last intercourse, which is collected in all DHS surveys \[[@B27],[@B28]\]. This group of techniques is based on the fact that mean duration between two successive acts of intercourse provides an estimate of the frequency of intercourse. The major difficulty with this approach is that the duration between two successive sex acts is a closed interval, while the available data -- duration since last intercourse -- is an open interval.
Slaymaker and Zaba \[[@B28]\] deal with this inconsistency by using survival analyses with an exponential decay function. The survival analysis estimates the daily probability of intercourse. The estimated annual number of sex acts is obtained by multiplying the average daily probability of intercourse by 365.
One of the main weaknesses of this approach is the assumption that daily probability of intercourse is constant and can be estimated with an exponential decay function. Since data on the actual distribution of the intervals between two successive sex acts are not available in DHS surveys, one cannot determine whether the exponential decay function provides a good fit for the data. Using a function that does not match the data well would introduce a very large error in the estimated annual number sex acts (and consequently in the estimated number of condoms used), rendering the results meaningless.
Methods for estimating the probability of condom use
----------------------------------------------------
As most DHS surveys only contain data on whether a condom was used in the respondent\'s last intercourse, we must assume that condom use at last sex is typical for the likelihood of condom use for a given subpopulation. Three different estimations for the likelihood of condom use are explored in this paper, two of which are based on data on condom use at last intercourse and one of which is based on the self-reported frequency of condom use.
### Method p~1~
For surveys that collected information on the frequency of condom use, this information can also be used to estimate the probability of condom use. Unfortunately, none of the DHS surveys asked direct questions about both the number of sex acts and the number of condoms used (for an example of a survey that collects such data, see \[[@B33]\]. However, some DHS surveys did ask respondents how frequently they used condoms. For example, the 1994 Zimbabwe DHS first established how often respondents had sex with their spouse and other partners in the past four weeks. Next, respondents were asked, \"Was a condom used on any of these occasions?\" Respondents who answered that a condom was used were asked, \"Was it each time or sometimes?\" Hence the frequency of condom use was coded as \"Yes, each time,\" \"Yes, sometimes,\" or \"Never.\" To obtain an estimate for the probability of condom use for each of these categories, we cross-tabulated this reported frequency of condom use against condom use in last intercourse. The results showed that 93% of men claiming to always use condoms reported using a condom in last intercourse. Similarly, 44% of those claiming to sometimes use condoms and 2% of those claiming to never use condoms reported that they had used a condom in last intercourse. Thus, we recoded the three categories for frequency of condom use among men as 0.93, 0.44, and 0.02. For women, the values were 0.94, 0.47, and 0.01, respectively. The probability of condom use was then calculated as the mean value for each of the sub-samples.
### Method p~2~
The first estimate of the probability of condom use simply equals the proportion of a sub-sample (by age and marital status) who reported using a condom at last intercourse. This estimate was also used by Collumbien et al. \[[@B24]\]. Information on condom use in last intercourse is available in all DHS surveys. For surveys that collected data on condom use at last intercourse by partner type, such as the 1998 Kenya and 1996 Tanzania DHS surveys, taking this information into account can refine the estimate of the probability of condom use.
### Method p~3~
An alternative measure of the probability of condom use equals the proportion of respondents who reported using a condom at last sex among those who had sex the previous day. This indicator has the advantage that it is less likely to be subject to recall errors. It also avoids the problem that condom use at last intercourse may be dependent on the time since last intercourse. However, this measure has the disadvantage that it tends to be less reliable because it is based on information from a much smaller number of observations (those reporting intercourse the day before the interview).
Estimating the annual number of condoms used
--------------------------------------------
We estimate the annual number of condoms used by multiplying the annual number of sex acts with the probability of condom use for each of the strata by age and marital status, as described in Equation 3. Because we have three different methods to estimate the annual number of sex acts and three methods to estimate the probability of condom use, up to nine estimates of the annual number of condoms used are provided, depending on the available data. Moreover, separate estimates were calculated using data from the female and male DHS surveys, as there are known gender differences in the reported frequency of intercourse and levels of condom use \[[@B30],[@B32],[@B34]\].
Results
=======
Reported condom sales and distribution
--------------------------------------
Figure [1](#F1){ref-type="fig"} shows trends in the annual number of condoms sold or distributed in Kenya, Nigeria, Tanzania, and Zimbabwe. Although these statistics represent the number of condoms sold or distributed to the trade (i.e., to distributors, wholesalers, and retailers), it is often assumed that they will mimic sales to consumers, because the trade is unlikely to re-stock unless there is sufficient consumer demand.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Annual number of condoms sold and distributed, by country
:::
:::
Figure [1](#F1){ref-type="fig"} reveals very erratic patterns in the number of condoms sold or distributed in each of the four countries. The most dramatic pattern is observed for Nigeria. The total number of condoms distributed in Nigeria increased from 13 million in 1989 to 42 million in 1990, but then declined to 14 million in 1992. Between 1992 and 1994, condom distribution increased rapidly to 83 million, and by 1995, Nigerian condom sales jumped to 227 million. However, the very next year the number of condoms distributed dropped back to 103 million and continued to decline to 68 million in 1998. In 1999, condom sales rapidly increased to 108 million. The trend in the number of condoms distributed in Kenya is equally erratic. In Kenya, the total number of condoms distributed increased rapidly from 17 million in 1989 to 39 million in 1992, to 97 million in 1995. However, from 1996 onward, the number of condoms distributed dropped dramatically, to reach only 12 million in 1998. By 1999, condom distribution jumped to 79 million. The number of condoms distributed in Tanzania and Zimbabwe is considerably lower, but also shows very large year-to-year fluctuations.
It is clear that these drastic fluctuations in the number of condoms sold or distributed do not reflect real differences in the level of condom use, as this would require major changes in behavior (and behavior is known to change very slowly). Since statistics on the number of condoms sold or distributed reflect sales to the trade, not consumers, it is highly likely that the observed fluctuations in the number of condoms distributed simply reflect fluctuations in condom inventory due to a stock-up of condoms at one or more levels of the distribution system, the addition of new condom outlets, and so on. For example, data from condom distribution surveys in Kenya indicate that the percentage of retail outlets that were selling socially marketed *Trust*condoms increased from 25% in 1998 to 32% in 1999. Similarly, the percentage of retail outlets selling public sector condoms increased from 2% to 6%. The percentage of retail outlets selling other brands stayed constant at 3% \[[@B35],[@B36]\]. Assuming that outlets sell only one type of condoms, the percentage of retail outlets selling any type of condom increased from 30% to 41%, which implies that that the total number of retail outlets that sell condoms may have increased by as much as 37% (= 41/30 \* 100) in just one year. Such an increase in the number of retail outlets that carry condoms would require a substantial increase in the number of condoms sold to the trade in order to fill the pipeline (i.e., to supply national and regional distributors, wholesalers, and retailers).
In addition, our estimates of the number of public sector condoms are not the actual number of public sector condoms distributed to the population, but rather the total number of condoms provided to each country by international donors. It is possible that many of these condoms are still stocked at Ministry of Health warehouses and similar distribution hubs, or at local health clinics. The actual number of public sector condoms that reach the hands of consumers is unknown. Therefore, the data that are available on the number of condoms that have been sold or distributed seem to provide an estimate of the total of number of condoms that were in circulation during the course of the year, rather than the number provided to consumers.
In other words, the current data on the number of condoms sold or distributed provide a very poor estimate of the actual number of condoms used. For example, as shown in Figure [1](#F1){ref-type="fig"}, condom distribution in Nigeria peaked at 227 million in 1995. However, condom distribution subsequently dropped to a level far below that of the period preceding the peak. This drop-off in sales to the trade between 1995 and 1997 suggests that some of the 227 million condoms sold to the trade in 1995 were not sold to consumers until 1996 or 1997, if not later. Hence, changes in condom sales do not necessarily indicate any changes in condom use. Measuring changes in the level of condom use requires either collecting data on retail sales, which is not feasible in most developing countries, or using sample surveys to measure the level of condom use.
Estimated annual number of sex acts
-----------------------------------
Table [2](#T2){ref-type="table"} summarizes the results of different estimates for the mean annual frequency of intercourse for both male and female samples in the six DHS surveys used. We first discuss the results from the 1994 Zimbabwe DHS survey, for which all three methods for estimating the per capita annual number of sex acts could be calculated. Hence, these data are ideal for comparing the estimate based on self-reported data, *F*~1~, with the two estimates based on the duration since last intercourse (*F*~2~and *F*~3~). Next, we discuss the results for the other surveys, for which only methods *F*~2~and *F*~3~could be estimated.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Estimated annual number of sex acts (mean number per sexually experienced respondent)
:::
Country Year Sex Marital Status N of Cases Proportion Currently Sexually Active Estimation Method
---------- ------ ------- ---------------- ------------ -------------------------------------- ------------------- ------ ------
Kenya 1998 Men Unmarried 1,644 66.1% -.- 4.8 6.2
Married 1,763 98.2% -.- 22.4 16.3
All 3,407 82.7% -.- 15.8 12.4
Women Unmarried 3,034 40.3% -.- 0.9 2.6
Married 4,847 93.5% -.- 16.5 9.2
All 7,881 73.0% -.- 13.2 7.8
Nigeria 1999 Men Unmarried 1,072 42.9% -.- 0.8 5.6
Married 1,608 92.0% -.- 3.4 7.2
All 2,680 72.4% -.- 2.7 6.8
Women Unmarried 4,002 34.5% -.- 2.2 3.5
Married 5,808 82.0% -.- 6.2 4.6
All 9,810 67.8% -.- 5.6 4.5
Tanzania 1996 Men Unmarried 985 43.0% -.- 13.8 8.7
Married 1,268 92.0% -.- 51.8 7.9
All 2,256 70.6% -.- 41.6 8.1
Women Unmarried 2,715 31.2% -.- 14.2 5.3
Married 5,404 86.2% -.- 49.7 5.5
All 8,120 67.8% -.- 44.2 5.4
1999 Men Unmarried 1,544 57.6% -.- 7.0 5.0
Married 1,998 98.1% -.- 48.9 15.6
All 3,542 80.5% -.- 35.9 12.3
Women Unmarried 1,421 47.0% -.- 7.7 3.6
Married 2,608 96.7% -.- 48.5 10.2
All 4,029 79.2% -.- 39.9 8.9
Zimbabwe 1994 Men Unmarried 1,126 53.3% 20.9 8.4 4.2
Married 1,015 99.3% 81.9 60.9 17.0
All 2,141 75.1% 59.4 41.6 12.3
Women Unmarried 2,349 36.5% 9.3 9.3 2.7
Married 3,777 94.9% 82.2 70.3 9.7
All 6,128 72.5% 68.1 58.5 8.3
1999 Men Unmarried 1,406 48.1% -.- 7.9 3.6
Married 1,203 99.4% -.- 57.9 23.3
All 2,609 71.8% -.- 40.4 16.4
Women Unmarried 2,354 38.9% -.- 2.4 2.4
Married 3,553 99.0% -.- 43.7 13.8
All 5,907 75.0% -.- 35.1 11.4
:::
The results from the 1994 Zimbabwe survey show that the three estimation methods yield very different estimates of the annual number of sex acts. Estimates based on the self-reported number of sex acts in the past four weeks (*F*~1~) give the highest estimates. Using this method, it is estimated that in 1994, sexually active unmarried males in Zimbabwe had 21 sex acts per year, while sexually active married men had 82 sex acts per year. For females, the number of sex acts is estimated at 9 per year for unmarried females and 82 for married females. This latter finding is fairly consistent with Brown (2002), who estimated the coital frequency for sexually active married women at 7.9 acts per month, which translates into 95 acts per year.
The second estimation method (*F*~2~), which is based on the proportion of respondents who reported having intercourse the day before the interview, results in an estimate of 8 sex acts per year for unmarried males, 61 for married males, 9 for unmarried females, and 59 for married females. Thus, this estimate consistently yields a lower estimate of the number of sex acts than the estimate based on the self-reported frequency of intercourse. This difference appears to be especially large for unmarried males.
The third estimation method (*F*~3~), which is based on a survival analysis using the assumption of a constant hazard, yields substantially lower estimates of the per capita annual number of sex acts. For unmarried males, the annual number of sex acts is estimated at only 4, while for married males it is estimated at 17. For females, the corresponding numbers are 3 and 10 per year, respectively. These estimates do not appear to be realistic.
For all other surveys examined here, we can also compare the estimates based on the proportion reporting intercourse the day before the survey (*F*~2~) and those based on the survival analysis with the assumption of a constant hazard (*F*~3~). The results confirm that this latter method consistently yields very low estimates of the number of sex acts. For example, among sexually active married males, the estimate of the annual number of sex acts ranges from 7.2 coital acts per year in the 1999 Nigeria survey to 23.3 in the 1999 Zimbabwe survey. For sexually active married females, the range is from 4.6 to 13.8, again in those same surveys. In other words, the results from the survival analysis using the assumption of a constant hazard suggest that in several countries, even married couples have intercourse less than once per month. Method F2 tends to yield higher estimates of the annual number of sex acts, but for both the 1998 Kenya and 1999 Nigeria surveys these estimates are also unrealistically low. In these latter cases, the low estimates are due to the fact that the number of respondents reporting last having intercourse the day before the survey is considerably lower than the number reporting last having intercourse two days ago.
The results based on the survival analyses appear unrealistic and are inconsistent with the published literature on the frequency of intercourse. For example, a study on coitus in sub-Saharan Africa estimates that the monthly coital frequency among sexually active married women ranges from 3.0 in Ghana to 8.1 for Rwanda \[[@B37]\], which corresponds with an annual frequency of 36 and 97 acts, respectively. Similarly, another study estimates the monthly coital frequency among married women at 6.1 act for Burundi, 3.0 for Kenya, and 5.7 for Uganda. Only Ghana has a substantially lower frequency of intercourse, at an average of 1.2 per coital acts per month \[[@B25]\]. The same study estimates that monthly coital frequency in Latin America ranges from 3.2 in Mexico to 8.0 in Brazil. A study on sexual activity among young women in Africa estimates the average number of sex acts in the past four weeks among women aged 15--24 in Kenya at 1.9 for the never married, and at 4.0 for the married. The corresponding data for Ghana are 0.7 and 1.0, respectively \[[@B29]\]. Hence, there is reason to believe that the results from the survival analysis are unreliable. (It is noteworthy that the results for Nigeria are substantially lower than those for the other countries, for both *F*~2~and *F*~3~, largely because a substantially lower percentage of respondents reported having intercourse the day before they survey. Since the percentage reporting intercourse on other days is more in line with the results from the surveys in other countries, we suspect that this inconsistency is the result of a coding error.)
It is important to note that the results of the survival analyses are greatly affected by the type of decay function selected. Preliminary analysis using a Weibull decay function yielded estimates of the annual number of sex acts that are roughly one and a half to two times as high as estimates based on the exponential decay function proposed by Slaymaker and Zaba \[[@B28]\]. Unfortunately, determining which decay function to use requires information on the distribution of the length of the interval between two successive coital acts, and such information is not available in the DHS surveys.
Probability of condom use
-------------------------
The estimates of the probability of condom use are shown in Table [3](#T3){ref-type="table"}. As before, the three estimates of the probability of condom use could be calculated only for the 1994 Zimbabwe survey. Moreover, since the self-reported frequency of condom use was coded as \"each time,\" \"sometimes,\" or \"never,\" we estimated the frequency on the basis of the proportion of each of these categories who reported using a condom in last intercourse. Thus, the estimates for *p*~1~and *p*~2~are nearly identical (although some differences exist when differentiating by marital status).
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Estimated probability of condom use per sex act
:::
Country Year Sex Marital Status N of Cases Estimation Method
---------- ------ ------- ---------------- ------------ ------------------- ------- -------
Kenya 1998 Men Unmarried 1,644 -.- 40.8% 40.3%
Married 1,763 -.- 9.1% 4.9%
All 3,407 -.- 21.1% 18.3%
Women Unmarried 3,034 -.- 17.2% 0.0%
Married 4,847 -.- 5.2% 3.0%
All 7,881 -.- 7.7% 2.4%
Nigeria 1999 Men Unmarried 1,072 -.- 39.2% 0.0%
Married 1,608 -.- 6.1% 9.2%
All 2,680 -.- 14.6% 6.9%
Women Unmarried 4,002 -.- 22.1% 7.9%
Married 5,808 -.- 2.9% 5.4%
All 9,810 -.- 5.8% 5.8%
Tanzania 1996 Men Unmarried 985 -.- 34.5% 15.9%
Married 1,268 -.- 5.5% 2.3%
All 2,256 -.- 13.3% 6.0%
Women Unmarried 2,715 -.- 16.1% 6.2%
Married 5,404 -.- 2.0% 1.0%
All 8,120 -.- 4.2% 1.8%
1999 Men Unmarried 1,544 -.- 33.1% 23.4%
Married 1,998 -.- 7.9% 3.2%
All 3,542 -.- 15.7% 9.5%
Women Unmarried 1,421 -.- 20.6% 7.5%
Married 2,608 -.- 3.8% 3.4%
All 4,029 -.- 7.3% 4.3%
Zimbabwe 1994 Men Unmarried 1,126 46.0% 53.6% 35.7%
Married 1,015 13.9% 12.1% 6.8%
All 2,141 25.8% 27.5% 17.5%
Women Unmarried 2,349 31.8% 30.7% 19.1%
Married 3,777 5.6% 5.9% 5.0%
All 6,128 10.7% 10.7% 7.7%
1999 Men Unmarried 1,406 -.- 65.6% 63.6%
Married 1,203 -.- 8.5% 5.1%
All 2,609 -.- 28.5% 25.5%
Women Unmarried 2,354 -.- 32.6% 19.7%
Married 3,553 -.- 4.4% 1.9%
All 5,907 -.- 10.3% 5.6%
:::
When we compare the different methods to estimate the likelihood of condom use we notice that in the overwhelming number of cases the estimates based on the proportion reporting condom use at last intercourse of those who reported sex on the day before the interview (*p*~3~) are lower than those based on the data from the last sex act (*p*~2~). For example, in the 1999 Tanzania survey, the proportion who used a condom in last intercourse is 15.7% for males and 7.3% for females. By contrast, of those who had sex the day before the interview, the proportion who used a condom is only 9.5% and 4.3%, respectively. In part, these low estimates of *p*~3~appear to stem from the fact that only a small number of survey respondents reported having intercourse the day before the interview. Consequently, there are some age groups where none of the respondents reported using a condom (not shown), which substantially lowers the estimate of the overall probability of condom use.
The results shown in Table [3](#T3){ref-type="table"} also indicate that the likelihood of having used condoms is substantially higher among unmarried than among married respondents. This finding is consistent with the literature \[[@B7],[@B28],[@B30],[@B32]\] and thus confirms that our stratification by marital status was necessary, as the two groups also substantially differ in frequency of intercourse.
As other authors also have noted, women tend to report a much lower likelihood of condom use than men \[[@B21],[@B31],[@B32]\]. For example, Table [3](#T3){ref-type="table"} shows that in the 1999 Zimbabwe survey 29% of men but only 10% of women reported using a condom in last intercourse. Similarly, in the 1998 Kenya survey, 21% of men but only 8% of women reported using a condom in last intercourse. These differences persist when differentiating by marital status.
It is noteworthy that some gender discrepancies in the probability of condom use would be expected because African men may have sexual partners who are substantially younger. If the age difference between partners explained the gender differential in the probability of condom use, then we would expect that the probability of condom use for males aged 30--34 should be closer to that of women aged 25--29 or 20--24. Several data sets show that these probabilities are indeed closer, but the differences remain very large \[[@B21],[@B31]\]. As most condoms are used in heterosexual sex acts, this discrepancy constitutes a serious problem when estimating overall condom use, because there is no way of verifying which of the two estimates provides the best estimate of the true probability of condom use.
Estimated annual number of condoms used
---------------------------------------
Table [4](#T4){ref-type="table"} shows the estimates of the total annual number of condoms used based on different combinations of estimates for the frequency of intercourse and the probability of condom use. To facilitate interpretation, the bottom panel of the table also provides the highest and lowest estimates. For comparison, we also added data on the reported number of condom sales in the survey year, and in the year prior to the survey.
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Estimated annual number of condoms used
:::
Estimation Method Kenya 1998 Nigeria 1999 Tanzania 1996 Tanzania 1999 Zimbabwe 1994 Zimbabwe 1999
----------------------- ---------------------- -------------- --------------- --------------- --------------- --------------- ------------
**Males**
F~1~Self-Reported p~1~Self-Reported -.- -.- -.- -.- 18,047,620 -.-
p~2~Last Intercourse -.- -.- -.- -.- 19,451,694 -.-
p~3~Previous Day -.- -.- -.- -.- 11,408,033 -.-
F~2~Previous Day p~1~Self-Reported -.- -.- -.- -.- 12,209,655 -.-
p~2~Last Intercourse 10,650,977 5,522,394 14,919,839 19,053,896 11,515,528 10,850,758
p~3~Previous Day 7,734,312 6,779,088 6,231,789 9,805,457 6,275,443 7,660,061
F~3~Survival Analysis p~1~Self-Reported -.- -.- -.- -.- 4,136,103 -.-
p~2~Last Intercourse 10,121,645 18,858,423 4,891,365 7,493,313 3,999,271 4,468,660
p~3~Previous Day 7,221,404 10,010,100 2,439,635 3,754,680 2,324,967 3,262,927
**Females**
F~1~Self-Reported p~1~Self-Reported -.- -.- -.- -.- 7,980,256 -.-
p~2~Last Intercourse -.- -.- -.- -.- 8,406,142 -.-
p~3~Previous Day -.- -.- -.- -.- 7,088,876 -.-
F~2~Previous Day p~1~Self-Reported -.- -.- -.- -.- 6,913,439 -.-
p~2~Last Intercourse 3,375,708 4,632,093 5,529,321 10,744,128 7,253,275 3,700,789
p~3~Previous Day 2,091,845 7,622,258 2,759,809 8,422,675 6,115,040 1,591,401
F~3~Survival Analysis p~1~Self-Reported -.- -.- -.- -.- 1,111,439 -.-
p~2~Last Intercourse 2,200,502 4,503,194 993,705 2,756,648 1,137,474 1,395,517
p~3~Previous Day 986,769 5,253,132 444,480 1,994,578 914,083 647,804
Highest Estimate 10,650,977 18,858,423 14,919,839 19,053,896 19,451,694 10,850,758
Lowest Estimate 986,769 4,503,194 444,480 1,994,578 914,083 647,804
Sales, Survey Year 11,797,536 108,444,464 41,629,132 45,024,836 38,316,656 71,432,882
Sales, Previous Year 13,516,931 67,629,732 51,030,840 53,409,352 63,778,992 35,751,329
:::
The results presented in Table [4](#T4){ref-type="table"} indicate that the methodologies yield radically different estimates of the total number of condoms used. This was anticipated, considering that our estimates of the frequency of intercourse and the probability of condom use also varied by estimation method. There are also very large differences between the estimates based on data from the female surveys and those from the male surveys. The bottom panel of Table [4](#T4){ref-type="table"} shows that the range of the estimates is very wide for all surveys. For example, in Kenya the high estimate of the total annual number of condoms used in 1998 is 10.7 million, while the low estimate is only 1.0 million. Similarly, for the 1999 Tanzania survey the highest estimate is 19.1 million while the lowest estimate is only 2.0 million.
It is unknown which of the estimates is most accurate. However, as we previously noted, the *p*~3~estimate (which is based on condom use among those who reported having intercourse the day before the survey) appears unreliable due to the small number of cases. In addition, the survival analyses yielded unrealistically low estimates of the frequency of intercourse (*F*~3~) that appeared inconsistent with the literature. Therefore, estimates that are based on these two factors are unlikely to be reliable. Table [4](#T4){ref-type="table"} confirms that estimates based on *F*~3~and *p*~3~usually yield the lowest estimates of the total number of condoms used.
When self-reported data are not available, estimates based on *F*~2~and *p*~2~are likely to be the most reliable. Data from the 1994 Zimbabwe survey confirm that the estimates based on the self-reported frequency of intercourse (*p*~1~) and the percentage who used a condom in last intercourse (*p*~2~) yield fairly similar results. This was anticipated, given that self-reported frequency of intercourse was coded as a categorical variable and subsequently quantified on the basis of the percentage who reported using a condom in last intercourse. Table [4](#T4){ref-type="table"} shows that estimates based on *F*~1~and *F*~2~are also fairly close.
Nevertheless, all survey-based estimates of the annual number of condoms used are substantially lower than the reported number of condoms sold for almost every country. The only exception is Kenya, where the high estimate of the total number of condoms used based on the 1998 Kenya DHS is fairly close to the number distributed (10.7 million vs. 11.8 million). For the other surveys, the reported number of condoms sold or distributed tends to be 2.5 to 3.0 times higher than even the highest survey-based estimate of the number of condoms used. Comparison with sales data from the previous year does not resolve these differences.
Conclusions
===========
The purpose of this paper was to estimate the annual number of sex acts and condoms used based on survey data, and to compare the latter with data on the annual number of condoms sold and distributed. The ability to estimate the number of sex acts from survey data would be a valuable tool for program managers, as it would enable them to estimate the number of condoms needed. Since the available data on condom sales and distribution measure the number of condoms supplied to the trade rather than to the consumer, survey estimates of the total number of condoms used could also help clarify to what extent data on the number of condoms supplied to the trade reflects actual consumer sales.
Analysis of the annual reported number of condoms sold and distributed reveals very erratic patterns. The large year-to-year differences in the total number of condoms distributed clearly do not reflect differences in the number of condoms sold to consumers, nor in the level of condom use, as this would imply major changes in behavior. The latter is unlikely to have occurred, since behavior is known to change very slowly. In other words, the large fluctuations in the number of condoms provided to the trade are likely to reflect fluctuations in condom inventory at various levels in the distribution chain. Because of this, the current data on the number of condoms sold and distributed say very little, if anything, about the number of condoms sold to consumers or about actual levels of condom use.
To estimate the annual number of condoms used from survey data, survey questionnaires would ideally ask respondents how often they had sex during a given reference period and how often they used a condom during that period. Considering that using very long reference periods (e.g., a year) is likely to cause recall errors, a shorter reference period is preferable. Of the DHS studies used in this paper, only one (Zimbabwe DHS-III, 1994) asked respondents about the frequency of intercourse during the four weeks preceding the survey. For the other surveys, the frequency of intercourse had to be estimated indirectly on the basis of the duration since last intercourse. Although older data on frequency of intercourse are available for some countries, such data may not provide reliable estimates of current behavior, as the HIV/AIDS crisis and other factors may have influenced sexual behavior.
If future surveys are to estimate the annual number of condoms used, then questions enquiring about the total number of sex acts and the total number of sex acts in a fixed time period should be added. For example, recent surveys in Zambia asked about the number of sex acts and the number of condoms used in the past week, which can easily be extrapolated to a one-year period \[[@B33]\]. Asking about the timing of the last two sex acts, rather than only the very last sex act, would also be recommended. This would provide data on the duration between two successive sex acts, which will improve estimation of the total number of sex acts using survival methodologies. Knowing the distribution of the time interval between successive sex acts would also enable researchers to identify a decay function that best fits the data, which will substantially increase the accuracy of the estimates.
The results of our survey analyses, which are based on DHS data currently available, show that the estimates of both the number of sexual acts and the number of condoms used vary enormously based on the estimation method used. For several surveys, the highest estimate of the annual number of condoms used is tenfold that of the lowest estimate. While some estimation methods can be disregarded because they yield results that are clearly not plausible, it is impossible to determine which of the remaining methods yield the most accurate results. Until the reliability of these various estimation methods can be established, estimating the annual number of condoms used from survey data will not be feasible.
To be able to verify the reliability of the estimates of the number of condoms used, it is necessary to have accurate data on the number of condoms sold and distributed to consumers. In developing countries, such is not feasible, in part due to the lack of standardized record-keeping, and because many condoms are distributed through informal retailers, such as street venders and hawkers, who are unlikely to keep records. For the purpose of testing the feasibility of the estimation methods, it may therefore be more productive to use data from developed countries where retail-level condom sales data are available (assuming such data are not proprietary). Alternatively, it may be possible to test the reliability of the estimates in developing countries, by obtaining the relevant sales data on a smaller scale (e.g., for one district only). However, sales data have the drawback that they do not provide information about the characteristics of the consumers. Consequently, sales data are unable to provide detailed information about program impact.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
DM conceived of the study and drafted the manuscript. RVR developed the study design and carried out the statistical analysis. Both authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1472-6963/5/5/prepub>
Supplementary Material
======================
::: {.caption}
###### Additional File 1
This file contains the background data for the calculations
:::
::: {.caption}
######
Click here for file
:::
Acknowledgements
================
This study was funded by the United States Agency for International Development (USAID), through a grant from the MEASURE *Evaluation*project (Activity \#3934). The authors are grateful to Ties Boerma and Emma Slaymaker for providing information about their own methodologies, to Jim Shelton, Beverly Johnston, and Nada Chaya for providing data on condom sales and distribution, to Gregory Fegan for literature searches and preliminary data analyses, to Susan Bergman for editing, and to John Stover for comments and suggestions on an earlier version of this paper.
|
PubMed Central
|
2024-06-05T03:55:52.009751
|
2005-1-15
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545997/",
"journal": "BMC Health Serv Res. 2005 Jan 15; 5:5",
"authors": [
{
"first": "Dominique",
"last": "Meekers"
},
{
"first": "Ronan",
"last": "Van Rossem"
}
]
}
|
PMC545998
|
Background
==========
With the sequencing of entire genomes and the exponential growth of sequence databases on the one hand, and the significant number of known folds compared to the putative number of possible folds in the fold space on the other hand, sequence-structure comparison is currently one main challenge of the post-genomic era. To this goal, 3D-environments were used by Eisenberg and coll. in the early 90s to build statistical potentials indicating the probability of finding each amino acid in a given structural environment as described by the secondary structure, the solvent accessibility and the polarity of neighboring atoms \[[@B1]\]. Such statistical potentials were successfully applied to protein fold recognition \[[@B1]-[@B4]\] or protein model evaluation \[[@B5],[@B6]\], and were shown to improve the quality of sequence-structure alignments \[[@B7]\].
Statistical potentials describing the propensity of a residue pair to be at a given spatial distance have proved successful as well \[[@B8]-[@B14]\], but are more difficult to use as information to guide sequence-structure alignments using dynamic programming. On the contrary, residue preferences for position-dependent structural environments are easily implemented in alignment programs \[[@B7],[@B15]\]. Recent improvements in this field were achieved by (i) optimizing definition and classification of the 3D-environments, and (ii) by constructing substitution matrices instead of residue preferences, i.e. taking into account the native residue type \[[@B15]-[@B17]\]. Indeed, it has been shown that amino acid substitutions are constrained by the structural environment, each environment displaying a distinct substitution pattern \[[@B18],[@B19]\]. The use of 64 distinct substitution matrices corresponding to different 3D environments based on secondary structure, solvent accessibility and hydrogen bonding, combined with structure-dependent gap penalty and with global or local alignment algorithms, provides good performance to the FUGUE software in fold recognition approaches \[[@B15]\]. In this paper we investigate the use of decision tree algorithms to automate and improve the classification of structural environments. The automation will allow easy adaptation to any particular selected data set, opening a way for the construction of various specific substitution matrices. Indeed, it appears that one problem in the use of statistical potentials for structure prediction is their lack of universality \[[@B13]\]. It may thus be worthwhile to derive potentials specific to prediction problems or to protein classes. The automated derivation proposed here will facilitate such developments.
In the first part of the work we focus on automatically building and evaluating structure-dependent substitution scores. The emphasis is given to the development of a method for automatic selection of the most informative classifications of 3D environments in order to set up a versatile method allowing easy compilation of structure-dependent substitution scores for any given set of proteins. In a second part, the method is applied to a specific protein class, the small disulfide-rich proteins.
Decision trees have attracted our attention for several reasons. Knowledge acquisition and description language are optimized automatically and do not require human expertise during the learning process. Thanks to the hierarchical organization of the inferred trees, classifications are obtained quickly and the chain of decisions leading to the prediction is explicit and can be easily interpreted (in contrast to artificial neural networks for example). Decision tree learning algorithms are also robust since they partition the data recursively. The first dichotomies near the tree root are then based on many examples and are therefore statistically reliable. To handle noise and uncertainty, the trees can be pruned during a post-processing step to remove possible misleading clusters at the bottom of the tree. The research field on this topic is well-established and the number of applications of decision trees to real world is huge.
Methods
=======
Structure-dependent substitution profiles
-----------------------------------------
Standard substitution matrices are deduced from multiple sequence alignments of similar sequences \[[@B20]-[@B22]\]. To derive structure-dependent substitution matrices, multiple sequence alignments are also needed as well as a description of the 3D structure at each position in the alignment \[[@B15],[@B17]\]. A schematic overview of the EvDTree method is displayed in Figure [1](#F1){ref-type="fig"}. Since we want to make sure that all residues at the same position in the alignment do share similar 3D structures, we will only use multiple alignments obtained from structural superimpositions (step 1 in Figure [1](#F1){ref-type="fig"}). From this, we extract all observed \"substitutions\" for each residue type and the corresponding structural environment (step 2 in Figure [1](#F1){ref-type="fig"}). The word \"substitution\" here is used to describe the residue replacement observed at an equivalent position in two structurally similar proteins. Then, for each residue type, the structural environments and associated substitutions are classified using a decision tree algorithm and substitution scores are computed from residue distributions observed in each cluster of the classification tree (step 3 in Figure [1](#F1){ref-type="fig"}).
Standard structure-dependent substitution matrices report the probability of 20 × 20 or 21 × 21 possible substitutions in a given structural environment \[[@B15]\]. In this work we classify 3D-environments and associated substitutions derived from alignments separately for each of the 21 residues (the 20 standard residues plus the half-cystine) using a decision tree algorithm (Figure [1](#F1){ref-type="fig"}, step 3 and Figure [2](#F2){ref-type="fig"}). As a result, we get several structure-dependent substitution profiles for each type of residue, that each indicates the relative probabilities of all 21 possible substitutions of one residue type in a given structural environment. Since the selected structural environments differ between residue types, the substitution profiles cannot be gathered into structure-dependent substitution matrices. As an example of how structural environments may differ between residues, a solvent-exposed environment might refer to a solvent accessibility \> 6% if the residue is a leucine, but to a solvent accessibility \> 33% if the residue is a glutamine.
Learning and test data sets
---------------------------
Several data sets of structure-structure superimpositions and the corresponding alignments are available \[[@B23]-[@B25]\]. We have selected the database of homologous structure alignments, HOMSTRAD \[[@B23]\] for constructing both the learning and the test data sets of sequence-structure alignments. Main HOMSTRAD strengths are (i) a selection of accurate protein structures, (ii) a combination of automatic procedures and of manual inspections that guarantee low data noise, and (iii) its successful use in the derivation of the structure-dependent substitution matrices used in FUGUE \[[@B15]\]. Moreover, to facilitate comparison between our method and FUGUE, we selected the very same learning set previously used by Mizuguchi et al. This subset consists of 177 families extracted from the HOMSTRAD database and does not contain membrane proteins. From this HOMSTRAD subset, a set of 2.5 million of observed substitutions were extracted, one substitution corresponding to a residue in a reference structure, and the corresponding residue observed at the same position in a structurally similar protein. Moreover, to remove non-structural constraints from the sequence-structure alignments, the following filters were applied:
\- Residues involved in a domain-domain or chain/chain interface, were excluded. Residues are considered to be involved in an interface when their solvent accessibility varies by more than 7 % when comparing the protein environment in the complex and in the isolated chain/domain. The cut-off value was taken from Mizuguchi et al \[[@B15]\], who used a similar filter to remove residues at a chain/chain interface.
\- Residues that are not correctly superimposed in the structural superimposition were also excluded. The superimposition was considered good enough when the deviation between the two corresponding alpha carbons is below 3.5 Å. We assume that larger deviations may correspond to incorrect structural superimposition for the particular residue even though other residues are correctly aligned. Large deviations may also imply significant modifications in the 3D-environment. Although this 3.5 Å criterion is sometimes too restrictive, it actually leaves enough data for robust statistical estimations while removing most of aligned amino acid pairs whose respective structural contexts are not superposable.
Application of the two above filters excluded about 20% of the initial substitutions leaving about 2 million substitutions for the learning process. This data set was split into (i) a learning data set containing 950 000 substitutions similar to the learning set used by Mizuguchi et al. (ii) a pruning data set containing 325 000 substitutions, and (iii) a test data set containing 355 000 substitutions. The learning data set has been in some cases filtered further based on the percentage of sequence identity between superimposed proteins, resulting in smaller sets of 500 000 (0--40% id) or 700 000 (0--60% id) substitutions, respectively.
\- Since we only work with three-dimensional structures, the oxidation state of any cysteine (free or disulfide bridged) is known. The symbol \'C\' refers to disulfide bridged cysteines (half-cystines), whereas the symbol \'J\' was used for free cysteines.
Structural descriptors
----------------------
Since the decision tree algorithm is able to automatically select the most discriminating structural descriptor at each classification step (see below) we do not need to empirically determine the \'best\' descriptors. In this work, twenty-three structural descriptors were provided to the classification algorithm. The secondary structure (ss1) was assigned to each residue according to STRIDE \[[@B26]\] into seven categories. Values are as follows: ss1 = 1, 2, 3, 4, 5, 6 and 7 for α-helices (H), 3~10~helices (G), π-helices (I), isolated bridges (B), extended conformations (E), turns (T) and coils (C), respectively. We also used a simpler 3-state description (ss2) deduced from the STRIDE assignment: ss2 = 1, 2 or 3 for helices (H or G), sheets (B or E) and coils (I, T, or C), respectively.
Hydrogen bonds were determined using the Hbond software \[[@B27]\]. Four different descriptors were used for different type of interactions: side-chain\...main-chain O atom (hb1), side-chain\...main-chain N atom (hb2), side-chain acceptor\...side-chain donor (hb3) and side-chain donor\...side-chain acceptor (hb4). For each interaction type, the number of interactions was used as the descriptor value. Here again a simpler description (lh) was also implemented that takes value of 0, 1, 2, or 3 if the side-chain of the residue makes no hydrogen bond, makes hydrogen bond(s) with side-chain atom(s), makes hydrogen bond(s) with main-chain atom(s) or makes hydrogen bonds with both side-chain and main-chain atoms, respectively.
Other structural parameters were obtained using the local program compilPDB \[J.G.\]. Beside the secondary structure, the local structure was also described by the Phi and Psi dihedral angles, and by Cα-Cα distances: d3 = Cα~i~- Cα~i+3~, d4 = Cα~i~- Cα~i+4~, d5 = Cα~i~- Cα~i+5~, d6 = Cα~i~- Cα~i+6~, d7 = Cα~i~- Cα~i+7~. Other descriptors were the buried surface area (bur), percent of accessibility (pac), contact area with carbon atoms (C), nitrogen atoms (N), oxygen atoms (O), sulfur atoms (S), positively charged atoms (pp), negatively charged atoms (nn), or polar atoms (pol).
For simplicity, these structural descriptors will now be called *s*~1~to *s*~23~.
It should be noted that some structural descriptors are correlated (e.g., the Phi and Psi dihedral angles versus the d3 and d4 alpha carbon distances). However, this descriptive redundancy is not a problem since it is eliminated during the tree construction where the most informative descriptors only are selected, as explained below.
Automated classification of structural environments using a decision tree algorithm
-----------------------------------------------------------------------------------
The native structural environments observed in the learning data set were classified for each of the twenty amino acids, plus the half-cystine, resulting in twenty-one independent decision trees (Figures [1](#F1){ref-type="fig"} and [2](#F2){ref-type="fig"}). The use of these decision trees is as follows: let (*a*(*k*), ***s***(k)) the position *k*in a protein for which we want to score substitutions, *a*(*k*) the residue type and ***s***(*k*) = (*s*~1~(*k*),\...,*s*~23~(*k*)) the structural environment description at this position. After the learning phase explained below, each tree node will be associated to particular structural descriptor *s*~*j*~and threshold *S*and will be linked by edges to two subnodes whose structural environments will be constrained respectively by the tests *s*~*j*~≤*S*and *s*~*j*~\>*S*. The classification of (*a*(*k*), ***s***(k)) will be obtained by selecting the decision tree corresponding to residue type *a*(*k*) and then by running through the tree from its root node to an appropriate leaf following at each node the edge whose test, *s*~*j*~≤*S*or *s*~*j*~\>*S*, is compatible with the value of the corresponding structural descriptor *s*~*j*~(*k*) (Figure [2](#F2){ref-type="fig"}). Contextual substitutions scores associated to the selected tree leaf, as explained in a further paragraph, will then evaluate each possible substitution of the amino acid *a*(*k*).
According to the standard data mining terminology, the predictive variables are therefore the native amino acid type and its associated structural descriptors and the dependent variable to be predicted is the substituted residue at this position. During the learning phase which we will now describe, the goal of the decision tree construction is to optimize the predictive power of the structural descriptor test chosen at each node and therefore to maximize the bias of the statistical distributions of the substituted residues associated to each subnode towards a few types of amino acids. Ideally, tree leaves should be associated to only one type of substituted amino acid, but this never happens in practice because of the tree depth limitation and the data set noise.
Let (*a*(*i*), ***s***(*i*), *b*(*i*)) be the *i*-th example of the whole learning data set where *a*(*i*) is a native residue, ***s***(*i*) = (*s*~1~(*i*),\...,*s*~23~(*i*)) is its structural environment description and *b*(*i*) is the substituted residue as observed in a structurally similar protein at the same position as *a*(*i*). The main steps of the decision tree construction from the learning data set are as follows (Figure [2](#F2){ref-type="fig"}):
1\. The decision tree for a given residue type *A*is initiated to a unique root node with an associated cluster *c*~0~= {*i*/ *a*(*i*) = *A*} grouping all examples with native residue type *A*.
2\. For each tree cluster *c*do :
a\. Test in turn each descriptor *s*~*j*~and each associated threshold *S*that creates possible dichotomies of *c*into two subclusters *c*~1~and *c*~2~. If *s*~*j*~has continuous values, 9 possible thresholds *S*are chosen to create dichotomies *c*~1~= {*i*∈*c*/*s*~*j*~(*i*)≤*S*} and *c*~2~= {*i*∈*c*/*s*~*j*~(*i*)\>*S*} corresponding to the 10^th^, 20^th^, \..., and 90^th^percentiles of the statistical distribution of the considered descriptor. If *s*~*j*~is restricted to a few discrete categories, all possible dichotomies *c*~1~= {*i*∈*c*/*s*~*j*~(*i*)==*S*} and *c*~2~= {*i*∈*c*/ *s*~*j*~(*i*)! = *S*} are created, where *S*is one of each possible value of *s*~*j*~.
b\. Select the optimal dichotomy from previous step which satisfies the tree constraints (see section (i) below) and minimizes the chosen splitting criterion (see section (ii) below).
c\. Insert the new clusters *c*~1~and *c*~2~as nodes in the tree by linking them to cluster *c*with respective edges labeled {*s*~*j*~(*i*)≤*S*} and {*s*~*j*~(*i*)\>*S*} or {*s*~*j*~(*i*) == *S*} and {*s*~*j*~(*i*)! = *S*}. The structural environment associated to a particular cluster will be defined by all edge labels from the tree root to the considered tree node or leaf (see figure [2](#F2){ref-type="fig"}).
3\. Finally, prune the tree according to the selected pruning method and pruning data set (see section (iii) below).
It should be noted the choice of the optimal descriptor at a given tree level will depend on both the amino acid identity of the native residue and each structural descriptor previously chosen as splitting criteria along the tree path that leads to the considered node.
Main parameters in the classification are (i) the tree constraints, (ii) the splitting criterion, and (iii) the tree pruning method.
### (i) Tree constraints
\- Tree depth: as the learning process goes deeper in the tree, more and more specific clusters are created. Beyond a certain depth, the chance that the corresponding rules can be applied to new examples outside the learning set drops significantly, resulting in an overfitting of the available data since deep clusters won\'t have enough associated examples to derive statistically significant distributions. Therefore, to avoid wasting time to partition the data into smaller and smaller clusters, maximum tree depths of 2 to 6 were tested.
\- Cluster cardinal: For the same reason as above, a minimum cardinal of examples was required for each cluster. We tested values between 200 and 1200 with increments of 200.
\- Tree balancing: A restriction on uneven distributions of samples among two clusters from the same parent was applied to prevent the creation of unbalanced trees which would require higher depth to fully partition the data. This restriction is achieved by the parameter *sim*~*cc*~measuring the cluster cardinal similarity between two subclusters obtained by splitting :
where, *n*~1~is the cardinal of the subcluster 1 and *n*~2~the cardinal of the subcluster 2.
### (ii) Three different splitting criteria were tested
\- The Gini criterion evaluates the probability that two randomly selected elements in a cluster correspond to two different types of residues \[[@B28]\]:
where *P*(*a*\|*c*) is the relative frequency of residue type *a*in cluster *c*.
To evaluate the quality of a given segmentation into several clusters, the splitting criterion is given by
where *n*~*c*~is the number of elements in the cluster *c*and *n*is the total number of elements in all clusters.
\- The Shannon entropy \[[@B29]\] tries to limit the distribution of elements of the same class among several clusters.
where *P*(*a*\|*c*) is the relative frequency of residue type *a*in cluster *c*.
\- We also used a specifically developed splitting criterion called the \"mean rank\" *MR*. Each class (residue type) in a cluster is ranked according to the number of elements of this class in the cluster (rank 1 is assigned to the most frequent residue type and rank 21 to the least frequent one). The mean rank *MR*evaluates the mean rank for a randomly selected element in the cluster. Low *MR*indicates clusters with only few well represented classes. Such clusters would correspond to structural environments that induce significant bias in the sequence and therefore strong structural constraints.
where *R*(*a*\|*c*) and *P*(*a*\|*c*) are the frequency rank and probability of the residue type *a*in the cluster *c*.
### (iii) Three different pruning methods were considered
\- The Pessimistic Error Pruning (PEP) \[[@B30]\] consists in recursively checking each cluster starting from the tree root and in cutting its corresponding subtree if this removal reduces the mean error rate estimated on the independent pruning test set by :
where *N*(*c*) is the number of examples assigned to the cluster *c*and *n*(*c*) is the number of occurrences of the most frequent amino acid in the cluster *c*. Let *C*be the father cluster from which *c*is derived in the tree, then *c*and its subtree will be removed if E(*c*)≥E(*C*).
\- The Mean Rank Pruning (MRP) has a principle similar to PEP, except that *c*will be removed if *R*(*c*)\<*R*(*C*) where *R*(*c*) and *R*(*C*) are respectively the mean ranks of the current cluster *c*and of its father cluster *C*averaged over the pruning test set.
\- The pessimistic Mean Rank Pruning (PRP) is a more stringent version of MRP using a confidence margin to prevent statistically biased clusters to be kept in the tree. The current cluster *c*will now be removed if *R*(*c*)+ σ *t*~80~\<*R*(*C*), where σ is the mean rank standard deviation over the pruning test set and the scaling factor *t*~80~= 1.82 corresponds to a 80% confidence level for a Gaussian distribution.
### Few other parameters were further optimized including
\- A mutation weight α = 1/*N*~*f*~inversely proportional to the total number of residues *N*~*f*~in each protein family *f*of the learning data set. This insures that all structural families have similar importance in the derivation of the substitution probabilities.
\- A mutation weight β = 25/*ide*inversely proportional to the percentage of identity *ide*between the two considered proteins. If *ide*\<25%, then the mutation weight is decreased to 1. This reduces the importance of substitutions observed in similar sequences and could be used later to specialize EvDTree on different kinds of applications involving different sequence similarities.
Residue specific environment-dependent substitution profiles
------------------------------------------------------------
Once trees have been constructed, statistical distributions of observed substitutions in each cluster are used to compute cluster-specific environment-dependent substitution profiles. As explained previously, the structural environment associated to a particular cluster will be defined by all edge labels from the tree root to the considered tree node or leaf.
The probability for the amino acid *a*in the 3D environment *s*to be substituted by amino acid *b*is
where  is the number of observed substitutions of amino acid *a*by amino acid *x*in the 3D environment *s*.
Smoothed probabilities *Q*(*b*\|*a*,*s*) are then calculated as
where *A*(*b*\|*a, s*) is the *a priori*distribution of Topham et al. \[[@B19]\]. Relative weights are calculated as
where  is the total number of occurrences of amino acid *a*in 3D environment *s*, *n*is the number of classes (21 in this case), and *σ*is a normalization constant. We used the value of 5 previously used by Topham et al. \[[@B19]\]. It should be noted that the weight of this \"a priori\" distribution is inversely proportional to the number of available examples  and is therefore maximum for undersampled substitutions.
Then the log odds scores are calculated as
where *P*(*b*) is the background probability of occurrence of amino acid *b*in the whole database.
These log-odds are calculated for each node cluster of each native amino acid tree.
Application and evaluation of environment-dependent and standard substitution scoring functions
-----------------------------------------------------------------------------------------------
To evaluate the EvDTree scoring function, each example of associated native residue, structural environment, substituted residue (*a*(*k*), ***s***(*k*), *b*(*k*)) from the test data set is classified by the tree corresponding to residue type *a*(*k*). Then the tree leaf corresponding to the structural environment *s*(*k*) is searched and its associated log-odds substitution scores are finally used to score the substituted residue *b*(*k*).
To compare the EvDTree substitution scores with other scoring methods, we have used the mean rank (*MR*) as the criterion to evaluate the quality of scoring functions. For each example in the test data set, the 20 possible substitutions are scored as indicated above, and the observed (real) substitution *b*(*k*) is ranked according to its score among all other possible substitutions. The mean rank over all examples in the test data set is indicative of how well the scoring function is able to recognize as probable the \"real\" substitutions. A MR of 1 would mean that the scoring function always gave the better score to the observed substitution. At the opposite, a MR of 10.5 would indicate that observed substitutions are scored randomly. The main advantage of the mean rank criterion is that it is fast to calculate and it is independent from the absolute values of the scores, therefore allowing comparisons between very different scoring functions. Similar criteria based on ranking were previously used to evaluate 1D-3D scoring functions \[[@B31]\].
The evaluation of environment-dependent substitution matrices requires computing the 3D environments the very same way they were computed when deriving the scores. Thanks to Dr K. Mizuguchi who provided us with all the necessary tools, we could include the FUGUE environment-dependent substitution matrices into our evaluation process.
To complement the *MR*evaluation, we also compared the performance of EvDTree with other scoring functions in sequence-structure alignments. To do this, 1000 sequence-structure alignments were selected from our test data set derived from the HOMSTRAD database. Each alignment was recalculated using a Smith and Watermann algorithm with several different substitution scoring functions. For each scoring function, the percentage of correctly aligned positions, according to the real alignments in the test data set, was compiled and used for comparisons. For each method, the gap opening (*Go*) and gap extension (*Ge*) penalties were optimized by comparing the alignments for several penalty combinations (*Go*= 2, 5, 10, 15, 20; *Ge*= 2, 5, 10, 15, 20).
Results and discussion
======================
Decision tree classifications
-----------------------------
Several learning data sets were compiled by filtering out observed protein substitutions (Figure [1](#F1){ref-type="fig"}) with sequence identity between superimposed proteins above thresholds of 40%, 60% or 80%. For each learning set, several combinations of parameters were tested for the construction of the EvDTree classifications and the calculation of the resulting structure-dependent substitution scores (Figures [1](#F1){ref-type="fig"} and [2](#F2){ref-type="fig"}). For each run, a set of 21 decision trees was built and the corresponding scoring function was evaluated on the test data set using the mean rank (*MR*) criterion as explained in Methods. Due to the amount of data, CPU time limitations did not allow systematic examination of all parameter combinations, and the best parameters were determined through a limited trial and error protocol.
The variation of the mean rank over sixteen different runs remained limited (6.74 \<*MR*\< 6.89) showing that the method is robust and is not critically affected by slight modifications of the parameters. The lowest *MR*(6.74) was obtained with the following protocol:
\- Minimum cardinal of any cluster = 600;
\- The minimal cluster cardinal similarity between two subclusters obtained by splitting of the parent cluster is
*sim*~*cc*~= 0.1.
\- Segmentation criterion: mean rank
\- The pruning method is MRP.
\- Examples in the learning set are weighted according to the number of residues in the protein family: α = 1/*N*~*f*~
\- No weighting is done in relation to the sequence identity (β = 25/*ide*)
\- Maximal sequence identity between superimposed proteins in the learning set = 60%
The values of the obtained mean ranks could, at first sight, appear rather high. However, it is worth noting that only substitutions were ranked in the evaluations, i.e. protein positions occupied by the same amino acid in the two structurally superimposed proteins were not considered in the evaluation process. They are included, however, in the statistics during the calculations of the substitution scores.
The evolution of the *MR*criterion along the learning process for alanine is shown in Figure [3](#F3){ref-type="fig"}. As expected, the MR decreases regularly when calculated on the learning set. On the other hand, when calculated on the test data set, the MR decreases in the first few learning steps, then increases in the following steps. The difference between the two curves in the last steps is due to overfitting, i.e. learning specific rules from the learning set that cannot be generalized, thus reducing the predictive power on the test data set. Pruning the tree using an independent pruning data set removes clusters with reduced predictive power resulting in a flat curve in the last steps.
Analysis of the EvDTree classifications
---------------------------------------
For each residue type, the maximal tree depth used was 6, leading to a maximal number of leaves 21 × 64 = 1344, each of them potentially leading to a substitution profile that corresponds to one line in a classical substitution matrix. In other words, the total amount of data corresponds to 64 distinct substitution matrices, although it is not possible to associate substitution profiles to matrices since each profile corresponds to a different structural environment. The overall amount of data is nevertheless, in principle, comparable to the 64 environment-dependent substitution matrices used in the FUGUE system. However, the tree pruning step removed a significant number of clusters that do not afford improved information, leaving only 111 environments each associated to a specific substitution profile (data available from <http://bioserv.cbs.cnrs.fr/HTML_BIO/EvDTree.html>).
Structural descriptors, thresholds and values used in the first dichotomies of the root clusters for each residue type are displayed in Table [1](#T1){ref-type="table"}. Our main interest in using a decision tree classification is that, in principle, optimal splitting parameters and associated thresholds or values are automatically selected for each residue. As an example, it has been suggested that different boundaries on the fraction of area buried should be used for different residue classes when determining if a residue is exposed or buried \[[@B17]\]. Several examples of this can indeed be found in the EvDTree classifications: the selected structural descriptor for splitting the root cluster of serine and alanine substitutions is the percent of accessible area (pac), but the threshold is 3% for serine, whereas it is 10% for alanine (Table [1](#T1){ref-type="table"}). The pac is also used as structural descriptor in the second dichotomy in the glutamic acid tree classification, with a threshold of 25%. The use of different accessibility thresholds by the decision tree algorithm fully supports previous observations by Rice and Eisenberg \[[@B17]\]. This observation also highlights the nice feature of decision trees that can be easily interpreted.
Analysis of the most discriminating structural descriptor selected in the learning process, i.e. the descriptor selected for the first dichotomy of the root cluster for each residue type (denoted c~0~in Figure [2](#F2){ref-type="fig"}), shows that the secondary structure is the most discriminating parameter for aspartic and glutamic acids, lysine, arginine and asparagine (Table [1](#T1){ref-type="table"}). Although contact polarity or solvent accessibility have been selected in subsequent dichotomies in most cases, it is clear that the substitution profile of charged residues primarily depends on the local structure. This result appears to be consistent with previous work by Gilis & Rooman \[[@B32]\] on the relative importance of local and non-local interactions in mutant stabilities. These authors showed that for solvent-exposed residues, the local structure is the most important factor, whereas distance potentials (i.e. 3D interactions) appear more suited to prediction of mutations in the protein core \[[@B32]\]. Here we show that substitution profiles for charged residues (which are largely solvent-exposed) mainly depend on the local structure. Another observation leads to a similar conclusion: for alanine and serine, the first selected structural parameter for splitting is the percent of accessibility (pac) and the most exposed resulting cluster is then split using secondary structure, whereas the most buried resulting cluster is split using the polarity of the protein environment (pol). These results confirm that, for solvent-exposed protein positions, the local structure is one main parameter that determines which amino acid can occupy this position.
Four substitution profiles (i.e. log-odds substitution scores for one residue type into one structural environment) are displayed in Figure [4](#F4){ref-type="fig"}. Comparison of the substitution profiles for alanine and aspartic acid in similar environments (exposed α-helix) reveals significant differences (Figure [4A](#F4){ref-type="fig"}). This observation is not trivial since it could be postulated that, except for functional residues, the probability that a residue *b*occurs in a structural environment *s*only depends on *s*but is independent of the observed residue *a*in structurally similar proteins. The fact that, for similar structural environment, substitution profiles vary with the native residue probably indicates that purely structural descriptions probably lack some essential information, possibly related to the evolution process. This observation also illustrates the limits of environment-dependent statistical potentials in which the native amino acid is not taken into account. As an example, using data in Figure [4A](#F4){ref-type="fig"}, substitutions to Met in exposed α-helices appear more likely than substitutions to Leu when the native residue is Ala but the reverse is true when the native residue is Asp. Such differences cannot appear in environment-dependent statistical potentials such as 3D-1D scores that only describe the relative preference of residues for particular structural environments \[[@B1]\].
On the other hand, the substitution profiles for leucines in different structural contexts also display significant differences (Figure [4B](#F4){ref-type="fig"}). Thus, substitutions Leu → Met are favored in exposed α-helical positions whereas substitutions Leu → Thr are favored in exposed non α-helical positions (Figure [4B](#F4){ref-type="fig"}). This observation is not unexpected since it is well-known that β-substituted residues do not like to be in α-helices. Nevertheless it shows that EvDTree was able to extract consistent knowledge on sequence-structure relationships and it confirms previous observations that substitution scores are indeed structure-dependent \[[@B18],[@B33]\], explaining why structure-dependent substitution matrices perform better than standard evolutionary matrices in fold recognition processes \[[@B1],[@B15],[@B17],[@B34]\].
Structural information improves prediction of substitution probabilities
------------------------------------------------------------------------
The detailed impact of structural information for correct prediction of substitution probabilities can be approached by comparing the EvDTree substitution profiles with the evolutionary substitutions matrices GONNET \[[@B20]\] and BLOSUM62 \[[@B35]\]. To this goal, the \"Mean Rank\" criterion has been used (the lower the Mean Rank, the better the scoring function; see Methods and Data). Results by residue type and averaged over all residues are shown in Table [2](#T2){ref-type="table"}. Comparison of EvDTree with the Gonnet and BLOSUM62 matrices shows that EvDTree performs clearly better on average, and individually for most residues. Therefore, the use of the structural information in EvDTree does improve the predictive power of substitution profiles versus structure-independent substitution matrices. Moreover, a substitution matrix was built from the tree clusters of EvDTree, i.e. before any structural information is taken into account. This matrix, referred to as EvDTree0, is simply derived from the structural superimpositions in the learning data set and is thus similar to other structure-derived substitution matrices \[[@B36]\]. Interestingly, EvDTree0 performs better than evolutionary matrices suggesting that, despite a lower amount of data, structure-derived alignments can provide data of higher quality than sequence alignments for derivation of substitution matrices.
The results in Table [2](#T2){ref-type="table"} show that EvDTree provides poorer evaluation than the evolutionary matrices for two residues, histidine and lysine, possibly due to an insufficient amount of data. It is also worth noting that histidine often participates in active sites or coordination sites, and the substitution probabilities may have been biased by this peculiarity. Filtering out the learning data set for coordination sites was performed by Shi et al \[[@B15]\], but such a filter was not implemented here.
More surprisingly, five residues (Gly, Met, Pro, Gln, and Thr) and the free cysteine (J) do not display evaluation improvement by using structural information (compare EvDTree0 and EvDTree). The latter remark means that for these five residues no structural descriptor permitted efficient splitting of the data. It is likely that for these residues new descriptors or descriptor combinations remain to be discovered. Nevertheless, on average, the structural information significantly improves the performance and EvDTree appears as a clearly better scoring function than evolutionary matrices in evaluation of sequence-structure alignments.
The EvDTree substitution profiles provide slightly better substitutions predictions
-----------------------------------------------------------------------------------
Comparison with structure-dependent substitution matrices obtained by other groups is not as simple as for standard substitution matrices, because we must make sure that we compute the structural environment exactly the same way that was used for generating the matrices. We were able to compare EvDTree with FUGUE, thanks to the FUGUE accessory programs kindly provided by Dr K. Mizuguchi.
As shown in Table [2](#T2){ref-type="table"}, the overall performance of EvDTree and FUGUE appear very similar. A comparison between these two scoring functions at different levels of sequence identity is displayed in Figure [5](#F5){ref-type="fig"}. EvDTree provides slightly better performances at sequence identity above 30% but similar results below 30% (Figure [5](#F5){ref-type="fig"}). The reason for this remains unclear, but might be a result of our filtering of the learning set that removed all positions were Cα deviates from more than 3.5 A. Further optimization of the EvDTree learning set would probably be necessary for use in fold recognition programs at low sequence identity. The observation that EvDTree performs at least as well as the scoring function of FUGUE but is computed in a fully automated manner opens the way for future potential applications. For example, we show below that EvDTree can easily optimize fold-specific scoring matrices specific of small disulfide rich proteins leading to improved substitution scores for this particular class of proteins.
Evaluation of EvDTree as scoring function in sequence-structure alignment
-------------------------------------------------------------------------
Although the Mean Rank test determines the ability of a scoring function to correctly evaluate structure-compatible sequences, it does not determine how well a scoring function would actually perform in a particular application, e.g. fold recognition. In this paper, we do not focus on a particular application, but rather on the method to automatically derivate structure-dependent substitution profiles, and the Mean Rank appears as a simple and efficient criterion to rapidly compare different learning parameterizations or scoring functions. To verify that the theoretical evaluations using the Mean Rank have some significance for future real applications, we compared EvDTree with other methods used as scoring functions in sequence-structure alignments. The percent of correctly aligned positions for different methods and for different sequence identity ranges is displayed in Figure [6](#F6){ref-type="fig"}. For each method, the best Gap opening and Gap extension penalties were roughly optimized by checking 25 combinations (*Go*= 2, 5, 10, 15, 20; *Ge*= 2, 5, 10, 15, 20). The results shown in Figure [6](#F6){ref-type="fig"} fully confirm previous analyses using the Mean Rank criterion: on average, better alignments are obtained using the structure-dependent scoring functions FUGUE and EvDTree, and EvDTree provides slightly better alignments than FUGUE above 30% of sequence identity but similar results at lower sequence identity. The slightly better accuracy of EvDTree for mid-range percentages of identities suggests that our approach could be particularly useful to improve sequence/structure alignments in homology modeling. The EvDTree learning algorithm could also be applied to fold recognition by optimizing scoring functions specific of particular protein families.
Application of EvDTree to a specific class of proteins, the small disulfide-rich proteins
-----------------------------------------------------------------------------------------
The strongest potential of EvDTree is its ability to adapt itself to any particular set of structures. As a test case, we have applied EvDTree to small disulfide-rich proteins. Small disulfide-rich proteins display several peculiar structural features: (i) due to their small size, a larger number of residues than usual are solvent-exposed, (ii) regular secondary structures are limited and the content in turns and loops is high, (iii) the hydrophobic core is largely constituted by the disulfide bridges that are responsible for the high stability despite the small size, and (iv) glycine, proline, and, of course, cysteine residues are more frequent than usual. With all these peculiarities, small disulfide-rich proteins are not well-suited to standard prediction methods and it has been shown that these proteins score poorly using the standard PROCHECK database \[[@B37]\]. Evaluations of non-specific substitution scoring functions on small disulfide-rich proteins using the mean rank criterion are reported in Table [3](#T3){ref-type="table"}. Comparison of values in Table [3](#T3){ref-type="table"} for EvDTree, Gonnet, BLOSUM62 and FUGUE with those in Table [2](#T2){ref-type="table"} clearly support the idea that small disulfide-rich score poorly when using standard databases and methods.
Thus, to test the ability of EvDTree to automatically adapt itself to a class of proteins with specific structural features, we have computed disulfide-rich specific substitution profiles (EvDTreeDS). For this, we have compiled a specific data set from the structural class \"small disulfide\" in the HOMSTRAD database. These data were complemented by data extracted from the KNOTTIN database \[[@B38]\]. The number of structural positions in the initial data set (about 40000) is clearly insufficient to divide this set into independent learning and test datasets. Therefore, a \"leave-one-out\" protocol was used. In this protocol, one protein family is excluded from the learning set and the resulting substitution scores are used to evaluate this family. This process is repeated for all protein families in the initial set. Also, due to the limited number of data, the tree pruning had to be performed using the learning data set, which is of course far less efficient than using an independent data set (Esposito, 1997), and resulted in a very limited pruning when compared to the general case.
Mean rank evaluations of the new, specific, EvDTreeDS substitution scores on small disulfide-rich proteins are reported in Table [3](#T3){ref-type="table"}. Despite the limited set of data, comparison of the efficiency of the EvDTreeDS specific substitution profiles with standard EvDTree clearly shows that the automated classification method was able to extract, at least in part, the specific features of the small disulfide-rich proteins.
Furthermore, the comparison of the structural descriptors used in the first partition of root clusters in EvDTree and EvDTreeDS classifications highlights interesting differences (Table [1](#T1){ref-type="table"}). First, when the solvent accessibility percentage is the first used descriptor in both trees (half-cystines and alanines), the thresholds retained by the learning algorithm are different for the disulfide-rich proteins which are small and whose residues, except half-cystines, are more exposed to solvent on average. Accordingly, the selected threshold is lower for half-cystines but higher for alanines.
Table [1](#T1){ref-type="table"} also reveals that most large hydrophobic residues (Leu, Phe, Tyr, His) switch their most discriminating descriptor from hydrophobic-hydrophilic measures in EvDTree (Pol, C) to secondary structure descriptors in EvDTreeDS (ss1 and ss2). This inversion can be interpreted by the reduction of the buried volume in the small disulfide-rich structures which cannot accommodate for large residues. This suggests that, in these proteins the stability is mainly due to disulfide bridges whereas the hydrophobic effect would be less crucial. Being more solvent exposed and less involved in hydrophobic packing, large hydrophobic residues might become more sensitive to local structure in small disulfide-rich proteins.
More surprisingly, we also notice that charged and polar residues (Asn, Asp, Glu, Lys, Arg) display the opposite switch, i.e. a secondary structure descriptor is used in EvDTree but a solvent accessibility descriptor is used in EvDTreeDS. Analysis of the EvDTreeDS classification suggests that these residues, when in more buried positions, are far more conserved than when in more exposed positions. We think that beside the disulfide bridge core, additional elements of stability often occur through specific hydrogen bonding networks between charged or polar residues and the backbone, rather than through hydrophobic packing. Typical examples of this are the conserved glutamic acid in position 3 of cyclotides \[[@B39]\], or the conserved aspartic acid in position 15 of squash inhibitors \[[@B40]\] which both participate in multiple hydrogen bonding with the backbone. This might explain, at least in part, why the distinction between exposed and partially buried charged residues is more critical in small disulfide-bridged proteins.
All these subtle modifications revealed by the EvDTree and EvDTreeDS classifications suggest that there is probably no universally optimal description language and that the choices and partitions of structural descriptors should be adapted to the class of proteins considered. Our new decision tree learning algorithm makes this fine tuning automatically from scratch whereas classical potentials are based on globally optimized description languages which may become suboptimal in specific contexts.
Conclusions
===========
We have described a new method, EvDTree, based on decision tree classification of structural environments to automatically construct structure-dependent substitution profiles from a set of sequence-structure alignments. The EvDTree method was shown to perform similarly to the successful environment-dependent substitution matrices used in FUGUE (Shi *et al.*, 2001). Interestingly, the tree-pruning step removed a significant number of structural clusters yielding an average tree depth of 4 instead of the six allowed levels. This is an indication that clusters at higher levels corresponded to the learning of specific sequence-structure relationships that could not be generalized (Figure [3](#F3){ref-type="fig"}). It may be expected that as more high quality data will become available, this effect could be reduced and higher levels of the decision tree will gain better performances.
In this work, we were interested in the development of a fully automatic method for the classification of structural environments and inference of structure-dependent substitution profiles. The evaluation of the intrinsic performance of the substitution profiles was primarily done on known sequence-structure alignments, using the mean rank of observed substitutions. We have shown that in this context, the EvDTree substitution profiles perform slightly better than other successful substitution matrices, and as such, the EvDTree matrices constitutes interesting elementary data for various applications. Moreover, comparison of the EvDTree substitution scores with other scoring functions for sequence-structure alignments led to results similar to the mean rank evaluation supporting the usefulness of the latter criterion.
One specific strength of the EvDTree method is its easy automatic adaptation to any specific data set. Here we have shown that it is possible to obtain structure-dependent substitution profiles specific of small disulfide-rich proteins with better predictive power than standard substitution scores. This approach could be easily extended to other specific protein classes such as coil-coils, membrane proteins, etc. as soon as enough structures are available for learning. Fold-specific substitution matrices have recently been proposed for protein classification \[[@B41]\]. The EvDTree approach opens the way for class-specific or fold-specific structure-dependent substitution scores for use in threading-based remote homology searches. Decision trees based on different learning sets and with different depths could be optimized depending on the available protein structures and sequences of the fold family considered.
The fact that, as stated above, the structural information did not yield better prediction for several residues in the EvDTree approach (Table [2](#T2){ref-type="table"}) suggests that improvements are still possible. To this end, 3D environments from the decision trees yielding poor performances should be determined in order to design more appropriate structural descriptors. It is tempting to speculate that using combined structural descriptors, e.g. (Phi, Psi) angle pairs which can delineate particular regions of the Ramachandran plot or (dCi,i+j, dCi,i+k) distance pairs which can introduce some super-secondary structural constraints, could increase the accuracy of the decision trees. Alternatively, the use of linear combinations of descriptors in decision tree induction algorithms have been reported and could be used for structural classifications \[42\].
Authors\' contributions
=======================
After an initial program from JG, JCG coded a new version of the software to incorporate a novel algorithm. All methods were implemented and tested by JCG. The whole work was conceived by JCG, JG and LC and was supervised by LC and JG. All authors read and approved the final manuscript.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Flow-chart of the EvDTree method.**1- Initial data are pairs of superimposed protein structures. For each position, the \"native\" residue in the reference protein, the \"substituted\" residue in the superimposed protein and structural parameters are tabulated. 2- The data are filtered and grouped according to the \"native\" residue type, resulting in root clusters for classifications. 3- For each residue type, a hierarchical classification is achieved using a decision tree algorithm. Substitution scores are computed for all resulting clusters from residue distribution in the cluster.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Decision tree classification of all observed substitutions for Leucines.**Similar trees are built for each of the 21 residue types, including half-cystine. The structural environment for a given cluster is defined by the edge labels along its path from the root cluster *c*~0~. For example, the nodes colored in gray indicate the partial classification path of a Leucine observed in a native structural environment whose descriptors *s*~23~, *s*~15~and *s*~10~verify *s*~23~= *pol*= 3%, *s*~15~= *pac*= 7% and *s*~10~= *d*~4~= 5.4A.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Evolution of the Mean Rank criterion averaged over Alanine residues at different tree depths.**Values are indicated for the evaluations of the learning set (filled diamonds), of the test set before pruning (filled triangles) and after pruning the decision tree (open squares).
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Substitution profiles for 4 different EvDTree clusters.**(A) Two profiles for different amino acids (Ala and Asp) in similar structural environments corresponding approximately to an exposed α-helix. Grey bars: Ala with pac \> 38 and ss1 = 1; Black bars: Asp with pac \> 20 and ss2 = 1. (B) Two profiles for the same amino acid (Leu) in different structural environments. Grey bars: Leu with pol \> 53 and ss1 = 1; Black bars: Leu with pol \> 53 and ss1! = 1.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
Comparison between EvDTree (grey) and FUGUE (black) at different levels of sequence identity in the test data set.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
**Comparison of sequence-structure alignments using several scoring functions.**EvDTree, white bars; Gonnet, hatched bars; BLOSUM62, grey bars, and FUGUE, black bars. Gap opening and gap extension penalties have been separately optimized for each scoring function.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Structural descriptors and associated values or thresholds used by EvDTree to partition the root cluster *c*~0~for each residue type into two subclusters (see Figure 2). The EvDTreeDS columns refer to the decision tree classification of a learning set containing only small disulfide-rich proteins (see text).
:::
**Amino acid**^**a**^ **EvDTree** **EvDTreeDS**
----------------------- ------------- --------------- -------- ----- --------- --------
C^**b**^ pac \<= 8 \> 8 pac = 0 \> 0
A^**b**^ pac \<= 10 \> 10 pac \<= 35 \> 35
I pol \<= 41 \> 41 C \<= 35 \> 35
L^**c**^ pol \<= 53 \> 53 ss1 = 6 ≠ 6
F^**c**^ C \<= 65 \> 65 ss2 = 2 ≠ 2
Y^**c**^ C \<= 65 \> 65 ss2 = 3 ≠ 3
H^**c**^ C \<= 35 \> 35 ss2 = 3 ≠ 3
V pol \<= 46 \> 46
M ss1 = 4 ≠ 4
W bur \<= 155 \> 155 N \<= 1 \> 1
N^**d**^ ss2 = 1 ≠ 1 pac \<= 49 \> 49
D^**d**^ ss2 = 1 ≠ 1 bur \<= 25 \> 25
E^**d**^ ss1 = 4 ≠ 4 pac \<= 54 \> 54
K^**d**^ ss2 = 2 ≠ 2 pac \<= 53 \> 53
R^**d**^ ss1 = 1 ≠ 1 bur \<= 101 \> 101
S pac \<= 3 \> 3
T pac \<= 49 \> 49
G ss1 = 7 ≠ 7
Q ss2 = 3 ≠ 3
P C \<= 7 \> 7
^a^Only dichotomies for which the structural information improves the Mean Rank evaluation are shown (Tables 2 and 3). ^b^Residues for which the same descriptor is used in EvDTree and EvDTreeDS. ^c^Residues for which a descriptor corresponding to the local structure has been used by the EvDTreeDS but not by the EvDTree learning process. ^d^Residues for which a descriptor corresponding to the local structure has been used by the EvDTree but not by the EvDTreeDS learning process.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Mean rank evaluation of the structure-dependent EvDTree substitution scores and comparison with other scoring functions.
:::
**Amino acid** **DT0^b^** **DT^a^** **GON^a^** **B62^a^** **Fug^a^**
---------------- ------------ ----------- ------------ ------------ ------------
**A** 7.93 **7.33** 8.21 8.34 7.51
**C** 10.62 10.60 11.01 11.86 **10.19**
**D** 6.48 **6.41** 6.85 6.59 6.48
**E** 6.83 **6.73** 6.92 6.95 6.79
**F** 6.66 **6.58** 7.04 6.86 7.01
**G** 7.14 7.14 7.16 7.20 **7.00**
**H** 8.18 8.05 8.29 7.98 **7.97**
**I** 5.65 **5.38** 5.88 5.72 5.58
**K** 6.97 6.93 6.88 6.86 **6.81**
**L** 6.71 **6.39** 6.58 6.69 6.52
**M** **5.68** **5.68** 6.01 5.84 5.87
**N** 6.82 **6.73** 7.20 7.03 6.90
**P** **7.15** **7.15** 7.58 7.80 7.33
**Q** **6.74** **6.74** 7.08 7.28 7.29
**R** 7.46 7.24 7.41 7.41 **7.21**
**S** 6.43 **6.32** 7.03 6.73 6.67
**T** **6.61** **6.61** 7.21 6.74 7.00
**V** 6.50 **6.28** 6.65 6.38 6.36
**W** 8.03 **7.61** 8.81 9.04 7.63
**Y** 8.38 8.11 9.15 8.82 **8.02**
**J** **7.12** **7.12** 8.17 7.48 7.16
**Average** 6.90 **6.74** 7.15 7.06 6.87
^a^Mean rank evaluation of the EvDTree (DT), Gonnet (Gon), Blosum62 (B62) and FUGUE (Fug) substitution scores. ^b^Mean rank evaluation of EvdTree0 (DT0), i.e. scores inferred from the root clusters before structural classifications. Bold numbers indicate the best values among the methods.
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Mean rank evaluation of small disulfide-rich proteins using standard and specific scoring functions.
:::
**Amino acid** **DT^a^** **DS0^a^** **DS^a^** **Fug^a^** **Gon^a^** **B62^a^**
---------------- ----------- ------------ ----------- ------------ ------------ ------------
**A** 8.16 8.59 **7.71** 8.39 8.85 8.83
**C** 9.84 11.60 **8.09** 10.42 11.8 11.72
**D** **6.85** 7.50 7.31 6.90 7.12 7.01
**E** **6.70** 7.57 7.22 6.89 6.88 6.89
**F** 8.58 8.77 **7.58** 10.14 9.79 9.86
**G** 8.03 7.74 **7.35** 7.48 7.40 7.46
**H** 7.99 8.04 **7.51** 8.00 7.72 8.09
**I** 7.74 8.55 **7.28** 7.82 8.36 8.19
**K** 7.47 8.07 7.56 7.49 7.59 **7.36**
**L** 9.13 9.41 **8.34** 9.25 9.70 9.90
**M** 9.74 9.04 **7.33** 8.95 9.16 8.87
**N** 7.57 8.14 8.09 **7.411** 8.40 8.02
**P** 8.20 8.41 **8.00** 8.11 8.70 8.56
**Q** **6.88** 7.94 7.81 7.35 7.38 7.38
**R** 7.93 8.20 **6.96** 7.84 7.99 8.11
**S** 7.56 **7.55** **7.55** 7.56 8.29 8.17
**T** 8.14 8.61 **7.63** 8.53 9.40 8.91
**V** 8.30 **7.69** **7.69** 7.82 8.76 8.48
**W** 8.78 7.46 **7.45** 8.20 9.03 8.82
**Y** 8.61 8.59 **7.53** 8.76 9.77 9.32
**Average** 7.90 8.27 **7.59** 8.02 8.44 8.38
^a^Mean rank evaluation using the EvDTree (DT), EvDTreeDS0 (DS0), EvDTreeDS (DS), FUGUE (Fug), Gonnet (Gon) and Blosum62 (B62) substitution scores. Bold numbers indicate the best values among the methods.
:::
|
PubMed Central
|
2024-06-05T03:55:52.015473
|
2005-1-10
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545998/",
"journal": "BMC Bioinformatics. 2005 Jan 10; 6:4",
"authors": [
{
"first": "Jean-Christophe",
"last": "Gelly"
},
{
"first": "Laurent",
"last": "Chiche"
},
{
"first": "Jérôme",
"last": "Gracy"
}
]
}
|
PMC545999
|
Background
==========
Transmembrane proteins are divided to date into two structural classes, the α-helical membrane proteins and the β-barrel membrane proteins. Proteins of the α-helical membrane class have their membrane spanning regions formed by hydrophobic helices which consist of 15--35 residues \[[@B1]\]. These are the typical membrane proteins, found in cell membranes of eukaryotic cells and bacterial inner membranes \[[@B1]\]. On the other hand, β-barrel membrane proteins, have their transmembrane segments, formed by antiparallel β-strands, spanning the membrane in the form of a β-barrel \[[@B2],[@B3]\]. These proteins are found solely in the outer membrane of the gram-negative bacteria, and presumably in the outer membranes of mitochondria and chloroplasts, a fact, perhaps, explained by the endosymbiotic theory \[[@B4]-[@B7]\]. Transmembrane protein topology prediction has been pursued for many years in bioinformatics, mostly focusing on the α-helical membrane proteins. One reason for that, is that α-helical transmembrane segments are more easily predicted by computational methods, due to the easily detectable pattern of highly hydrophobic consecutive residues, and the application of simple rules as the \"positive-inside rule\" \[[@B8]\]. On the other hand, another reason is the relative abundance of α-helical membrane proteins compared to that of the β-barrel membrane proteins. This discrepancy, is present in both the total number of membrane proteins in complete genomes, an also in the datasets of experimentally solved 3-dimensional structures. Currently, the number of structures of outer membrane proteins known at atomic resolution raises rapidly, due to improvements in the cloning and crystallization techniques \[[@B9]\]. This, fortunately, gave rise to an increase of the number of prediction methods and the online available web-predictors. The first computational methods that were deployed for the prediction of the transmembrane strands were based on hydrophobicity analyses, using sliding windows along the sequence, in order to capture the alternating patterns of hydrophobic-hydrophilic residues of the transmembrane strands \[[@B10],[@B11]\]. Other approaches included the construction of special empirical rules using amino-acid propensities and prior knowledge of the structural nature of the proteins \[[@B12],[@B13]\], and the development of Neural Network-based predictors to predict the location of the Cα\'s with respect to the membrane \[[@B14]\]. The major disadvantages of these older methods, were the limited training sets that they were based on, and the reduced capability to capture the structural features of the bacterial outer membrane proteins, especially when it comes to sequences not having similarity with the proteins of the training set. During the last few years, other more refined methods, using larger datasets for training, appeared. These methods, include refined Neural Networks (NNs), \[[@B15],[@B16]\], Hidden Markov Models (HMMs) \[[@B17]-[@B21]\] and Support Vector Machines (SVMs) predictors \[[@B22]\]. Some of these methods are based solely on the amino acid sequence and others use also as input evolutionary information derived from multiple alignments. Other popular methods such as the method of Wimley \[[@B23]\] and BOMP \[[@B24]\] do not explicitly report the transmembrane strands, but instead they are oriented towards genome scale discrimination of β-barrel membrane proteins.
In this work, we evaluate the performance of the available prediction methods to date. Using a non-redundant dataset of 20 outer membrane β-barrel proteins, with structures known at atomic resolution, we compare each predictor in terms of the per-residue accuracy (using the correctly predicted residues, and the Mathews correlation coefficient \[[@B25]\]) and that of the strands\' prediction accuracy measured by the segments overlap measure (SOV) \[[@B26]\]. We also report the number of the correctly predicted topologies (i.e. when both strands localization and orientation of the loops are correctly predicted). We conclude, that the recently developed Hidden Markov Model methods HMM-B2TMR \[[@B17]\], ProfTMB \[[@B21]\] and PRED-TMBB \[[@B20]\], perform significantly better than the other available methods. We also conclude that the prediction accuracy is affected significantly, if the full sequences (including long N-terminal and C-terminal tails and the signal peptide) are used for input and not only the transmembrane β-barrel domain. This finding is again more profound when referring to the NN and SVM predictors, since the regular grammar of the HMMs maps successfully the model topology to the proteins\' modular nature. Finally, we developed a consensus prediction method, using as input the individual predictions of each algorithm, and we conclusively show that this approach performs better, in all the measures of accuracy, compared to each individual prediction method separately. Although consensus methods have proven to be more accurate in the past, in the case of α-helical membrane proteins \[[@B27]-[@B29]\] and also for secondary structure prediction of globular, water soluble proteins \[[@B30]-[@B32]\], this is the first time that such a method is applied to β-barrel outer membrane proteins.
Results and discussion
======================
The results obtained from each individual algorithm, on the test set of the 20 proteins are summarized in Table [1](#T1){ref-type="table"}. It is obvious that all of the methods perform worse for the measures of per-segment accuracy in the case of full-length sequences. On the other hand, for measures of per-residue accuracy, most of the methods perform better in the case of full-length sequences, a fact already mentioned in \[[@B21]\]. This is explained, considering the fact that when using full-length sequences, more non-transmembrane residues are predicted correctly, thus increasing the fraction of correctly predicted residues and the correlation coefficient. Furthermore, when ranking the different methods PRED-TMBBposterior performs better, followed by HMM-B2TMR and ProfTMB. PRED-TMBBnbest, performs slightly worse than PRED-TMBBposterior in terms of per-residue accuracy and SOV, but is inferior to the other top-scoring HMMs in terms of the correctly predicted topologies. In order to assess the statistical significance of these observations and draw further safe conclusions, we should rely on a statistical analysis of the results obtained.
The MANOVA test (Table [2A](#T2){ref-type="table"}) yields a highly significant p-value for both the 2 independent variables (p \< 10^-4^). This means, that there is truly a difference in the vector of the five measured attributes across the different methods and the type of sequence that we use as input. By including in the model the interaction term between the two factors, we get a marginally insignificant p-value (p = 0.0619), indicating that some of the methods behave differently with input sequences of different type. Examining each one of the attributes independently (Table [3A](#T3){ref-type="table"}), we observe that the type of the input sequence does not influence significantly the effect on all the measures of per-residue accuracy (correctly predicted residues and the correlation coefficient, p-values equal to 0.9444 and 0.0224 respectively) but, instead, influences a lot the per-segment measures such as SOV (p \< 10^-4^), correctly predicted topologies (p = 0.0193) and correct barrel size (p = 0.0001). In all cases, the type of the method is a highly significant factor (p \< 10^-4^), reflecting the fact that there are true differences in the performance of the methods. The interaction term in the ANOVA is significant only for the SOV measure (p = 0.0272), and marginally significant for the correctly predicted residues (p = 0.402). However, these results do not provide us with a clue as to which method performs better (or worse) than the others; it states that one or more methods depart significantly from the mean. The ranking of the methods has to be concluded by observing Table [1](#T1){ref-type="table"}.
In order to discover the statistically significant differences between the methods, we proceeded by grouping the methods according to the type of the algorithm they utilize. This way, we grouped together the HMM-based methods (HMM-B2TMR, PRED-TMBB, ProfTMB and BETA-TM) and the NN and SVM-based methods (TMBETA-NET, B2TMPRED, PSI-PRED and TBBPred). Thus, instead of having a factor with 8 levels describing the methods, we now have a factor with 2 levels (HMM and not HMM). The MANOVA test (Table [2B](#T2){ref-type="table"}) once again yields a statistically significant result, for both the 2 factors (p \< 10^-4^) and the interaction term (p = 0.0025), giving us a clear indication that the visually observed superiority of the HMM-based methods has a statistically significant justification. The statistically significant interaction of the 2 factors, furthermore suggests that the decrease in some of the measured attributes when submitting full-length sequences, is smaller (if anything) for HMM-based methods than for the NN and SVM-based ones. In fact, considering the three top-scoring HMM methods, we observe that the per-segment measures are not influenced from the type of the input sequence whereas the per-residue measures are significantly increased with full-length sequences as input, reflecting the fact that more non-transmembrane residues are correctly predicted, as noticed already in \[[@B21]\]. Considering each one of the measures of accuracy with ANOVA (Table [3B](#T3){ref-type="table"}), the type of the method is a highly significant factor in all of the tests, and the type of the input sequence highly significant for the per-segment measures of accuracy. The interaction term is highly significant for SOV (p = 0.0011) and marginally insignificant for correctly predicted residues (p = 0.052).
These findings suggest, that the HMM-based predictors perform better, on average, than the NN and SVM-based methods, in almost all of the measured attributes. We should mention here, that the difference between HMM and NN/SVM methods is larger for the measures of per-segment accuracy than for per-residue accuracy. Even the simplest and less accurate HMM-based method, BETA-TM, that uses single sequence information compares favorably to the refined NN/SVM methods that use profiles derived from multiple alignments. As a matter of fact, only B2TMPRED, which uses a dynamic programming algorithm to refine the prediction, predicts more accurately than BETA-TM the correct topology and/or the barrel size of the proteins, but still cannot reach the accuracy of the other HMM-based methods. Furthermore, the HMM-based methods are not influenced significantly whether full-length sequences or just the β-barrel domains are submitted for prediction. Interestingly, the NN/SVM methods, often falsely predict the signal peptide sequences as transmembrane strands in the precursors whereas HMMs do not. This observation is consistent with the theory regarding the nature of HMM and NN-based methods. Thus, it is consistent with the fact that the regular grammar of the HMMs can capture more effectively the temporal variability of the protein sequence and map successfully the proteins\' modular nature to a mathematical sound model. Therefore, it is not surprising that also for α-helical membrane proteins\' topology prediction the best available predictors are those based on HMMs \[[@B33]\]. On the other hand, NN methods are more capable of capturing long-range correlations along the sequence. This results to the correct identification of an isolated strand, but since the β-barrel proteins follow strict structural rules, the modular nature of the barrels is captured more effectively by HMMs. NNs may often falsely predict isolated transmembrane strands in non-barrel domains or predict strands with a non-plausible number of residues or even barrels with an odd number of strands. From a structural perspective, it is also of great interest to consider that the repetitive structural domains of β-barrels are the β-hairpins whereas the α-helical membrane proteins counterparts are the isolated hydrophobic helices often connected by loop regions of arbitrary length.
These observations, will have a significant impact not only on isolated predictions for one or few proteins, but also on predictions for sequences arising from genome projects where one expects to have the precursor sequences. Thus, predictions on such sequences will be more reliable, when obtained from HMM-predictors rather than NN and SVM-based ones. However, the performance of even the best currently available predictors are not as good as the predictions obtained for α-helical membrane proteins \[[@B33]\]. This is somewhat expected, and has a simple interpretation considering the grammatical structure of the short amphipathic transmembrane β-strands as opposed to the longer and highly hydrophobic transmembrane α-helices \[[@B1]\].
One issue that was not possible to investigate statistically is that of the use of evolutionary information in the form of profiles derived from alignments. It is well known, that the inclusion of information arising from alignments, increases significantly the performance of secondary structure prediction algorithms \[[@B34]\]. This was exploited in the past, in the case of α-helical membrane protein prediction \[[@B35],[@B36]\], and it was investigated thoroughly in a recent work \[[@B37]\]. However, for β-barrel membrane proteins there is not such a clear answer. The authors of the methods that use evolutionary information \[[@B15],[@B17],[@B21]\] justified their choice showing that the inclusion of alignments as input, improves the performance of their models up to 18%. Furthermore, we showed here that NN-based methods, using multiple alignments (B2TMPRED) perform significantly better, compared to similar methods that are relying on single sequences (TMBETA-NET). However, the top scoring HMM method, PRED-TMBB, performs comparably to the other HMM methods that are using evolutionary information, even though it relies on single sequence information. This finding may be explained considering the choice of the training scheme for PRED-TMBB, since it is the only method trained according to the CML criterion, and with manually curated annotations for the transmembrane strands. However, it raises an important question as to whether the prediction accuracy, could be improved more by using evolutionary information, or not. Future studies on this area will reveal if improvements in the prediction could arise by combining evolutionary information with appropriate choice of training schemes, or if we have eventually reached a limit of the predictive ability for β-barrels membrane proteins, and we depend only on the advent of more three-dimensional representative structures.
Comparing the performance of individual methods, one has to keep in mind several important aspects of the comparison. From the one hand, the limited number of β-barrel membrane proteins known at atomic resolution, resulted in having a test set, that includes some (or all) of the proteins used for training each individual method or a close homologue. This does not imply that the comparison of the methods is biased (regarding the ranking), but that the absolute values of the measures of accuracy may be influenced. Thus, when it comes to newly solved structures, we may expect somewhat lower rates in the measures of accuracy for all methods examined. On the other hand, when comparing the results of the individual methods, as they appear in the original publications, we observe some discrepancies. These arise, mainly due to the fact, that when reporting results of a prediction method, the authors usually report the measures of accuracy obtained in the jackknife test (leave one out cross-validation test). Furthermore, the authors of the individual methods report the measures of accuracy obtained using as input different types of sequences, and comparing using as observed different annotations for the transmembrane strands. For instance, other authors report measures of accuracy obtained from the β-barrel domain of the proteins, others from the sequences deposited in PDB, and others report also the results from precursor sequences. As for the observed transmembrane strands used for comparisons, most of the authors used the annotations for the strands found in PDB, and only PRED-TMBB used manually annotated segments that resemble better the part of the strand inserted into the lipid bilayer. The last observation, partly explains the better prediction accuracy obtained by PRED-TMBB, mainly in the measures of per-residue accuracy (correctly predicted residues and correlation coefficient).
One important result of this study is the development of the consensus prediction method, for predicting the transmembrane strands of β-barrel membrane proteins. Even though consensus prediction has been proved to be a valuable strategy for improving the prediction of α-helical membrane proteins \[[@B27],[@B29],[@B38]\], no such effort has been conducted before, for the case of transmembrane β-barrels. A consensus of all of the available methods, does not improve the prediction accuracy compared to the top-scoring methods, indicating that there is a considerable amount of noise in the individual predictions, originating mainly from the low-scoring methods. However, when using the three top-scoring HMM methods (PRED-TMBB, HMM-B2TMR and ProfTMB) along with one or more of the best performing NN/SVM methods (B2TMPRED, TBBPred-SVM, TBBPred-NN and TBBPred-Combined) we get impressive results, outperforming the top-scoring methods in almost all measured attributes. As it is obvious from Tables [1](#T1){ref-type="table"} and [4](#T4){ref-type="table"}, the consensus prediction method performs better than each one of the individual predictors. The improvement ranges from a slight improvement around 1% for the correctly predicted residues and correlation coefficient, up to 4% for SOV and 15% for the correctly predicted topologies. We should note that these particular results were achieved using PRED-TMBBposterior, ProfTMB, HMMB2TMR, B2TMPRED and TBBPred-NN, but other combinations of the aforementioned methods perform similarly (Table [4](#T4){ref-type="table"}). This large improvement in the measures of per-segment accuracy is an important finding of this study.
However, in the web-based implementation of the consensus prediction method, we allow the user to choose at will the methods that will be used for the final prediction. This was decided for several reasons: Firstly, for a newly found protein, we might have larger variations on the predictions, and we could not be sure if the choice of different algorithms will give better results or not. Secondly, the different predictors are not sharing the same functionality and availability. For instance, some predictors respond by e-mail (B2TMPRED, PSIPRED), most of the others by http (PRED-TMBB, BETA-TM, TMBETA-NET etc), and others may be downloaded and run locally (ProfTMB, PSIPRED), whereas one of the top-scoring methods (HMM-B2TMR) is available as a commercial demo only, requiring a registration procedure. These facts, forced us not to have a fully automated server (but instead we require the user to cut \'n paste the predictions) but also to allow flexibility on the chosen methods, and let the user decide alone which methods he will use. For this reason, we also give to the users the opportunity to provide, if they wish, custom predictions. This way, a user may choose to use another method, that will come up in the future, or, alternatively, to use manually edited predictions.
Conclusions
===========
We have evaluated the currently available methods, for predicting the topology of β-barrel outer membrane proteins, using a non-redundant dataset of 20 proteins with structures known at atomic resolution. By using multivariate and univariate analysis of variance, we conclude that the HMM-based methods HMM-B2TMR, ProfTMB and PRED-TMBB perform significantly better than the other (mostly NN-based) methods, in both terms of per-residue and per-segment measures of accuracy. We also found, a significant decrease in the performance of the methods when full-length sequences are submitted for prediction, instead of just the β-barrel domain. However, the HMM-based methods are more robust as they were found largely unaffected by the type of the input sequence. This is an important finding that has to be taken in account, not only in the cases of single proteins\' predictions, but mostly in cases of predictions performed on precursor sequences arising from genome projects. Finally, we have combined the individual predictors, in a consensus prediction method, that performs significantly better even than the top-scoring individual predictor. A consensus prediction method is for the first time been applied for the prediction of the transmembrane strands, of β-barrel outer membrane proteins. The consensus method, is freely available for non-commercial users at <http://bioinformatics.biol.uoa.gr/ConBBPRED>, where the user may choose which of the individual predictors will include, in order to obtain the final prediction.
Methods
=======
Data sets
---------
The test set that we used has been compiled mainly with consideration of the SCOP database classification \[[@B39]\]. In particular, all PDB codes from SCOP that belong to the fold \"Transmembrane beta-barrels\" were selected, and the corresponding structures from the Protein Data Bank (PDB) \[[@B40]\] were obtained. For variants of the same protein, only one solved structure was kept, and multiple chains were removed. The structure of the β-barrel domain of the autotransporter NalP of *N. meningitidis*\[[@B41]\] was also included, which is not present in the SCOP classification although it is clearly a β-barrel membrane protein. The sequences have been submitted to a redundancy check, removing chains with a sequence identity above a certain threshold. Two sequences were considered as being similar, if they demonstrated an identity above 70% in a pairwise alignment, in a length longer than 80 residues. For the pairwise local alignment BlastP \[[@B42]\] was used with default parameters, and similar sequences were removed implementing Algorithm 2 from \[[@B43]\]. The remaining 20 outer membrane proteins constitute our test set (Table [5](#T5){ref-type="table"}).
The structures of TolC \[[@B44]\], and alpha-hemolysin \[[@B45]\], were not included in the training set. TolC forms a trimeric β-barrel, where each monomer contributes 4 β-strands to the 12-strand barrel. Alpha-hemolysin of *S. aureus*is active as a transmembrane heptamer, where the transmembrane domain is a 14-strand antiparallel β-barrel, in which two strands are contributed by each monomer. Both structures are not included in the fold \"transmembrane beta-barrels\" of the SCOP database. In summary, the test set (Table [5](#T5){ref-type="table"}), includes proteins functioning as monomers, dimers or trimers, with a number of transmembrane β-strands ranging from 8 to 22, and is representative of the known functions of outer membrane proteins to date.
In order to investigate the effect of the full sequence on the different predictors, we conducted two sets of measurements. In the first place, all proteins were submitted to the predictors, in their full length. We chose not to remove the signal peptides, considering the fact that completely unannotated sequences, mostly originating from genome projects, are most likely to be submitted to predictive algorithms, in their pre-mature form. Of the 20 sequences constituting our set, 4 belonging to the family of TonB-dependent receptors, namely FhuA \[[@B46]\], FepA \[[@B47]\], FecA \[[@B48]\] and BtuB \[[@B49]\] posses a long (150--250 residues) N-terminal domain that acts as a plug, closing the large pore of the barrel. This domain is present in all four of the structures deposited in PDB. One of the proteins of our dataset, OmpA possesses a long 158 residue C-terminal domain falling in the periplasmic space, which is absent from the crystallographically solved structure \[[@B50]\]. Finally, the Secreted NalP protein, possesses a very long, 815 residues in length, N-terminal domain that is being transported to the extracellular space passing through the pore formed by the autotransporter β-barrel pore-forming domain, of which we have the crystallographically solved structure \[[@B41]\]. For the second set of measurements, for all proteins constituting our dataset we extracted only the transmembrane β-barrel domain. In the case, of long N-, or C-terminal domains mentioned above, we retained only the last or first 12 residues, respectively.
Even in the structures known at atomic resolution, there is not a straightforward way to determine precisely the transmembrane segments, since the lipid bilayer itself is not contained in the crystal structures. This is the case for both α-helical and β-barrel membrane proteins. There are, however a lot of experimentally and theoretically derived sources of evidence, suggesting that the lipid bilayer in gram-negative bacteria, is generally thinner than the bilayer of the inner membrane or those of a typical cell membrane of an eukaryote. Thus, it is believed that the outer membrane possesses an average thickness around 25--30 Å, a fact mainly explainable by its lipid composition, average hydrophobicity and asymmetry \[[@B51]\]. The annotations for the β-strands contained in the PDB entries, are inadequate since there are strands that clearly extend far away from the bilayer. Some approaches have been used in the past, to locate the precise boundaries of the bilayer, but they require visual inspection of the structures and human intervention \[[@B23],[@B52]\]. In order to have objective and reproducible results, we used the annotations for the transmembrane segments deposited in the Protein Data Bank of Transmembrane Proteins (PDB\_TM) \[[@B53]\]. The boundaries of the lipid bilayer in PDB\_TM have been computed with a geometrical algorithm performing calculations on the 3-dimensional coordinates of the proteins, in a fully automated procedure.
Prediction methods
------------------
The different freely available web-predictors, evaluated in this work, along with the corresponding URLs are listed in Table [6](#T6){ref-type="table"}. OM\_Topo\_predict, is the first Neural Network-based method trained to predict the location of the Cα\'s with respect to the membrane \[[@B14]\]. Initially, the method was trained on a dataset of seven bacterial porins known at atomic resolution, but later it was retrained in order to include some newly solved (non-porin) structures <http://strucbio.biologie.uni-konstanz.de/~kay/om_topo_predict2.html>. B2TMPRED is a Neural Network-based predictor that uses as input evolutionary information derived from profiles generated by PSI-BLAST \[[@B15]\]. The method was trained in a non-redundant dataset of 11 outer membrane proteins, and uses a dynamic programming post processing step to locate the transmembrane strands \[[@B54],[@B55]\]. HMM-B2TMR, is a profile-based HMM method, that was trained for the first time on a non-redundant set of 12 outer membrane proteins \[[@B17]\] and later (current version) on a larger dataset of 15 outer membrane proteins \[[@B55]\]. This method also uses as input profiles derived from PSI-BLAST. It was trained according to a modified version of the Baum-Welch algorithm for HMMs with labeled sequences \[[@B56]\], in order to incorporate the profile as the input instead of the raw sequence, whereas for decoding utilized the posterior decoding method, with an additional post-processing step involving the same dynamic programming algorithm used in B2TMPRED \[[@B55]\]. We should note, that HMM-B2TMR is the only method that currently is available as a commercial demo only, requiring a registration procedure. PRED-TMBB is a HMM-based method developed by our team \[[@B19]\]. Initially, it was trained on a set of 14 outer membrane proteins \[[@B19]\] and later on a training set of 16 proteins \[[@B20]\]. It is the only HMM method trained according to the Conditional Maximum Likelihood (CML) criterion for labeled sequences, and uses as input single sequences. The prediction is performed either by the Viterbi, the N-best algorithm \[[@B57]\] or \"a-posteriori\" with the aid of a dynamic programming algorithm used to locate both the transmembrane strands and the loops. In this work, we chose to use both N-best and \"a-posteriori\" decoding, and treat them as different predictors. This was done, since the two alternative decoding algorithms, follow an entirely different philosophy, and in some cases yield different results. BETA-TM, is a simple HMM method trained on 11 non-homologous proteins using the standard Baum-Welch algorithm \[[@B58]\]. It also operates on single sequence mode, and the decoding is performed with the standard Viterbi algorithm. ProfTMB is the last addition to the family of profile-based Hidden Markov Models \[[@B21]\]. It also uses as input evolutionary information, derived from multiple alignments created by PSI-BLAST. It is trained using the modified Baum-Welch algorithm for labeled sequences whereas the decoding is performed using the Viterbi algorithm. Its main difference with HMM-B2TMR, PRED-TMBB, BETA-TM and other previously published, but not publicly available HMM predictors \[[@B18]\], is the fact that it uses different parameters (emission probabilities) for strands having their N-terminal to the periplasmic space, and other for those having their N-terminal to the extracellular space. Furthermore, it uses different states for the modeling of inside loops (periplasmic turns) with different length. TMBETA-NET is a Neural Network based predictor using as input single sequence information \[[@B16]\]. This method uses a set of empirical rules to refine its prediction, in order to eliminate non-plausible predictions for TM-strands (for instance a strand with 3 residues). TBBpred is a predictor combining both NNs and SVMs \[[@B22]\]. The NN-based module also uses evolutionary information, derived from multiple alignments, whereas the SVM-predictor uses various physicochemical parameters. The user may choose one of the methods, or combine them both. The authors of the method have shown, that combining the predictions obtained by NNs and SVMs, improves significantly the prediction accuracy \[[@B22]\]. For the evaluation of the performance and for the Consensus Prediction, we chose to use all three options, in order to investigate which one performs better. Finally, we evaluated the prediction of the transmembrane strands, obtained from a top-scoring general-purpose secondary structure prediction algorithm. This was done, in order to investigate systematic differences in the prediction of the transmembrane β-strands, but also because experimentalists continuously use such algorithms in deciphering assumed topologies for newly discovered β-barrel membrane proteins \[[@B59]-[@B61]\]. For this purpose, we have chosen PSI-PRED, a method based on Neural Networks, using multiple alignments derived from PSI-BLAST for the prediction, that has been shown to perform amongst the top-scoring methods for secondary structure prediction \[[@B62]\]. Other, equally successful methods such as PHD \[[@B63]\], perform similarly but they are not considered here.
Measures of accuracy
--------------------
For assessing the accuracy of the prediction algorithms several measures were used. For the transmembrane strand predictions we report the well-known *SOV*(measure of the segment\'s overlap), which is considered to be the most reliable measure for evaluating the performance of secondary structure prediction methods \[[@B26]\]. We also report the total number of correctly predicted topologies (*TOP*), i.e. when both the strands\' localization and the loops\' orientation have been predicted correctly, and the correctly predicted barrel size (*BS*), i.e the same with the correctly predicted topologies, but allowing for one strand mismatch \[[@B20]\]. As measures of the per residue accuracy, we report here both the total fraction of the correctly predicted residues (*Q~β~*) in a two-state model (transmembrane versus non-transmembrane), and the well known Matthews Correlation Coefficient (*C~β~*) \[[@B25]\].
Statistical analysis
--------------------
The measures of accuracy mentioned earlier are the dependent variables that we wish to compare. We treat each prediction on each protein as an observation, and as independent variables we use the type of the submitted sequences (*TYPE*) that could be either the full precursor sequence or the transmembrane barrel domain only, a factor with two categories, and the individual predictive method (*METHOD*), which has 11 categories. Furthermore we tried to group the methods to those based on a Hidden Markov Model and those that were not. This factor (*HMM*) was evaluated later, in order to assess the impact of the type of the prediction method. The formal way to assess the overall statistical significance is to perform a two-way multivariate analysis of variance (MANOVA) \[[@B64]\]. For the evaluation of the statistical significance we evaluated the Wilk\'s lambda, but the results are not sensitive to this choice since other similar measures (Hotelling-Lawley trace, Roy largest root e.t.c) gave similar results. A statistical significant result, for both the 2 factors (*TYPE*, *METHOD*), will imply that the vector of the measured attributes varies significantly across the levels of these factors. We also included into the models, the interaction term between the two factors (*TYPE\*METHOD*or *TYPE\*HMM*). This was necessary in order to investigate, the potential differences of the dependent variables in the various combinations of the independent variables. For instance, a significant interaction of TYPE with HMM, will indicate that the effect of the input sequence will be different on the two types of methods.
Having obtained a significant result from the MANOVA test, we could use a standard 2-way analysis of variance (ANOVA) for each of the dependent variables, in order to be able to confirm which one of the measured attributes, varies significantly across the two factors. In the ANOVA models, we also included the interaction terms. In all cases, statistically significant results were declared those with a p-value less than 0.05. We report for the ANOVA and MANOVA models, the test statistic and the corresponding p-value, for the fitted models (including the interaction term).
The consensus prediction method
-------------------------------
In order to produce a combined prediction, we have two alternatives: One is to use some kind of ensemble Neural Network, or, alternatively, to summarize the individual predictions using a consensus method. Ensemble Networks show a number of significant advantages over the consensus methods \[[@B65],[@B66]\], but suffer for the limitation that each individual predictor has to be available, every time that a request is made. Since we are dealing with web-based predictors, and we do not have the option to have local copies of each predictor installed, this could be disastrous, thus, the consensus method is the only available and reliable solution.
Suppose we have an amino acid sequence of a protein with length *L*, denoted by:
**x**= *x*~1~, *x*~2~,\..., *x*~*L*~,
and for each residue *i*we have the prediction of the *j*~*th*~predictor (*j*= *1, 2, \..., 7*)
where, 
Thus, we can define a per-residue score *S*~*i*~by averaging over the independent contributions of each predictor:
This way, we can obtain a consensus prediction score for the whole sequence,
This score is capable of yielding inconsistent predictions, such as a strand with 3 residues for example. For this reason it is then submitted to a dynamic programming algorithm, to locate precisely the transmembrane strands. The algorithm is essentially the same used by \[[@B19]\], with the major difference being the fact that it considers only two states (transmembrane vs. non-transmembrane). It optimizes the predicted topology, according to some predefined parameters, imposed by the observed structures. We also force the algorithm to consider as valid only topologies with an even number of transmembrane strands, as those observed in the crystallographically solved structures. Having determined the number of the transmembrane strands, the final choice of the topology is based on the consideration of the length of the predicted loops. As it has already been mentioned for the 3-dimensional structures, the periplasmic loops have significantly lower length than the extracellular ones, thus by comparing the total length of the two alternative topologies, we decide for the final orientation of the protein.
Authors\' contributions
=======================
PGB conceived of the study, performed the collection and analysis of the data and drafted the manuscript, TDL participated in data collection, implemented the consensus algorithm and designed the web interface and SJH supervised and coordinated the whole project. All authors have read and accepted the final manuscript.
Acknowledgements
================
PB was supported by a grant from the IRAKLEITOS fellowships program of the Greek Ministry of National Education, supporting basic research in the National and Kapodistrian University of Athens. We thank the University of Athens for financial support.
Figures and Tables
==================
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Obtained accuracy of the predictors, in a test set of 20 outer membrane proteins
:::
**METHOD** **TYPE** **SEQUENCE** **Q~β~** **C~β~** **SOV** **Correctly predicted topologies** **Correctly predicted barrel size**
--------------------------------------------------------- ----------- -------------- ----------- ----------- ----------- ------------------------------------ -------------------------------------
HMM-B2TMR HMM barrel 0.737 0.557 0.836 15 17
precursor 0.790 0.600 0.813 14 16
ProfTMB HMM barrel 0.734 0.537 0.818 14 17
precursor 0.777 0.575 0.784 12 16
PRED-TMBBpost HMM barrel **0.818** **0.630** **0.886** **14** **19**
precursor **0.842** **0.637** **0.852** **14** **16**
PRED-TMBBnbest HMM barrel 0.818 0.629 0.877 12 17
precursor 0.849 0.637 0.856 11 13
TBBPred-comb NN+SVM barrel 0.702 0.428 0.664 0 0
precursor 0.701 0.424 0.496 0 0
TBBPred-nn NN barrel 0.735 0.466 0.672 0 1
precursor 0.726 0.432 0.496 0 1
TBBPred-svm SVM barrel 0.744 0.458 0.721 1 3
precursor 0.744 0.426 0.535 0 0
B2TMPRED NN barrel 0.723 0.498 0.738 7 9
precursor 0.709 0.466 0.551 0 0
TMBETA-NET HMM barrel 0.697 0.415 0.698 3 8
precursor 0.663 0.353 0.515 0 4
BETA-TM NN barrel 0.690 0.395 0.691 1 2
precursor 0.663 0.322 0.497 0 1
PSI-PRED NN barrel 0.731 0.484 0.690 0 0
precursor 0.756 0.495 0.569 0 0
HMM-B2TMR, ProfTMB, PRED-TMBBpost, B2TMPRED, TBBPred-nn CONSENSUS barrel **0.819** **0.641** **0.924** **18** **20**
precursor **0.849** **0.660** **0.874** **15** **18**
For an explanation of the measures of accuracy see the *Materials and Methods*section. Abbreviations: PRED-TMBBpost: PRED-TMBB method with posterior decoding, PRED-TMBBnbest: PRED-TMBB method with NBest decoding, TBBPred-nn: The Neural Network module of TBBPred, TBBPred-svm: The SVM module of TBBPred, TBBPred-comb: TBBPred, combining the Neural Network and SVM modules. The performance of the best individual predictor, and the best available consensus obtained are highlighted with bold.
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Multivariate Analysis of Variance (MANOVA) using as dependent variables the vector of the 5 measures of accuracy.
:::
**A.** **Wilk\'s Λ** **df1** **df2** **F** **p-value**
------------------ --------------- --------- --------- ------- -------------
**overall** 0.1981 105 2029 7.59 \<10^-4^
**type** 0.8455 5 414 15.13 \<10^-4^
**method** 0.2582 50 1891 13.08 \<10^-4^
**type\*method** 0.8541 50 1891 1.33 0.0619
**B.**
**overall** 0.4511 15 1193 26.58 \<10^-4^
**type** 0.8609 5 432 13.96 \<10^-4^
**hmm** 0.5441 5 432 72.40 \<10^-4^
**type\*hmm** 0.9585 5 432 3.74 0.0025
A. Model that includes as independent variables the individual methods (11 factors), the type of the sequence (barrel/precursor) and their interaction term. B. Model that includes as independent variables the type of the method (HMM/not-HMM), the type of the sequence (barrel/precursor) and their interaction term. We report the Wilk\'s lambda statistic (Wilk\'s Λ), the degrees of freedom of the numerator (df1), the degrees of freedom of the denominator (df2), the F statistic (F) and the corresponding p-value (p-value).
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Univariate Analysis of Variance (ANOVA) using each time as dependent variable each one of the 5 measures of accuracy.
:::
**Q~β~** **C~β~** **SOV** **Correctly predicted topologies** **Correctly predicted barrel size**
------------------ ---------- ------------- --------- ------------------------------------ ------------------------------------- ------------- ------- ------------- -------- -------------
**A.** **F** **p-value** **F** **p-value** **F** **p-value** **F** **p-value** **F** **p-value**
**overall** 15.8 \<10^-4^ 13.55 \<10^-4^ 13.33 \<10^-4^ 19.07 \<10^-4^ 27.34 \<10^-4^
**type** 0 0.9444 5.25 0.0224 56.86 \<10^-4^ 5.51 0.0193 14.97 0.0001
**method** 31.26 \<10^-4^ 26.97 \<10^-4^ 20.25 \<10^-4^ 38.49 \<10^-4^ 54.14 \<10^-4^
**type\*method** 1.93 0.0402 0.96 0.4758 2.05 0.0272 1.01 0.4318 1.77 0.0645
**B.**
**overall** 27.13 \<10^-4^ 32.43 \<10^-4^ 58.18 \<10^-4^ 72.27 \<10^-4^ 123.71 \<10^-4^
**type** 0.06 0.8144 3.49 0.0625 45.22 \<10^-4^ 4.33 0.0379 12.28 0.0005
**hmm** 77.59 \<10^-4^ 91.52 \<10^-4^ 113.97 \<10^-4^ 212.4 \<10^-4^ 358.84 \<10^-4^
**type\*hmm** 3.8 0.052 1.79 0.1822 10.83 0.0011 0.01 0.9428 0.1 0.7502
A. Model that includes as independent variables the individual methods (11 factors), the type of the sequence (barrel/precursor) and their interaction term. B. Model that includes as independent variables the type of the method (HMM/not-HMM), the type of the sequence (barrel/precursor) and their interaction term. We report the F statistic (F) of the ANOVA test and the corresponding p-value (p-value).
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Obtained accuracy of the consensus predictions, in the test set of 20 outer membrane proteins
:::
**METHOD** **TYPE** **SEQUENCE** **Q~β~** **C~β~** **SOV** **Correctly predicted topologies** **Correctly predicted barrel size**
------------------------------------------------------------------------------------------------- ----------- -------------- ----------- ----------- ----------- ------------------------------------ -------------------------------------
PRED-TMBB, ProfTMB, HMM-B2TMR CONSENSUS barrel 0.771 0.596 0.877 17 19
precursor 0.818 0.628 0.86 15 18
PRED-TMBB, ProfTMB, HMM-B2TMR, B2TMPRED CONSENSUS barrel 0.790 0.616 0.896 17 19
precursor 0.832 0.641 0.865 15 18
PRED-TMBB, ProfTMB, HMM-B2TMR, TBBPred-nn CONSENSUS barrel 0.809 0.635 0.917 18 20
precursor 0.839 0.653 0.867 15 18
PRED-TMBB, ProfTMB, HMM-B2TMR, TBBPred-svm CONSENSUS barrel 0.809 0.629 0.906 15 19
precursor 0.847 0.658 0.882 15 18
PRED-TMBB, ProfTMB, HMM-B2TMR, TBBPred-comb CONSENSUS barrel 0.791 0.607 0.894 17 20
precursor 0.833 0.648 0.859 15 18
PRED-TMBB, ProfTMB, HMM-B2TMR, TBBPred-nn/svm CONSENSUS barrel 0.824 0.638 0.92 17 19
precursor 0.85 0.647 0.871 13 17
PRED-TMBB, ProfTMB, HMM-B2TMR, B2TMPRED, TBBPred-nn/svm CONSENSUS barrel 0.825 0.637 0.927 17 18
precursor 0.854 0.652 0.876 15 17
PRED-TMBB, ProfTMB, HMM-B2TMR, B2TMPRED, TBBPred-nn CONSENSUS barrel **0.819** **0.641** **0.924** **18** **20**
precursor **0.849** **0.660** **0.874** **15** **18**
PRED-TMBB, ProfTMB, HMM-B2TMR, B2TMPRED, TBBPred-comb CONSENSUS barrel 0.807 0.625 0.907 17 19
precursor 0.845 0.658 0.868 15 18
PRED-TMBB, ProfTMB, HMM-B2TMR, B2TMPRED, TBBPred-svm CONSENSUS barrel 0.819 0.637 0.910 15 19
precursor 0.853 0.659 0.880 14 18
PRED-TMBB, ProfTMB, B2TMPRED, TBBPred-svm/nn CONSENSUS barrel 0.829 0.642 0.923 17 18
precursor 0.851 0.648 0.861 15 16
PRED-TMBB, ProfTMB, B2TMPRED, TBBPred-svm, TBBPred-nn, HMM-B2TMR, TMBETA-NET, PSI-PRED, BETA-TM CONSENSUS barrel 0.808 0.582 0.851 11 13
precursor 0.844 0.604 0.841 12 13
We report the consensus of all the available methods, and the ones that were obtained using the 3 top-scoring HMMs combined in various ways with some of the top-scoring NN/SVM methods. The best results are highlighted with bold. For abbreviations see also Table 1.
:::
::: {#T5 .table-wrap}
Table 5
::: {.caption}
######
The non-redundant data set of 20 β-barrel outer membrane proteins used in this study.
:::
**Protein name** **Number of β-strands** **PDB ID** **Reference** **Organism**
------------------ ------------------------- ------------ --------------- -----------------------------
NspA 8 1P4T \[67\] *Neisseria Meningitidis*
OmpX 8 1QJ8 \[68\] *Escherichia coli*
Pagp 8 1MM4 \[69\] *Escherichia coli*
OmpA 8 1QJP \[50\] *Escherichia coli*
OmpT 10 1I78 \[70\] *Escherichia coli*
OpcA 10 1K24 \[71\] *Neisseria Meningitidis*
Nalp 12 1UYN \[41\] *Neisseria Meningitidis*
OmpLA 12 1QD5 \[72\] *Escherichia coli*
Porin 16 2POR \[73\] *Rhodobacter capsulatus*
Porin 16 1PRN \[74\] *Rhodopseudomonas blastica*
OmpF 16 2OMF \[75\] *Escherichia coli*
Osmoporin 16 1OSM \[76\] *Klebsiella pneumoniae*
Omp32 16 1E54 \[77\] *Comamonas Acidovorans*
Phosphoporin 16 1PHO \[78\] *Escherichia coli*
Sucrose porin 18 1A0S \[79\] *Salmonella typhimurium*
Maltoporin 18 2MPR \[80\] *Salmonella typhimurium*
FhuA 22 2FCP \[46\] *Escherichia coli*
FepA 22 1FEP \[47\] *Escherichia coli*
FecA 22 1KMO \[48\] *Escherichia coli*
BtuB 22 1NQE \[49\] *Escherichia coli*
:::
::: {#T6 .table-wrap}
Table 6
::: {.caption}
######
The available predictors, used for predicting the transmembrane strands of β-barrel outer membrane proteins.
:::
Method Reference Type TM Strands TM Strands + Orientation Discrimination URL
----------------------- ------------ -------- ------------ -------------------------- ---------------- ---------------------------------------------------------------------------
B2TMPRED \[15\] NN x \- \- <http://gpcr.biocomp.unibo.it/cgi/predictors/outer/pred_outercgi.cgi>
HMM-B2TMR (1) \[17\] HMM x x \- <http://gpcr.biocomp.unibo.it/biodec/> (1)
OM\_Topo\_predict (2) \[14\] NN x x \- <http://strucbio.biologie.uni-konstanz.de/~kay/om_topo_predict2.html> (2)
PRED-TMBB \[19, 20\] HMM x x x <http://bioinformatics.biol.uoa.gr/PRED-TMBB/>
ProfTMB \[21\] HMM x x x <http://cubic.bioc.columbia.edu/services/proftmb/>
TBBpred \[22\] NN+SVM x \- x <http://www.imtech.res.in/raghava/tbbpred/>
BETA-TM \[58\] HMM x x \- <http://dblab.sejong.ac.kr:8080/barrel/index.html>
TMBETA-NET \[16\] NN x \- \- <http://psfs.cbrc.jp/tmbeta-net/>
PSI-PRED \[62\] NN \- \- \- <http://bioinf.cs.ucl.ac.uk/psipred/>
We list the name of the predictor, the reference paper, the type of the method (HMM, NN or SVM), whether it predicts the transmembrane strands, the full topology (TM strands+orientation) and if they are capable of discriminating between β-barrel membrane proteins from non-β barrel membrane proteins.
\(1) HMM-B2TMR is available as a commercial demo only.
\(2) The OM\_Topo\_predict web server was not operational, at the time when this research was conducted.
:::
|
PubMed Central
|
2024-06-05T03:55:52.020772
|
2005-1-12
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC545999/",
"journal": "BMC Bioinformatics. 2005 Jan 12; 6:7",
"authors": [
{
"first": "Pantelis G",
"last": "Bagos"
},
{
"first": "Theodore D",
"last": "Liakopoulos"
},
{
"first": "Stavros J",
"last": "Hamodrakas"
}
]
}
|
PMC546000
|
Background
==========
Early exposure to dietary and supplementary vitamin D has been predicted to be a risk factor for later allergy and asthma \[[@B1]\]. Supported by *in vitro*\[[@B2]\] and *in vivo*studies \[[@B3]\], also epidemiological studies \[[@B4],[@B5]\] report a positive association between supplementary vitamin D use and later allergy \[[@B6]\].
Vitamin D has been used for many years in various doses and preparations to prevent rickets, a disease usually induced by poor dietary calcium intake and sun deprivation. It seems that widespread \"historical\" rickets in industrial countries was also a genetic disease. A formal twin analysis yielded a 91% concordance rate in monozygotic twins compared to 23% in dizygotic twins \[[@B7]\]. Also a very recent study of baseline gene expression in lymphoblastoid cell-lines found the expression of at least four vitamin D related genes as a heritable trait \[pers. comm. Monks 2004\], which also makes a genetically determined vitamin D sensitivity likely. It may be speculated that common rickets is the low sensitivity form (in the absence of proper endogenous vitamin D production), and allergy the high sensitivity form (in the presence of high oral vitamin D exposure).
The active vitamin D metabolite 1,25(OH)~2~D~3~binds to nuclear vitamin D receptor (VDR), which exists from under 500 to over 25,000 copies / cell in many human tissues including thymus, bone marrow, B and T cells and lung alveolar cells \[[@B8]\]. The gene for VDR was cloned in 1988, it consists of 9 exons with at least 6 isoforms of exon 1 and spans 60--70 kb of genomic sequence (Fig. [1](#F1){ref-type="fig"}) \[[@B8]\]. The VDR is also a first-order positional candidate as nearly all asthma and allergy linkage studies found linkage on chromosome 12q \[[@B9],[@B10]\]. While an own study of a single *Fok1*restriction site (that alters the ATG start codon in the second exon of the VDR) in asthma families did not find an association \[[@B11]\], positive association of several VDR variants with asthma has been shown in the meantime by two U.S. \[[@B12]\] as well as one Canadian study \[[@B13]\].
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Structure of the vitamin D receptor \[40, 41\]. Upper: LD blocks in Caucasians \[14\] where blocks \"C\" and \"A\" extend to both sides. In Africans block \"C\" is split into 3 parts, \"C1\", \"C2\" and \"C3\". Middle: Exon structure including several SNP variants examined in about 100 disease-association studies. Lower: Aligned protein domains, DNA binding, hinge and ligand binding region including phosphorylation sites.
:::
:::
So far, dbSNP catalogued 117 SNPs in the VDR when a resequencing approach of the VDR published in June 2004 found 245 SNPs \[[@B14]\]. In this study three LD blocks were localized. Block \"A\" at the 3\' end of exon 9 spans approximately 10.5 kb. VDR exons 3 -- 9 are situated in block \"B\", which spans 40.8 kb. A 5.7 kb LD-breaking spot separates blocks \"A\" and \"B\", while blocks \"B\" and \"C\" are separated by a 1.3 kb LD-breaking spot; this region also includes VDR exon 2 and the commonly studied *FokI*SNP. All three LD blocks have now been covered by additional SNPs in our asthma family sample.
Results
=======
The sample analyzed here consisted of 951 individuals from 224 pedigrees contributing 11,383 genotypes.
Mean pedigree size was 4.4; the number of phase-known individuals ranged from 221 to 305 (with the exception of rs2853563) in affected, and from 24 to 41 in the unaffected, children. Except for rs2853563, the minor allele frequency always exceeded 19%. In two families both parents had asthma, in 82 only one parent had asthma and in 140 families both parents were disease-free.
Of the markers tested, only marker rs2239186 showed a slightly reduced transmission ratio in asthmatic children (0.8, P = 0.073). In unaffected children three markers showed significantly altered ratios: 0.5 for hCV2880804, and 2.0 and 1.9 in rs1989969 and rs2853564, respectively. Excess transmission was generally more pronounced in unaffected children (Fig. [2](#F2){ref-type="fig"}), however, the three significant associations would also not persist if adjusted for multiple testing with the method described by Bonferroni.
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Excess transmission of VDR-SNPs in asthma families. Blue bars indicate transmission to affected children, grey bars transmission to unaffected children.
:::
:::
The TDT~DS~test as a global family test seems to capture the information both from affected and unaffected children, with four SNPs being at least marginally significantly associated at the 5% level. As unaffected children might be the younger children (that still have not developed a phenotype), the age distribution of affected and unaffected children was also compared (Fig. [3](#F3){ref-type="fig"}) but no difference was found. Affected and non affected children, however, show several other differences, all known as risk factors and symptoms for asthma: There are more boys in the affected group (57,6% vs. 36,1%), they are more often exposed to indoor environmental tobacco smoke (43.0% vs. 31.7%) and they are suffering more frequently from eczema (43,0% vs. 26,9%). Their average log(IgE) values are higher (7,7 kU/l vs. 5,8 kU/l), their forced 1 second capacity FEV1 is lower (2387 ml vs. 2625 ml) together with the forced vital capacity (2876 vs. 2998).
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Age distribution in children with, blue box (left), and without asthma, grey box (right).
:::
:::
Bronchial hyper-reactivity (BHR), measured as quantitative trait locus (QTL) by using the slope of the dose-response curve in a standardized methacholine challenge protocol, also indicated an association with the same three SNPs identified in unaffected sibs. Further dichotomizing BHR by comparing the upper versus all other quartiles also yielded significant associations (hCV2880804, P = 0.005, rs1989969 P = 0.001, rs2853564 P = 0.001). The sum score of all specific IgE serum levels (RAST) was associated with markers rs1989969 (P = 0.021) and rs2853564 (P = 0.018), while total IgE was not found to be associated.
Table [2](#T2){ref-type="table"} summarizes the LD structure of all SNPs in the German families. LD was generally low, except for three SNPs on block \"B\". SNP rs2853563 probably does not interrupt block \"B\", as might be concluded from table [2](#T2){ref-type="table"}, as this marker has a very low minor allele frequency (table [1](#T1){ref-type="table"}). The overall low LD was very similar in the Swedish and Turkish subgroups of our families and are in line with recently published data \[[@B12]-[@B14]\]. The association seen for the two SNPs at the 3\'-terminal region is probably influenced by the high LD between these two SNPs.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Transmission disequilibrium of 13 VDR-SNPs in affected and unaffected children. P-values \< 0.05 are indicated in bold.
:::
asthma no asthma
------------- ------------ -------- ------------ ------------ ---------- -------- ----------- ------ ------- ----- ----- ------ ----------- -----------
LD block \* dbSNP allele chr12 (bp) geno-types freq (%) T UT T/UT P T UT T/UT P P TDT~DS~
A hCV2880804 C 47954855 918 28 170 178 1,0 0,668 18 33 0,5 **0,049** 0,055
\-- rs2228570 T 47977944 724 35 159 150 1,1 0,609 21 12 1,8 0,117 0,079
? rs3819545 C 47981753 912 38 186 191 1,0 0,797 25 31 0,8 0,423 1,000
B rs3782905 G 47982914 934 32 201 174 1,2 0,163 26 27 1,0 0,891 0,388
? rs2853563 G 48248398 661 97 20 24 0,8 0,547 \-- \-- \-- \-- \--
B rs731236 C 48251417 923 37 202 206 1,0 0,843 25 29 0,9 0,586 0,767
B rs1544410 A 48252495 896 37 190 186 1,0 0,837 23 26 0,9 0,668 1,000
B? rs2239185 C 48257219 903 49 198 214 0,9 0,431 29 24 1,2 0,492 0,650
? rs987849 C 48267336 825 45 164 169 1,0 0,784 28 18 1,6 0,140 0,173
B rs1540339 A 48269986 916 36 189 187 1,0 0,575 25 27 0,9 0,782 0,706
? rs2239186 C 48282070 929 19 116 145 0,8 0,073 20 18 1,1 0,746 0,189
C rs1989969 T 48290670 922 43 204 198 1,0 0,765 39 20 2,0 **0,013** **0,025**
C rs2853564 C 48291147 920 44 196 201 1,0 0,802 38 20 1,9 **0,018** **0,021**
\*according to reference \[15\]
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
D\' matrix of 13 VDR-SNPs in German parents. D\' values \> = 0.69 are indicated in bold.
:::
rs2228570 rs3819545 rs3782905 rs2853563 rs731236 rs1544410 rs2239185 rs987849 rs1540339 rs2239186 rs1989969 rs2853564
------------ ----------- ----------- ----------- ----------- ---------- ----------- ----------- ----------- ----------- ----------- ----------- -----------
hCV2880804 0,01 0,02 -0,04 0,05 0,01 0,01 -0,02 -0,01 0,05 0,02 -0,27 -0,27
Rs2228570 0,02 -0,03 0,01 0,06 0,02 -0,01 -0,03 -0,03 -0,10 0,12 0,13
Rs3819545 -0,53 0,08 -0,41 -0,41 0,26 0,25 **0,85** 0,58 -0,02 -0,02
Rs3782905 -0,08 0,53 0,55 -0,44 -0,45 -0,45 -0,33 0,05 0,05
Rs2853563 -0,14 -0,16 0,20 -0,10 0,19 -0,01 -0,07 -0,08
Rs731236 **0,98** **-0,76** **-0,69** -0,46 -0,32 0,07 0,05
Rs1544410 **-0,76** **-0,69** -0,46 -0,33 0,06 0,05
Rs2239185 **0,85** 0,28 0,33 -0,06 -0,05
Rs987849 0,22 0,40 -0,02 -0,01
Rs1540339 0,58 -0,07 -0,07
Rs2239186 -0,11 -0,11
Rs1989969 **0,98**
:::
Finally, a 4-locus haplotype was constructed of those SNPs associated in the TDT~DS~. Again, no significant transmission distortion was found in affected children, while one haplotype showed a 5.1-fold over-transmission in unaffected children (P = 0.009).
Discussion
==========
This study addresses a previously described association of VDR SNPs with asthma and related phenotypic traits. Although a preferential transmission of vitamin D receptor variants to children with asthma could not be confirmed, it raises the possibility of a protective effect in unaffected children. Although the effect is rather moderate, several SNPs support this association. If any, the association is probably more related to RNA turnover than to a structural modification of the receptor as there was no association with LD block \"B\" that codes for the translated exons. Gene expression can be varied over a 100-fold range by subtle modifications of the 3\'-terminal sequence \[[@B15]\], which will requires further research on the function of these allelic variants in target tissues.
This notion is partially in contrast with a previous study of 7 VDR SNPs in the CAMP (Childhood Asthma Management Program) study of 582 nuclear families where SNP rs7975232 (akin ApaI) in intron 8 showed a highly significant effect \[[@B12]\]. A confirmation study by the same group in a case-control sample of the Nurses\' Health Study NHS \[[@B12]\] also associated asthma with rs3782905 (intron 2, P = 0.02), rs2239185 (intron 3, P = 0.02) and rs731236 (Ile352Ile, P = 0.03). A study of 223 independent Canadian families reported six out of twelve SNPs to be associated with asthma (rs3782905, rs1540339, rs2239182, rs2239185, BsmI, ApaI, TaqI), most of these on block \"B\". As none of these SNPs was giving rise to an amino acid change the authors speculated about an intronic regulatory SNP or one or more functional variants at the 3\' end of the VDR locus.
It is unlikely that increased or decreased vitamin D sensitivity is simply mediated by a genetic variation in the VDR. Vitamin D requires several enzymatic steps to be activated, transported and degraded; receptor signalling requires several co-factors and all of these may contribute additive or multiplicative effects on vitamin D sensitivity. The co-activator retinoic acid receptor RXR may itself affect Th1 and Th2 development \[[@B16]\]. Other transcription factors involved are SRC/p160, CBP and p300 \[[@B17]\]. The DRIP complex attaches to the VDR/RXR complex and binds to vitamin D responsive elements (VDRE) through histone acetyltransferase activity. Importantly, some of the vitamin D regulated genes are also located in allergy linkage regions. The renal 1-α-hydroxylase (12q13) and the 24-hydroxylase (20q13), as well as RXR (6p21), are all positional candidates. RXR has already been tagged by a SNP in a previous study \[[@B18]\] and also 3 SNPs in CYP24A1 (rs751089, rs2296241 and rs2248137) were significantly associated with asthma (*unpublished own observation*). CYP24A1 is particular interesting as it is the major enzyme of the degradation pathway that showed a 97-fold increase after vitamin D treatment of rats \[[@B19]\] or 12-fold increase in a human colon cancer cell line \[[@B20]\].
The number of genes regulated by vitamin D has been recently extended beyond those genes with known VDRE promoter motifs (calbindin, PTH, PTHRP, ITGB3, OC, GH, osteopontin, osteocalcin, c-fos, IL-2Rβ, NFκB, sCD23) to another 150 up- and down-regulated regulated genes \[[@B19]-[@B23]\]. The interaction of genetic variants in the VDR and other positional, as well as functional, candidate genes is therefore a current research topic. So far, only two gene associations have been published: Vitamin D binding protein (GC\*2 and GC\*1F) was associated with an increased risk for COPD \[[@B24],[@B25]\] and osteopontin (OPN C8090T and T9250C) with increased total IgE in asthmatic patients \[[@B26]\].
In addition to its biological context, this study also has some implications on the statistical analysis as the rather low number of unaffected sibs seemed to contribute to a few positive associations. Despite enormous efforts to map complex genetic diseases, SNP association studies are often lacking power \[[@B27]\]. Linkage studies in affected sib pairs have been preferably used to map complex diseases to chromosomal regions \[[@B9]\], while \"no substantial study of normal sib-pairs has been undertaken, making this family of surveys one of the largest undertaken in the absence of controls\" \[[@B28]\] although unselected affected sib pairs tend to share more than half of their alleles \[[@B29]\]. This omission is even more remarkable as discordant sib pairs (DSP) have been shown to be a more powerful alternative \[[@B30]\]. Risch proposed to test DSP in the top ten and bottom ten percent distribution of quantitative traits, as pairs with intermediate values (between the 30^th^and 70^th^percentiles) did not provide much information for linkage analysis \[[@B31]\]. Although the DSP concept was appealing from a theoretical standpoint, it turned out that there are disadvantages for practical reasons. It requires a large amount of individuals to be screened, which might be the reason that the DSP approach has not received the expected attention except for a few studies (for a summary see \[[@B31]\]).
Transmission disequilibrium testing of unaffected child -- parent trios originating from families with another two affected offsprings, may be a powerful alternative. As there is a strong ascertainment bias of these families toward a genetic risk, as well as a disease causing environmental factor, being unaffected is an extreme phenotype (\"being sane in an insane world\"). The transmission to unaffected children can be seen as an independent cross-match to the transmission to affected children. The high power of testing unaffected sibs has already been predicted on theoretical grounds \[[@B32]\]. The number of DSPs required to achieve 80% power (with a difference in the allele frequency of 15% and λ~s~of 3.2) has been estimated to be approx. 250 \[[@B33]\]. A sample size of 1,500 families was estimated by including two affected and one unaffected children \[[@B34]\]. In this study already 50 DSPs were sufficient to show a significant distortion in the allele transmission. Non-paternity as a reason for discordant traits \[[@B35]\] is unlikely as nearly all families were included in a previous genome-wide scan.
Conclusions
===========
The transmission disequilibrium in unaffected sibs in otherwise multiple-affected families seems to be a powerful test. A preferential transmission of vitamin D receptor variants to children with asthma could not be confirmed but raises the possibility of a protective effect in unaffected children.
Methods
=======
Study population
----------------
The German asthma sib pair families were collected in 26 paediatric centres in Germany and Sweden for an initial genome-wide linkage scan. In these families at least two children were required with confirmed doctor-diagnosed asthma, while prematurity or low birth weight of the children were excluded, along with any other severe pulmonary disease. All affected children should have after their 3^rd^birthday a history of at least three years of recurrent wheezing and should not have any other airway disease diagnosed. Unaffected siblings were also sampled if they were at least 6 years old, eligible for pulmonary function testing and did not have doctor-diagnosed asthma.
On the first home visit a complete pedigree of the family was drawn and information collected in a questionnaire. Participants were examined for several associated phenotypes. Pulmonary function tests were performed by forced flow volume tests and bronchial challenge was done by methacholine.
Briefly, pulmonary function tests were performed by forced expiration in a sitting position using a nose-clip. Forced flow volume tests were performed until three reproducible loops were achieved. Of these the trial with the maximum sum of FVC and FEV1.0 was used for the analysis. Bronchial challenge with methacholine was done with increasing doses of 0, 0.156, 0.312, 0.625, 1.25, 2.5, 5, 10, 25 mg/ml during 5 consecutive breaths with 14 mg delivered from a de Vilbiss 646 nebulizer chamber by using a breath-triggered pump ZAN 200 (Zan, Oberthulpa, Germany). The provocation was stopped either with the occurrence of symptoms or a fall of 20% from the baseline FEV1.0 and the slope of the dose-response curve calculated.
Total IgE was determined with an ELISA (Pharmacia Diagnostics, Uppsala, Sweden). The allergens tested were birch (betula verruscose) ALK SQ108, hazel (corylus avellana) ALK SQ113, the herbs ribworth (plantago lanceolata) ALK N342 and mugwort ALK SQ312, mixed grass ALK SQ299, dust mite dermatophagoides farinae ALK SQ 504, dermatophagoides pteronyssimus ALK SQ 503, cat dander ALK SQ555 and dog dander ALK SQ 553, and fungi (aspergillus fumigatus ALK N405 and alternaria alternata ALK N402), which were bought in one batch and stored at +5°C until analysis.
The original family collection \[[@B36]\] has been expanded since 1994 to a larger sample, of which 218 families were available in 2002 for a complete genome scan. The consecutive families were tested with the same protocol as described earlier \[[@B36]\] except that the time-consuming methacholine challenge protocol was omitted. Excluded from the final sample were individuals with incomplete data, missing DNA samples, identical twins and all probands with more than 2 non-segregating out of 408 microsatellite markers. Another 6 families could be included in this analysis, resulting in a final sample of 224 families. Each study participant, including all children, signed a consent form. All study methods were approved in 1995 by the ethics commission of \"Nordrhein-Westfalen\" and again in 2001 by \"Bayerische Landesärztekammer München\".
DNA preparation and genotyping
------------------------------
DNA was isolated from peripheral white blood cells using Qiamp (Qiagen, Germany) or Puregene isolation kits (Gentra Systems, Minneapolis, MN, USA). From the VDR we selected 16 SNPs where 13 could finally be analyzed. They had to be polymorphic, complete, received a high calling score, passed paternity checks and were in Hardy-Weinberg equilibrium. The following three markers have been excluded from the analysis: rs2239179 (contaminant in mass spectrum peak), rs2853559 (Hardy-Weinberg disequilibrium, paternity errors) and rs797523 (typing error in primer sequence).
Genotyping was performed using MALDI-TOF mass spectrometry of allele-specific primer extension products (Table [3](#T3){ref-type="table"}) generated from amplified DNA sequences (MassARRAY, SEQUENOM Inc., San Diego, CA, USA). Primers were obtained from Metabion GmbH (Planegg-Martinsried, Germany).
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Primer used for genotyping
:::
SNP left primer right primer extension primer stop mix
------------ ---------------------------------- --------------------------------- ----------------------- ----------
hCV2880804 ACGTTGGATG-CTGGGATACTTCTGAGAGTG ACGTTGGATG-TTTTCCCTGGAAAGTTTGGG CCATTGTCCTGGTATAACCA ACG
rs2228570 ACGTTGGATG-TCAAAGTCTCCAGGGTCAGG ACGTTGGATG-AGACCTCACAGAAGAGCACC CCCTGCTCCTTCAGGGA ACG
rs3819545 ACGTTGGATG-AATGGTGGTTACTGCAGCTC ACGTTGGATG-GAAACCAGTCCTCTGTCATG TAGGTTCGGTCTTTGGCT ACG
rs3782905 ACGTTGGATG-GGGTCTCAAATTCTTAATGAG ACGTTGGATG-AAACTAGCAGAAAGAGGCAG GTGGGAGGGAGTGCTGA ACT
rs2853564 ACGTTGGATG-CTGCATTGCTCCTGACTTAG ACGTTGGATG-AGGTAGCTTAGCTCTGAGTC TTTCTGCAACCCTAAGCC ACT
rs731236 ACGTTGGATG-TGTGCCTTCTTCTCTATCCC ACGTTGGATG-TGTACGTCTGCAGTGTGTTG CGGTCCTGGATGGCCTC ACT
rs1544410 ACGTTGGATG-TAGATAAGCAGGGTTCCTGG ACGTTGGATG-AATGTTGAGCCCAGTTCACG AGCCTGAGTATTGGGAATG ACG
rs2239185 ACGTTGGATG-CAATTCCAGTCACATCTCGG ACGTTGGATG-CCTGTGTGACATTTACACCC CCCTCCTCTGTCTTCAC ACT
rs987849 ACGTTGGATG-GAATAGTGCCTTATAGATAG ACGTTGGATG-AGCTAGAAGTTCTGGTGATC GAAATATTCGTAATGCTGGAT ACT
rs1540339 ACGTTGGATG-TCACACACATTCTCAGTGGG ACGTTGGATG-TTTGCAGAGGCTGTCTTCTC GTTGGTGCCCACCCTAA ACG
rs2239186 ACGTTGGATG-GTCCACAGTGACTATAGACC ACGTTGGATG-AAGAAGGAGAAGCAGGCATC CAGGGGTGGAAGAAGAGGAG ACT
rs1989969 ACGTTGGATG-TGTATGCAGAGCTTAGCAGG ACGTTGGATG-TTTCAGAGGTCAGAGGTGAC GTCAGAGGTGACATCCAG ACT
rs2853564 ACGTTGGATG-CTGCATTGCTCCTGACTTAG ACGTTGGATG-AGGTAGCTTAGCTCTGAGTC TTTCTGCAACCCTAAGCC ACT
:::
Data handling and statistical analysis
--------------------------------------
Clinical data and genotypes were transferred to a SQL 2000 database by using Cold Fusion 4.0 scripts. Statistical analyses were performed using R 1.8.1 \[[@B37]\] by accessing the database with the RODBC module. For each SNP, the distribution of genotypes in pseudo-controls created by combining the parental alleles not transmitted to asthma children was tested by a χ^2^-test as well as the transmission to their unaffected sibs. An extension of the classical TDT \[[@B34]\] was also implemented that incorporates effectively both affected and unaffected children (TDT~DS~). The standardized linkage disequilibrium coefficient D\' and the correlation coefficient R^2^were calculated for each pair of SNPs in parents by using the R package \"genetics\". All analyses were cross-checked by using SIBPAIR software. Haplotypes were estimated using TDTPHASED in the UNPHASED package \[[@B38]\] and transmission to affected and unaffected children tested separately. For phase-certain haplotypes a conditional logistic regression model was used, corresponding to the probability of the offspring conditional upon the parents. When phase was uncertain, unconditional logistic regression on the full likelihood of parents and offspring was used instead (see \[[@B39]\] for the formulations of these likelihoods). Transmitted haplotypes were compared to all untransmitted haplotypes, equivalent to the haplotype-based haplotype relative risk, while an EM algorithm was used to obtain maximum-likelihood estimates of case and control parental haplotype frequencies under both null and alternative hypotheses.
Abbreviations
=============
VDR vitamin D receptor
VDRE vitamin D responsive element
SNP single nucleotide polymorphism
MALDI-TOF matrix assisted laser desorption ionisation -- time of flight
DSP discordant sib pairs
TDT transmission disequilibrium
LD linkage disequilibrium
QTL quantitative trait locus
BHR bronchial hyperreactivity
FEV1 forced volume during the 1st second
Author\'s contribution
======================
The author developed the idea presented in this paper, initiated the study, applied for funding, developed protocols, participated in the clinical survey, planned the laboratory analysis, did the statistical analysis, and drafted the report. This manuscript contains no patient identifiable information.
Acknowledgments
===============
I thank all participating families for their help and the members of the clinical centers for their work: R. Nickel, K. Beyer, R. Kehrt, U. Wahn (Berlin), K. Richter, H. Janiki, R. Joerres, H. Magnussen (Grosshansdorf), I. M. Sandberg, L. Lindell, N.I.M. Kjellman (Linkoeping), C. Frye, G. Woehlke, I. Meyer, O. Manuwald (Erfurt), A. Demirsoy, M. Griese, D. Reinhardt (München), G. Oepen, A. Martin, A. von Berg, D. Berdel (Wesel), Y. Guesewell, M. Gappa, H. von der Hardt (Hannover), J. Tuecke, F. Riedel (Bochum), M. Boehle, G. Kusenbach, H. Jellouschek, M. Barker, G. Heimann (Aachen), S. van Koningsbruggen, E. Rietschel (Köln), P. Schoberth (Köln), G. Damm, R. Szczepanski, T. Lob-Corzilius (Osnabrück), L. Schmid, W. Dorsch (München), M. Skiba, C. Seidel, M. Silbermann (Berlin), A. Schuster (Düsseldorf), J. Seidenberg (Oldenburg), W. Leupold, J. Kelber (Dresden), W. Wahlen (Homburg), F. Friedrichs, K. Zima (Aachen), P. Wolff (Pfullendorf), D. Bulle (Ravensburg), W. Rebien, A. Keller (Hamburg) and M. Tiedgen (Hamburg). I wish to thank M. Hoeltzenbein who helped start the study and J. Altmüller who managed the second part of the study, G. Schlenvoigt and L. Jaeger for IgE determination; former group members L. Thaller, G. Fischer, T. Illig, N. Klopp, C. Vollmert and M. Werner and doctoral students H. Gohlke, N. Herbon and G. Dütsch for their contribution in previous genotyping projects; P. Lichtner for administrative help; A. Jendretzke from Sequenom Inc. for technical support; M. Bahnweg, A. Luze, C. Braig and B. Wunderlich for excellent laboratory work. C. Braig did the genotyping reported here. The asthma family study was funded by BMBF 07ALE087, Deutsche Forschungsgemeinschaft DFG WI621/5-1, DFG FR1526/1 and National Genome Network 01GS0122 (until 30th June 2004). I was funded by GSF FE 73922. Finally I wish to thank E. André and M. Emfinger for proof-reading of the manuscript; E. Hyppönen and B. Raby for helpful discussions on various aspects covered in this paper.
|
PubMed Central
|
2024-06-05T03:55:52.027490
|
2005-1-15
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546000/",
"journal": "BMC Genet. 2005 Jan 15; 6:2",
"authors": [
{
"first": "Matthias",
"last": "Wjst"
}
]
}
|
PMC546001
|
Background
==========
In the past decade, the demand for DNA sequence data has driven the transformation of sequencing from a research activity into a manufacturing process. High-throughput sequencing facilities are focused on establishing automated procedures that maintain long read length and high overall success rates. It is neither practical nor economical to test each and every DNA template before sequencing \[[@B1]\]. Sequencing centres, therefore, monitor sequencing success on a larger scale referencing overall pass rates and average read lengths, typically in terms of Phred 20 bases \[[@B2]\]. The percentage of \"sporadic sequence dropouts\" or failed reads that inevitably occur within a pool of high quality data is often overlooked and rarely examined. Failed reads can be a result of numerous variables ranging from pipeline methodology employed to the nature of samples being sequenced. A Failure Mode Analysis (FMA) strategy was developed to determine the likely causes of sporadic unsuccessful sequence reads. We systematically examine these failed reads in the context of a high-throughput sequencing pipeline to establish the mode and frequency of each type of failure. The standard production pipeline at Canada\'s Michael Smith Genome Sciences Centre (BCCRC, British Columbia Cancer Agency, Vancouver, Canada) has a capacity to generate over 3.6 million reads per year. As of December 8, 2004, we have generated 1,263,904,347 Q~20~bases using our 384-well culturing, DNA preparation, and cycle sequencing procedures. The average Q~20~read length of data generated in the past 12 months (December 2003 to December 2004) from various library types and vector systems is 751 bases. The present study was undertaken to provide insight into the causes of sequencing failures and possible corrective actions. Although our pipeline uses exclusively ABI 3700 and 3730XL automated sequencers, these results should be applicable, in principle, to the improvement of other high throughput sequencing platforms.
Results
=======
We generated 9,216 reads from 2,304 clones selected randomly from two cDNA libraries. For each of the two libraries, 1,152 bacterial colonies containing cDNA inserts were picked and arrayed into 384-well microtiter plates (Figure [1](#F1){ref-type="fig"}). To verify loss of DNA due to handling or equipment mishaps (i.e. clogged capillary or tip), each microtiter plate was cultured in duplicate and replicates were processed using the same instrument model but on different physical units where available. A resulting 4,608 reads were generated for the 5\' end using the M13Reverse (5\'-CAGGAAACAGCTATGAC-3\') primer and 4,608 reads were generated from the 3\' end using the M13 Forward (5\'-TGTAAAACGACGGCCAGT-3\') primer. The average Q~20~read length for the entire data set was 771 bases, average pass rate was 87% which was calculated as a percentage of sequencing reactions yielding a minimum of 600 Phred 20 bases. Figure [2](#F2){ref-type="fig"} illustrates a break down of Q~20~read lengths from the full data set. The analysis methodology employed to determine the failure mode of each trace is outlined in Figure [3](#F3){ref-type="fig"}. 1,172 reads (13%) represent the failed portion of the data set (Q~20~\< 600) for further analysis to determine failure mode. The electropherograms from the 1,172 failed reads were evaluated and subsequently categorized into failure mode categories. 64 of these reads were yielded from sequencer capillaries that were clogged and therefore were removed from further analysis and categorized into the \"Blocked capillary\" failure mode. The remaining 1,108 traces were further classified into nine additional failure mode categories including Low signal strength, Mixed clone with vector sequence, Mixed clone- no vector sequence, Low signal to noise ratio, Excess dye peaks, Hardstop, Repetitive sequence, Homopolymer stretch, and Poly A tail. Results and trace characteristics used to classify each read are as described in Table [1](#T1){ref-type="table"}. Eight of the classifications described in Table [1](#T1){ref-type="table"}, except \"Low signal strength\", are final failure mode categories and contain 74% of the total reads.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Failure mode categories Failed wells were distributed into each category based on observational data taken during sequencing pipeline procedures and manual evaluation of electropherogram traces.
:::
Failure Mode Trace characteristic No. of sequencing reactions Percent of all failed wells (Q~20~\< 600)
--------------------------------- ----------------------------------------------------------------------------------------------------------------------------------- ----------------------------- -------------------------------------------
Blocked capillary Noisy or no data with a low signal intensity value (\<100). Verified with capillary control results. 64 5.5
Low signal strength\* Noisy or no data with a low signal intensity value (\<100) that is very close to or falls below the instruments detectable limit. 310 26.5
Mixed clone w/ vector sequence Clean vector sequence followed by noisy data immediately after the cloning site. 137 11.7
Mixed clone, no vector sequence Noisy data throughout the trace with sufficient signal intensity. 27 2.3
Low signal to noise ratio Discernable sequence peaks with strong intensity background noise. 22 1.9
Excess Dye peaks Large dye front usually followed by noisy data. 10 0.9
Hardstop Abrupt end to good sequence. 2 0.2
Repetitive Sequence Long stretch of repetitive DNA sequence that is followed by slippage in sequence or noisy data. 17 1.5
Homopolymer stretch Long stretch of a single nucleotide followed by slippage in sequence or noisy data. 99 8.4
Poly A Tail Stretch of Ts (template A) followed by slippage in sequence or noisy data. 484 41.3
Total 1172 100.0
\* Preliminary failure mode, further broken down to final failure modes in Table 2.
:::
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Process pipeline.**Observational checks within the pipeline are shaded in grey. Absence of bacterial colonies, no-grows, and unusual observations are recorded on logsheets then entered into the FMA database. A. Verification of the colony picking procedure to ensure that all original clones are accounted for in the source microtiter plates. B. To further confirm A, we stamp a replicate of each microtiter plate containing transformed bacterial cultures onto agar plates. The resulting pattern of colonies is examined to determine presence or loss of DNA in each well of the source microtiter plate. C. Every 384-well culture plate is visually examined for presence of DNA after bacterial DNA culturing. D. Agarose gel electrophoresis is used to evaluate presence and quality of prepared and purified template DNA. E. During sequencing reactions, all volume additions are visually verified and manually adjusted using a single channel pipettor where necessary.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Average read length breakdown**Distribution of read length (Q~20~bases) for full data set of 9,216 reads. Results were divided into 100 bp bins, failed reads (Q^20^\<600 bp) make up 13% (1,172 reads) of overall reads. Reads with Q^20^\<100 bp make up the largest proportion of failed reads but contributes to only 4% of overall data set.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Analysis pipeline**Analysis methodology used to determine failure modes. FM= Failure Mode.
:::
:::
We identified various failure mode trends by examining process versus template-related failures. Figure [4](#F4){ref-type="fig"} shows 62.4% of reads with Q~20~\<100 bp fail as a result of process-related problems while 75.5% of reads in the Q~20~: 500--599 bp bin fail due to template-related characteristics. The proportion of failed reads due to process-related problems are more abundant in traces with lower average read lengths while the opposite trend showing template-related failures increasing with increased average read length is true. A breakdown of failure modes for each 100 bp Q~20~bin is shown for process-related failed reads in Figure [5A](#F5){ref-type="fig"} and template-related failures in Figure [5B](#F5){ref-type="fig"}.
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Distribution of overall process vs template failed reads**Percentage of process versus template-related failures for each 100 bp Q~20~read length bin. Numbers in each column represent the number of reads in each bin category. Process-related failure modes are more prevalent in lower average Q~20~read lengths. Template-related failure modes are more prevalent in higher average Q~20~read lengths.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Process failure mode distribution**(A) Process-related failure mode distribution. Most of the process-related failed reads have a Q~20~\< 100 with the primary mode of failure being \"Low signal, no DNA,\" or no template DNA in the sequencing reaction (confirmed by agarose gel electrophoresis). (B) Template-related failure mode distribution. Most of the template-related failed reads are a result of \"Poly A Tail,\" or 3\' poly A stretch that leads to slippage in sequence data. Reads with a \"Low signal 5\', 3\' template characteristic\" failure mode contribute to the largest proportion of data with read lengths Q~20~\< 100.
:::
:::
The most common mode of failure due to process was \"low signal, no DNA\", there were 146 reads in this category, where no template DNA was present in the sequencing reactions. This was confirmed by agarose gel electrophoresis. 145 of the reads resulted in read lengths less than 100 bp. From these 146 reads, 128 original clones failed to grow in the source glycerol stock microtiter plates. The remaining 18 clones were lost during the DNA preparation/purification procedure. Mixed clone reads containing vector sequence was the second most prevalent process-related failure mode. The 137 reads in this category were distributed broadly across the Q~20~: 100--599 bp range. The more successful reads yielded electropherograms with significantly stronger signal strength from one reaction product compared to the other.
The most common failure mode resulting from template-related characteristics was \"Poly A tail\" or failed reads with attenuated sequence due to poly A tails, there were 484 reads in this category. The distribution of these 484 reads is skewed towards the Q~20~\>300 bp bins. Failures due to \"Low signal, 5\', 3\' template characteristics\" also contribute to a significant number of template-related failures. These 116 reads were Q~20~\<100 bp and make up the majority failure mode within that Q~20~bin.
Our sample set was made up of cDNA clones and therefore contain some 3\' template biases inherent in the sequencing of cDNAs. In a best effort to obtain a distribution representing randomly generated end sequence from various library types, we remove all failed reads resulting from 3\' template attributes. The two failure modes targeted for removal are \"Poly A tail\" and the 3\' reads within the \"Low signal, 5\', 3\' template characteristics\" category. 484 and 94 reads were removed from each respective category representing 49.3% of all failed reads. The resulting template-related failure mode distribution is represented in Figure [6](#F6){ref-type="fig"}. The process-related distribution of failed reads (Figure [5A](#F5){ref-type="fig"}) remains the same. Failed reads due to homopolymer stretches other than those resulting from poly A tails, make up the prominent failure mode within this new template-related failure distribution with 99 reads. The distribution of these reads is weighted towards Q~20~\>300.
::: {#F6 .fig}
Figure 6
::: {.caption}
######
**Template failure mode distribution**Template-related failure mode distribution excluding reads failing due to poly A tail and 3\' template-related attributes. A majority of failed reads are a result of \"Homopolymer\" or single nucleotide repeat sequence leading to slippage in sequence data, these failed reads are skewed towards read lengths Q20\>300.
:::
:::
We further analyzed the 310 reads (26.5% of failed) in the preliminary \"Low signal strength\" category in Table [1](#T1){ref-type="table"} to more finely determine each failure mode (Figure [2](#F2){ref-type="fig"}, Table [2](#T2){ref-type="table"}). We removed the 146 reads and classified them as process-related, \"low signal, no DNA\" failures as described above. For the remaining 164 reads, DNA was found to be present but there was no evidence from gel images for excess DNA. It is therefore unlikely that the low signal failures were due to overloading of capillaries with excess carryover template DNA.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Breakdown of \"Low signal strength\" reads Reads from the preliminary \"Low signal strength\" failure mode category (310 reads) are further categorized into finer failure mode classifications. Failed reads from each unique clone are grouped together where possible (excluding reads that do not confirm presence of DNA on evaluation agarose gel) to determine mode of failure.
:::
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Groupings of failed reads No. of occurrences Process associated failures (No. of reads) Template associated failures (No. of reads) Failure mode
---------------------------------------- -------------------- -------------------------------------------- --------------------------------------------- -------------------------------------------------
F1/F2/R1/R2 4 0 16 Low signal, 5\' and 3\' template characteristic
F1/F2/R1 1 1 2 1\. Low signal, DNA lost during precipitation\
2. Low signal, 5\' template characteristic
F2/R1/R2 1 1 2 1\. Low signal, DNA lost during precipitation\
2. Low signal, 3\' template characteristic
R1/R2 43 0 86 Low signal, 3\' template characteristic
F1/F2 5 0 10 Low signal, 5\' template characteristic
F1/R1 1 2 0 Low signal, DNA lost during precipitation
No available clone pairing (singleton) 44 44 0 Low signal, DNA lost during precipitation
No DNA in agarose gel 146 146 0 Low signal, no DNA
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
:::
The 164 reads that appear to have sufficient template DNA are grouped by source clone for further analysis. Four possible reads may have failed for each unique clone: 5\' replicate \#1 (F1), 3\' replicate \#1 (R1), 5\' replicate \#2 (F2), and 3\' replicate \#2 (R2). There were 49 occurrences where two reads failed from one source clone, 2 instances where three reads from the same clone failed and 4 occurrences where all four reads failed from the same source clone. These results are shown in Table [2](#T2){ref-type="table"} and Table [3](#T3){ref-type="table"} summarizes the overall breakdown of 310 reads in the Low Signal Strength category. A binomial test of our data set indicates that the likelihood of 2, 3, or 4 reads failing from the same source clone is 1.9 × 10^-3^, 2.2 × 10^-5^, and 9.7 × 10^-8^, respectively. Thus, it is unlikely that more than one read failed per clone by chance.
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Summary of \"Low signal Strength\" reads Summarizing results of the 310 \"Low signal strength\" reads in Table 2.
:::
Low Signal Strength Failure Mode No. of reads
---------------------------------------------- --------------
Low signal, no DNA 146
Low signal, 5\', 3\' template characteristic 116
Low signal, DNA lost during precipitation 48
Total 310
:::
Groupings with two failed reads from the same end of the template, \"F1/F2\" and \"R1/R2\" showed trace data with noisy background or no signal trace therefore there was insufficient information to determine actual template characteristic leading to the problematic results. With the low probability of seeing two failed traces from one unique clone, we conclude that these failures are most likely a result of template DNA with 5\' or 3\' ends that are problematic for sequencing. Possibilities include secondary structure, mutated priming sites, long poly A tail, or other homopolymer stretches near the priming sites on the template.
The \"F1/R1\" grouping contained failed reads from both ends of the same replicate but successful end reads from the second replicate. This would indicate that there were no clone related attributes on either end of the template to prevent good sequence thus the failed reads were assessed a process-related failure due to loss of DNA during the precipitation process.
The \"F1/F2/R1/R2\" grouping indicates that no passing sequence data were obtained from 5\' or 3\' ends of both replicates from the original source clone, all 16 reads showed presence of DNA at each observational check in the process pipeline. The probability that 4 reaction products from a single clone were lost during precipitation is extremely improbable and it is more likely that the original clones had attributes that prevented successful sequencing from either ends. We therefore assign 5\' and 3\' template-related failure modes for the 16 reads in this grouping.
The final groupings of \"F1/F2/R1\" and \"F2/R1/R2\" contain 6 reads in total. Each include failed reads from both ends of one replicate and a single failed read from the second replicate. We maintain the same logic for our failure mode evaluation by assessing both a template-related failure to the 2 failed reads from the same replicate and a process-related loss of DNA during precipitation failure for the single read from the second replicate.
There were 44 reads in the singleton category where only a single read (F1, F2, R1, or R2) failed from each clone. These failures cannot be attributed to the template as three other reads were successfully sequenced from the same source clone. Failure mode was therefore likely to be process related and due to loss of DNA during the precipitation wash steps.
Discussion
==========
Having established a stringent pass criteria of 600 Phred 20 bases or greater, we observed a broad range of failure modes distributed among the failed reads. Reads less than 100 bp did not yield useable data and were usually a result of process-related failure modes. Process-related mishaps in the pipeline would include missed colony picks, blocked capillaries on sequencers, faulty tips on automated liquid handling devices, or lost DNA template during precipitation. The sequences generated have low average read lengths, or in cases where DNA is totally lost, zero read length. 88% of reads within the largest \"process\" related failure mode category, \"low signal, no DNA\", were a result of failed cultures at the beginning of the process pipeline. The absence of DNA was verified early on by failed growth in the source microtiter plates and further confirmed by failed growth in the stamped replicate agar plate. Failures are perhaps due to missed bacterial colony picks and adjustments such as recalibration or refining the morphology criteria of the colony picker should help. The second most prevalent failure mode in the \"process related\" category, \"Mixed clone with vector sequence\" (30% of reads in this classification), is likely a result of two side by side colonies being picked into the same well or through contamination originating from the use of automated liquid handlers. These problems might be alleviated by increasing the stringency of the proximity criteria on the colony picking device to ensure sufficient distance exists between two bacterial colonies before they are chosen for picking. As we use new tips on the instruments daily, increasing the length or number of wash cycles on the automated liquid handlers between volume transfers may help in reducing carry over contamination from one plate to another. This will need further investigation. Cross contamination between adjacent wells can also arise with the mishandling of plates, or when tips are plunged too deeply into the wells, displacing the volume into adjacent wells. Adjusting the liquid handler by reducing the depth and speed of tips entering each well may help prevent the occurrences of cross contamination within a plate. Despite the presence of mixed clones within one well, traces often yielded discernable sequence with read lengths broadly distributed up to the Q~20~: 500--599 bp bin. This is likely the result of a presence in greater amount of one reaction product over the second leading to proportionately stronger signal strength.
Template-related failures often still yielded reads of several hundred base pairs. Template characteristics that might contribute to failed or truncated reads could be poly A tails, homopolymer stretches, or other highly repetitive regions. Problems in sequencing these types of regions are well cited in the literature and are dependent on chemistry and methodology used in sequencing. These problems are best avoided by using an anchored primer for first strand synthesis of cDNA as the first step in library construction. \"Poly A tail\" was the single largest cause of failed sequence in both the template-related failure category as well as the entire data set, yet the reads make up the majority in the higher failed read length bins. In our regular production pipeline, we often circumvent this problem by using a combination of a 3\' primer to resolve sequence immediately following the poly A coupled with an oligo(dT)~23~N (N = A, G, C) anchored primer to resolve the 3\' ends of cDNA inserts.
Conclusions
===========
The FMA pipeline described here was tailored to our high-throughput 384-well automated sequencing pipeline but many of the components in this platform are shared within the high-throughput sequencing community such as the alkaline lysis procedure to prepare and purify DNA, DNA sequencing using Big Dye chemistry, ethanol precipitation to clean up reaction products and equipment such as the Genetix QPIX, Beckman Coulter Biomek FX, and Applied Biosystems 3730 xl DNA analyzer. For this reason our FMA methods can be readily adapted to analyze other similar sequencing platforms. Extending our present study to other library types commonly sequenced in the high-throughput community, such as shotgun or Serial Analysis of Gene Expression (SAGE) should offer further information regarding template specific failures. This failure mode analysis provides information on distinguishing between process and template-related attributes that may lead to downstream failed sequence. It can therefore be a useful tool used to audit overall sequencing procedures and identify key problematic steps in a process pipeline. Making proper adjustments to the pipeline based on the results will likely result in increased efficiency, enhanced data quality, and decreased cost.
Methods
=======
Samples are processed in duplicate, with quality assurance and observational checks to account for status of each read after every procedure leading to the final sequence read. These recorded observations facilitate downstream systematic analysis of failed sequence data. The regular core production pipeline was not altered for this study, as the purpose is to assess sequence failure modes under ordinary conditions in a high-throughput sequencing environment. An overview of the observational checkpoints within the FMA pipeline is outlined in Figure [1](#F1){ref-type="fig"}. Template DNA is considered present in all wells up to the completion of DNA sequencing unless absence is indicated by no growth of bacterial culture or blank lane on gel. Once cycle sequencing has completed, the reaction products are precipitated, pelleted, then washed. As it is very difficult to qualitatively asses the presence of DNA accurately after the precipitation procedures other than by sequencing, we draw conclusions regarding loss of sequencing reaction products during precipitation by a process of elimination. Any failed reads that result from a process-related loss of DNA and have no prior observations indicating absence of DNA are attributed to the precipitation procedure.
Transformation and colony picking
---------------------------------
One microliter of ligation mix from each of two *Populus trichocarpa*cDNA libraries were transformed by electroporation into 40 μl of *E. coli*DH10B T1 resistant cells (Invitrogen). Transformed cells were recovered using 1 mL of SOC medium (prepared in house) and plated onto 22 cm × 22 cm agar plates (Genetix) containing 2xYT agar and 100 μg/μl Ampicillin. Agar plates were incubated overnight at 37°C for 14 hours. Bacterial colonies were picked from the agar plates and arrayed into 384-well microtiter plates (Genetix) containing 60 μl of 2xYT medium + 7.5% glycerol (made in house) using the Genetix QPIX automated colony picker (Genetix). A total of six 384-well microtiter plates, three plates for each of the two cDNA libraries, were picked. The plates were incubated overnight at 37°C for 16 hours then each microtiter plate was inspected for wells that contained no growth. The positions of all failed cultures were recorded. To further verify growth in each well after the incubation period, a disposable 384-well replicator was used to stamp bacterial culture from each 384-well microtiter plate onto a new 22 cm × 22 cm agar plate containing 2xYT agar and 100 μg/ul ampicillin. Agar plates were incubated overnight at 37°C for 14 hours then inspected the next day for colonies from every well. The positions of all failed growths were recorded.
384-well culturing and DNA purification of plasmid clones
---------------------------------------------------------
Two microliters of bacterial culture was transferred from the 384-well microtiter plate into a 240 μl 384-well deep well diamond plate (Axygen) containing 60 μl of 2xYT medium and 100 μg/ml ampicillin using a 384-well slotted inoculator (V&P Scientific). This was done in duplicate to create two sets of six 384-well deep well inoculated diamond plates. Inoculated plates were sealed with AirPore™ tape (Qiagen) and placed into a 37°C shaking incubator (New Brunswick Scientific C25 Incubator Shaker) at 350 rpm for 18 hours. After the incubation period, cultures were removed and each plate inspected for growth and contamination. All failed cultures were recorded. Cultures from both replicates were then placed onto separate multi-tube floor vortexers (VWR) at maximum speed for approximately 5 minutes until all cells were resuspended. Cultures were stored at 4°C until ready for DNA preparation. DNA was prepared using alkaline lysis \[[@B3]\] with the following modifications that have been implemented for the standard GSC template production pipeline. Culture blocks from both replicates were removed from the 4°C refrigerator and mixed using a multi-tube floor vortex (VWR) for 5 minutes at maximum speed (or until all cells appeared resuspended). A Titertek MapC2 liquid handling device was used to dispense 60 μl of Lysis Buffer (Qiagen Buffer P2). After 5 minutes of lysis, 60 μl of Neutralization Buffer (Qiagen Buffer P3) was added. Plates were tape sealed (Edge biosystems clear tape) and mixed on a multi-tube vortex at maximum speed for 2 minutes prior to centrifugation at 4250 × g for 45 minutes in a Jouan KR422 centrifuge. 120 μl of lysate were transferred from pelleted culture blocks into 240 μl 384-well deep well diamond plates containing 90 μl per well 100% isopropanol using a 384-well Hydra pipetting instrument (Robbins Scientific). Destination plates were sealed (Edge biosystems clear tape) and mixed by inversion, followed by centrifugation at 2830 × g for 15 minutes in an Eppendorf 5810R centrifuge (Brinkmann Instruments). After centrifugation the isopropanol was decanted, the DNA pellet washed with 50 μl 80% ethanol using a Robbins 384-well Hydra, and the plates left to dry upright for three hours on the benchtop. DNA pellets were resuspended in 10 mM Tris-HCl, pH = 8 containing 10 μg/ml RNase A (Qiagen) and mixed for 1 minute at maximum speed on a multi-tube vortexer. Plates were briefly centrifuged at low speed, stored at 4°C overnight, then transferred to a -20°C freezer until required for DNA evaluation and sequencing reactions.
DNA evaluation
--------------
DNA preparations were evaluated by agarose gel electrophoresis. A 1.5 μl aliquot of purified DNA was combined with 1.5 μl of bromophenol blue loading buffer (0.21% bromophenol blue; 12.5% ficoll) and 2 μl was loaded onto a 1.2% agarose gel. Samples were loaded using a 12-channel loader (Hamilton) beside 3 ng of 1 kb plus DNA marker (Invitrogen). Gels were run at 120 volts for 90 minutes in TAE (Tris/Acetate/EDTA) buffer followed by staining for 35 minutes in SybrGreen Nucleic Acid stain (Cambrex). Gels were scanned using a Fluorimager 595 (Molecular Dyanmics) scanner. The image was visually examined for genomic DNA, as well as presence and quality of DNA. All empty lanes and observations were recorded and entered into the FMA database.
DNA sequencing
--------------
DNA Sequencing reactions were assembled in 384-well clear optical reaction plates (Applied Biosystems) using a Biomek FX workstation (Beckman-Coulter). In each 5 μl reaction (total volume) the following were added: 3 μl of purified plasmid DNA (\~45 ng/μl), 0.26 μl of sequencing primer (5 pmol/μl, Invitrogen), 0.43 μl of 5X reaction buffer (Applied Biosystems Big Dye Terminator 5X Sequencing Buffer), 0.77 μl of Ultrapure water (Gibco), and 0.54 μl of BigDye v.3.1 ready reaction mix (Applied Biosystems). Each well of the reaction plate was visually inspected for appropriate volumes after both reaction mix and DNA addition. Volumes were manually adjusted using a single channel pipet (Gilson) where required and all observations were recorded. Sequence data were obtained using universal M13 Forward (5\'-TGTAAAACGACGGCCAGT-3\') and M13 Reverse (5\'-CAGGAAACAGCTATGAC-3\') primers on each set of replicate plates. Thermal cycling was performed on PTC-225 thermal cyclers (MJ Research) with parameters of 35 cycles at 96°C for 10 seconds, 52°C for 5 seconds using M13 Forward primer or 43°C for 5 seconds using M13 Reverse, 60°C for 3 minutes, followed by incubation at 4°C. Reaction products were precipitated by adding 2 μl of 125 mM EDTA (pH8) and 18 μl of 95% ethanol per well followed by centrifugationat 2750 × g for 30 minutes in an Eppendorf 5810R centrifuge. The EDTA/ethanol was immediately decanted and reaction products washed with 70% ethanol. The 384-well cycle plates were allowed to dry inverted for 15 minutes. Samples were resuspended in 10 μl of Ultrapure water and analyzed using a 3730XL DNA analyzer (Applied Biosystems). The performance of each capillary on the four DNA analyzers used in this experiment were validated using one 384-well control plate for each instrument. Each 384-well plate contained our in-house control standard, a full-length human cDNA clone obtained from the I.M.A.G.E. Consortium (I.M.A.G.E. ID \#3609158, Lawrence Livermore National Laboratories). Blocked capillaries were recorded into the FMA database and traces originating from these capillaries flagged. Sequence data were evaluated using PHRED software \[[@B2]\] (v.0.020425.c) and the chromatograms were viewed using a java applet based on \'ted\' \[[@B4]\] -- a publicly available trace file viewer. A relational database \"FMAdb\" was created using MySQL for flexible querying of results. The FMAdb was populated with sequence data, plus process observations such as absence of bacterial colonies and no-grows, trace evaluations and information, as well as equipment and sequence run details.
Authors\' contributions
=======================
GSY was responsible for design and execution of this study, plus analysis of data and drafting of the manuscript. RAH conceived of the study and directed the analysis. JMS and SAB generated the data described in the study and participated in study design. DS developed several of the key protocols used in this study. MB provided quality assurance of the experimental procedures and participated in study design. MAM participated in the establishment of the production pipeline and read the manuscript and provided comments.
Acknowledgements
================
We thank Steven Ralph and Joerg Bohlmann, University of British Columbia, Vancouver, BC, Canada for providing us with the two cDNA libraries used in this experiment. We thank the staff at Canada\'s Michael Smith Genome Sciences Centre, British Columbia Cancer Agency, Vancouver, BC, Canada for their technical and administrative assistance. Especially to the members of the Sequencing staff who contributed to the generation of data. MAM and RAH are Michael Smith Foundation for Health Research scholars.
|
PubMed Central
|
2024-06-05T03:55:52.031782
|
2005-1-4
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546001/",
"journal": "BMC Genomics. 2005 Jan 4; 6:2",
"authors": [
{
"first": "George S",
"last": "Yang"
},
{
"first": "Jeffery M",
"last": "Stott"
},
{
"first": "Duane",
"last": "Smailus"
},
{
"first": "Sarah A",
"last": "Barber"
},
{
"first": "Miruna",
"last": "Balasundaram"
},
{
"first": "Marco A",
"last": "Marra"
},
{
"first": "Robert A",
"last": "Holt"
}
]
}
|
PMC546002
|
Background
==========
Hematophagy, blood-feeding, is a behavior exhibited by most arthropod vectors of human pathogens. In anautogenous mosquitoes, the female generally feeds to repletion on a single blood meal and then proceeds to use this nutrition as the basis for the development of a batch of eggs. The cycle of host seeking, blood feeding, egg development, and oviposition is generally called the gonotrophic cycle, a term coined by Beklemishev in 1940 \[[@B1]\]. For most mosquitoes living in optimal field or laboratory conditions, this cycle requires about forty-eight hours and involves a complex series of biological events, including peritrophic matrix formation, blood digestion, oocyte development, vitellogenesis, and excretion.
Digestion of the proteinaceous blood meal is required for oocyte development and vitellogenesis, and consequently these are coordinated processes. Multiple hormones interact to alter tissue states and to activate genes involved in these processes. The two hormones juvenile hormone (JH) and 20-hydroxyecdysone (20-E) are most fundamental to ovarian development. Within several days after emergence of female mosquitoes from the puparium, juvenile hormone (JH) stimulates the separation of ovarian follicles from germaria and limited growth of the ovarian follicle to its pre-vitellogenic resting state \[[@B2]\]. JH also confers competence to fat body cells and ovarian follicles for uptake of ecdysteroidogenic hormone (OEH). Then, in response to a blood meal, gonadotrophins are released from cerebral neurosecretory cells and cause the ovaries to become OEH-responsive \[[@B3]\]. OEH stimulates the ovaries to secrete ecdysone, the precursor to 20-E, as well as 20-E during vitellogenesis \[[@B4]-[@B6]\]. Fat body cells take up ecdysone, convert it to 20-E and use it to activate transcription of vitellogenin genes \[[@B7]\], the genes encoding the major egg-yolk proteins, as well as a large number of other genes, many of whose products will be incorporated into eggs \[see \[[@B8],[@B9]\] for reviews\].
Prior to the blood meal, female mosquitoes access sugars for nutritional sustenance. During the first several hours following a blood meal, the mosquito undergoes physiological changes in addition to hormonal ones. Acquisition of a blood meal stimulates midgut proteolytic activity such that approximately 80% of the protein content is digested within one day \[[@B10]-[@B14]\]. Serine proteases including trypsins and chymotrypsins are responsible for the majority of endoproteolytic activity \[[@B11],[@B12],[@B15]\]. The role of trypsins in blood digestion has been well documented in *Aedes aegypti*, and more recently it has been investigated in *An. gambiae*. Despite the digestive proteolysis peak at 24 hours post blood meal, digestive enzymes exhibit two phases of transcription \[[@B16],[@B17]\]. In *Ae. aegypti*there are three trypsins, early trypsin, which is constitutively expressed prior to blood feeding and two late trypsins which are blood induced. These two types of trypsins are also found in *An. gambiae*. The *An. gambiae*trypsin family includes seven genes clustered within 11 kb on chromosome 3R, in division 30A, that encode five functional proteins \[[@B18]\]. Trypsins 1 and 2 are both induced by a blood meal and exhibit similar expression profiles. In contrast to Trypsins 1 and 2, Trypsins 3, 4, and 7 are constitutively expressed in unfed females \[[@B18]\]. Trypsins 3 and 7 are down-regulated following a blood meal and not expressed again at levels detectable by RT-PCR until 28 hours post blood meal \[[@B18]\]. In addition to the trypsins, three chymotrypsin genes have been isolated and characterized in *An. gambiae*, two of which are located in tandem on chromosome 2L, in division 25D \[[@B19],[@B20]\]. Both of these genes, AnChym 1 and 2, are expressed in the midgut by 12 hours post blood meal and their transcripts are abundant until 48 hours, as determined by PCR, unlike the levels of Trypsins 1 and 2 that have decreased dramatically by this time \[[@B19]\]. In contrast, the other characterized chymotrypsin, AgChyL, exhibits transcript level changes more similar to those of Trypsins 3--7 \[[@B20]\].
Two types of exopeptidases, carboxypeptidases and aminopeptidases, have been characterized in Anopheline mosquitoes. Edwards et al. \[[@B21]\] cloned a carboxypeptidase that was rapidly induced in *An. gambiae*midguts following blood meal ingestion. Multiple aminopeptidases have been isolated from hematophagous insects, and it has been suggested that they may play different roles in digestion \[[@B22]-[@B25]\]. Additional enzymes including glycosidases and lipases are also required for the digestion of non-proteinaceous blood constituents \[[@B26],[@B27]\].
In addition to dramatic changes in physiology, blood feeding also induces changes in mosquito morphology. Following gut distension by blood ingestion, midgut epithelial cells secrete a Type I peritrophic matrix (PM) that is continuous along the length of the midgut \[[@B14],[@B28],[@B29]\]. Prior to the blood meal, the midgut epithelial cells contain high concentrations of apically located, morphologically granular, secretory vesicles. Presumably these apical granules contain precursors of the peritrophic matrix: as early as an hour after the adult female has taken a blood meal, they are no longer detectable \[Staubli et al., 1966, as cited in \[[@B30]\]\]. In *An. gambiae*, the PM can be visualized by electron microscopy as early as 12 hours PBM and it is fully formed by 48 hours PBM \[[@B28],[@B31]\]. The PM is a biochemically complex structure containing not only chitin and other proteoglycans, but as many as 20--40 different proteins \[[@B31]-[@B33]\]. However, only one gene encoding a peritrophic matrix protein has been cloned in *An. gambiae*\[[@B34]\]. The exact functions of the PM remain unknown, but it has been suggested that this semi-permeable porous structure may function as a restrictive layer protecting the midgut epithelium from proteolytic digestive enzymes, from haematin crystals that form following hemoglobin breakdown and as a barrier to blood-borne pathogens including bacteria and malaria parasites \[reviewed in \[[@B35]\]\].
Once the adult mosquito acquires a blood meal, she spends approximately 48 hours converting about 20% of it into egg constituents \[[@B36]\], using another fraction of it to support the intense biosynthetic activities of this period and defecating the rest. Oogenesis in the mosquito ovary actually begins post-eclosion but oocyte growth attenuates at a resting stage until blood meal ingestion. Once reinitiated, egg development continues until oviposition. Successful egg production not only requires ovarian events for development and maturation of oocytes, but also synthesis of yolk constituents, both protein and lipid, in the fat body, followed by their uptake by oocytes and storage for later use during embryogenesis. Collectively, the events of yolk synthesis, uptake and storage constitute the process of vitellogenesis \[[@B30]\]. Vitellogenesis and oogenesis require the coordination of molecular events in at least these two different abdominal tissues, the fat body and ovary. Based on morphological and physiological criteria, the ovarian cycle can be divided into four phases: 1) Pre-vitellogenic, 2) Initiation, 3) Trophic, and 4) Post-trophic Phase \[[@B30]\].
The meroistic ovary of *An. gambiae*contains approximately 50 functional egg-production structures, the ovarioles. Each ovariole is comprised of two parts, a distal germarium and a vitellarium proximal to a common oviduct through which eggs will pass as they are laid. In the germarium, mitosis of the primordial germ cells creates a syncitium with an oocyte and seven nurse cells interconnected by intracellular bridges, or ring canals as a result of incomplete cytokinesis. Both the germ cell and the nurse cells are surrounded by a somatically derived follicular epithelium \[[@B37]-[@B39]\]. The first pre-vitellogenic phase is completed within three days of eclosion and ends with the separation of these follicles from the germaria and entry into the vitellaria. At the end of this phase, oocytes may have undergone some growth but then arrest until events initiated by acquisition of a blood meal cause them to become competent for ovarian vitellogenic events.
Ingestion of a blood meal reinitiates ovarian development and follicle growth resumes. In *Ae. aegypti*and *Anopheles albimanus*, this period appears variable, lasting 3--10 and 8--16 hours, respectively, and ends with the initiation of vitellogenin synthesis \[[@B30]\]. In the next two days, during the trophic phase, the mosquito generates large amounts of vitellogenin, the secreted precursor to the major yolk protein vitellin. In addition to vitellogenin, the developing oocytes also accumulate other proteins, and lipids from the hemolymph, as well as ribosomes and mRNAs synthesized in the syncitial nurse cells. These latter constituents are transported to the germ cell through the ring canals connecting the oocytes and nurse cells by a process of cytoplasmic streaming \[[@B40]\]. Following delivery, several maternal mRNAs become localized within the oocyte. These maternal transcripts are fundamental for dorsal/ventral and anterior/posterior patterning of the embryo that will develop from the oocyte. This pattern of deposition and the patterning of the eggshell also depend on a complex signaling process involving both the somatic cells of the follicular epithelium and the oocyte.
Once oocyte growth has ceased, vitellogenin synthesis terminates. This signals the onset of the post-trophic phase. During this time, the oocytes mature and eggshell structures begin to develop. The chorion, part of the eggshell, is secreted by the follicular epithelium and contains two layers, the first secreted, inner endochorion and the later secreted, outer exochorion \[[@B30]\]. It is the endochorionic layer that will harden and melanize after oviposition.
Specialization of eggshell structures necessitates communication between cells. The RAS 1 signaling cascade is an important means of communication during the processes of oocyte and eggshell patterning, as it is during eye development and differentiation of structures late in embryogenesis \[[@B41],[@B42]\]. During patterning in *Drosophila*, developing oocytes produce the TGFα protein Gurken that binds to the epidermal growth factor receptor (EGFR), a receptor tyrosine kinase (RTK) localized to the posterior follicle cells, to initiate RAS 1 signaling. Downstream from the activation of this RTK, the GTP-binding protein RAS 1 initiates a series of enzymatic events propagated successively by three protein kinases, RAF, MAPK, and MAPK kinase (MEK), resulting in the translocation of nuclear factors and possibly the concomitant reorganization of the cytoskeleton \[reviewed in \[[@B43]\]\]. Thus, a cascade of events leads to the establishment of the posterior follicle cell fate. The posterior follicle cells then signal back to the germ cells. This results in the reorganization of the oocyte cytoskeleton, and regulates the localization of anterior/posterior determinants. Similar to the eggshell, the oocyte also undergoes dorsal-ventral patterning. Following patterning of the follicle cells, maternal gene products are regulated by the Toll signaling pathway to generate a transcription factor gradient that will spatially regulate activity of specific zygotic genes within the fertilized oocytes \[[@B44]\].
Vitellogenic events in the fat body have also been divided into phases: 1) Pre-vitellogenic, 2) Vitellogenic, and 3) Termination. The pre-vitellogenic phase in the fat body coincides with the pre-vitellogenic phase of the ovarian cycle. During this phase, RNA synthesis increases in the fat body and the rough endoplasmic reticulum and the Golgi complex proliferate to prepare for the production of vitellogenin. At the start of the vitellogenic phase, the release of mosquito hormones initiated by digestion signal the onset of vitellogenesis \[[@B45]\]. Synthesis of large amounts of vitellogenins is facilitated by the large quantities of biosynthetic machinery generated during pre-vitellogenic stages, but also depends on the presence of multiple vitellogenin genes (Romans, unpublished). Following synthesis, vitellogenin is released into the hemolymph and eventually diffuses through channels between the cells of the follicular epithelium, whereupon it is accumulated by the oocytes by a process of receptor-mediated endocytosis in clathrin-coated pits \[[@B8]\]. When vitellogenesis has ceased, during the termination phase, the biosynthetic machinery in the fat body is degraded via a lysosomal pathway, at least in *Ae. aegypti*\[[@B46]\].
Thus, blood feeding initiates a complex series of physiological events in at least three tissues that are integrated by the actions of JH, 20-E and peptide hormones. These events may be required for parasite development; they certainly can be modulated by the presence of parasites \[[@B47],[@B48]\] and may provide points of intervention for mosquito control. Microarray analysis provides a tool to study global expression patterns of thousands of genes simultaneously. By comparing the level of transcription of a gene over time between two states, *e.g*. blood-fed *vs*. sugar-fed, an expression signature for each gene can be defined in response to blood feeding. Consequently, these expression patterns may indicate how these genes are regulated and interact, and also the biological processes in which the act. In this study we performed microarray analysis of genes in female mosquito abdomens during the first 48 hours after a blood meal. We have implicated many of these genes in different processes stimulated *de novo*by blood feeding. The elucidation of the expression profiles of abdominal genes will provide a broadened basis for understanding vector-parasite interactions. Our study certainly provides insights into the physiology of the malaria vector *Anopheles gambiae*.
Results
=======
Array composition
-----------------
Microarray analysis was conducted on 3057 cDNA clones generated from three different adult female *An. gambiae*mosquito abdomen-derived cDNA libraries to elucidate major patterns of gene expression through 48 hours post ingestion of a blood meal. Arrays were constructed from triplicate spotted negative controls (purified water, 3 × SSC with no DNA, and empty wells), positive controls for blood-fed samples consisting of 3 clones whose ESTs corresponded to rat (*R. norvegicus*) α and β hemoglobin chains, and PCR-amplified fragments obtained from 1132, 721, and 1204 clones randomly picked from the sugar-fed (harvested after 30 hours at 19°C), rat blood-fed (harvested 30 hours PBM at 19°C), and *P. berghei*infected rat blood-fed (harvested 30 hours PBM at 19°C) abdomen libraries, respectively (Table [1](#T1){ref-type="table"}). Approximately 84% of PCR-amplified fragments were visualized on ethidium bromide stained 1% agarose, 1 × TBE gels prior to spotting (data not shown). Of these PCR-amplified fragments, 2219 clones (87% of electrophoresed PCR products) were represented by a single defined band (Table [2](#T2){ref-type="table"}).
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Microarray composition
:::
***Controls*** **Clones**
--------------------------------------------------------------------------------------------- ------------
**Negative Controls** **108**
**Positive Controls** **3**
**Libraries**
Sugar-fed Adult Female (incubated 30 hours at 19°C) Abdomen library 1132
Blood-fed Adult Female (incubated 30 hours PBM at 19°C) Abdomen library 721
*Plasmodium berghei*Blood-fed Adult Female (incubated 30 hours PBM at 19°C) Abdomen library 1204
Total 3168
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Appearance of PCR product following gel electrophoresis
:::
**Total Clones** **Singlet** **Doublet** **Smear** **No Product Visible**
--------------------------------------------------------------------------------- ------------------ ------------- ------------- ----------- ------------------------
Amplification of 2558 Clones 2558 2098 (82%) 117 (5%) 6 (\<1%) 337 (13%)
Re-amplification of 183 Clones previously amplified with \"No Product Visible\" 183 121 (66%) 14 (8%) 0 48 (26%)
Cumulative 2558 2219 (87%) 131 (5%) 6 (\<1%) 202 (8%)
Note: PCR-amplified fragments were visualized on ethidium bromide stained 1% agarose, 1 × TBE gels. Approximately 13% of PCR products could not be visualized following the first amplification due to product yield below the threshold of ethidium bromide detection; 54% of these (183) were re-amplified.
:::
ESTs corresponding to these spotted cDNAs were screened for mitochondrial contamination, filtered based on sequence trace file quality, and assembled (EST clustered) using the DNAstar Seqman II software (DNAstar, CA) (Table [3](#T3){ref-type="table"}). The high quality ESTs clustered into 491 contigs (consensus sequence generated from ≥2 overlapping ESTs) and 1415 singletons (ESTs with no sequence similarity to any other EST in the assembly) for a total of 1906 unique transcripts (Table [4](#T4){ref-type="table"}).
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
EST composition of array
:::
**Number** **Percentage**
---------------------------- ------------ ----------------
High Quality Sequence Data 2707 88.56%
Poor Quality Sequence Data 131 4.28%
Mitochondrial DNA 222 7.25%
Total 3060 100%
Note: EST analysis includes the ESTs of the positive controls (*Rattus norvegicus*hemoglobin chains).
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Putative transcripts represented on the array following EST assembly
:::
**Number** **ESTs Represented**
------------ ------------ ----------------------
Contigs 491 1292
Singletons 1415 1415
Total 1906 2707
Note: EST assembly includes the ESTs of the positive controls (*Rattus norvegicus*hemoglobin chains).
:::
Microarray and bioinformatic analyses
-------------------------------------
Global patterns of greater than two-fold up-regulation or down-regulation for these cDNAs were established by comparing transcript levels in blood-fed *An. gambiae*adult females at ten time points during and post ingestion of a blood meal to the levels in sugar-fed females. First strand cDNA was generated from total RNA collected at 5 min and 30 min after initiation of blood feeding and at 0, 1, 3, 5, 12, 24, and 48 hr post-blood meal. All cDNA populations were labeled and hybridized to arrays. For each PCR-amplified insert, Cy3 and Cy5 fluorescent dye levels were measured from 3 replicate spots on each of 50 arrays to generate average signal intensities, and an expression ratio depicting transcript fold change between sugar-fed and blood-fed mosquitoes calculated. Following quality control filtering and normalization, 456 cDNAs and the rat β-hemoglobin gene, the positive control, were expressed more than twofold above or below control, sugar-fed levels at one or more of the 10 blood-feeding time points. Following EST analysis, the 456 cDNAs were found to represent 413 unique mosquito transcripts, 10 of which were present in more than one set. This anomaly is due to EST clustering of alternatively spliced transcripts with different expression patterns. More unique transcripts are up-regulated than down-regulated in response to blood feeding, while 10% of them are both up-regulated and down-regulated over the time course of this study: 192 are up-regulated at least twofold, 173 are down-regulated at least twofold, and 48 are down-regulated and up-regulated. Bioinformatic analyses of these 413 unique transcripts showed that all sequences shared sequence identity with the *An. gambiae*genome (Table [5](#T5){ref-type="table"}), 90% of which shared sequence identity with an entry in Nr of dbEST (Table [5](#T5){ref-type="table"}). In this analysis Blast hits with an E value ≤1 × 10^-4^were considered significant.
::: {#T5 .table-wrap}
Table 5
::: {.caption}
######
Sequence similarity of consensus sequences
:::
**Blastn (WGS *An. gambiae*)** **Blastx (Nr)** **Blastn (Nr)** **Blastn (dbEST)** **Unannotated** **Total**
------------------------ -------------------------------- ----------------- ----------------- -------------------- ----------------- -----------
3 Positive Controls No Significant Hit 3 3
403 Unique Transcripts 403 240 8 112 43 403
Early Genes 144 77 2 45 20 144
Middle Genes 130 78 6 36 10 130
Late Genes 139 93 2 31 13 139
Total 413 248 10 112 43 413
:::
Microarray gene clustering and principal components analysis
------------------------------------------------------------
The behaviors of the gene products identified as at least twofold up/down-regulated were grouped into three sets using *k*-means clustering (Figure [1](#F1){ref-type="fig"}) and named according to the time of their induction during the 48-hour time course following the initiation of blood feeding. Set 1, hereafter referred to as the \"Early Genes\", contains 144 unique transcripts derived from 152 cDNAs, which are expressed mainly during the early time points (Table [6](#T6){ref-type="table"}, Figure [1A](#F1){ref-type="fig"}). The majority of these genes are appear induced at least twofold more abundantly than in sugar-fed mosquitoes during the first five minutes of blood feeding. Many of these transcripts remain induced until 1-hour PBM, although some remain induced until 5 hours PBM. After 5 hours post blood meal, the majority of Early Genes is down-regulated and they remain down-regulated even 48 hours after blood meal ingestion. A small subset of the Early Genes shows a variant pattern of gene expression in which the transcripts are up-regulated from the first 5 minutes of blood uptake through 1 hour PBM followed by a repression in expression from 3 to 24 hours PBM and then a greater than twofold induction at 48 hours PBM. The 130 unique transcripts represented by 147 cDNAs in Set 2, the \"Middle Genes\", follow a more dynamic pattern of gene expression than the Early Genes (Table [6](#T6){ref-type="table"} and Figure [1B](#F1){ref-type="fig"}). Most Middle Genes are down-regulated in blood-fed versus sugar-fed mosquitoes until 3 hours PBM followed by an increase in expression commencing at 5 hours PBM and peaking between 12 and 24 hours PBM. Subsequently, Middle Genes are down-regulated to initial transcript abundances by 48 hours PBM. Also, in a behavior largely exhibited by the Middle Genes, approximately 40% of genes are down-regulated when the mosquitoes completed feeding and left the rat (0 hours PBM). Set 3, the \"Late Genes\" contains 139 unique transcripts, 157 cDNAs, which are either down-regulated or constitutively expressed until 12 to 16 hours PBM after which they are up-regulated and, in contrast to the Middle Genes, continue to be highly expressed even at 48 hours PBM (Table [6](#T6){ref-type="table"}, Figure [1C](#F1){ref-type="fig"}).
::: {#F1 .fig}
Figure 1
::: {.caption}
######
Gene trees displaying the microarray generated expression profiles of abdomen-derived cDNAs in blood-fed compared to sugar-fed adult female mosquitoes during the following times: 5 and 30 minutes during blood meal (DBM), 0, 1, 3, 5, 12, 16, 24, and 48 hours post blood meal (PBM). *k*-means clustering of all genes up-regulated and down-regulated at least two-fold during at least one of the ten time points generated three sets of genes. These *k*-means-derived groups of genes were hierarchically clustered for visualization and include Set 1 designated the Early Genes (A), Set 2 designated the Middle Genes (B), and Set 3 designated the Late Genes (C). Each gene is represented by a single row of colored boxes; each time point is represented as a single column of colored boxes. The expression scale is represented as a gradation of color ranging from 5 fold induced genes indicated by saturated red to 2.5 fold repressed genes indicated by saturated green.
:::
:::
::: {#T6 .table-wrap}
Table 6
::: {.caption}
######
Bioinformatic analysis of two-fold expressed genes
:::
**Early Genes** **Middle Genes** **Late Genes** **Total**
--------------------------------------------------------------------------------- ----------------- ------------------ ---------------- -----------
Twofold Expressed cDNAs (represented by consensus sequences \>100 bp in length) 152 147 157 456
Replicate Consensus Sequences within *k*-means Sets 8 17 18 43
Unique Transcripts 144 130 139 403\*
Unique Transcripts included within PCA Analysis 82 69 98 249
\*The complete data set (Early, Middle and Late genes combined) contains 413 unique transcripts. However, ten transcripts are present in two different gene sets and were counted twice as a result.
:::
Principal components analysis (PCA) of these 413 unique differentially expressed transcripts was also conducted to support the *k*-means defined sets. Following PCA, each transcript was plotted in a scatter plot comparing the PCA 1 and PCA 2 values. The three *k*-means-defined sets of genes did not overlap on these scatter plots. Thus the PCA results support classification of these transcripts into three groups. For each of the *k*-means defined groups, genes that had a PCA 1 or PCA 2 value greater than 0.5 or less than -0.5 were plotted on a parallel coordinates display (Figures [2](#F2){ref-type="fig"} and [3](#F3){ref-type="fig"}). The resulting data sets contained 82, 69, and 98 unique transcripts representing Early Genes, Middle Genes, and Late Genes, respectively (Table [6](#T6){ref-type="table"}).
::: {#F2 .fig}
Figure 2
::: {.caption}
######
Scatter plot of the first two components for each gene that is either up-regulated or down-regulated at least two-fold during one time point in the course of this experiment. The Early Genes, Middle Genes, and Late Genes are colored red, blue and green, respectively. The genes colored in grey include Early, Middle and Late genes that did not have a value greater than 0.5 or less than -0.5 of both PCA 1 and PCA 2 components.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
Parallel coordinates display of expression profiles of differentially expressed PCA-filtered genes from the Early Genes (A), Middle Genes (B), and Late Genes (C). X-axes correspond to a successive time point. Y-axes denote the ratio of fluorescent intensities of blood-fed to sugar-fed samples at each time point for each gene in Panels (A), (B), and (C). Plotted genes had a PCA 1 value greater than 0.5 or less than -0.5, or a PCA 2 value greater than 0.5 or less than -0.5. In Panel (D), the activity levels of juvenile hormone (JH) and ecdysone (20-E) are plotted on a similar parallel coordinate graph (modified from Dhadialla and Raikhel 1994).
:::
:::
Gene annotation and gene ontology assignments
---------------------------------------------
To identify *An. gambiae*genes whose products are involved in related processes, the EST consensus sequences of the transcripts differentially expressed in these 3 patterns were annotated using sequence similarity and categorized using the molecular functions listed by the Gene Ontology Consortium (GOC) and the biological processes defined by Holt et al. \[[@B49]\]. Gene annotations for all 413 at least twofold differentially expressed gene products are given in the Supplementary Table S1. They were then categorized into 9 major categories with 31 subdivisions (Table [7](#T7){ref-type="table"}). 48% of the genes could not be annotated and therefore were categorized as \"Unknown\". The three most numerous categories containing annotated gene products were \"Metabolism\", \"Protein Synthesis\", and \"Egg Production\".
::: {#T7 .table-wrap}
Table 7
::: {.caption}
######
Functional annotation of AS represented by microarray expression group
:::
**Early** **Middle** **Late**
------------------------------------------------------------------------ ----------- ------------ ----------
**Metabolism**
Simple/Complex Carbohydrate Metabolism and Transport 3 0 2
Oxidative Phosphorylation 8 5 1
Lysosomal Enzymatic Digestion 0 0 2
Protein Digestion 4 4 3
Protein Modification, Metabolism, Transport and Localization 5 12 8
Amino Acid and Derivative Metabolism and Transport 1 8 2
Nucleobase/Nucleoside/Nucleotide/Nucleic acid Metabolism and Transport 4 1 3
Fatty Acid/Lipid Metabolism and Transport 1 0 2
Vitamin/Vitamin Derivative/Cofactor Metabolism and Transport 2 0 0
Xenobiotic Metabolism and Transport 1 0 2
Total 29 30 25
**Transport**
Ion Transport 3 3 0
Receptor-mediated Endocytosis 1 0 2
Total 4 3 2
**Protein Synthesis**
Transcription and mRNA Processing 2 5 6
Translation 11 10 4
Protein Folding 3 4 4
Total 16 19 14
**Cellular Processes**
Cell Cycle 0 3 3
Cellular Proliferation 0 0 2
Chromatin Assembly/Disassembly 0 0 5
Apoptosis 1 0 0
Senescence 1 0 0
Total 2 3 10
**Egg Production**
Vitellogenesis/Oogenesis/Embryogenesis 2 10 7
Melanization 0 1 0
Total 2 11 7
**Cellular Communication**
Signal Transduction 1 1 2
Cell-cell Signaling 3 0 1
Total 4 1 3
**Intra-/Extra-cellular Architecture Maintenance**
Structural 3 4 2
Muscle-related 1 0 0
Cell Adhesion 1 1 0
Cytoskeleton Organization and Biogenesis 1 0 2
Total 6 5 4
**Response to Stress/External Stimulus**
Response to Oxidative Stress 2 1 3
Immune/Defense Response 2 2 3
Total 4 3 6
**Unknown**
Total 77 55 68
:::
During the 48 hours PBM, the majority of gene products involved in metabolism were up-regulated Early and Middle Genes. Largely different metabolic biological processes were up-regulated in Early vs. Middle and Late Genes. More than half of the Early metabolic gene products, 20/29 unique transcripts, appear to be involved in carbohydrate metabolism, oxidative phosphorylation, and protein digestion. In contrast, 80% of the metabolic genes, 24/30 unique transcripts, represented in the Middle Genes contribute to various processes in protein digestion and metabolism, and metabolism of amino acids and their derivatives. One third of the Late genes involved in metabolism, 8/25 unique transcripts, are involved in protein metabolism. Five of these annotated sequences, ASs 368, 807, 1179, 279, and 922, encode products involved in post-translational modification.
Reflecting the necessity of biosynthetic machinery in cell maintenance and growth, and probably also the highly conserved nature of proteins involved in housekeeping functions, the protein synthesis category contained the second largest number of genes functioning in a known process. 69%, 11/16 unique transcripts and 53%, 10/19 unique transcripts, respectively of the Early and Middle Genes, in this category are involved in translation. Approximately 25%, 5/19 unique transcripts, and 40%, 6/14 unique transcripts, of the protein synthesis genes represented among Middle and Late genes, are required for transcription and mRNA processing. This result seems almost paradoxical because transcription and mRNA processing necessarily precede translation.
A number of biological processes were related by their involvement in nuclear events or the overall activity of the cell. There were three times as many genes involved in cellular processes among the Late genes than in the other two sets combined. Although all the genes involved in the cell cycle are Middle and Late Genes, the most strikingly up-regulated cellular process genes were those involved in chromatin assembly/disassembly. 5 of the 10 cellular process Late genes (ASs 1136, 59, 592, 1011, and 1792) are involved in maintenance of chromatin structure, a biological process represented only in the Late genes. These Late genes include both histones and high mobility group proteins.
The majority of significantly up and down regulated genes appear to function in egg production, either in the development and maturation of oocytes or in the fat body synthesis of products that will be deposited in oocytes. Almost 90% of them, 17/19 unique transcripts, are Middle and Late genes. These genes are described in detail in the Discussion section. In contrast, half of the genes categorized as cellular communication genes, 4/8 unique transcripts, are Early genes. The majority of the cellular communication gene products in the combined sets of Middle and Late genes, are involved in different signal transduction pathways. Additionally, almost half of the intra-/extracellular architecture maintenance genes are Early genes. This category includes a wide variety of gene products such as peritrophin, both muscle-related and cytoskeletal actins, α-catenin and β-integrin. The Middle and Late genes in this category were mainly structural and included two peritrophins (ASs 13 and 642).
The biological process categorized as transport included not only the movement of ions such as zinc, sodium and potassium, but also transport of molecules via receptor-mediated endocytosis. All three transport gene products, ASs 1336, 1605, and 432, in the Middle genes are responsible for the movement of ions. In contrast, in the Early and Late genes, several gene products (ASs 1974, 1071, and 2086) appear to be involved in receptor-mediated transport via clathrin-coated vesicles.
A number of genes responding to oxidative stress (6 genes in total) were found in all three sets of genes indicating that they are transcribed throughout the 48 hours PBM. Seven additional gene products most probably involved in immunity, a response to external stress, were found among these three gene sets.
qRT-PCR analysis
----------------
Expression profiles of eight selected genes and the RP S7 control gene were confirmed using a quantitative real-time PCR strategy (Figure [4](#F4){ref-type="fig"}). Transcript levels for each of the eight genes were quantified using SYBR Green technology and differences in their expression between sugar-fed and blood-fed mosquitoes at 0, 5, 12, 24 and 48 hours PBM determined. Although the magnitudes of the changes in transcript abundances of all the genes whose expression levels were quantified by both techniques differed between the techniques, the changes in direction of expression, whether positive or negative, remained consistent for the majority of them. In addition, the overall patterns of expression exhibited by the three sets of genes were also apparent in the expression profiles created by qRT-PCR analysis. For the two Early genes, microarray analysis overestimated transcript levels between 2- and 30-fold more than qRT-PCR analysis. In contrast, for the majority of the Middle and Late gene expression measurements, microarray analysis underestimated transcript abundances relative to qRT-PCR analysis.
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Comparison of microarray and qRT-PCR gene expression profiles for selected genes. X-Y plots were generated from the ratio of transcript levels in the blood-fed adult female mosquitoes to the transcript levels in the sugar-fed adult female mosquitoes for eight selected genes at five time points including 0, 5, 12, 24 and 48 hours post blood meal (PBM). Genes were randomly selected from the three sets and included ASs 648 and 1786 from the Early Genes (A), ASs 996, 1158, 1949, and 679 from the Middle Genes (B), and ASs 12 and 1357 from the Late Genes (C). Triangles indicate fold expression data generated by qRT-PCR analysis; circles indicate fold expression data generated by microarray analysis. A horizontal line connecting either the diamonds or circles illustrates each gene expression profile. For each gene, the expression profiles created by qRT-PCR and microarray analysis are indicated using the same color.
:::
:::
Discussion
==========
We have determined the gene expression patterns of 3,068 abdomen-derived cDNAs from adult female *An. gambiae*mosquitoes representing 1906 unique transcripts were determined in the first two days following ingestion of a blood meal by microarray analysis. 413 unique transcripts were shown to be up-or down-regulated at least twofold in blood fed mosquitoes relative to sugar-fed mosquitoes at one or more of the following times: 5 min and 30 min following initiation of blood feeding and 0, 1, 3, 5, 12, 24, and 48 hr post-blood meal. These transcripts were clustered into three sets with different temporal patterns of expression that may reflect the major hormonal changes occurring within the mosquito during a gonotrophic cycle. These differentially expressed gene products were annotated putatively using sequence similarity searches and categorized by biological process to identify the major events occurring post blood meal ingestion in the female mosquito.
Multiple hormones interact to alter tissue states and to activate genes involved in the female mosquito\'s digestion of a blood meal, in oocyte development and in vitellogenesis. The three sets of differentially transcribed genes discerned in this study, the Early, Middle and Late genes may reflect differential hormonal responsiveness. After acquisition of a blood meal, the transcript levels of the Early genes which were abundant during blood feeding showed general declines until 24 hours PBM, after which a subset of transcript levels began to rise again. Expression of Early genes may be linked to the relatively high titers of JH present at the beginning of the first gonotrophic cycle and may then be repressed as a result of declining JH titers or of increasing 20-E. The expression of Middle gene transcripts followed an expression pattern that reflects the titers of 20-E: levels sharply increased by 12 hours PBM, remained stable or increased only slightly until 24 hours PBM, and then declined rapidly. OEH secreted by median neurosecretory cells stimulates the ovaries to secrete 20-E during vitellogenesis and the activity of this hormone begins to rise by 3--5 hours, peaks between 12 and 24 hours, and then declines to baseline levels by 48 hours post blood ingestion \[[@B4]-[@B6]\]. In contrast to the Middle genes, most Late gene transcripts exhibited baseline, steady state levels until 12 hours PBM after which they were induced at least twofold and continued to exhibit increased transcript levels at 48 hours PBM. This increase in transcript levels mirrored the increase in JH titer observed by 48 hours post blood meal ingestion \[[@B50]\]. These results suggest that Middle genes products may be ecdysone-responsive whereas Late genes products may be JH-responsive.
Based on their finding that the *Drosophila minidiscs*gene product showed a primary response to JH, Dubrovsky et al. \[[@B51]\] have suggested that JH may transcriptionally regulate genes encoding maternally inherited products. Our Late gene, AS 806, shares sequence similarity with *minidiscs*. Additionally, the majority of mosquito gene products showing sequence similarity to maternally active *Drosophila*genes are categorized as Late genes. Whether transcription of Early genes is directly influenced by levels of JH or 20-E, cannot be determined easily because neither hormone is present at high levels during the first few hours following blood meal ingestion. Regardless of expression pattern, the gene products in each set reflect a diverse array of processes occurring in the female mosquito within 48 hours following initiation of blood feeding. The major processes initiated in response to blood feeding including digestion, peritrophic matrix formation, oogenesis and vitellogenesis, are discussed below with emphasis on the likely roles of particular gene products.
Digestion
---------
Digestion of the two different food sources, nectar sugars and blood, requires changes in the types of enzymes present within the digestive tract of the mosquito. The numbers of genes associated with sugar and protein metabolism within each set of genes may reflect the switch from sugar to protein metabolism. For instance, there are increases in transcript abundance of genes involved in carbohydrate metabolism and oxidative phosphorylation in the Early genes. However, both Early and Middle gene categories are enriched in genes involved in protein digestion. This result stems from the involvement of certain gene products in blood meal digestion that initiate a signaling cascade resulting in the up-regulation of other related proteolytic enzymes. The Middle genes contain the majority of gene products involved in amino acid metabolism, a process that follows protein digestion, whereas by 48 hours PBM, the time at which the majority of Late genes are induced, there is a generalized decrease in digestive enzyme transcripts.
Considering that blood contains large quantities of protein, the mosquito requires a variety of proteolytic enzymes to digest the recently acquired meal. In the present study, 11 genes were identified whose products are most likely required for protein digestion. These include 5 previously characterized digestive enzyme genes, two trypsins, a chymotrypsin, a serine protease and a carboxypeptidase. The majority of these digestive enzyme genes were transcribed at levels greater than twofold induction after 6 hours PBM. The *An. gambiae*Trypsins 1 and 2 are both induced by a blood meal and exhibit similar expression profiles although Trypsin 1 is expressed at higher levels. Muller and coworkers, using an RT-PCR strategy, showed that Trypsin 2 mRNA is present at 8, 12, 24 and 28 hours post blood meal \[[@B18],[@B52]\]. In our study, the Trypsin 2 gene (AS 569) also exhibited increased transcript abundance at 12 hours PBM with maximal expression occurring at 24 hours PBM, but these levels decreased by 48 hours. In contrast to Trypsins 1 and 2, Trypsins 3, 4, and 7 are constitutively expressed in non-bloodfed females. By 4 hours following a blood meal, levels of Trypsin 4 become undetectable by Northern and RT-PCR analysis and do not reappear until 20 hours PBM \[[@B18]\] and unlike Trypsins 1 and 2, Trypsin 4 reaches maximal expression by 48 hours, near the end of the gonotrophic cycle. Our study identified two clones corresponding to Trypsin 4 (AS 568) but the two cDNAs exhibited different expression patterns. One (Accession no. CD747033) reached maximal transcript abundance at 48 hours PBM, the expected expression pattern. However, the other cDNA (Accession no. CD747029) was expressed at high levels prior to 6 hours PBM and also reached at least twofold increased levels by 48 hours PBM. These cDNAs may not have clustered together for technical reasons intrinsic to Seqman II, or they may correspond to alternatively spliced variants of the same gene. The sequence alignment showed 97% nucleotide sequence identity between the two ESTs over the region common to both. However, the CD747033 EST is only 253 bp in length and it is possible that the full-length cDNA represents an alternatively spliced transcript. In addition to Trypsins 2 and 4, we identified a trypsin-like serine protease (AS 648) which shared the greatest amino acid similarity with Trypsin 4. However, this serine protease exhibited highest nucleotide identity to a different region of chromosome 3R than that containing all previously identified digestive trypsins. This gene product was induced greater than twofold within 6 hours PBM and was repressed during the height of digestion. Unlike Trypsins 3--7, it was not expressed at higher levels at 48 hours PBM. It is possible that this trypsin is not involved in digestion but in another proteolytic process that is down-regulated following a blood meal.
Barillas-Mury *et al*. \[[@B53]\] demonstrated that the early trypsin activity is essential to the transcription and subsequent expression of the late trypsins in *Ae. aegypti*. *An. gambiae*Trypsins 3--7 may indirectly activate transcription and increase the expression of Trypsins 1 and 2, the major endoproteolytic enzymes required for blood meal digestion \[[@B18]\]. In both *Ae. aegypti*and *An. gambiae*, these early expressed digestive enzymes are presumed to act as signal transducers causing transcriptional up-regulation of the late expressed ones \[[@B14],[@B53]\]. Thus the digestive process is regulated by an elaborate biphasic expression pattern of serine proteases. Additional evidence suggests that not only tryptic peptides but cleaved amino acids may serve as systemic signals regulating subsequent processes. In contrast to the trypsin by-products, cleaved amino acids may also function as negative regulators of food intake. AS 1158 shared weak sequence similarity with the *Drosophila pumpless*protein, a larval fat body-expressed enzyme involved in glycine catabolism. In the fruit fly, larvae expressing the *pumpless*mutation are unable to pump food from the pharynx to the esophagus \[[@B54]\]. These mutant animals do not feed, neither do they upregulate genes normally involved in responses to starvation. Because feeding amino acids to wild type larvae phenocopied effects of the *pumpless*mutation, Zinke *et al*. \[[@B54]\] proposed that amino acids released from the fat body normally act as signals for cessation of feeding.
In addition to the trypsins, three chymotrypsin genes have been characterized in *An. gambiae*\[[@B19],[@B20]\]. The expression of two of these digestive enzymes, AnChym 1 and 2, has been localized to the mosquito midgut by analysis of Northern blots \[[@B19]\]. RT-PCR showed that both chymotrypsin genes are expressed at 12 hours PBM and are abundant until 48 hours PBM, unlike transcripts of Trypsins 1 and 2 which have decreased dramatically by this time \[[@B19]\]. The cDNA representing chymotrypsin 2, AS 99, was categorized as a Late gene since maximal transcript levels were achieved after 24 hours PBM. In contrast, the other characterized chymotrypsin, AgChyL, exhibits changes in transcript abundance which is more similar to those of Trypsins 3--7. mRNA is present in non-blood fed females, detectable until 8 hours post blood meal after which mRNA can no longer be measured until 48 hours PBM \[[@B20]\]. The cDNA corresponding to this chymotrypsin-like serine protease, AS 994, was induced more than twofold prior to the peak of digestion and clustered with the Early Genes. In addition to the aforementioned chymotrypsins, our study identified a previously uncharacterized chymotrypsin (AS 2243) also located on chromosome 2L but in a different region from both AnChym1 and AnChym2. Exhibiting an expression pattern different from both AnChym2 and AgChyL, this gene product was characterized as a Middle Gene with maximal transcript levels achieved between 12 and 24 hours PBM and a return to baseline by 48 hours PBM, similar to the expression patterns of Trypsins 1 and 2.
Edwards *et al*. \[[@B21]\] investigated expression levels of *An. gambiae*carboxypeptidase A following blood meal ingestion. Northern blot analysis indicated that levels of Carboxypeptidase A mRNA rose rapidly to a ten-fold increase within 3--4 hours following a blood meal, then dropped to the pre-induction state by 24 hours PBM. We identified a carboxypeptidase gene located on chromosome 2L (AS 1742) that is transcribed in a manner similar to that of carboxypeptidase A. However another cDNA representing a carboxypeptidase (AS 44) exhibited a radically different expression pattern. Transcripts were present at low levels 1--5 hours PBM but increased more than twofold between 12 and 24 hours PBM, a pattern that resembled the enzymatic activity in *An. stephensi*observed by Jahan *et al*. \[[@B13]\], namely a rapid increase until 12 hours PBM, with a peak at 24 hours, followed by a steady decline over the next day.
In contrast to the *An. gambiae*carboxypeptidase A, the levels of aminopeptidase peaked around 30 hours in *An. stephensi*\[[@B12]\]. Additionally, Lemos *et al*. \[[@B14]\] recorded peak aminopeptidase activity at 24 hours PBM in *An. gambiae*. We identified two aminopeptidases with at least two-fold increased expression during the 48 hours PBM. The first aminopeptidase, AS 340, reached peak transcript levels at 24 hours PBM, showing a similar expression pattern to the enzyme activity levels reported by Billingsley and Hecker \[[@B12]\] and Lemos *et al*. \[[@B14]\]. In contrast, the other aminopeptidase, AS 430, showed amino acid similarity to the *Ae. aegypti*aminopeptidase N. This aminopeptidase was classified as a Late gene, due to increased transcript levels at 24 hours PBM but maximal levels were not achieved until 48 hours PBM. Jahan *et al*. \[[@B13]\] documented two different kinetic profiles of aminopeptidase enzymes in *An. stephensi*depending on whether the enzyme was soluble or membrane-associated. The soluble aminopeptidase exhibited a kinetic profile similar to AS 340 and to that presented by Billingsley and Hecker \[[@B12]\] with peak activity at 24 hours PBM.
Geering \[[@B55]\] had suggested that phospholipase activity plays a role in blood digestion in *Ae. aegypti*although no conclusive evidence was demonstrated. However, Geering and Freyvogel \[[@B56]\] demonstrated that lipolytic activity increased 15 hours after blood feeding. Of the three gene products that are characterized as being involved in Fatty Acid/Lipid Metabolism and Transport, two, ASs 1177 and 997, encoding an acetate-CoA ligase and a fatty acid binding molecule, respectively, are expressed at more than twofold greater abundance at 24 hours PBM. These gene products may be involved in fatty acid degradation of blood meal components and the transport of these lipids to the oocytes.
The erythrocyte membrane contains a number of glycoproteins. It is therefore possible that enzymes normally associated with carbohydrate metabolism of nectar meals could also be involved in blood digestion. Almost all of the Simple/Complex Carbohydrate Metabolism and Transport genes identified in our study as being at least twofold upregulated were categorized as Early genes. Within the first 6 hours PBM, their transcripts are present at higher levels than in sugar-fed females and thereafter, they decrease steadily until 12 and 24 hours PBM. This expression profile does not exclude these genes from having a role in RBC glycoprotein metabolism. Several glycosidases are present within the midgut of *An. stephensi*, either associated with the lumen or with epithelial lysosomes \[[@B12]\]. The enzymatic activity of α-glucosidase, the major midgut glycosidase in *An. stephensi*, increased from 6 hours PBM to maximal levels by 24 hours and decreased to basal levels by 36 hours PBM in the anterior midgut. The transcript abundance of the α-glucosidase, AS1786, characterized in this study followed a different pattern than the enzymatic activity of *An. stephensi*α-glucosidase. It showed greatest amino acid sequence similarity to *Drosophila melanogaster*gene product CG8690 α-glucosidase and was categorized as an Early Gene with minimal transcript levels occurring at and after 24 hours PBM.
Peritrophic matrix formation
----------------------------
Shen and Jacobs-Lorena \[[@B34]\] characterized the *An. gambiae*peritrophic matrix protein Peritrophin 1 (Ag-Aper1) by analysis of Northern blots, and demonstrated that transcripts were present 6 hours PBM, increased by 12 hours, and remained elevated between 24--48 hours PBM. The present study identified several genes encoding proteins with a chitin-binding domain (InterPro ID IPR002557: Chitin binding Peritrophin-A) similar to the one found in Peritrophin 1. The Early gene AS 928 contains the Peritrophin-A chitin binding domain and maps to chromosome 3L *in silico*, corresponding to agCP10685. Another Early gene, AS 13, shows high identity to Peritrophin 1. A Middle gene, AS 516, is expressed by 6 hours PBM but does not reach maximal levels until 24 hours PBM. This gene product maps *in silico*to chromosome 2L, exhibits a similar transcript profile, and also shares 98% amino acid identity with Peritrophin 1. It is not clear why two sets of ESTs, both identified as Peritrophin 1, should exhibit different transcription patterns unless they are derived from differentially regulated genes. Another Middle gene, AS 1164, contains the Peritrophin-A chitin binding domain in addition to a prenyl-group binding site (InterPro ID IPR001230: CAAX box). This Middle gene may not be involved in peritrophic matrix formation but in some other process coinciding with protein digestion. Two other cDNAs, Accessions CD746211 and CD746202, both in AS 13, could also be localized to chromosome 2L. However, their ESTs exhibited greatest nucleotide identity to the gene predicted as ENSANGG00000020776 located 4 kb 3\' to Ag-Aper1. These two cDNAs clustered as Early and Late genes respectively, and may represent alternatively spliced gene products.
As early as an hour after the adult female has taken a blood meal, secretory vesicles previously present in the apical brush border of midgut epithelial cells are no longer detectable \[Staubli et al., 1966 as cited in \[[@B30]\]\]. These apical granules presumably contain precursors of the peritrophic matrix. The Middle and Late peritrophin gene products may be packaged into vesicles in preparation for a subsequent blood meal. In contrast, the Early peritrophin genes may be transcribed in response to blood meal acquisition and their products used immediately in the formation of the peritrophic matrix.
Ovarian cycle and oogenesis
---------------------------
The extensive literature on genes involved in *Drosophila*ovarian development and early embryogenesis opens windows into interpreting our *An. gambiae*microarray results and understanding mosquito egg development. The majority of *An. gambiae*genes upregulated at least twofold following a blood meal appear to function in egg production. The only gene possibly involved in oogenesis during the early phases of the ovarian cycle is the Early gene AS 670. It shares sequence similarity with *peter pan*, a *Drosophila*gene product required during oogenesis. Oocytes in *peter pan*mutants often have an incorrect number of associated nurse cells, suggesting that the *peter pan*protein influences the separation of cells within the germaria \[[@B57]\]. The identification of other genes involved in the early stages of mosquito oogenesis may be facilitated by the construction of cDNA libraries from the abdomens or ovaries of recently blood fed female mosquitoes.
Several differentially expressed gene products found in the present study may be involved in the formation of ring canals, structures necessary for the delivery of maternal factors to oocytes. In particular, bulk transfer of cytoplasmic content from nurse cells to oocytes depends on actin structures \[[@B40]\]. A Middle gene product, AS 679, shares sequence similarity to the *Drosophila*gene CG13388 encoding the protein kinase anchor protein 200, *Akap200*. *Akap200*protein localizes to ring canals during oogenesis, regulates protein kinase C activity, and controls their morphology \[[@B58]\]. Late gene product AS 1317 shows sequence similarity to the *Drosophila pendulin*gene product, encoded by CG4799. This gene product is also required for assembly of fully functional ring canals. *pendulin*encodes an importin-α2, a protein necessary for the localization of the *kelch*gene product, CG7210. *kelch*encodes an actin organizer without which the ring canals become occluded and nurse cell-oocyte cytoplasmic transport is inhibited \[[@B59],[@B60]\]. Though we found an apparent *pendulin*gene, we did not find *kelch*. A Late gene product, AS 578, is the homolog of *Cdc42*, which encodes a small monomeric RHO GTPase involved in signal transduction. Rohatgi *et al*. \[[@B61]\] suggested that *Cdc42*protein most likely links signal transduction to the actin cytoskeleton in *Xenopus*. In *Drosophila*ovaries, mutations in *Cdc42*caused nurse cells to deflate and coalesce, and inhibited transfer of nurse cell cytoplasm to oocytes in late stage egg chambers \[[@B62]\].
*Drosophila*nurse cells transcribe the *bicoid*anterior determinant gene and the resulting mRNA is transported to the anterior region of developing oocytes via polarized microtubules \[[@B63]\]. *bicoid*does not appear to be an anterior determinant in other insects, but other genes important for its localization are conserved. The Early gene AS 2047 shares similarity with the *Drosophila cornichon*gene, CG5855. In *Drosophila*, *cornichon*is required for formation of a functional microtubular cytoskeletal scaffold used to transport *bicoid*mRNA and the posterior group *oskar*gene product to their proper location within the embryo \[[@B64]\]. The Late gene product, AS 1044 exhibits greatest sequence similarity to *D. virilis exuperantia*, a gene whose product is also required for proper *bicoid*mRNA localization *D. melanogaster*\[[@B65],[@B66]\]. The Late gene AS 2222 is putatively identified as the *An. gambiae*homolog of *Drosophila Notch*. In *Drosophila*, *Notch*signaling regulates a large number of ovarian events beginning with cyst development in the germarium and extending through oogenesis \[[@B67]\]. The mechanisms by which *Notch*signaling activates transcription of its target genes are reviewed by Barolo and Posakony \[[@B68]\]. Since we identified *Notch*as a Late gene, its activities may be more restricted in *An. gambiae*, and/or reflect fundamental differences in ovarian biology. AS 1391, a Late gene product, shares sequence similarity with *Drosophila Rab-protein 11*. This *Drosophila*small monomeric RAB GTPase is also involved in the polarization of the microtubules for the organization of the posterior pole and for *oskar*localization there \[[@B69]\].
To regulate the progress of oogenesis and embryogenesis, stored maternal mRNAs are translationally repressed during early oocyte development. The Middle gene product AS 2031 shares sequence similarity with the *Drosophila*gene product *Bicaudal C*, a RNA binding protein that may play a role in translational silencing of maternal mRNAs in addition to its role in eggshell patterning \[[@B70]\]. Mutations in *Bicaudal C*result in premature translation of *oskar*mRNA before it has reached the posterior region of the oocyte \[[@B71]\]. The Middle gene product AS 1490 is the putative homolog for the *Drosophila*gene product *vasa*(CG3506), an ATP dependent helicase involved in pole plasm assembly that may also be involved in translational modification of maternal mRNAs \[[@B72]\]. The Late gene AS 453 shares sequence similarity with the *Drosophila cup*protein (CG11181). *cup*protein interacts with *nanos*, the posterior determinant, and a translational regulator of the gap gene *hunchback*mRNA during oogenesis, although the exact function of the *cup*protein still remains unknown \[[@B73]\]. A DEAD box protein encoded by *vasa*also influences oocyte differentiation and the development of the *Drosophila*embryo body plan via translation of *oskar*, *nanos*, and g*urken*during oogenesis \[[@B74]-[@B77]\]. In amphibians, several mRNA binding proteins have been identified that are only present in oocytes and not post cleavage embryos \[[@B78],[@B79]\]. One *An. gambiae*Middle gene product, AS 2449, shared sequence similarity with the *Xenopus laevis*poly(A)-specific ribonuclease and also has mRNA binding motifs, thus it may also repress translation of mRNAs in embryos.
Maternal nurse cells not only provide the biosynthetic machinery and mRNA needed for oocyte axis determination, but also many transcripts and proteins required for zygotic development through the cellular blastoderm stage. The Middle gene product AS 1032 shares sequence similarity with *nop5*, encoding a maternally derived product of the *Drosophila*CG10206 gene, a component of the small nucleolar ribonucleoprotein (snoRNP) complex involved in rRNA processing \[[@B80]\]. The Late gene product AS 806, referred to above in the context of its possible regulation by JH, shares sequence similarity with the *Drosophila minidiscs*gene product, an amino acid transporter. In *Drosophila*ovarian nurse cells, JH induces the expression of *minidisks*and its transcripts are most likely transferred to the oocyte during nurse cell cytoplasmic streaming \[[@B51]\].
Similar to the oocyte, the eggshell undergoes dorsal-ventral patterning. Crucial to this process is the correct placement of the oocyte relative to the maternal somatic follicle cells. In *Drosophila*, the localization of the oocyte depends on cadherin-associated adhesion \[[@B81]\]. The Late gene product AS 1890 is the homolog of α-catenin, the CG17947 gene product. A cytoskeletal anchor protein, α-catenin is required for positioning of the oocyte relative to the posterior follicle cells during germ cell rearrangement in *Drosophila*\[[@B81]\]. The RAS 1 signaling cascade is an important means of cell communication during embryo and eggshell patterning \[[@B41],[@B42]\]. The Middle gene product AS 657 is weakly similar to the *Drosophila Star*protein, a RAS 1 enhancer involved in the EGF receptor signaling pathway, either upstream or in parallel to EGFR, during formation of the embryonic ventral midline. *Star*encodes a single pass transmembrane protein that may be involved in the processing of *gurken*protein. The DNA damage checkpoint *14-3-3epsilon*protein also participates in RAS 1 signaling, normally functioning downstream or in parallel to RAF, but upstream of transcription factors. The Middle gene AS 106 shares sequence similarity with *Drosophila 14-3-3epsilon*. The *14-3-3epsilon*protein is also capable of binding to a large number of other proteins in a phosphorylation-dependent manner. One of its functions may be to alter the cell cycle by binding Cyclin B and appears to have homologs in most if not all eukaryotes \[[@B82],[@B83]\]. The *An. gambiae Cyclin B*homolog, AS 1357, grouped as a Late gene.
Mutation screens in *Drosophila*have led to the identification of a number of other gene products that may be involved in RAS 1 signaling. *TppII (tripeptidyl-peptidase II)*, and *smt3*(SUMO) were discovered in a search for lethal mutations that could enhance a weak RAS 1 eggshell phenotype \[[@B84]\]. The Middle gene product AS 1268 shares sequence similarity with *Drosophila tripeptidyl-peptidase II*, the CG3991 gene, encoding a serine protease that degrades neuropeptide signals \[[@B85]\]. The Late gene product AS 922 is the homolog of *Drosophila smt3*, CG4494, whose product is ubiquitin-like protein that may tag proteins for nuclear localization or retention in the cytoplasm \[[@B86],[@B87]\]. *smt3*protein may modulate activity of transcription factors in the follicle cells downstream of EGFR activation. However, we feel that *smt3*is likely to be a minor player in RAS 1 signaling in the events following blood ingestion in the mosquito, because its mRNA reaches maximal expression after 12--24 hours PBM, unlike the other gene products we identified as potentially influencing RAS 1 signaling.
*smt3*protein may also play a role in *Toll*signaling. This signal transduction pathway is known to be necessary for dorsal/ventral patterning of the *Drosophila*embryo. *smt3*protein binds the NFκB homolog *dorsal*protein and targets this Rel transcription factor for migration to the nucleus \[[@B88],[@B89]\]. Bhaskar *et al*. \[[@B88]\] demonstrated that *smt3*conjugation to the *dorsal*protein enhanced its transcriptional activity. *smt3*protein may play other roles in the cell by altering the interactions of septins, cytoskeletal proteins involved in cytokinesis \[[@B90],[@B91]\]. In *Drosophila*, septins have been found in the cytoplasm of nurse cells and at the baso-lateral surfaces of follicle cells \[[@B92]\]. These results suggest even more pleiotrophc roles for the *smt3*gene product in oogenesis. We also found that the Late gene AS 2034, a homolog of *Drosophila Aos1*, the CG12276 gene, the *smt3*(SUMO) activating enzyme, was also expressed at least twofold more abundantly during the height of *smt3*expression. This result reinforces the importance of *smt3*in the events occurring between 24--48 hours PBM.
In addition to genes regulating the polarity of the embryo and eggshell, genes involved in cellular growth and differentiation were differentially expressed during the 48 hours PBM ingestion. AS 337 and 495 shared sequence similarity with the *Ae. aegypti*ornithine decarboxylase antizyme, a protein that modulates polyamine synthesis. The homologous *Drosophila ornithine decarboxylase antizyme*gene, formerly known as *gut feeling*, has been shown to be important in developing oocytes. It is one target of *Sex lethal*which encodes an RNA binding protein that regulates mRNA splicing and the mitotic events in early germ cells via regulating Cyclin B \[[@B93]\]. The Late gene product AS 2073 shares sequence similarity with the *Drosophila polo*CG12306 gene product, a protein kinase required for cytokinesis and another regulator of Cyclin B \[[@B94]\]. The Early gene AS 1972 shows identity with the *Drosophila black pearl*CG5268 gene product. This protein contains DnaJ domains implying that it is necessary for cellular growth \[[@B95]\]. Northern blot analysis of *black pearl*RNA from various developmental stages showed two transcripts with greatest expression in *Drosophila*embryos 0--6 hours old \[[@B95]\], the stages in which DNA replication recurs most rapidly. The Late gene AS 2268 shares sequence similarity with the *Drosophila Imaginal disc growth factor4*(*Idgf4*), a mitogen with a non-functional chitinase domain. Transcripts of *Idgf4*are detected in *Drosophila*nurse cells, oocytes, and in the yolk cytoplasm of early embryos \[[@B96]\]. The Middle gene product AS 2273 shares sequence similarity with *Drosophila β Integrin*. *An. gambiae*β Integrin may interact with the Middle gene AS 985 product to promote somatic cell adhesion and cell migration during oogenesis and embryogenesis. This is due to the similarity of the AS 985 gene to *Drosophila Receptor of activated protein kinase C*, RACK1. RACK1 can bind a number of different signaling and cell adhesion molecules including the activated form of protein kinase C (PKC), Src family kinases, and β Integrins \[[@B97]-[@B99]\]. Cox *et al*. \[[@B100]\] demonstrated that, in a mammalian system, RACK1 organizes focal adhesions and directional cell migration via its Src-binding site. Mahairaki *et al*. \[[@B101]\] found that the *An. gambiae*β Integrin gene was expressed at highest levels 48 hours PBM, whereas we found that the β integrin homolog reached at least twofold increased expression by 24 hours PBM.
A number of genes have been implicated in the development of the egg shell structures. Our screen does not appear to have identified any homologs of the several endochorionic structural proteins characterized in *Ae. aegypti*\[[@B102],[@B103]\]. This was unexpected because Northern blot analysis had indicated that transcription of the vitelline membrane proteins 15a-1, 15a-2, 15a-3 was induced rapidly between 10 and 24 hours PBM, reached maximal levels between 30 and 40 hours PBM, and decreased to baseline levels between 50 and 60 hours PBM \[[@B102],[@B103]\]. Our study also did not identify a *Dopa decarboxylase (Ddc)*gene, Ddc is an enzyme involved in the tyrosine metabolic pathway necessary for eventual chorion melanization in *Ae. aegypti*, and other melanization events. The gene is up-regulated in response to blood meal with transcripts initially detectable by 12 hours PBM, and maximal levels achieved between 24 and 48 hours PBM \[[@B104]\]. However, we did identify a gene encoding another enzyme involved in tyrosine metabolism. AS 1340, a Middle gene product, shared sequence similarity with the *Ae*. *aegypti*Dopachrome conversion enzyme \[[@B105]\]. This enzyme is required for processing of dopachrome to melanin. It is interesting that its mRNA is constitutively expressed in *Ae. aegypti*females, but becomes upregulated when they are infected with *Dirofilaria*\[[@B105]\]. Since insect melanins can be produced via any of three intermediates, Dopa, Dopamine, or Dopachrome, it may be that *An. gambiae*differs from *Ae. aegypti*in the substrate metabolized to produce chorionic melanin.
Several Middle and Late genes encoding antioxidants were upregulated at least twofold 12--48 hours PBM. The Middle gene, AS 2033, a glutathione S-transferase D3, and the three Late genes, ASs 1684, 35, and 2156, encoding glutathione S-transferase 1--6 class theta, and homologs of *Drosophila*thioredoxin and *Ae. aegypti*2-Cys thioredoxin peroxidase, may have roles in regulating reactive oxygen species that can be produced from the highly reactive quinones which are normally cross-linked into melanin immediately after they are formed.
Ovarian cycle and vitellogenesis
--------------------------------
Paramount to the development of the embryo is the massive accumulation of vitellogenin by the oocyte. In *An. gambiae*there is a small, polymorphic tandem array of vitellogenin genes and a single dispersed vitellogenin gene, all located on Chromosome 2R in division 18B (P. Romans and M. Sharakhova, unpublished observations). Vitellogenin mRNA becomes detectable by Northern blot analysis by 8 hrs PBM, though it is detectable earlier by RT-PCR, increases dramatically by 12 hours, reaches maximal levels by 24 hours, and declines to undetectable levels by 48 hours PBM \[[@B47]\]. Our microarray study identified three cDNAs, all Middle gene products and greater than twofold induced, as vitellogenin gene homologs. Two of the ESTs were not conjoined during EST assembly because they represented non-overlapping 5\' and 3\' ends of the Vg1 gene. The third EST included the more closely resembled the sequence of the dispersed vitellogenin gene (P. Romans and A. Dana, unpublished). As expected, all three vitellogenin clones exhibited expression profiles similar to the overall pattern previously described \[[@B47]\].
Following synthesis in the fat body, vitellogenins are released into the hemolymph. Eventually, they diffuse through channels between the cells of the follicular epithelium and are accumulated by the oocyte by receptor-mediated endocytosis in clathrin-coated pits \[[@B8]\]. The increased number of gene products involved in receptor-mediated endocytosis before and after the height of vitellogenin gene transcription, 12--24 hours PBM in this study, may reflect a preparation for the increase in receptor-mediated endocytosis when the oocytes are accumulating vitellogenins and other yolk constituents during the trophic phase of the ovarian cycle. These genes included an Early gene, AS 1974, similar to the *Drosophila*Adaptin subunit, *AP-1σ*, CG5864, and the Late gene, AS 2086, homolog of another *Drosophila*clathrin-associated protein, *AP-50*, CG7057.
When vitellogenesis has ceased, the biosynthetic machinery in the fat body is degraded in lysosomes \[[@B46]\]. In *Ae. aegypti*, the lysosomal cathepsin D-like aspartic protease (AeLAP) exhibited a similar transcription profile to vitellogenin \[[@B106]\]. Cho *et al*. \[[@B107]\] also identified a Cathepsin B-like thiol protease, vitellogenic Cathepsin B or VCB, which is secreted from the fat body with a peak at 24 hours PBM and incorporated into oocytes. It appears to be involved in the degradation of vitellin in embryos. The Middle gene AS 996, a Cathepsin B, shares identity with this *Ae. aegypti*protein, exhibits the same expression profile, and may be its homolog. At approximately 30 hours PBM, 6 hours after peak production of vitellogenin, the activity of four other lysosomal enzymes, arylsulfatase A, acid phosphatase-1, β-galactosidase, and Cathepsin D, has dramatically increased to reach maximal levels by 36--42 hours PBM \[[@B108],[@B109]\]. The two Late gene products, ASs 1254 and 2231, were identified putatively as the lysosomal enzymes, acid phosphatase-1 and Cathepsin F, respectively. These genes also may be involved in the termination phase of vitellogenesis. Cathepsin F is necessary for oocyte growth in a teleost fish and has been suggested to be associated with yolk protein processing \[[@B110]\]. It will be a very interesting example of gene co-evolution, should processing of vitellogenins, proteins conserved between egg-laying vertebrates and non-Brachyceran insects, actually be accomplished by similarly conserved cathepsins.
Conclusions
===========
Holt *et al*. \[[@B49]\] performed the first genomic-scale study of hematophagy in *An. gambiae*by identifying 168 ESTs that differed in statistical abundance between cDNA libraries made from adult female mosquitoes fed on sugar and 24 hours PBM. Ribeiro \[[@B111]\] extended this study by describing an additional 267 such genes. We have expanded on these studies by identifying additional 359 ESTs and by examining virtually a complete first gonotrophic cycle experimentally. In addition, we found 18 ESTs present in our microarray and Ribeiro\'s \[[@B111]\] studies. All but one (AS 205) showed the same expression patterns at 24 hours PBM (Table [8](#T8){ref-type="table"}). These highly synchronous expression profiles of those ESTs further validate that experimental microarray and *in silico*data can complement each other. However, our study is unique in that we have determined the temporal patterns of expression of the genes we identified. The observed similarities between the gene expression patterns and production of the two principal insect hormones suggest that gene transcription may be influenced by changes in JH titers as well as by 20-E levels, a phenomenon that has been well-studied in the context of *Drosophila*metamorphosis and in *Ae. aegypti*vitellogenesis. Future analysis may reveal genes co-regulated via the same promoters. Indeed, this now appears possible for organisms whose genomes have been sequenced \[[@B112]-[@B114]\]. As new regulatory sequences are identified, the arsenal of transcriptional regulators to drive their tissue- and stage-specific gene expression will be increased. We expect that this increased promoter availability will supplement current vector-control strategies.
::: {#T8 .table-wrap}
Table 8
::: {.caption}
######
List of genes differentially expressed\* in female *A. gambiae*at 24 hours post-blood meal in both microarray and *in silico*(Ribeiro, 2003) gene expression studies.
:::
**AS ID** **Ensembl ID** **Microarray** ***in silico*** **Molecular Functions** **Biological Processes**
----------- ---------------- ---------------- ----------------- ------------------------- ---------------------------
99 agCP3123 Up Up enzyme Protein Digestion
996 agCP14019 Up Up enzyme Egg Development
2222 agCP8969 Up Up unknown Unknown
1949 agCP12846 Up Up unknown Unknown
1317 agCP8818 Up Up transporter Transport
516 agCP3409 Up Up binding Structural
1044 agCP3927 Up Up unknown Egg Development
995 agCP5701 Up Up enzyme Protein Digestion
230 agCP2518 Up Up nutrient reservoir Egg Development
180 agCP1111 Up Up unknown Unknown
2207 agCP15442 Up Up transporter Ion Transport
2256 agCP2731 Up Up unknown Unknown
2243 agCP3610 Up Up enzyme Protein Digestion
205 agCP5849\*\* Down Up unknown Unknown
86 agCP6049 Down Down unknown Unknown
553 agCP11425 Down Down transporter Oxidative phosphorylation
2123 agCP11416 Down Down transporter Transport
642 agCP8191 Down Down structural molecule Cuticle biosynthesis
\*; These genes displayed at least 2-fold up- or down-regulation relative to the control.
\*\*: AS 205 shows discrepancy between microarray and *in silico*expression data.
Up: up-regulation; Down: down-regulation.
:::
Great progress has been made in the annotation of the *An. gambiae*genome, culminating in the public announcement of the genome sequence in 2002 and its subsequent updates. Yet, although we have identified 413 differentially expressed gene products, we could not assign almost half of them to a biological process. Of the 200 \"Unknowns,\" 43 unique transcripts shared no significant identity with sequences in the Nr and dbEST databases. The genes corresponding to these transcripts may be identified following the second gene build of the *An. gambiae*genome. Functional studies using microarray analysis verified by qRT-PCR must confirm *in silico*predicted annotations and provide biological information about gene products. Many of the gene products identified in this study share sequence similarity with *Drosophila*proteins. Much of the information generated by studies of fruit fly cell biology and development may also apply to mosquitoes, although it will be more difficult to test in *An. gambiae*, since it is not easily manipulated genetically. This study underscores the importance of ongoing functional studies including tissue-specific expression profiling using microarray analysis and qRT-PCR. Understanding how the events following blood feeding are related to each other on a molecular level will provide a more comprehensive picture of this unique behavior and may also delineate new vector-control strategies.
Methods
=======
Microarray chip fabrication
---------------------------
Three cDNA libraries were constructed from abdomens of adult female *An. gambiae*which had been sugar-fed (harvested 30 hours post-eclosion), rat blood-fed (harvested 30 hours PBM), and *P. berghei*-infected rat blood-fed (harvested 30 hours PBM), all at 19°C (Dana, unpublished PhD thesis). Clones from all three libraries were subjected to PCR-based insert amplification using λTriplEx2 vector specific primers (3\' LD Amplimer Primer 5\'-ATACGACTCACTATAGGGCGAATTGGC-3\'; 5\' LD Amplimer Primer: 5\'-CTCGGGAAGCGCGCCATTGTGTTGG-3\'). Amplification reactions contained 1.0 μL eluted phage, 0.03 pmol of each primer, 1 × Taq Polymerase Buffer (Invitrogen), 3 mM MgCl~2~, 1 mM of each dNTP, and 0.2 U Taq Polymerase (Invitrogen), in a total volume of 100 μL. Reactions were conducted in 96-well plates on a Perkin-Elmer 9700 Thermocycler using the following cycling conditions: initial denaturation at 95°C for 5 min, followed by 35 cycles of denaturation at 94°C for 30 s, annealing/elongation at 70°C for 2 min, and a final elongation step at 68°C for 3 min. Samples of all PCR products were electrophoresed on 1% agarose, 1 × TBE gels and visualized by ethidium bromide staining. PCR products were purified on a Beckman Biomek FX using Montage PCR 96 Cleanup kits (Millipore), eluted in 100 μL of water, evaporated overnight and the pellets resuspended in 30 μL of 3 × SSC microarray spotting buffer. A total of 3060 resuspended cDNA inserts and 108 controls were spotted in triplicate on CMT-Gaps II slides (Corning, NY) using the Affymetrix Arrayer 417 at 19 -- 20°C and relative humidity between 50 -- 60%. Slides were post-processed by baking at 80°C for three hours, incubation in 1% SDS for 2 min, in 95°C purified water for a further 2 min, and then plunged 20 times into 100% ethanol kept at -20°C and air-dried via centrifugation at 500 RPM for 5 min.
Microarray target preparation and hybridization
-----------------------------------------------
Total RNA was extracted from blood-fed and sugar-fed whole adult female mosquitoes of the malaria susceptible 4Arr strain, 5--7 days post eclosion, using Trizol (Molecular Research Center, Inc) according to the manufacturer\'s directions. Mosquitoes were blood-fed on anesthetized white rats and maintained under conditions similar to those for sugar-fed mosquitoes, 25°C with 80% humidity and a 12-h light/dark cycle with available 20% sucrose solution, until collection. Fully-engorged mosquitoes were identified visually and harvested at the following 10 time-points: 1) 5 min after initiation of blood feeding during the acquisition of the blood meal (DBM), 2) 30 min DBM, 3) 0 hr post-blood meal (PBM), immediately after they ceased feeding on the rat, 4) 1 hr PBM, 5) 3 hr PBM, 6) 5 hr PBM, 7) 12 hr PBM, 8) 16 hr PBM, 9) 24 hr PBM, and l0) 48 hr PBM. For each blood-fed sample, total RNA was extracted from batches of approximately 10--15 females. Total RNA was also extracted from batches of 100 sugar-fed females for reference samples. RNA samples were then treated with 1.0 μL DNase I (Life Science Technology) according to manufacturer\'s instructions. Following DNase I treatment, total RNA was re-extracted with Trizol. First strand cDNA synthesis and labeling with Cyanine 3 (Cy3) or Cyanine 5 (Cy5), were performed on 15 μg of total RNA from each sample using the Genisphere 3DNA Array 50 kit according to the manufacturer\'s protocol. Hybridizations were conducted following the two step protocol recommended by the manufacturer: 1) cDNA hybridization to the amplified cDNA probes spotted on the slides, 2) hybridization of 3-DNA fluorescent dendrimers (Genisphere) to cDNAs via the capture sequences incorporated into them during first strand synthesis. All cDNA and fluorescent dye hybridizations were performed in a volume of 50 μL using the formamide-based hybridization buffer provided by the manufacturer. The cDNA hybridizations were performed at 45°C overnight. The slides were then washed according to the 3DNA Array 50 kit protocol and air dried by centrifugation for 3 min at 800 RPM. The 3-DNA hybridizations were performed at 53°C for 2 hours as described above, except that 0.5 mM DTT was added to the first two wash solutions to protect the fluorochromes from oxidation. Five replicate slides were generated for each of the ten time points for a total of 50 hybridized and labeled slides. These included two dye-swap experiments performed to eliminate dye fluorescence bias. Pilot experiments conducted with total RNA from the same sample labeled with both Cy3 and Cy5, self-self hybridization, indicated that there was no dye labeling bias following data normalization (data not shown).
Microarray data acquisition and statistical analysis
----------------------------------------------------
Following hybridization and washing, microarray slides were scanned successively at 532 and 635 nm using the Affymetrix 428 Array Scanner. Raw signal intensities were acquired using the adaptive circle algorithm and spot intensities quantified using the Jaguar 2.0 segmentation and data analysis software (Affymetrix, CA). Average signal intensities were normalized using the Loess curve for intensity dependent normalization followed by a per gene median normalization using the Genespring 5.1 software (Silicon Genetics, CA). Signal intensities were filtered such that only gene products exhibiting a raw signal intensity value greater than 300 pixels in one channel and greater than two-fold expression difference between the sugar-fed and blood-fed samples from at least one time point hybridized to the same array were utilized in further analysis. Gene expression level measurements falling outside one standard deviation from the mean signal intensity of each gene product calculated from the five replicates were excluded from further analysis. As an additional quality control, only genes whose PCR amplified products migrated as a single band in agarose gel electrophoresis and that generated high quality sequences for use in EST assembly were analyzed.
Gene products that were induced or repressed at least twofold during blood feeding were initially clustered hierarchically using the Genespring software to determine the user-defined number of centroids (clusters) to be used in *k*-means clustering (data not shown). From this preliminary analysis it was determined that three major clades existed and the genes were clustered using Genespring software using a *k*-means clustering algorithm with a centroid number of 3 and the Pearson Correlation distance metric. Finally, an independent analysis using principal components analysis (PCA) was conducted on the genes induced or repressed at least twofold, using the Genespring software.
qRT-PCR
-------
Transcript levels of several selected genes were measured using SYBR dye technology (Applied Biosystems, CA) and quantitative real-time PCR (qRT-PCR) analysis in order to validate microarray data,. The Primer Express Software v. 1.5 (Applied Biosystems, CA) was used to design the following primers to nine genes: the two Early Genes agCP4871 (AS 648; Forward 5\'-TGATTCGTGCCAGGGTGAT-3\'; Reverse 5\'-CACCACACCAACAAGGACATC-3\') and CG8690 (AS 1786; Forward 5\'-GCTGACTTTGAGCGGTTGG-3\'; Reverse 5\'-CACAAAGTCCATGATCACCTTCA-3\'), the four Middle Genes agCP8064 (AS 679; Forward 5\'-TGGCGAGGTCGATCAGCTA-3\'; Reverse 5\'-CATTATCGCCATCGTTGTGTTG-3\'), agCP12846 (AS 1949; Forward 5\'-TTTGTGGTTCGGTATCGATCTG-3\'; Reverse 5\'-CGAGCACTTTGGCGAACTTC-3\'), CG7758 (AS 1158; Forward 5\'-CACGGTTGGCATTTCGAAC-3\'; Reverse 5\'-GCAGCTGTGCGAACACCA-3\'), and agCP14019 (AS 996; Forward 5\'-GTCGGGCGATTCCAATGA-3\'; Reverse 5\'-TGTAACCGGGCTGGCAAA-3\'), and the two Late Genes agCP14623 (AS 12; Forward 5\'-CGGCAAATCGGTTCAGCT-3\'; Reverse 5\'-TGAATCGGTGCCTTGCG-3\') and agCP2112 (AS 1357; Forward 5\'-CCTGCATGAAGGTGGAATGA-3\'; Reverse 5\'-TTGCCAAGCTCTCCCAACAC-3\'), and the ribosomal protein S7 (RP S7) gene control (Forward 5\'-CATTCTGCCCAAACCGATG-3\'; Reverse 5\'-AACGCGGTCTCTTCTGCTTG-3\'). RP S7 was used as an internal control since its expression is constitutive during blood-feeding \[[@B115]-[@B118]\]. All amplifications and fluorescence quantification were performed using an ABI 7700 Sequence Detection System and associated Sequence Detector Software v. 1.7 (Applied Biosystems, CA). Standard curves were generated using 10-fold serial dilutions of genomic DNA (ranging from 0.0116 to 116 ng per reaction). These qPCR reactions were performed in duplicate in a total volume of 25 μL containing 12.5 μL of SYBR green PCR Master Mix, 300 nmol of each primer, and nuclease free water (Gibco, UltraPURE) using the following conditions; 50°C for 2 min, then denaturation at 95°C for 10 min followed by 45 cycles of denaturation at 95°C for 15 s, annealing and extension at 60°C for 1 min. qRT-PCR reactions for quantification of transcript levels were conducted using 50 ng of first strand cDNA prepared from RNA samples isolated for the microarray analysis. The abundance of each transcript in an RNA sample was estimated from the corresponding gene\'s standard curve and normalized against RP S7 transcript abundance in the same RNA sample.
Authors\' contributions
=======================
AND carried out the cDNA library construction, microarray fabrication, data analysis, and drafted the manuscript. YSH performed the microarray experiment, data analysis, and helped draft the manuscript. MKK performed the qRT-PCR experiments and MEH annotated the ESTs. BWH constructed the microarray genechips and carried out microarray data acquisition. NFL sequenced the cDNA library and JRH maintained and provided mosquito samples throughout the project. PR assisted AND and YSH to draft the manuscript and reviewed it. FHC (P.I.) initiated and supervised the project. All authors read and agreed on the final version of this manuscript.
Supplementary Material
======================
::: {.caption}
###### Additional File 1
Supplementary Table S1 in a Microsoft Excel format where gene annotations for all 413 at least twofold differentially expressed gene products are given.
:::
::: {.caption}
######
Click here for file
:::
Acknowledgements
================
We are grateful to Dr. M. Ferdig of the University of Notre Dame for his constructive discussion throughout the work. We also acknowledge the technical support of Hannah Kim who performed cDNA selections for microarray fabrication. This project was supported by grants U01-AI48846 and R01-AI44273 from NIH/NIAID to F.H.C. Additional support was provided by a grant to P.R. from the Natural Sciences and Engineering Research Council of Canada.
|
PubMed Central
|
2024-06-05T03:55:52.034955
|
2005-1-14
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546002/",
"journal": "BMC Genomics. 2005 Jan 14; 6:5",
"authors": [
{
"first": "Ali N",
"last": "Dana"
},
{
"first": "Young S",
"last": "Hong"
},
{
"first": "Marcia K",
"last": "Kern"
},
{
"first": "Maureen E",
"last": "Hillenmeyer"
},
{
"first": "Brent W",
"last": "Harker"
},
{
"first": "Neil F",
"last": "Lobo"
},
{
"first": "James R",
"last": "Hogan"
},
{
"first": "Patricia",
"last": "Romans"
},
{
"first": "Frank H",
"last": "Collins"
}
]
}
|
PMC546003
|
Background
==========
Phosphorylation and dephosphorylation of proteins are major mechanisms mediating signal transduction throughout the cell and are intimately involved in the regulation of cell growth, physiology, differentiation, and death. Phosphorylation is accomplished by means of kinases which when stimulated by an afferent signal transmit the signal via phosphate transfer to the next site in a pathway. In some cases phosphoprotein-protein interactions take place that modulate signal transduction, e.g. by revealing previously sequestered phosphoacceptor sites in one or both of the interacting proteins, thus creating branch points in pathways. Critical questions exist regarding the identification of the *true in vivo*substrates of kinases, identification of phosphotyrosine interaction domains, and mapping the radiation of these protein interactions throughout extremely complex networks. Clearly new technologies capable of accelerating the processes for defining the interactions between kinases and their substrates and modulators would be of great value.
Two highly productive approaches have been the determination of optimal substrate motifs favored by individual kinases, by various combinatorial peptide library approaches and, the use of antibodies to study phosphorylated peptide motifs (reviewed in \[[@B1],[@B2]\]). Synthetic peptides have played a long and useful role in characterizing kinase substrate sequences, particularly for the ser/thr family, which is now seen to consist of a few distinct category types, basophilic, acidophilic and proline directed. Protein tyrosine kinases, on the other hand, are less well defined by their natural substrates but make more use of docking intermediaries to perform the task of substrate recognition. Nevertheless, optimal substrates have been found which can then aid in the search for the identity of natural or *in vivo*targets and inhibitors of the kinase \[[@B3],[@B4]\]. While capable of assessing mixtures of very large numbers of random peptides, combinatorial methods require deconvolution strategies, which can be time-consuming, and technically demanding.
A second search strategy for functional peptides employs arrays of spatially addressable peptides that can be tested *in situ*, accelerating the deconvolution process when the number of combinations is, or becomes more limited. Peptide arrays capable of displaying diverse functions including kinase substrate activity have been successfully produced by two methods: *in situ*synthesis on planar membranes or arrays of pins \[[@B5]-[@B10]\], or attachment of preformed peptides as performed in a variety of microarray printing procedures. While the existing synthetic methods are capable of producing large numbers of peptides in good purity, none are fully automatable. They require manual intervention between each synthesis cycle and thus are not totally automatable. Peptide synthesis in microtiter plate wells would allow the use of fully automated robotic handlers. Further, arrays of peptides produced in a microtiter plate format, which is an industry standard for numerous types of high throughput analytical procedures, could also be tested in automated multiplex fashion.
We present data here demonstrating the applicability of an automated system for peptide design and synthesis in microtiter plates to the production of peptides and phosphopeptides. We further demonstrate the capability of these peptide arrays to be recognized correctly by specific phospho-motif antisera and to serve as kinase substrates.
Results and discussion
======================
(Phospho)peptide synthesis method
---------------------------------
Previous work led to the development of a system for the automated synthesis of peptide arrays on the inner surfaces microtiter plate wells \[[@B11]-[@B13]\], (Figure [1](#F1){ref-type="fig"}). However, direct characterization of the synthesized peptides did not become realizable until the recent availability of modern high-sensitivity mass spectrometers. In this method polymethylpentene (TPX) microtiter plates are activated by oxidation with nitric acid, and made functional for peptide synthesis by condensation with poly(D-lysine) (n = 100), which serves as a free-floating support or polymeric handle containing an extended array of amino groups. Each lysine side-chain then serves as an initiation point for synthesis. The synthesized peptides are extended from the well surface by their attachment through their peptide carboxyl-termini, on a molecular tether of average estimated length of 50 lysine subunits. This simply assumes that each polylysine molecule is attached to the surface through one bond at its midpoint. It is expected that multiple attachment bonds between the surface and a single polylysine chain can also form but would be minimized by the 2000-fold molar excess of polylysine over attachment capacity used during the polylysine coating step. Model experiments have shown that the polylysine helical structure would be maintained in solution \[[@B14],[@B15]\] for most, if not all cases. If multiple attachment bonds were formed then distal ends would prefer the helical form while the sequences between attachment points should also prefer the helical form up to the limits imposed by torsional constraints, and depending on the closeness of the attachment positions. Thus, the peptides are well positioned spatially, to interact with a variety of macromolecules such as antibodies and structures as large as a cell surface. The lengths and sequences of the peptides are programmable; the total elapsed cycle time to extend each peptide by one residue for all 96 wells is 1 hour. The process has been optimized with respect to activation conditions, length and composition of the polymeric tether and conditions for the handling and storage of the pre-diluted amino acid derivatives and condensing agents. Stabilities exceeding two years have been achieved for all reagents used. The capability of these peptides to be recognized by antibodies, leading to the identification of sequences and structures of the immunoreactive domains of viral proteins and biological response modifiers has been previously been demonstrated \[[@B11]-[@B13]\].
Reproducibility and quality of synthetic peptides
-------------------------------------------------
To evaluate the fidelity and reproducibility of peptide syntheses and ability of the synthetic peptide arrays to serve as specific targets in sequence defined molecular affinity interactions, a model system was chosen to provide known test parameters. The EGF receptor system was selected since it provided commercially available monoclonal antibodies with documented specificity for an activation state associated autophosphorylation site (pY1173) and known sequence (NAEpYLRV). Testing was begun with the monoclonal antibody 9H2 produced against a peptide containing the NAEpYLRV sequence. An array of alternating NAEYLRV and NAE(pY)LRV peptides was prepared in a microtiter plate consisting of 12 8-well strips. Monoclonal 9H2 antibody ELISA was performed using three of the strips from the middle section of the plate. It was found that the antibody reacted strongly with the peptide wells containing phosphorylated tyrosine but not with the non-phosphorylated peptide wells or control wells without peptide (Figure [2](#F2){ref-type="fig"}). For the three strips ELISA means and (standard deviations) for the phosphotyrosine peptides were 3.62(0.20), 3.64(0.25), 3.67(030) and for the tyrosine peptides were 0.071(0.003), 0.068(0.007), 0.076(0.004). For wells containing no peptide, the values for all three strips were 0.063(0.0006). The coefficients of variation for all replicate sets ranged between 5% and 8%.
For evaluation of the fidelity and authenticity of phosphorylated and non-phosphorylated peptide products wells containing NAEYLRV and NAEpYLRV were prepared as before except that a cleavable linker was added to the polylysine matrix before peptide synthesis. The synthesized peptides were then cleaved from the surface with TFA using conditions under which the protecting groups were removed from the amino acid side-chains but not from tyrosine phosphates. When the released peptide products were concentrated and analyzed by MALDI-TOF-TOF mass spectrometry it was found that both peptides yielded essentially monodisperse m/z of the predicted molecular weights, 1036.56 for NAEYLRV and 1170.62 for NAEpYLRV (Figure [3A](#F3){ref-type="fig"} and [3B](#F3){ref-type="fig"}, respectively; \[bis(dimethylamino)phosphono\]-tyrosine species shown). These data demonstrate high coupling efficiency at each step of synthesis and stability of the activated amino acids throughout the process.
Redundancy of an autophosphorylation site antibody epitope in the EGFR cytoplasmic domain
-----------------------------------------------------------------------------------------
Since many of the twenty tyrosines found in the EGFR cytoplasmic domain are known to serve as substrate or SH2 binding site for other tyrosine kinases, it was decided to test the specificity of the 9H2 antibody. To examine this question, overlapping peptide arrays covering the transmembrane and cytoplasmic domains of the EGF receptor were prepared. One array contained only phosphotyrosine and the other only tyrosine. The arrays consisted of 92 peptides, each of which was 21 amino acids in length and overlapped by 15 amino acids. In the EGFR sequence arrayed, there are 20 unique occurrences of tyrosine-containing peptide sequences. In the array, each such sequence appears three to four times in progressively overlapping fashion. Clone 9H2 antibody showed high reactivity (from 9 to 57 times background) with eight out of the twenty phosphotyrosine-containing sequences (Figure [4](#F4){ref-type="fig"}, also \[see [Additional file 1](#S1){ref-type="supplementary-material"}\]). In general, each reactive peak was associated with three to four appearances of the identifiable tyrosine as its position moved progressively along the overlapping peptide sequences, suggesting consistent reliable synthesis throughout the synthetic process. None of the non-phosphorylated tyrosine-containing peptides showed any comparable reactivity with the monoclonal antibody 9H2 although a barely detectable level of antibody reactivity with all peptides containing tyrosine could be seen. There was no detectable reactivity against peptides not containing tyrosine \[see [Additional file 1](#S1){ref-type="supplementary-material"}\]. Steric hindrance by the attachment matrix did not appear to be a significant problem in the recognition of reactive peptide sequences separated by just one, and two amino acids from the peptide carboxyl terminus (\[see [Additional file 1](#S1){ref-type="supplementary-material"}\], array positions 81, 91)
To assess the significance of these primary cross-reactivities a consensus table was constructed from 9H2 antibody reactive and non-reactive sequences (Figure [5](#F5){ref-type="fig"}). A strong preference for hydrophobic amino acids in the Y+1 position is readily apparent with leucine the most preferred appearing in 6 of 8 peptides, followed by isoleucine and valine, both appearing once. At the Y-1 position glutamic acid was the most preferred, appearing in 4 of 8 peptides, followed by arginine, asparagine, aspartic acid, and glutamine with one appearance each. Thus the preference appears to be primarily for glutamic acid but other hydrophilic amino acids were also accepted. There did not appear to be any discernable pattern of preference at the remaining peptide positions. The least common denominator among positive peptides therefore appears to be E/(R, N, D, Q)-p[Y]{.underline}- L/(V, I), or E-pY-L in its predominant form. Peptide sequence \# 84 \[see [Additional file 1](#S1){ref-type="supplementary-material"}\] which lacked an acidic or hydrophilic amino acid at the pY-1 position was still strongly reactive demonstrating the strong contribution of a hydrophobic amino acid in the Y+1 position. However, since the peptide at position Y730 containing A(pY)V was not reactive (\[see [Additional file 1](#S1){ref-type="supplementary-material"}\], Figure [5](#F5){ref-type="fig"}), 9H2 binding appears to involve more than the pY-hydrophobic sequence alone.
Three of the 9H2 cross-reactive phosphotyrosines are clustered within the C-terminal end of the receptor (992--1173), are known autophosphorylation sites and are known to be recognized by proteins containing Group III SH2 binding domains \[[@B16]\] (e.g. p85, phospholipase C~γ1~, the tyrosine phosphatases, Figure [5](#F5){ref-type="fig"}), that similarly recognize phosphotyrosines with hydrophobic amino acids at the Y+1 position. Since some of the 9H2 cross-reactive epitopes are not associated with phosphoacceptor activity it suggests that the phosphoacceptor site specificities are more stringently controlled than the 9H2 epitope or that phosphoacceptor activity simply has not been demonstrated yet. A similarity in the processes for recognition of specificity determinants within the deduced epitope of the 9H2 autophosphorylation site antibody and optimal substrate motif found for the EGF receptor \[[@B16]-[@B18]\], E**EEE[Y]{.underline}F**ELV, may also exist.
Twelve of the twenty tyrosines present in the EGFR cytoplasmic domain were not reactive with 9H2. None of the twelve conformed to the deduced 9H2-epitope motif, although six out of 12 were involved with other aspects of kinase signal transduction (Figure [5](#F5){ref-type="fig"}). As confirmation that this was not a result of failure to incorporate phosphotyrosine in the negative peptides a plate array containing all of the EGFR phosphotyrosine peptides was constructed and tested against a group of commercial phosphotyrosine antibodies prepared in various ways. These results are shown in Figure [6](#F6){ref-type="fig"}. Antibodies 4G10, AB8076, and PT101L were prepared using immunogens which were chemically modified or haptenized and not known to be sequence restricted and all reacted strongly with all of the EGFR phosphopeptides. Antibody RDI-egfract-1 was prepared against the activated EGF receptor and known to be activation specific but not known to be active against linear peptides. These results add support to the specificity of the epitope deduced above and contribute additional useful information and reagents, namely confirmation of the pan-specific nature of the phosphotyrosine antibodies as used in the microtiter plate peptide array system
Tyrosine kinase activity on synthetic peptide substrates in microtiter plate arrays
-----------------------------------------------------------------------------------
To characterize the quantitative aspects of substrate phosphorylation by c-Src kinase, a panel of peptides based on known substrate specificities of c-src and related enzymes were synthesized in microtiter plate wells. Each peptide extended from the polylysine backbone by an ε-amino side-chain and by an additional C-terminal Cys unit. Thus, all of the peptides were equally available to the enzyme, with reduced steric hindrance and a uniform presentation. Subsequently, 90 μL of reaction buffer containing varying amounts of c-Src kinase were incubated for 20 min at 30°, according to the manufacturer\'s instructions. The wells were then washed with distilled water and assayed for the presence of phosphotyrosine by ELISA using a mixture of the broadly cross-reacting phosphotyrosine antibodies previously described in Figure [6](#F6){ref-type="fig"}. The peptide substrates **E**E**IYGEF**F \[[@B17]\] (Src,1) and YIYGSFK \[[@B19]\] (Src,2) have been shown separately by somewhat different combinatorial methods to have relatively potent activity for protein tyrosine kinases and are shown to be reactive here as well (Figure [7](#F7){ref-type="fig"}). (Src, 1) was phosphorylated to a greater extent than (Src, 2) and showed no decrease of reactivity even at 0.1 Unit of enzyme, the lowest concentration used. The (Src, 1) variant peptides were chosen so that validation could be made by direct comparison with the highly oriented peptide chip system recently described \[[@B20]\]. The (Src, 1) variant peptide (-E) made by truncation of the N-terminus still showed good reactivity, although much lower than the longer (Src,1) peptide, and the (-Y) peptide in which tyrosine was exchanged for phenylalanine showed no reactivity as expected and required. The slight c-ABL tyrosine kinase substrate peptide IYAAPKKK \[[@B17]\] reactivity at the highest c-Src input and negative protein kinase A substrate peptide LRRASLGC \[[@B21],[@B22]\] activity are consistent with the level of cross-reactivities expected between familial and nonfamilial kinases. Three of the peptide substrates showed dose-response characteristics conforming, by nonlinear regression, to a one site binding model for (-E), (Src, 2), and (ABL), with R^2^equal to 0.98, 0.99, and 0.92. Activity of the most reactive peptide, (Src, 1) was greater than expected based on quantity of enzyme used in previously published work \[[@B23],[@B24]\] and was at least 30 times more reactive than its truncated form (-E) \[[@B20]\]. By extrapolation, c-Src activity would be detectable at concentrations as low as 0.01 Units of enzyme using the (Src, 1) substrate.
Conclusions
===========
The method for production of synthetic peptide solid-phase arrays in microtiter plates described here is capable of making high quality peptides, as seen by mass spectrometry of the released unfractionated products. The peptide synthesis method is completely automated and has been greatly simplified by the use of standard automated liquid handlers and the use of activated amino acid solutions that may be prepared and stored in advance and added just once at the beginning of the synthetic process. In the present demonstration, 96 well microtiter plates were used; but, 384, 1536, or containers of any well density compatible with the solvents and liquid handler may be used. The solid phase peptide arrays produced on a polylysine backbone were found to be of very high density, provide very low levels of nonspecific binding and steric hindrance, and participate effectively in a variety of biochemical reactions.
The strategy described here for the preparation of solid phase synthetic peptide arrays in microtiter plate wells for use in multiplexed assays offers many advantages in the study of protein kinases, particularly in a research environment. In a research environment combined cycles of hypothesis generation and testing, with assay flexibility, speed, quantitative accuracy and precision are of greater concern than in large scale screening applications of large numbers using limited, previously selected variables. The industry standard microtiter plate format ensures compatibility with a vast number of assay platforms and the polylysine backbone with its extended three-dimensional display provides a highly efficient, sterically unhindered, and extremely low background display of the peptide products.
Using 96-member peptide arrays of 21-mers created in less than 24 hours, we have shown that the peptide array synthesis provided a highly reproducible model for a tyrosine peptide, EGFR Y1173 and its phosphorylated counterpart. Using a monoclonal antibody prepared against a synthetic peptide representation of the Y1173 EGF receptor autophosphorylation site, we have provided evidence that, unexpectedly, the deduced epitope, E/H~L~-pY-L/H~B~(where H~L~- is hydrophilic and H~B~is hydrophobic) is highly redundant within the cytoplasmic domain. Three of the eight antibody reactive sites have been previously identified as autophosphorylation sites and are recognized by Group III SH2 domain proteins (Figure [5](#F5){ref-type="fig"}) that have similar specificity patterns. Furthermore, the EGFR substrate sequence EEEEYFELV, derived by combinatorial peptide optimization \[[@B17]\], resembles the E-Y-(hydrophobic) motif found in these studies. Thus there is a consistent linkage between a subset of EGFR phosphotyrosine sequences recognized by the 9H2 antibody, a subset of sequences autophosphorylated by EGFR kinase, and EGFR autophosphorylation sites recognized by the Group III SH2 domain. There is a clear parallel between the 9H2 peptide epitope and the peptide substrate specificity of the EGFR catalytic activity \[[@B16]\]. Songyang has further suggested that the catalytic and SH2 domains of PTKs may have converged to recognize similar sequences. So, questions regarding which site (or sites) the antibody actually recognizes on stimulated EGF receptor molecules and what other parallels might exist between 9H2 binding, SH2 binding, and catalytic substrate selection become of interest. At the cell protein level, 9H2 is specific by Western blot for the stimulated EGF receptor. Furthermore, there are preliminary data (Saxinger, unpublished) suggesting that PDGF peptide sequences conforming to the 9H2 binding sequence of (hydrophilic)-pY-(hydrophobic) deduced from EGFR phosphotyrosine peptides, appear to be recognized differently by 9H2. While nine of the twenty-seven PDGFR phosphotyrosine sequences satisfied the (hydrophilic)-pY-(hydrophobic) sequence definition, and could be expected to be as reactive as those in EGFR, only one showed comparable reactivity. Thus, the binding determinants of 9H2 are more complex than the simple epitope deduced from Figure [5](#F5){ref-type="fig"}. An intriguing possibility is that the 9H2 epitope may be a fairly simple one but that its appearance, or access to it has been limited or distorted in specific ways by spatially adjacent sequences or structures that have evolved to create opportunities for exploitation in biologically specific processes.
We have also successfully demonstrated the use of microtiter plate peptide arrays in faithfully reproducing the known substrate phosphorylation specificities of c-Src protein kinase. In these studies a broadly reactive phosphotyrosine antibody ELISA detected phosphorylation. Although a mixture of antibodies that were not known to be sequence restricted was used, the possibility exists that some tyrosine-containing peptides could become phosphorylated and be recognized less well than others. Therefore in studies where the need for precise quantitation outweighs the convenience and safety considerations of ELISA, incorporation of radioisotopic phosphate would provide an alternative. The microtiter plate format with reactants bound to the well surface would provide a well-contained and safe vehicle for washing and subsequent measurement of radioactivity.
Thus, the microtiter plate array system is well suited to the study of protein kinase substrates, antigens, related binding molecules, and inhibitors since these all can be quantitatively studied at a single uniform, reproducible interface. For applications requiring larger numbers of solid phase peptides, the synthetic process can easily be transferred to more powerful workstations such as the Biomek FX platform in which many plates with higher well densities can be synthesized simultaneously and conventional particle based substrates for peptide synthesis can be manipulated using filter plates or magnetic devices available for this system. Moreover, the current capacity of approximately 150--380 pMoles/well can be considered large enough for preparative or analytical applications coupled with mass spectrometric analyses, such as affinity-based proteomic screening for ligand-protein interactions. In this and other applications, such as assessment of protease activity where strict isolation of adjacent components is required, microtiter plate wells enjoy a significant mechanical advantage over two-dimensional spotting or other synthesis methods.
Methods
=======
Materials
---------
Solvents and reagents for peptide synthesis were obtained as synthesizer grade from Applied Biosystems (Foster City, CA). Specifically, these were N-methylpyrollidone (NMP), 1 M N-Hydroxybenzotriazole (HOBT) in NMP, 1 M dicyclohexylcarbodiimide (DCC) in NMP, diisopropylethylamine (DIPEA), acetic anhydride, and trifluoroacetic acid (TFA). NMP from this vendor was consistently free of basic impurities and was stored over Molecular Sieve (4A) after opening. NMP solutions were stored at -20C and allowed to attain room temperature before opening. FMOC-protected α-amino acids were purchased from Penninsula Laboratories (San Carlos, CA) and side-chain substitutions were, Asn(Trt), Asp(OtBu), Cys(Acm) or Cys(Trt), Gln(Trt), Glu(OtBu), His(Trt), Lys(Boc), Ser(tBu), Thr(tBu) and Tyr(tBu). FMOC-Arg(Pbf), FMOC- \[bis(dimethylamino)phosphono))-tyrosine\] \[[@B25]\], and FMOC Rink amide (linker p- \[(R,S)-a-\[1-(9H-Fluoren-9-yl)-methoxyformamido\]-2,4-dimethoxybenzyl\]-phenoxyacetic acid) were purchased from NovaBiochem. Phenol, thioanisole, ethanedithiol (EDT), triisopropylsilane (TIPS), and carbonyldiimidazole (CDI) were obtained from Aldrich Chemical Co. (Milwaukee, WI). Poly(D-Lys•HBr) (dp 100) and Poly(L-Lys•HBr) (dp 100) were obtained from Sigma-Aldrich (St. Louis, MO). TPX (polymethylpentene) 8-well strips, non-sterile, non-tissue culture treated, were obtained from Costar on special order (Cambridge, MA).
(Phospho)Peptide synthesis
--------------------------
Automated preparation of solid phase synthetic peptide arrays in microtiter plates was performed as described previously \[[@B13]\] (Saxinger, US Patent 6031074, Feb. 29, 2000). Carboxyl functional groups are formed on the hydrocarbon surface of Costar TPX microtiter plate strip wells by oxidation with 70% nitric acid for two hours at 65°C or for two weeks at room temperature. Polylysine chains are then attached to the surface by condensation using 0.05 M carbonyldiimidazole in NMP for 30\' at 20°C. Poly(L-Lys•Hbr) (PLL) or poly(D-Lys•Hbr) (PDL) (1 mg/ml in 90% NMP-10% water) was neutralized with DIPEA and reacted with the CDI-activated surface for 1 hour at 20°C and followed by 4°C overnight. Plates were rinsed twice with water and twice with methanol, air-dried, wrapped in plastic cling-wrap and stored at room temperature. Peptide synthesis next takes place by the sequential addition of Nα-FMOC amino acids activated with DCC/HOBT as in conventional peptide synthesis with the growing peptide chain covalently attached to the polylysine chain through its carboxyl terminus. Stock solutions of FMOC amino acids were prepared in advance by dissolving 6 mmoles of each in 12 mL of NMP containing a 10% molar excess of HOBT, and stored at -20°C. The solutions could then be used repetitively for at least three months. Prior to peptide synthesis, the required amounts of amino acid were diluted to 0.1 M in NMP, mixed with an equimolar volume of 0.1 M DCC in NMP and allowed to react for thirty min in 2 mL cryovials. The 20 mixtures were then placed in a 24 tube rack on the Biomek 1000 workstation tablet for use during subsequent peptide synthesis cycles. Synthesis automation is achieved through software that receives input designating the peptide desired in each of the 96 microtiter plate wells and creates output files to indicate the quantity of reagents needed and a set of Beckman Biomek 1000 arrays.bio files, each member of the set directing the distribution of amino acids in one of the sequential chain extension cycles. An automated repetitive set of reagent transfers and washing steps is then recycled using a new arrays.bio file for each cyle until the peptide synthesis is complete. After synthesis the peptides are usually capped by reaction with acetic anhydride (1 mL in 10 mL of NMP containing 0.1 mL of DIPEA per plate) before side-chain deprotection and testing, in multiplex fashion, for reactivity. Side chain protecting groups were generally removed with: 10 mL of TFA +0.75 g of crystalline phenol + 0.5 mL of purified water + 0.5 mL of thioanisole +0.25 mL of ethanedithiol per plate, incubated for three hours in a sealed container, washed with ether, air-dried in a chemical fume hood, and stored at -20°C in plastic wraps. For peptides not containing Met, Trp, or Cys, TFA containing TIPS and H~2~O was used (95%, 2.5%, and 2.5%, respectively).
Peptide synthesis capacity of microtiter plate wells
----------------------------------------------------
TPX microtiter plates were oxidized at either room temperature or 60°C and reacted with polylysine. Peptide quantities were estimated by two different methods. In the first, the polylysine amino groups were modified with the cleavable linker HMPB \[4-(4-Hydroxymethyl-3-methoxyphenoxy)-butyric acid\] and FMOC-L-Cys(Acm) was coupled to the HMPB-substituted support as described above. Terminal amino groups were deprotected with 20% piperidine in NMP, washed with NMP and derivatized with dabsyl chloride. The dabsyl-amino acid derivative was released with 95% TFA, harvested from multiple wells by serial transfer, and measured by spectrophotometric scanning in a Beckman DU65 using an extinction coefficient of Dabsyl Chloride = 6.5 × 10^4^in TFA, λ~max~= 495). In the second method, FMOC-L-Ala was coupled directly to the polylysine supports. The FMOC protecting group was released in 20% piperidine, harvested from multiple wells by serial transfer, and quantitated by fluorimetric scanning in a Perkin Elmer LS50B using E(301) = 7800 (1 mM/ml). Values for spectral constants and coefficients were determined from standards dissolved in the cleavage solvent. The solid phase capacity was 380 pMoles/well and the releasable capacity was 150 pMoles/well.
Mass Spectrometry of synthetic array peptides
---------------------------------------------
A polylysine coated TPX microtiter plate was modified by preliminary reaction with FMOC-Rink Amide linker (Novabiochem) to allow peptide cleavage after synthesis using standard DCC/HOBT coupling conditions, as described above. All peptides were initiated with C-terminal Cys(Acm) to provide a constant initiation and cleavage environment. Peptides were released and deprotected by incubation for 3 hours at room temperature in 100 μL of TFA containing water, ethanedithiol and phenol, or TFA containing water and triisopropylsilane. An additional 10 uL of water and 16 hours of incubation time are required for deprotection of \[bis(dimethylamino)phosphono))-tyrosine\] \[[@B25]\]. Solutions of released peptides were concentrated by vacuum centrifugation and 0.25 μL of sample was co-crystallized with 0.25 μL of α-cyano-4-hydroxycinnamic acid in 50% ACN, 1% trifluoroacetic acid and spotted directly on a stainless steel matrix-assisted laser desorption ionization (MALDI) plate. Mass spectra were acquired using an Applied Biosystems 4700 MALDI-TOF/TOF mass spectrometer (Applied Biosystems, Foster City, CA). MALDI mass spectra were externally calibrated (\<20 ppm) using a standard peptide mixture.
Antibody recognition of microtiter plate arrayed peptides
---------------------------------------------------------
Goat anti-mouse IgG (phosphatase conjugated) was purchased from Kirkegaard & Perry (Gaithersburg, MD) and Upstate Biotechnology (Lake Placid, NY). Murine monoclonal antibody 9H2, purchased from Upstate Biotechnology was prepared against a synthetic peptide containing the EGFR autophosphorylation sequence at Y1173 and was certified by the manufacturer to be specific for EGF-stimulated A431 cells by Western blot. Murine antisera prepared against phosphotyrosine were 4G10 (Upstate), ab8076 (Abcam, Ltd.), and PT01L (Oncogene Research Products). Antiserum RDI-egfract-1 was prepared against isolated tyrosine phosphorylated (activated) EGF receptor from EGF-challenged murine L-cells (Research Diagnostics, Inc.). All sera were used according to the manufacturer\'s recommendations and assayed by indirect ELISA. The extent of reaction was determined using a phosphatase assay kit from Kirkegaard & Perry.
Kinase assay
------------
Peptide array substrate evaluations were performed using p60^c-src^protein-tyrosine kinase, Cat\# PK03 (Oncogene Research Products) according to the instructions in the manufacturer\'s insert. In sequence, all wells received 30 μL of Kinase Assay Buffer (0.05 M HEPES, pH 7.5 + 0.1 mM EDTA + 0.015% BRIJ 35), 30 μL of appropriately diluted p60^c-src^in Kinase Dilution Buffer (0.1 mg/mL BSA + 0.2% β-mercaptoethanol), 30 μL of ATP mix (0.03 M MgCl~2~+0.15 mM ATP) in Kinase Assay Buffer. Plates were incubated at 30°C for 30 min, rinsed with distilled water and assayed for the presence of phosphotyrosine as described above. Each unit of p60^c-src^enzyme catalyzes the incorporation of one pMole of phosphate into tyrosyl residues.
List of abbreviations used
==========================
(9H2): antibody prepared against a synthetic peptide containing the EGFR autophosphorylation site at Y1173
carbonyldiimidazole (CDI)
dicyclohexylcarbodiimide (DCC)
diisopropylethylamine (DIPEA)
EGF receptor (EGFR)
ethanedithiol (EDT),
N-Hydroxybenzotriazole (HOBT)
N-methylpyrollidone (NMP)
phosphotyrosine kinase (PTK)
Polymethylpentene (TPX)
trifluoroacetic acid (TFA)
triisopropylsilane (TIPS),
Amino acid and polypeptide abbreviations were in accordance with IUPAC-IUB recommendations.
Authors\' contributions
=======================
WCS designed the strategy for peptide synthesis, participated in the design of the strategy for peptide synthesis validation, carried out the peptide synthesis, testing, functional analyses, and drafted the manuscript. TPC participated in the design of the strategy for peptide synthesis validation, carried out the mass spectrometry analyses and interpreted the ms results. DJG participated in the design of the strategy for peptide synthesis validation. TDV participated in the design of the strategy for peptide synthesis validation and coordinated the mass spectrometry analyses.
Supplementary Material
======================
::: {.caption}
###### Additional File 1
EGRF (pY) peptide scanning array Vs 9H2 Antibody ELISA. EGF receptor phosphotyrosine and tyrosine overlapping peptide array sequences and ELISA test results for each peptide array.
:::
::: {.caption}
######
Click here for file
:::
Acknowledgements
================
We acknowledge Dr. Robert Wiltrout, NCI, for encouragement and support during this work and Dr. Stephen Shaw, NCI, for encouragement and critique of the manuscript. This work was supported in part by Federal funds from the National Cancer Institute, National Institutes of Health, under Contract NO1-CO-12400 (T.P.C. and T.D.V.). By acceptance of this article, the publisher or recipient acknowledges the right of the United States Government to retain a nonexclusive, royalty-free license and to any copyright covering the article. The content of this publication does not necessarily reflect the views or policies of the Department of Health and Human Services, nor does mention of trade names, commercial products, or organization imply endorsement by the United States Government.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Microtiter plate peptide array chemistry**Costar TPX microtiter plate strip well surfaces were oxidized with nitric acid, coated with polylysine and peptide synthesis is performed as described in Methods. Attachment of the polylysine chains to the wells and initiation of peptide synthesis takes place through the lysine ε-amino groups. The predicted polylysine helical segment shown is stabilized primarily through backbone H-bonding and synthesized peptides are depicted by red spheres.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Reproducibility and specificity of multiplexed EGFR peptide synthesis and ELISA testing**Peptides were synthesized as microtiter plate array strips using either tyrosine or phosphotyrosine. Three replicate strips were used, each with no peptide in row A and alternating phosphotyrosine and tyrosine peptides in rows B to H. The strips were tested for reactivity with clone 9H2 antibody as described in Methods. ELISA results (x-axis) are plotted against tyrosine or phosphotyrosine replicates for each strip (y-axis). Standard deviations varied between 5% and 10% of the means. Data are displayed without any correction.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Mass spectrometric characterization of EGFR phosphopeptides produced in microtiter plate arrays**Panel A illustrates the mass spectrum of the NAEYLRV derivative (\[M+H\]^1+^ion at *m/z*1037.56) and panel B for the NAEpYLRV derivative (\[M+H\]^1+^ion at *m/z*1170.62). Peptides were released by incubation for 3 hours in TFA-H~2~O-triisopropylsilane, concentrated *in vacuo*and reconstituted for MALDI-TOF-TOF analysis. Under these conditions the bis(dimethylamino) phosphate protecting groups of phosphotyrosine were preserved and mass spectra were improved.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Distribution of 9H2 phosphotyrosine epitopes in EGFR**Overlapping peptide arrays were made and acetylated in microtiter plates using human EGF receptor residues 622--1186 (Swissprot database locus P0053 with leader sequence removed). Each array contained exclusively tyrosine or phosphotyrosine peptides and was composed of 92 peptides each of which was 21 amino acids long and offset from its neighbor by 6 amino acids. Array peptide wells were reacted in multiplexed ELISA format with antibody 9H2 as described in Methods. Data are plotted without any corrections. Peptide sequences and raw data are listed \[see [Additional file 1](#S1){ref-type="supplementary-material"}\].
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Pattern of recognition of arrayed EGFR peptides by autophosphorylation site antibody (9H2)**Tyrosine sequence positions correspond to human EGF receptor residue positions (Swissprot P0053, leader sequence removed). Amino acids are color coded for acidic, basic, hydrophilic, and hydrophobic as red, blue, green and brown. \*. For (phospho)peptide array compositions and quantitative results \[see [Additional file 1](#S1){ref-type="supplementary-material"}\]. Footnotes: 1. autophosphorylation site \[26-28\], 2. PLC-gamma SH2 domain binding\[29\], 3. GRB2/SH2 domain binding \[30\], 4. Shc, SHP1, PLC-gamma SH2 domain binding \[31, 32\], 5. Src phosphorylation site \[33-35\],6. AP2 \[36, 37\], 7. Cbl, SH2 domain binding \[38-40\], 8. Shc SH2 domain binding \[31, 32, 41\].
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
**Reactivity of all EGFR phosphotyrosine peptides with a panel of pan-specific phosphotyrosine monoclonal antibodies**Phosphotyrosine peptide sequences selected from the arrays used in Figure 4 \[see [Additional file 1](#S1){ref-type="supplementary-material"}\] were synthesized in a new microtiter plate array and are identified on the y-axis. Each peptide and an empty control well were tested with each of the four antisera identified in the Figure legend, 4G10, ab8076, PT01L, and RDI-egfract-1 by standard ELISA as described in Methods. Data are plotted without any corrections
:::
:::
::: {#F7 .fig}
Figure 7
::: {.caption}
######
**Microtiter plate array peptides serve as substrates for p60^c-src^with specificity and concentration dependence**Eight-well strips bearing six substrate peptide wells and two control wells were robotically synthesized *ab initio*, in microtiter plate format. All of the peptide substrates were initiated with Cys(Acm) to provide additional extension from the polylysine backbone and a common attachment site. The Cys-SH protecting group (Acm) remains in place and all peptides were acetylated at the amino terminus. Duplicate substrate strips were reacted with each 90-μL dilution of c-Src enzyme containing 3, 1, 0.3, and 0.1 Units of enzyme. The wells were reacted for 20\' at 30°, washed with distilled water and assayed by antibody ELISA for the presence of phosphotyrosine using a mixture of antibodies described in Figure 5. Nonlinear regression plots were computed in Prism and error bars represent the means ± 1 SD, which ranged between 3% and 7% of mean values. Iterations for (ABL) did not converge. All data points were corrected by subtraction of values obtained from wells with no peptide.
:::
:::
|
PubMed Central
|
2024-06-05T03:55:52.043097
|
2005-1-12
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546003/",
"journal": "BMC Immunol. 2005 Jan 12; 6:1",
"authors": [
{
"first": "Carl",
"last": "Saxinger"
},
{
"first": "Thomas P",
"last": "Conrads"
},
{
"first": "David J",
"last": "Goldstein"
},
{
"first": "Timothy D",
"last": "Veenstra"
}
]
}
|
PMC546004
|
Background
==========
Cells have evolved surveillance mechanisms to ensure the fidelity of gene expression. One such mechanism, nonsense-mediated mRNA decay (NMD), was discovered about twenty years ago in yeast \[[@B1]\] and then described in human inherited diseases caused by nonsense or frameshift mutations \[[@B2]-[@B4]\], which introduce premature termination codons (PTCs). Contrary to what would be predicted based on the nature of the mutation (a premature translational arrest), the resulting nonsense mRNAs rarely code for truncated protein products and are rather rapidly degraded \[[@B5]\]. Hence, NMD was first envisaged as a mean to protect cells against the effects of deleterious truncated proteins, with potential dominant-negative effects or a gain of function. Moreover, it seems that NMD has not solely evolved under the pressure of nonsense mRNAs originating from mutations but it also monitors PTC-containing transcripts arising from abnormalities in gene expression \[[@B6]\]. NMD plays a role in normal cellular development as examplified by the production of functional TcR and Ig genes. During lymphocyte maturation, these genes are subjected to extensive rearrangements and somatic mutation events. Approximately two-thirds of the rearranged genes are not in the proper translational reading frame and the resulting transcripts are down-regulated by NMD, ensuring that only functional TcR and Ig genes are expressed \[[@B7],[@B8]\]. More recently, it was suggested that NMD could play a role in the regulation of gene expression. This was first suspected following the identification of unproductive splicing products of the SRp20 and SRp30b proteins or the ribosomal proteins L3, L7a, L10a and L12 in *Caenorhabditis elegans*\[[@B9],[@B10]\]. The coupling of alternative splicing, which generates transcripts containing PTCs and NMD, which degrades these transcripts, enables the negative regulation of gene expression \[[@B9],[@B10]\]. This system, termed RUST (for Regulated Unproductive Splicing and Translation, \[[@B11]\]) is also used in humans \[[@B12]-[@B14]\]. It was suggested that alternative splicing leading to NMD might prove to be a common mechanism of autoregulation of many splicing factors \[[@B12],[@B15],[@B16]\]. For example, the polypyrimidine tract binding protein (PTB), which generally acts as a splicing repressor, downregulates its own expression by repressing exon 11 inclusion in the mature mRNA \[[@B12]\]. The resulting alternative PTB mRNA lacking exon 11 contains a PTC and is subjected to NMD \[[@B12]\]. This negative autoregulation prevents the accumulation of PTB, and therefore the inappropriate processing of its targets \[[@B12]\]. This coupling of alternative splicing and NMD seems to be a rather common mechanism, as *in silico*analyses show that 35% of EST-suggested alternative transcripts contain PTCs \[[@B17]\]. Alternative splicing is thought to occur in 30--60% of human genes \[[@B18]\], and in addition to expanding proteome diversity, it may play a role in gene expression regulation by generating PTC-containing alternative isoforms. Interestingly, 10 to 30% of nonsense transcripts can escape NMD and are further immune to degradation \[[@B19]\]. Whether such transcripts code for proteins with a physiological function is unknown.
Fumarylacetoacetate hydrolase (FAH, E.C. 3.7.1.2) is the last enzyme of the tyrosine catabolic pathway. A deficiency in FAH causes hereditary tyrosinemia type I (HTI; OMIM 276700), the most severe disease of the pathway \[[@B20]\]. This inherited metabolic disorder is characterized by severe hepatic and renal dysfunctions often resulting in death in the first years of life if untreated. HTI displays phenotypic heterogeneity with both chronic and acute forms \[[@B20],[@B21]\]. The *fah*coding gene located on chromosome 15 in the q23-q25 region \[[@B22]\] spans over 35 kb and contains 14 exons \[[@B23]\]. Forty-seven mutations have been identified so far in the *fah*gene, including 7 nonsense mutations \[[@B24]-[@B26]\]. While characterizing the effects of the W262X nonsense mutation on FAH mRNA metabolism, we identified two alternative transcripts, del100 and del231 in a HTI patient homozygous for W262X. These transcripts are found in normal cells and thus are not due to the nonsense mutation *per se*. Interestingly, del100 has skipped exon 8 and as a consequence, the reading frame is shifted, with the appearance of several new PTCs. This transcript is therefore likely subjected to NMD, as suggested by a block of translation by cycloheximide. However, the amount of nonsense transcript which escapes NMD seems to be sufficient to produce a protein of 31-kDa, detected in several human tissues. This report suggests that NMD may allow for the production of low amounts of protein.
Results
=======
Identification of two alternative transcripts of the *fah*gene
--------------------------------------------------------------
The W262X mutation is a G-\>A transversion located in exon 9 of the *fah*gene at nucleotide position 786 and is frequent in the Finnish population \[[@B27]\]. No protein was detected in the liver or the fibroblasts of homozygous patients \[[@B28]\]. Consistent with this type of mutation, we demonstrated that W262X mRNAs are degraded by NMD in the cytoplasm \[[@B29]\]. While studying the decay of W262X nonsense FAH transcripts in lymphoblastoid cell lines, we repeatedly observed two additional RT-PCR products in homozygous W262X/W262X cells (Figure [1A](#F1){ref-type="fig"}). Purification and sequence analysis of these two products revealed that they were alternative transcripts of the *fah*gene. The first one, del100, lacks exon 8 (Figure [1A](#F1){ref-type="fig"}) and the second one, del231, which is less abundant, lacks both exons 8 and 9 (Figure [1A](#F1){ref-type="fig"}). In del100, the skipping of exon 8 from the mature transcript causes a shift in the reading frame and as a consequence, several new PTCs appear, the first one being located at the 3\' end of exon 10, at the new amino acid position 270 (Figure [1A](#F1){ref-type="fig"}). The G786A mutation does not code for a stop codon in del100 but causes the replacement of a glycine by a glutamate residue (Figure [1A](#F1){ref-type="fig"}). Del231 does not show further disruption of the open reading frame downstream of the deletion and is predicted to code for a shorter FAH-like protein (about 34-kDa) missing the region encoded by exons 8 and 9 (Figure [1A](#F1){ref-type="fig"}).
Del100 and del231 were first identified in the liver of a patient harboring another mutation (Q279R), which weakens the donor splice site of exon 9 \[[@B30]\]. Because the W262X mutation is located in the same exon and that in some cases a nonsense codon can affect splicing \[[@B31]\], we wondered whether these two transcripts were due to an effect of the W262X mutation on a cis-acting splicing element in exon 9. To test this hypothesis, RT-PCRs were performed using primers spanning the exon7-exon9 junction or the exon7-exon10 junction to specifically amplify del100 and del231 respectively. As shown in Figure [1B](#F1){ref-type="fig"}, both del100 and del231 transcripts were detected by this method, in homozygous mutant cells (W262X/W262X; Figure [1B](#F1){ref-type="fig"}, middle panel), as well as in normal cells (wt/wt; Figure [1B](#F1){ref-type="fig"}, middle panel). The identity of these amplification products was verified by sequencing (data not shown). A similar result was found in various human cell lines (Figure [1B](#F1){ref-type="fig"}, right panel). Indeed, del100 and del231 were amplified in fibroblasts, HeLa cells (Figure [1B](#F1){ref-type="fig"}, right panel) and in human liver (Figure [1B](#F1){ref-type="fig"}, right panel), the tissue where FAH is the most expressed \[[@B32]\]. Moreover, del100 was amplified in two HTI cell lines (Figure [1B](#F1){ref-type="fig"}), which harbor either a splice mutation in intron 12 (IVS12/IVS12; \[[@B33]\]) or two nonsense mutations in exon 13 (E357X/E364X; \[[@B33]\]). Del231 was not detected in these two cell lines as expected, since these mutations introduce PTCs either following exon 12 skipping (IVS12/IVS12) or due to the two nonsense mutations themselves (E357X and E364X), that likely target the nonsense transcripts to the NMD pathway.
Altogether, these data strongly argue in favor of del100 and del231 being minor alternative transcripts of the *fah*gene, rather than resulting from the presence of the W262X mutation.
The del100 transcript is translated into a protein
--------------------------------------------------
The identification of two minor alternative transcripts of the *fah*gene raised the question whether they resulted from errors of the splicing apparatus and were unproductive alternative transcripts or whether they could produce protein products with potential physiological roles. There is presently no reported indication for the existence of additional FAH isoforms. The DEL231 open reading frame is identical to FAH, except for the missing region encoded by the skipped exons 8 and 9 and corresponding to amino acids 203 to 280 (Figure [2A](#F2){ref-type="fig"}). The open reading frame of the del100 transcript is identical to FAH and del231 from the ATG start codon to amino acid 202. However, the last 67 amino acids of the putative DEL100 protein encoded by exons 9 and 10 are completely different, due to the shift in the reading frame following exon 8 skipping (Figure [2A](#F2){ref-type="fig"}). To find out if del100 was translated into a protein, we raised an antiserum against the last 67 amino acids of the putative DEL100 protein and used it to search for the presence of this protein in different adult human tissues. These tissues were obtained after an autopsy and only one sample per tissue was tested, due to the difficulty to obtain them. As shown in Figure [2B](#F2){ref-type="fig"} (middle panel), a cross-reacting band was present in heart, liver, kidney, spleen, suprarenals and bladder. The DEL100 protein has an apparent molecular weight of 31-kDa, consistent with the value of 29.7-kDa calculated from its sequence. Interestingly, the expression level of the DEL100 protein varied between the different tissues and it differed from that of FAH (Figure [2B](#F2){ref-type="fig"}, top panel). For example, FAH was barely detected in the spleen, whereas the expression level of the DEL100 protein was the highest in this tissue (Figure [2B](#F2){ref-type="fig"}). A monoclonal antibody directed against the N-terminal part of the FAH protein was used to detect both FAH and DEL100 in the tissues where FAH is the less expressed (Figure [2C](#F2){ref-type="fig"}). DEL100 was barely detected in the spleen, suggesting that the protein is synthezised in very low amounts.
The specificity of the signal was verified by adsorbing the antiserum on the purified C-terminus of DEL100, used for the mouse immunization. As shown in Figure [3A](#F3){ref-type="fig"} (top panel), the affinity-purified antiserum still recognized the 31-kDa protein, whereas the non-adsorbed fraction did not show any cross-reactivity (Figure [3A](#F3){ref-type="fig"}, bottom panel). The protein of low molecular weight, which is recognized by the non-adsorbed antiserum in the spleen, is unspecific background (Figure [2B](#F2){ref-type="fig"}, middle panel and Figure [3A](#F3){ref-type="fig"}, bottom panel; indicated by a star). A DEL100 protein with the Myc tag was synthesized in an *in vitro*transcription-translation assay and immunoprecipitated using an anti-Myc antibody. The anti-DEL100 antiserum recognizes the immunoprecipitated DEL100-Myc protein, further demonstrating its specificity (Figure [3B](#F3){ref-type="fig"}). Altogether these results suggest that the del100 alternative transcript is translated into a protein of 31-kDa whose expression in different tissues differs from that of FAH.
The del100 transcript seems to be subjected to NMD
--------------------------------------------------
The skipping of exon 8 in del100 causes a change in the reading frame and as a consequence, several new stop codons appear (different from the W262X mutation). To verify if the nonsense del100 transcript was subjected to NMD, lymphoblastoid cells were treated with cycloheximide (Figure [4](#F4){ref-type="fig"}) an inhibitor of translation. Stabilization of nonsense transcripts following such a treatment suggests that they are degraded through the NMD pathway \[[@B34]\]. The effectiveness of the treatment was previously verified on the full-length W262X containing transcript \[[@B29]\] and an example is given in Figure [4A](#F4){ref-type="fig"} (upper panel). The full-length transcript was up-regulated in the homozygous cell line (Figure [4A](#F4){ref-type="fig"}; W262X/W262X) but remains unaffected in wild-type cells (Figure [4A](#F4){ref-type="fig"}; wt/wt). The same treatment was used to determine the fate of the del100 transcript (Figure [4A](#F4){ref-type="fig"} and [4B](#F4){ref-type="fig"}). A stabilization of this alternative transcript was observed when the FAH transcripts were amplified from exons 6 to 14 in wild-type and homozygous cells (indicated by a star in Figure [4A](#F4){ref-type="fig"}). The same result was obtained using a specific amplification of the del100 transcript (Figure [4A](#F4){ref-type="fig"}, second panel and Figure [4B](#F4){ref-type="fig"}). The amount of del100 following cycloheximide treatment increased about 5-fold the level observed in untreated cells (Figure [4B](#F4){ref-type="fig"}). These results suggested that del100 is indeed subjected to NMD in homozygous cells and in normal cells as well. In contrast del231, which does not contain any PTC as a result of the skipping of both exons 8 and 9, seemed relatively unaffected by the cycloheximide treatment as expected (Figure [4C](#F4){ref-type="fig"}) and does not seem to be subjected to NMD. This result confirmed those obtained in E357X/E364X or IVS12/IVS12 HTI fibroblasts (Figure [1B](#F1){ref-type="fig"}), where del231 was undetectable probably because of the introduction of PTCs in these transcripts and their targeting to the NMD pathway.
Discussion
==========
Del100 and del231 were originally identified while studying the impact of NMD on hereditary tyrosinemia type I. During the characterization of the effects of the W262X mutation on FAH mRNA metabolism \[[@B29]\], we detected two minor alternative transcripts. If due to the W262X mutation, they should only be produced when the nonsense mutation is present, i.e. in heterozygous (W262X/wt) and homozygous (W262X/W262X) cell lines. However, by using specific primers for each transcript, we found that both del100 and del231 are produced in normal lymphoblastoid cells (Figure [1B](#F1){ref-type="fig"}) and in normal human liver. Del100 was also present in two different HTI cell lines harboring different nonsense mutations (E357X/E364X) or a splice mutation (IVS12+5g-\>a) in exon 13 or intron 12 respectively. These data argue in favor of del100 and del231 resulting from alternative splicing pathways rather than from a W262X-associated altered splicing mechanism. We suggest that both transcripts result from a weak definition of exon 8 (Figure [5](#F5){ref-type="fig"}). Indeed, exon 8 is subjected to many alterations as a result of splice mutations in the region encompassing exons 6 to 9. For example, due to a splice mutation in intron 6 (IVS6-1g-\>t; \[[@B30],[@B35]\]), a cryptic acceptor site in exon 8 is activated or exon 8 is skipped \[[@B30],[@B35]\]. Del100 and del231 were also identified in the case of the Q279R mutation, a splicing mutation that weakens the donor splice site of exon 9 \[[@B30]\]. Figure [5](#F5){ref-type="fig"} presents with a model that could explain these observations: we suggest that in the major splicing pathway, intron 8 is removed before introns 7 and 9, leading to a splice intermediate that contains the merged exons 8 and 9. Both exons are subsequently defined as a single exon. The order of intron removal is an important determinant of the outcome of splice-site mutations and could explain some unusual alterations, like the skipping of contiguous exons, as strongly suggested by studies of the COL1A1 and COL5A1 splice mutations \[[@B36],[@B37]\]. Altogether, these data suggest that at least del231 may arise through a minor splicing pathway due to an error-prone splicing apparatus because of the weak definition of exon 8 and the order of intron removal. Del100 could originate from a second minor splicing pathway, in which exon 8 is skipped alone because of its weak definition (Figure [5](#F5){ref-type="fig"}).
Del100 and del231 are thus the first cases of alternative splicing for the *fah*gene. However, it remained to see whether they were unproductive splice isoforms or whether they could code for protein isoforms. The del231 transcript retains an unchanged open reading frame when compared to FAH. The putative DEL231 protein would be similar to FAH except for the lower molecular weight (about 35-kDa), due to the missing region encoded by exons 8 and 9. We have been unable to detect a protein species of the size that could correspond to DEL231 using an antibody against full-length FAH. Whether this reflects the absence of such a protein or its presence in a very low amount undetectable with the presently available antibodies remains unknown. The latter explanation seems plausible since the del231 transcript, although not subjected to NMD, is much less abundant that the full-length FAH transcript or del100, as it is barely detected in W262X cells with the RT76 and RT025 primers (Figure [1](#F1){ref-type="fig"}) and the number of PCR cycles needed for its visualization is higher than for del100 when using the specific primers.
The structure of the putative DEL100 protein in the N-terminal part is identical to that of FAH. But due to exon 8 skipping, the reading frame is very different in the last 67 amino acids. DEL100, a 31-kDa protein, was detected in different human tissues using an antiserum raised against the specific C-terminal part of the putative protein. The antiserum is specific for the DEL100 protein and does not cross-react with FAH. In addition, the cross-reacting 31-kDa band was lost after adsorbing the antiserum against the purified peptide used for the immunization and the antiserum recognizes an *in vitro*translated DEL100-Myc protein, demonstrating its specificity. Thus the del100 transcript seems to direct the synthesis of a protein. This result is surprising because this transcript contains PTCs and seems to be subjected to NMD, as shown by a block of translation following a cycloheximide treatment. Interestingly, FAH and DEL100 have converse expression patterns in the human tissues examined. This suggests a post-transcriptional regulation of the expression of the two proteins, since the two transcripts originate from the same pre-mRNA. Alternative splicing, a highly regulated process, which can be developmental-stage or cell-specific, could be responsible for this difference of expression. For example, exon 8 may be more prone to skipping in the spleen given the concentration of specific *trans*-acting splicing regulators. This could be a way to downregulate the level of FAH transcript, by producing an alternative transcript, which is further eliminated by NMD. Indeed recent *in silico*analyses and observations on splicing factors have suggested that NMD, when coupled to alternative splicing, could regulate gene expression \[[@B10]-[@B12],[@B17]\]. Del100 could be another example of such a coupling of alternative splicing and NMD. In such a case, DEL100 would not be expected to play any function in the cell.
Very low levels of proteins can sometimes have enormous effects. Interestingly, 10--30% of nonsense transcripts escape NMD and when associated with polysomes are stable \[[@B19]\]. Is the coupling of alternative splicing with NMD in order to degrade the unproductive isoform the only option? An alternative, as proposed by Neu-Yilik *et al*. \[[@B38]\], may be that NMD could function in quantitatively controlling the expression of low amounts of protein. In this view, the PTC-containing del100 transcript may produce a protein with a physiological, although still unknown, function in the cell. The FAH structure contains a C-terminal part of 300 residues, which presents a novel arrangement of β-strands and plays a functional role in Ca^2+^binding, dimerization and catalysis of its substrate, fumarylacetoacetate \[[@B39]\]. Many of the residues encoded by exon 8 are part of the β-strands and residue 233 serves to bind the Ca^2+^\[[@B39]\]. The DEL100 protein, which lacks these residues, is thus very unlikely to function in catalyzing the hydrolytic cleavage of carbon-carbon bonds. While the function of DEL100 is unknown at this time, it may have a function in tyrosinemia. Indeed, not all mutations affecting the *fah*gene will similarly affect the DEL100 protein. For example, mutations affecting exons downstream of exon 10 will affect FAH production but not that of the DEL100 protein. This might be reflected in the phenotypic heterogeneity observed in HTI patients \[[@B21]\]. Preliminary computer analyses of DEL100 motifs using Proscan at PBIL suggest that it contains a putative DNA-binding motif (RVFLQNLLSvSQARLR with 89% similarity found to the consensus sequence). Whether DEL100 can function as a regulating factor remains unknown.
Conclusions
===========
NMD was first envisaged as a mechanism to prevent the accumulation of faulty transcripts, arising from mutations or processing abnormalities. Recent *in vivo*observations and *in silico*analyses have suggested a new role of NMD in gene expression regulation, when coupled to alternative splicing. We report here the identification of an alternative nonsense transcript of the *fah*gene, which despite being subjected to NMD, produces a protein in different human tissues. This provides an interesting starting point for the analysis of the role of NMD in the regulated productive splicing and translation.
Methods
=======
Cell culture
------------
The lymphoblastoid cell lines were established from lymphocytes of a HTI patient and his parents as described in Tremblay and Khandjian \[[@B40]\]. Cells were maintained in RPMI-1640 supplemented with 15% fetal bovine serum. The other human cell lines, HeLa (cervix) and normal fibroblasts, were cultured in DMEM 10%. Fibroblasts of other HTI patients (WG1647, mutation IVS12/IVS12; and WG1922, mutation E357X/E364X \[[@B33]\]) obtained from the Montreal Children\'s Hospital (C.R. Scriver) were maintained in DMEM 10% FBS. In translation inhibition experiments, lymphoblastoid cells were treated with 100 μg/ml cycloheximide 3 hours prior to RNA extraction (see below).
RT-PCR analysis
---------------
Total RNA was extracted from 5·10^6^cells with Trizol reagent (Gibco-BRL). RNA from human normal liver was extracted using the RNAqueous kit (Ambion). 1 μg RNA was reverse transcribed using an oligo(dT) and Stratascript (Stratagene). FAH cDNA was amplified from exons 6 to 14 using the following primers: RT76 (5\'-CGT GCC TCC TCT GTC GTG-3\') and RT025 (5\'-GGG AAT TCT GTC ACT GAA TGG CGG AC-3\'). Sense primers were designed to specifically amplify del100 and del231. RT84 (5\'-TGG AGC TGG AAA TGC ACG-3\') spans the exon 7 to exon 9 junction, whereas RT85 (5\'-TGG AGC TGG AAA TGG ACC-3\') spans the exon 7 to exon 10 junction. Amplification of the alternative transcripts with RT84 or RT85 was performed using HotStart (Qiagen) and PCR conditions were optimized in order to minimize nonspecific hybridization of the primers. Moreover, for each amplification (the FAH transcripts and RAR), the kinetic of the reactions were performed and the number of cycles used for each PCR was in the exponential phase.
Analysis of the cycloheximide treatment was done as previously described \[[@B29]\].
Production of an antiserum against the DEL100 protein
-----------------------------------------------------
An antiserum against the C-terminal part of the DEL100 protein was raised in mouse. The antigen is the C-terminal part (the last 67 amino acids) of the DEL100 protein and is different from FAH or the DEL231 protein. FAH cDNA was amplified from exon 9 to exon 14 using the primers hFAHsstermdel100 (5\'-CGG GAT CCC TGC AGC ACG AGA CAT TCA GAA GTG G-3\') and RT025 with Expand High Fidelity. The PCR product was inserted into pET30a (Novagen) at the BamHI and EcoRI sites, in order to express the reading frame of the C-terminal part of the DEL100 protein. The His-Tag fusion protein used for immunization was purified by affinity chromatography on a Ni-NTA column (Qiagen) in denaturing conditions with 6 M urea.
Preparation of the anti-hFAH monoclonal antibody
------------------------------------------------
The anti-hFAH monoclonal antibody was raised against the N-terminus of the protein. The 161 residue peptide was obtained by cutting the pET30a-FAH vector \[[@B41]\] with the NcoI and XhoI restriction enzymes (New England Biolabs), overhangs were then filled using T4 DNA polymerase (New England Biolabs) and the vector was ligated with the T4 DNA ligase (New England Biolabs). The peptide was expressed in the GJ1158 strain of *Escherichia coli*as previously described \[[@B41]\] and purified by affinity chromatography on a Ni-NTA column under denaturing conditions (6 M urea). The purified peptide was injected into BALB/c mice and hybridomas were prepared according to the procedures described in \[[@B42]\].
Western blot analysis
---------------------
Cells were harvested and lysed in 1 × SDS sample buffer (62.5 mM Tris-Hcl, pH 6.8, 2% SDS, 2.5% 2-β mercaptoethanol, 75 mM DTT, 10% glycerol and 0.005% Bromophenol blue). Human tissues obtained at autopsy and stored at -70°C until used \[[@B32]\] were homogenized in 10% (w/v) 0.01 M K-phosphate buffer (pH 7.3) and centrifuged for 20 min at 15,000 g. The supernatant was used for immunoblot assay. Samples were electrophoresed on SDS-15% polyacrylamide gels and proteins transferred to a nitrocellulose membrane. Antiserum against the DEL100 protein was used at a dilution of 1:20,000 against purified proteins or 1:1,000 against human tissues. FAH was detected using the polyclonal antibody \#488 (1:25,000) as described previously \[[@B32]\] or using a monoclonal antibody directed against the N-terminal part of the FAH protein (dilution 1/500). Protein loading was verified by using a monoclonal antibody against β-actin (dilution 1/400; Neomarker).
Tests of the specificity of the anti-DEL100 antiserum
-----------------------------------------------------
In some experiments, the mouse antiserum against the DEL100 protein was adsorbed on the recombinant His-tag C-terminal protein blotted on nitrocellulose. After an overnight incubation at 4°C, the non-adsorbed fraction was removed and conserved for further characterization. The adsorbed antibody fraction (affinity purified) was eluted using 1 ml of glycine-HCl 0.1 M, pH 2.8 and the pH immediately neutralized by adding 100 μl of 1 M K~2~HPO~4~, pH 8.2.
The del100 cDNA was obtained by RT-PCR on total RNA extracted from W262X/W262X cells using ND1 (5\' CCC AAG CTT CAG CAT GTC CTT CAT CCC GGT GG 3\') and ND2 (5\' TGC TCT AGA TTT ATT TGT CAC TGA ATG GCG G 3\'). The amplification products were cloned into pDrive (Qiagen) and different clones were sequenced. One clone containing the del100 cDNA was used for further cloning. It was amplified using 5\'del100-Eco (5\' GGA ATT CCA GCA TGT CCT TCA TCC 3\') and ND2. The amplified fragment was digested with EcoRI and XbaI and ligated into EcoRI-XbaI-digested pcDNA3-myc RANGAP, replacing the insert coding for RANGAP. pcDNA3-mycRANGAP was kindly provided by Dr M. J. Matunis (Johns Hopkins University, Baltimore, MD) The construct was used for coupled *in vitro*transcription-translation using the TNT coupled reticulocyte lysate system (Promega) according to the manufacturer\'s recommendations. 25 μl of the reaction were used for immunoprecipitation using the anti-Myc antibody as follows: the anti-Myc (1/100) was incubated 1 hour with protein A-sepharose beads (Sigma). The antibody was next immobilized on the beads using 20 mM dimethyl pimedilate (Sigma) in 0.2 M borate sodium (pH 9.0) for 30 min at room temperature. The reaction was stopped by washing the beads twice in 0.2 M ethanolamine and incubation in this solution for 2 hours at room temperature. The antigen (25 μl of the *in vitro*translated DEL100-Myc protein) was incubated with the beads for two hours at 4°C in a dilution buffer containing 10 mM Tris-HCl pH 8.0, 1 mM EDTA and 10% glycerol. The immunoprecipitated protein was eluted by adding 25 μl of SDS loading buffer (62.5 mM Tris-HCl, pH 6.8; 2% SDS; 2.5% β-mercaptoethanol; 75 mM DTT; 10% glycerol and 0.01% bromophenol blue). Samples were electrophoresed on SDS-15% polyacrylamide gels and proteins transferred to a nitrocellulose membrane. The anti-Myc was used at a dilution of 1/2,000 and the anti-DEL100 antiserum at a dilution of 1/1,000.
List of abbreviations
=====================
CHX, cycloheximide; ESE, exonic splicing enhancer; FAH, fumarylacetoacetate hydrolase; HTI, hereditary tyrosinemia type I; Ig, immunoglobulin; NMD, nonsense-mediated mRNA decay; PCR, polymerase chain reaction; PTB, polypyrimidine tract binding protein; PTC, premature termination codon; RAR, retinoic acid receptor; RT, reverse transcription; RUST, regulated unproductive splicing and translation; TcR, T-cell receptor.
Author\'s contributions
=======================
ND carried out the experiments and wrote the manuscript. AM participated in the design of the study. JFBL participated in the cloning of Del100 into pcDNA3. AB raised the mAb directed against the N-terminal part of the FAH protein. RMT participated in the design, coordination of the study, and in the writing of the manuscript.
Acknowledgements
================
This work was supported by a grant from the Canadian Institutes of Health Research (CIHR) to RMT. ND received studentships from the Government of Canada (International Program) and from le Centre de Recherche sur la Fonction, la Structure et l\'Ingénierie des Protéines (CREFSIP). JFBL is supported by a studentship from NSERC. We thank Dr M. Salo (Tempere, Finland) for providing lymphocytes from Finnish patients and their parents, S. Tremblay and Dr E. W. Khandjian (CHUQ, Pavillon St-François D\'Assise, Québec) for establishing the lymphoblastoid cell lines, Dr A. C. Scriver (Montreal Children\'s Hospital) for the HTI fibroblasts and Dr A. Darveau (Université Laval, Québec) for a control lymphoblastoid cell line (T19) and his help with the production of the antibody directed against the N-terminal of FAH.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Del100 and del231 result from minor alternative splicing pathways and not from the presence of the W262X mutation.**Total RNA extracted from different normal cell lines or cells of HTI patients was subjected to RT-PCR. Controls for FAH, del100 and del231 are displayed on the left to show that the PCR reactions are in the linear range. (A) FAH amplification from exons 6 to 14 in the lymphoblastoid cell lines. The two additional products (del100 and del231) seen in the homozygous cell line (W262X/W262X) are depicted below the gels. Exons are represented by boxes and introns by lines. Del100 has skipped exon 8 and as a consequence, the reading frame is shifted (W262X becomes G229E). A first PTC is located at the 3\'end of exon 10 (270X). Del231 has skipped both exons 8 and 9, but without any change in the reading frame. (B) Primers designed to specifically amplify del100 and del231 span exon 7 to 9 or exon 7 to 10 junctions respectively (depicted in blue). Full-length, del100 or del231 cDNAs were cloned in pRC/CMV and were amplified to verify primer specificity to each alternative transcript. Both transcripts are present in normal cells (wild-type lymphoblasts or fibroblasts and HeLa cells) and human liver. Del100 is amplified in HTI cells with a splice mutation in exon 12 (IVS12/ IVS12) or two nonsense mutations in exon 13 (E357X/E364X). FAH was also amplified as a control in the different human cell lines.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Del100 transcript encodes for a protein expressed in various tissues.**(A) Schematic representation of the putative DEL100 and DEL231 proteins. FAH and DEL231 proteins are identical, except for the region encoded by exons 8 and 9, which is missing in DEL231. DEL100, due to the skipping of exon 8 and the resulting change in the coding frame, differs from FAH at its C-terminal end. The black bar below the DEL100 protein represents the region used to raise the antiserum. (B) 30 μg proteins of a 10% (w/v) tissue homogenate were separated on a 15% SDS-polyacrylamide gel. Proteins were transferred to a nitrocellulose membrane and blotted using a rabbit anti-FAH (top panel; dilution 1:25,000) or a mouse anti-DEL100 antiserum (middle panel; dilution 1:1,000). A band with an apparent molecular weight of 31-kDa is detected by the anti-DEL100 antiserum in all tissues tested. The star indicates non-specific signal. The membrane was blotted against β-actin to control for protein loading. (C) The 31-kDa band (indicated by a red arrow) was detected in spleen using a monoclonal antibody directed against the N-terminal part of the FAH (green arrow).
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Specificity of the anti-DEL100 antiserum.**(A) The mouse antiserum was affinity purified on the recombinant C-terminal of DEL100 and used to probe the same membrane as in Figure 2B. The adsorbed fraction (top panel), specific of the C-terminal of DEL100, still recognizes the 31-kDa protein. The non-adsorbed fraction (bottom panel) shows no cross-reactivity to the protein. The star indicates a non-specific signal. (B) The DEL100-Myc protein was synthesized in a rabbit reticulocyte lysate and labeled with ^35^S (translation panel). The tagged protein is indicated by an arrow. The band below is the protein synthesized without the Myc tag. The anti-Myc was immobilized on protein A-sepharose and incubated with *in vitro*translated DEL100-Myc. Effectiveness of the immunoprecipitation was verified by blotting using the anti-Myc (1/2,000) antibody. The anti-DEL100 antiserum (1/1,000) recognizes the *in vitro*translated protein in the input (10% of the reaction) and the immunoprecipitate (IP, indicated by an arrow).
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Del100 is subjected to NMD.**Control cells (wt/wt) and the homozygous cell line from the Finnish patient (W262X/W262X) were treated with 100 μg/ml cycloheximide (CHX) for 3 hours. Total RNA was extracted and subjected to RT-PCR. RAR serves as a control for RNA quantity in each sample. Control amplifications (showing an increasing number of cycles) for del100, del231 and RAR are displayed on the left to show that, with the conditions used, each PCR reaction is in the exponential phase. (A) \[α^33^P\]-dATP was incorporated during the PCR reaction and the products loaded on a 6% acrylamide gel. The gel was used to directly expose an X-ray film and the signal was quantified using the NIH Image 1.2 software. FAH was amplified from exons 6 to 14 (the del100 transcript which is detected using these primers is indicated by a star). Del100 and del231 were amplified using the specific primers RT84 and RT85. (B) Quantification of the amount of del100 mRNA in 3 different experiments, normalized to RAR levels. The error bars represent standard deviations. Del100 is stabilized between 4- to 7-fold in homozygous cells (W262X/W262X) and normal cells (wt/wt) following the cycloheximide treatment (yellow bars: - cycloheximide; purple bars: + cycloheximide). (C) del231 is relatively unaffected by translation inhibition.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Suspected splicing pathways of the region encompassing exons 6 to 10.**Schematic representation of exon 8 and its splice sites is given on the top of the figure. The effectiveness of the donor and acceptor splice sites were determined using the webgene program <http://itba.cnr.it/webgene/> and is indicated above the sequence. A cryptic acceptor splice site is used in the case of the IVS6-1g-\>t mutation (described in \[35\]). Exons are represented by boxes and introns by lines. The thick arrow is the major splicing pathway, whereas thin arrows are minor alternative splicing pathways. We suggest that del231 could be explained by the fact that intron 8 is spliced out before introns 7 and 9, leading to an intermediate in which exons 8 and 9 are defined together. Because of exon 8 weak definition, a minority of transcripts eliminates both exons 8 and 9. Del100 could be produced by another minor splicing pathway, in which exon 8 is skipped alone.
:::
:::
|
PubMed Central
|
2024-06-05T03:55:52.046649
|
2005-1-7
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546004/",
"journal": "BMC Mol Biol. 2005 Jan 7; 6:1",
"authors": [
{
"first": "Natacha",
"last": "Dreumont"
},
{
"first": "Antonella",
"last": "Maresca"
},
{
"first": "Jean-François",
"last": "Boisclair-Lachance"
},
{
"first": "Anne",
"last": "Bergeron"
},
{
"first": "Robert M",
"last": "Tanguay"
}
]
}
|
PMC546005
|
Background
==========
The concept of a direct, one-to-one association between a sensory stimulus and a motor response has been strongly influential in neuroscience \[[@B1]\]. Such associations may be quite complex; for instance, monkeys can learn visuomotor mappings based on arbitrary rules \[[@B2]-[@B4]\]. But from a mechanistic point of view, it is their flexibility which is remarkable. Humans and other mammals react to a given stimulus in drastically different ways depending on the context \[[@B1],[@B5]-[@B7]\]. What is the neural basis for this? How do current goals, recent events, and other environmental circumstances gate or route immediate sensory signals to generate an adequate action?
Gain control is a common mechanism by which neurons integrate information from multiple modalities or sources \[[@B8],[@B9]\]. Gain-modulated neurons typically have a sensory receptive field, but in addition, their overall excitability depends on some other modulatory parameter. A classic example are the neurons in parietal area 7a, whose activity can be described by the product of a gain factor, which is a function of the gaze angle, and the response profile of the visual receptive field \[[@B10],[@B11]\]. That is, gaze direction determines the amplitude of their stimulus-dependent responses. According to theoretical studies, gain-modulated responses are useful for performing a class of mathematical operations known as coordinate transformations \[[@B12]-[@B16]\]. For example, by combining multiple eye-centered inputs that are gain modulated by gaze direction, a downstream neuron can generate a response that depends on the location of a stimulus relative to the body \[[@B12]-[@B14]\]. Experimental studies have reported gain changes due to a wide range of proprioceptive signals, such as gaze direction \[[@B10],[@B11],[@B17]\], eye and head velocity \[[@B18]\] and arm position \[[@B19],[@B20]\]. Modulations relevant to attention-centered \[[@B21]-[@B23]\] or object-centered representations \[[@B24],[@B25]\] have also been documented.
Interestingly, all of these examples deal with the same problem -- spatial localization -- but the computations that can be effectively carried out through gain-modulated responses are much more general \[[@B13],[@B16],[@B26]\]. In particular, here I show that modulating the activity of a population of neurons is equivalent to turning on and off different subsets of neurons. Thus, the modulation can be thought of as a switch that can activate one of many possible sensory networks, each instantiating a different sensory-motor map. Crucially, the modulatory signal itself does not have to provide any spatial information; it can be a recent instruction or some other aspect of the current behavioral context. Examples of choices between multiple sensory-motor maps determined in a context-dependent manner include speaking in one language or another, and the ability of musicians to interpret a musical score depending on the clef and key signature at the beginning of each stave. But the same principles also apply in more simplified settings, such as behavioral tasks where a given stimulus is arbitrarily associated with two or more motor responses, depending on a separate instruction \[[@B4],[@B27]-[@B29]\]. For instance, the shape of a fixation point may be used to indicate whether the correct movement should be a saccade toward a spot of light or an antisaccade away from it \[[@B30]\]. What all of these cases have in common is a functional reconnection between visual and motor networks that must occur very quickly and without explicit spatial guidance from the context information.
Using theoretical and computer-simulation methods, I show that this type of functional switching can be achieved through contextual modulation regardless of how the context is encoded -- whether continuously or discontinuously -- and independently of the discriminability of the stimuli. The results are presented using neural network models of hypothetical behavioral tasks similar to those used in experiments with awake monkeys. A report with a different example was published previously \[[@B31]\].
Results
=======
All model networks discussed below have the same general, two-layer architecture \[[@B14]-[@B16]\]. A first layer of gain-modulated (GM) neurons drives a second layer of output or motor neurons through a set of feedforward connections, with each GM unit projecting to all output units. In each trial of a task, the GM neurons are activated by the sensory and context signals, and a motor response is generated by the output neurons (see Methods). Each model proceeds in three steps. First, the GM and the desired output responses are specified according to the task. Then, synaptic weights are found that, across all stimulus and context combinations, make the driven output responses as close as possible to the desired ones. Finally, the network is tested in multiple trials in which the GM neurons drive the output units. Model performance is measured by comparing the resulting, driven pattern of motor activity in each trial with the desired, pre-specified one. The first task, with only two contexts, serves to illustrate the analogy between gain modulation and a switch.
Switching between saccades and antisaccades
-------------------------------------------
In the antisaccade task, a stimulus appears briefly at position *x*along the horizontal and the subject responds by making an eye movement (Fig. [1](#F1){ref-type="fig"}). There are two possible contexts or conditions. In the first one, the movement should be to the location where the stimulus appeared, *x*; in the second one, the movement should be to the mirror-symmetric point, -*x*. Both condition and stimulus location vary across trials. The color of the fixation spot (or any other arbitrary cue) may serve to indicate which condition applies in each trial \[[@B30]\].
Examples of model GM responses chosen for this task are shown in Fig. [2](#F2){ref-type="fig"}. These neurons simply respond to visual stimuli presented at different locations; however, they are also sensitive to the context. Each graph shows the mean firing rate of one unit as a function of *x*, with one curve for each of the conditions (red and green traces). These tuning curves are bell-shaped because Gaussian functions were used to define them (see Methods). Because context affects the gain of the responses, for any given cell, the two curves differ only in their amplitudes. The context that produces the highest gain is the preferred one. The maximum and minimum gains for each neuron are model parameters that can be between 0 and 1. The four GM units in Fig. [2](#F2){ref-type="fig"} illustrate various degrees of modulation. The case of full modulation (maximum gain = 1, minimum gain = 0) depicted in Fig. [2a](#F2){ref-type="fig"} corresponds to a neuron that is switched on and off by context: in its preferred condition it is highly active, whereas in its non-preferred condition it is fully suppressed.
First consider what happens if the first layer of a model network is composed of two populations of such switching neurons. One population is active in context 1 and the other in context 2. This is illustrated in Fig. [3a](#F3){ref-type="fig"}. The rectangle encloses the responses of all model neurons (60 GM and 25 output units) in a single trial of the antisaccade task. The firing rates of the GM neurons are in color. The two populations (red and green) have opposite context preferences but identical sets of sensory tuning functions. The black dots are the responses of the driven output neurons. Their center of mass (Equation 19), which in this case is the same as the location of the peak, is interpreted as the target location for an impending saccade. The network performs accurately in the four trials shown in the column, since the encoded movement location is equal to *x*for saccades (context 1) and to -*x*for antisaccades (context 2). It is easy to see why such a network can implement two entirely independent sensory-motor maps: each population has its own set of synaptic connections driving the downstream motor neurons, and the maps are kept separate because the two populations are never active at the same time.
Figure [3d](#F3){ref-type="fig"} shows the corresponding matrix of synaptic connections. To interpret this figure, notice that GM units 1--30 are the ones that prefer context 1 (red dots in Fig. [3a](#F3){ref-type="fig"}), whereas units 31--60 prefer context 2 (green dots in Fig. [3a](#F3){ref-type="fig"}). Preferred stimulus locations are arranged in increasing order for both populations. Units 1--30 generate direct saccades, so their connections are aligned with the motor neurons; that is, GM neuron 1 excites output neuron 1 most strongly, GM neuron 2 excites output neuron 2 most strongly, etc. Thus, in context 1, stimuli to the right generate movements to the right. In contrast, the GM units that generate antisaccades are connected in the reverse order: GM neuron 31 excites output neuron 25 most strongly, GM neuron 32 excites output neuron 24 most strongly, and so on. Thus, in context 2, stimuli to the right result in movements to the left.
The model correctly produces saccades in context 1 and antisaccades in context 2. Furthermore, this scheme for switching sensory-motor maps as a function of context would also work for any two maps driven by the two populations. This model switches maps successfully because the GM neurons are themselves switched on and off by context, so this case is trivial. However, the main result in this section is that a network of partially modulated GM neurons has exactly the same functionality. The more rigorous statement is this: for a discrete number of contexts and everything else being equal, a network of partially modulated neurons can generate the same mean downstream responses as a network of switching neurons. Figure [3](#F3){ref-type="fig"} illustrates this equivalence: identical output activity profiles are generated when all GM neurons are fully suppressed in their non-preferred context (Fig. [3a](#F3){ref-type="fig"}), when all are partially modulated by the same amount (Fig. [3b](#F3){ref-type="fig"}), and when the modulation varies randomly across cells (Fig. [3c](#F3){ref-type="fig"}). These three cases require different sets of synaptic connections between GM and output layers, but this is simply because the GM responses vary across cases. In particular, note the dark blue diagonal bands in Figs. [3e,3f](#F3){ref-type="fig"}, compared to Fig. [3d](#F3){ref-type="fig"}. They correspond to negative weights needed to subtract out activity that is irrelevant to a particular context. For instance, in the direct saccade trials of Fig. [3b](#F3){ref-type="fig"}, the responses of the antisaccade-preferring neurons should be cancelled, and viceversa. The new negative weights combined with larger positive weights achieve this.
The key point is that, under relatively mild conditions, partial and full modulation lead to the exact same repertoire of switchable sensory-motor maps (the difference lies in their accuracy, as discussed below). The formal proof is presented in Appendix A. This result is interesting because it provides an intuitive interpretation of gain modulated activity: modulations that may seem small at the single-unit level may produce drastically different output responses due to their collective effects, the result being as if different sensory populations had been turned on and off.
Partial versus maximum gain modulation
--------------------------------------
The equivalence between networks of neurons that switch across contexts and networks with partial modulation is subject to an important condition and a qualification.
The key condition is that the modulation factors that determine the gain of all the neurons with similar stimulus selectivities must be linearly independent across contexts (Appendix A). In practice, one way to achieve this is to include all relevant combinations of sensory and contextual preferences. For instance, if there are two neurons that respond maximally when *x*= 5, the condition is fulfilled for that pair if one neuron prefers context 1 and the other context 2. As long as this independence constraint is satisfied, there is great flexibility in the actual amount of modulation; it does not need to be 100%, as with a full switch.
The qualification, however, is also critical, because a network of partially modulated GM neurons is not exactly the same as one composed of switching neurons: in most functionally relevant cases, partially modulated neurons are effectively noisier. In general, variability plays an important role in the performance of these networks. No fluctuations were included in the simulations of Fig. [3](#F3){ref-type="fig"}, so performance was virtually perfect. But the magnitude of the error between correct and encoded movement directions increases depending on the amount of noise that is added to the GM responses, and as the difference between the minimum and maximum gains diminishes, the impact of noise typically goes up. This is shown analytically in Appendix C and is illustrated in Fig. [4](#F4){ref-type="fig"}.
Two measures of noise sensitivity are plotted in Fig. [4](#F4){ref-type="fig"}. The first one is the standard deviation of a single output response across trials with identical stimulus and context. This number, *σ*~*R*~, quantifies the variability of single neurons. Figure [4a](#F4){ref-type="fig"} plots *σ*~*R*~as a function of *γ*, which is the minimum gain of the GM neurons (the maximum is 1). When *γ*= 0, the GM neurons are fully suppressed in their non-preferred context; when *γ*= 1, the GM responses are identical in both contexts. The three curves are for three levels of noise. Their order shows that, as expected, higher noise in the input layer always produces higher variability in the output. For each data point, the synaptic weights were set so that the average firing rates of the output neurons, as functions of stimulus location and context, were always the same (Appendix B). Thus, for all *γ*values, the average profile of motor responses for *x*= -15 and *x*= 10 looked exactly like those in Fig. [3](#F3){ref-type="fig"}. The monotonically increasing curves in Fig. [4a](#F4){ref-type="fig"} indicate that the variability of the output rates goes up with *γ*, as predicted theoretically (Appendix C).
The second measure of noise sensitivity is *σ*~*CM*~, which estimates the error between the desired movement location and the center of mass of the output population, which is considered the encoded movement location (Equations 19, 20). Thus, *σ*~*CM*~quantifies the variability of the network. Figure [4b](#F4){ref-type="fig"} shows that *σ*~*CM*~also increases with *γ*, reaching a saturation level. This error saturates because, in contrast to the individual neuron responses, the encoded movement location is restricted to a limited range of values, so its variance cannot grow above a certain limit.
Figures [4c](#F4){ref-type="fig"} and [4d](#F4){ref-type="fig"} show the same measures of variability but when the synaptic weights are computed using the standard, optimal algorithm (see Methods). For each value of *γ*, the optimal algorithm considers both the mean and the variance of the output responses \[[@B32],[@B33]\], striking a balance between them that, overall, minimizes the average squared difference between the driven and the desired output rates (Equation 11). Therefore, in Figs. [4c,4d](#F4){ref-type="fig"}, the mean output responses are not quite the same for all data points; in particular, for *x*= -15 and *x*= 10 there are small differences compared to the curves in Fig. [3](#F3){ref-type="fig"} (data not shown). This method markedly reduces the variability of the individual output neurons relative to the case where only the mean values are considered. It also produces a modest decrease in *σ*~*CM*~(compare Figs. [4b](#F4){ref-type="fig"} and [4d](#F4){ref-type="fig"}). However, it does not change the main effect: the error in the encoded location still grows monotonically with *γ*.
Note that, as explained in Appendix C, *γ*\> 0 does not always produce higher variance in the output, compared to *γ*= 0. For instance, if the sensory-motor maps in the two contexts are the same, the optimal strategy is to activate both populations of GM neurons simultaneously, i.e., to use *γ*= 1. This is simply because the average of two noisy responses with equal means is better than either of them. In general, however, switching is relevant precisely when the sensory-motor maps are different, as in Figs. [3](#F3){ref-type="fig"} and [4](#F4){ref-type="fig"}, in which case weaker modulation (higher *γ*) results in higher output variability.
In conclusion, as the modulation becomes weaker, the performance of the network typically becomes less accurate, even though the average output responses may be close or identical to those obtained with maximum modulation. In Fig. [4](#F4){ref-type="fig"}, this becomes more of a problem when the minimum gain *γ*is above 0.6 or so, at which point *σ*~*CM*~is about twice that observed with full modulation. These results were obtained using the same *γ*for all GM neurons, but almost identical curves were produced when *γ*varied randomly across cells and the results were plotted against its average value.
Continuous vs discontinuous context representations
---------------------------------------------------
The possible contexts encountered by an organism could be numerous and diverse, so it is not clear how the brain might encode them. There are at least two distinct ways: as separate, discrete states, or as points along a smooth, continuous space. What would be the difference in terms of the functionality of the remapping networks studied here? This is investigated next, using a generalization of the antisaccade task referred to as the scaling task.
The scaling task is very much like the antisaccade task, except with more contexts. The subject\'s response should be an eye movement toward a location determined by the position of the stimulus, *x*, and a scale factor, *y*; the movement should be toward the point *xy*. When *y*= 1, the movement is simply a saccade toward *x*; when *y*= -1, the movement is an antisaccade toward -*x*; when *y*= 0.5, the movement should be to a point halfway between fixation and the location of the stimulus, and so on. To begin with, five possible conditions are considered, corresponding to scales of -1, -0.5, 0, 0.5 and 1.
Figure [5](#F5){ref-type="fig"} shows the responses of four GM units in this task plotted as functions of the position of the stimulus along the horizontal. A family of five curves, one per condition or scale factor, is drawn for each unit. As in the previous task, the shape of these curves is constant across conditions, because of the multiplicative interaction between stimulus- and context-dependent factors. The neurons in Figs. [5a,5b](#F5){ref-type="fig"} encode the context in a discontinuous way, because the order in which they prefer the five scales was set randomly (see Methods). Thus, for each unit, the order of the colors in Figs. [5a,5b](#F5){ref-type="fig"} is random. In contrast, the neurons in Figs. [5c,5d](#F5){ref-type="fig"} encode context smoothly; their response amplitudes decrease progressively as the current scale *y*differs from each cell\'s preferred scale. All units in the figure have approximately the same minimum gain, *γ*≈ 0.5.
Differences between these two coding strategies can be observed in Fig. [6](#F6){ref-type="fig"}. This figure shows the performance of two versions of the network model, each with 900 GM cells, in four trials of the scaling task. In the first version, illustrated in Figs. [6a-6d](#F6){ref-type="fig"}, context is encoded discontinuously, as in Figs. [5a,5b](#F5){ref-type="fig"}. The GM firing rates are color-coded, ordered according to their preferred stimulus locations (x-axis) and preferred scales (y-axis). In each trial, the GM rates form a band of activity centered on the location of the stimulus. The most intense responses are somewhat clustered, although high firing rates are scattered throughout the band. The band occurs because the responses vary smoothly as functions of stimulus location, and the scatter in the y-direction is due to the random order in which each neuron prefers the contexts; such scatter would be present even without noise. The output neurons have profiles of activity (black traces) with the highest peak located near the intended movement target. The small wiggles and secondary bumps are due to noise. The performance of the network is accurate, however: the encoded movement is close to the intended one for all combinations of stimulus location and scale factor (Figs. [6a-6d](#F6){ref-type="fig"}, red vs black vertical lines). The second version of the model, illustrated in Figs. [6e-6h](#F6){ref-type="fig"}, is almost identical to the first, except that context is encoded continuously, as in Figs. [5c,5d](#F5){ref-type="fig"}. Now the the activation pattern that emerges is clearly localized, centered on the current stimulus and context values. Performance is similar for the two networks, both having *σ*~*CM*~≈ 0.6.
Figures [7a,7b](#F7){ref-type="fig"} evaluate the performance of these two models across a wider range of parameters. The graphs show *σ*~*CM*~as a function of the number of GM neurons for three levels of noise. In all cases, the error decreases approximately as  -- a sign that noise is what limits the accuracy of the system. This is consistent with the virtually perfect performance obtained with zero noise. With the five selected contexts, results are almost identical for the continuous and discontinous cases.
Robustness and generalization
-----------------------------
There are two aspects of these networks that could vary depending on how context is encoded. The first is their robustness. In addition to random variations in the GM responses, there could be fluctuations in other elements of the circuits, in particular, the synaptic connections. Thus, a key question is how finely-tuned these connections need to be in order to obtain accurate performance. The answer: not very much. The networks tolerate considerable alterations in synaptic connectivity. This is illustrated in Figs. [7c,7d](#F7){ref-type="fig"}, which show *σ*~*CM*~as a function of the number of GM neurons in networks in which the connections were corrupted. For these plots, the connections were first set to their optimal values, as in the standard simulations, but then 25% of them, chosen randomly, were set to zero. To generate the same range of output firing rates, all remaining connections were divided by 0.75, but no further adjustments were made. Performance was then tested. Compared to the results with unaltered weights (Figs. [7a,7b](#F7){ref-type="fig"}), performance is evidently worse, but the disruption is not catastrophic; in particular, the error still goes down with network size. The increase in error is most evident when the noise is relatively low. Random weight deletion was used for these simulations because it is a rather extreme form of weight corruption, but other manipulations generated similar results.
The second important issue about these networks is their capacity to generalize. So far, the models have been tested with the same stimuli and contexts used to set the connections, but what happens when new stimuli or contexts are presented? This is where partial modulation and a smooth organization of response properties make a difference. First consider the model in which scale is encoded discontinuously. Its performance in generalization is shown in Fig. [7e](#F7){ref-type="fig"}. For this graph, only 8 stimulus locations, in combination with the 5 possible scales, were used to calculate the synaptic weights. That is, only 8 evenly-distributed values of *x*were used in Equation 3, giving a total of 40 combinations of stimulus and context. However, the network was tested with all 151 combinations of 31 stimulus locations (between -15 and +15) and 5 scales. Accuracy is practically the same as in the original simulations (Fig. [7a](#F7){ref-type="fig"}), where the 31 stimuli and 5 scales were used both for setting the connections and evaluating performance. The same scales had to be used in both cases because, given the discontinuous encoding, the gain factors for other scales could not be interpolated or inferred.
In contrast, in the continuous case, generalization can be tested in both the sensory and modulatory dimensions; the GM responses can be obtained for any combination of stimulus location and scale, because both quantities are defined analytically (Equations 3 and 6). Results are shown in Fig. [7f](#F7){ref-type="fig"}. For this graph, 8 stimulus locations and 8 scale factors were used to set the connections. The network was then tested on 31 stimulus locations and 31 scales within their respective ranges. Performance is slightly better than in the standard condition in which identical combinations of 31 stimulus locations and 5 scales were used throughout (Fig. [7b](#F7){ref-type="fig"}).
In summary, this task requires somewhat more complex GM neurons than the antisaccade task, because there are more contexts. In the discontinuous case, the basic intuition for why the model works is the same as in the previous task: with the proviso that they are effectively noisier, partially modulated neurons are equivalent to switching neurons, which can trivially establish independent sensory-motor maps. However, the key advantage of a continuous neural representation of context over a discontinuous one is that it allows a network to perform accurately on combinations of stimulus and context that have not been explicitly encountered before. By its very definition, such continuous encoding requires partial modulation. Therefore, although partial modulation is typically detrimental for switching between discrete contexts (relative to full switching), it is highly advantageous when context is parameterized by a continuous variable, because it serves to generalize.
Remapping based on ambiguous stimuli
------------------------------------
In the scaling task, all stimuli and contexts are unambiguous, but in many real-life situations and experimental paradigms, motor actions are preceded by perceptual processes that involve the interpretation or analysis of sensory information. That is, specific actions (e.g., pressing a left or right button) are often based on ambiguous information (e.g., whether on average a group of flickering dots moved to the left or to the right). In theory, switching between maps should be independent of the perceptual component of a task (Appendix A).
To investigate this, consider the orientation discrimination task illustrated in Fig. [8](#F8){ref-type="fig"}. In each trial, a bar is presented and the subject must determine whether it is tilted to the left or to the right. The judgement is indicated by making an eye movement either to a left or a right target. Discrimination difficulty varies depending on orientation angle *x*. The task is most difficult when *x*is near 0° and the bar is nearly vertical, but it becomes easier as *x*approaches ± 45°. This is also a remapping task because the association between bar orientation and correct target is not unique: the color of the fixation spot determines whether left and right targets correspond to bars tilted to the left (*x*\< 0) and to the right (*x*\> 0), respectively, or viceversa. There is also a no-go condition, which gives a total of three.
The GM cells in this case are tuned to stimulus orientation. The response curves are not shown, but have a single peak, as in Figs. [5a,5b](#F5){ref-type="fig"} -- the difference is that the sensory variable is orientation, which varies from -90° to +90°, and that there are only three conditions, three values of *y*(see Methods). The order in which each GM cell prefers the three contexts is set randomly, so context is encoded discontinuously.
The responses of the model output units are shown in Figs. [9a-9h](#F9){ref-type="fig"}. In no-go trials (Figs. [9g,9h](#F9){ref-type="fig"}), all neurons fire near their baseline rates, as prescribed (Equation 9). Thus, in this condition the network ignores the stimuli. In go trials, however, the profile of output responses has peaks at -10 and +10, which are the only two target locations in this task. In contrast to the activity profiles seen in previous tasks, here there never is a unique peak, even with zero noise (Figs. [9a,9c,9e](#F9){ref-type="fig"}). Instead, the relative amplitude of the two peaks varies as a function of bar orientation. The difference in the amplitudes of the two hills of activity decreases as the bar becomes more vertical, thus reflecting the difficulty of the task. Without any noise, the largest peak is always located at the correct target, but with noise the amplitudes vary across trials and errors are produced (Fig. [9d](#F9){ref-type="fig"}).
To quantify the performance of the network in this case, the generated movement was set equal to the location of the tallest hill of activity. This always corresponded to one or the other target location, +10 or -10, so each trial could be scored as either correct or incorrect. The assumption here is that a profile of activity with two peaks, as in Figs. [9c,9d](#F9){ref-type="fig"}, can be converted into a profile with a single peak, such that the smaller hill of activity is erased. Networks with recurrent connections organized in a center-surround fashion can do just that \[[@B26],[@B34]-[@B36]\]. So, if such lateral interactions were added to the output layer of the network, only the largest hill of activity would remain. Equivalently, the responses of the output neurons could serve as inputs to an additional, third layer that performed the single-target selection \[[@B34]\]. Either way, given that this is a plausible operation, it is reasonable to simply consider the location of the largest peak to determine the evoked movement.
Based on this criterion, the performance of the network is shown in Figs. [9i,9j](#F9){ref-type="fig"}, which plot the probability or fraction of movements to the target on the right as a function of stimulus orientation *x*. These are essentially neurometric curves -- psychometric curves computed from neuronal responses -- and indeed have the sigmoidal shape that is characteristic of many psychophysical measurements. Figure [9i](#F9){ref-type="fig"} shows the results for condition 1, in which bars with *x*\> 0 correspond to movements to the right; Fig. [9j](#F9){ref-type="fig"} shows the results for condition 2, in which the association is reversed and bars with *x*\> 0 correspond to movements to the left. The gray curves are best fits to the simulation data points. The fits have two parameters, the center point, or bias (indicated by dashed lines), and a second parameter that determines the steepness at the center point and is inversely proportional to the discrimination threshold (see Methods). Without noise, performance is virtually perfect (not shown), in which case the bias and threshold are zero and the neurometric curve becomes a step function. However, both quantities increase in magnitude as noise is increased, producing the observed sigmoidal curves.
The presence of a bias might be surprising, given the symmetry of the network. However, the bias depends on the number of trials used to estimate the probabilities. If each orientation were tested an infinite number of times, the data points in Figs. [9i,9j](#F9){ref-type="fig"} would line up perfectly along continuous curves. The discrimination thresholds of those curves would not be significantly different from those shown, but their biases would be zero. With finite samples, a bias in the neurometric curve is inevitable.
Figures [10a,10b](#F10){ref-type="fig"} show the bias and discrimination threshold as functions of network size for three levels of noise. Both quantities decrease with network size, so in this sense, the network is just as effective as that for the scaling task. Because large numbers of trials were used, the bias is about an order of magnitude smaller than the threshold. Figures [10c,10d](#F10){ref-type="fig"} plot the results when the synaptic connections in the network are corrupted by deleting 25% of them at random, as in Figs. [7c,7d](#F7){ref-type="fig"}. This manipulation leaves the discrimination threshold virtually unchanged, but increases the bias by about an order of magnitude, making it comparable to the threshold. This bias is a true limitation of the network; it does not decrease with more trials. Figures [10e,10f](#F10){ref-type="fig"} show performance during generalization, as in Fig. [7e](#F7){ref-type="fig"}. In this case, only the two extreme orientations, -8° and +8°, were used to set the connections (in combination with the three possible conditions). The network was then tested on the standard set of 64 orientations. A true bias also appears in this case. It stays lower than the threshold, which remains essentially unchanged.
In summary, although the ambiguity of the sensory information is reflected in the motor responses, it does not interfere with the context-dependent selection mechanism.
Discussion
==========
Gain modulation as a switch
---------------------------
The above results demonstrate that contextual modulation could serve to select one of many associations or maps between sensory stimuli and motor responses. Indeed, a key insight is that modulating the gain of a neural population is, in a sense, equivalent to flipping a switch that turns on or off specific subpopulations of neurons. This explains why networks of GM neurons can generate large changes in downstream responses -- even all-or-none changes, as in go vs no-go conditions (Figs. [9a-9h](#F9){ref-type="fig"}) -- although their own activity may vary rather subtly. In this framework there is a distinction between the selection process and the sensory representations. The capacity to switch depends on the collection of gain factors, whereas the space of possible functions of the stimulus that can be computed downstream is determined primarily by the sensory tuning curves (Appendix A). A weaker modulation typically increases the sensitivity to noise of the resulting motor responses (Appendix C), but otherwise, partial modulation can achieve the same sensory-motor map selection as maximal, all-or-none modulation. This is why the mechanism works across a large variety of tasks and representations that involve some type of switch.
In a landmark paper, Pouget and Sejnowski \[[@B13]\] studied the capacity of GM networks for coordinate transformations using the concept of basis functions. A group of functions of *x*form a basis set when any arbitrary function of *x*can be computed as a linear superposition of those functions in the group; sines and cosines of are a well known example. The function of *x*typically associated with a neuron\'s response is its tuning curve -- its firing rate measured as a function of *x*. Pouget and Sejnowski showed that, starting with two networks that form separate basis sets for *x*and *y*, a network of GM neurons comprising all possible combinations (i.e., pairwise products) of those two sets would form a basis set for functions that depend simultaneously on *x*and *y*. This means that any function of *x*and *y*can be computed from the resulting GM responses. This was a crucial result, because it provided a rationale for generating such a combined representation. However, it assumed that both the sensory and modulatory variables are continuous and that, taken independently, the sets of *x*- and y-dependent tuning curves both form true basis sets.
The present results relax some of these assumptions and provide a complementary point of view. When the modulatory quantity *y*varies discretely, each of its values corresponds to computing a different function of the stimulus *x*. Furthermore, the *x*- dependent tuning curves determine what functions of *x*or sensory-motor maps can be computed downstream, but there is no requirement for them to form a strict basis set. As mentioned, the discontinuous case fits better with the idea of switching between various possible maps, as if separate populations of neurons were turned on and off. This approach also highlights two important characteristics of these networks, that the modulation factors need to be nonlinear functions of context (Appendix A), and that the sensitivity to noise depends on the magnitude of the modulation (Appendix C).
Relation to other models
------------------------
An important property of networks of GM neurons is that the output units read out the correct maps using a simple procedure, a weighted sum \[[@B13]-[@B15]\]. As a consequence, the overall strategy of these networks can be described as follows: the input data are first projected onto a high-dimensional space, and the responses in this space are then combined through much simpler downstream units that compute the final result -- in the present case, *x*and *y*are the inputs and the high-dimensional space is composed of the GM responses. Interestingly, such expansion into an appropriate set of basis functions \[[@B13],[@B37]\] is the central idea of many other, apparently unrelated models. For instance, this scheme is a powerful technique for tackling difficult classification and regression problems using connectionist models \[[@B33]\]. It also works for calculating non-trivial functions of time using spiking neurons \[[@B38]\]. This strategy might constitute a general principle for neural computation \[[@B37]\]. In addition, these networks are capable of generalizing to new stimuli and are quite resistant to changes in the connectivity matrix, so they don\'t require exceedingly precise fine-tuning.
The problem of high dimensionality
----------------------------------
A crucial requirement for the above scheme of projecting the data onto a suitable set of basis responses is to cover all relevant combinations of sensory stimuli and modulatory signals in the GM array \[[@B13]-[@B15]\]. It is the potentially large number of such stimulus-context combinations that may pose a challenge for these networks, a problem sometimes referred to as the curse of dimensionality \[[@B33]\]. In terms of the antisaccade task, for example, the context could be signaled by the shape or color of the fixation spot, the background illumination of the screen, a sound, or simply by past events, as would happen if the experiment ran in blocks of saccade and antisaccade trials. Each one of these potential cues would need to have a similar modulatory effect on the sensory responses, and it is not clear how the brain could establish all the necessary connections for this. Part of the problem is that we don\'t know how many independent dimensions there are -- independence being the crucial property. For instance, the model for the antisaccade task has two contexts and requires two populations of switching neurons. More neurons are needed to deal with the version of the task that has five scales or contexts, but the number of necessary neurons does not keep growing endlessly; if the modulatory terms are chosen appropriately, a relatively small number of neurons can generalize to any scale, in effect generating an infinite number of sensory-motor maps. Of course, the key is that these are not independent, so the network can generalize. Thus, the scheme might work with realistic numbers of neurons if the number of independent context dimensions is not exceedingly large, but estimating this number is challenging.
Another possibility is to have a relatively small number of available gain modulation patterns controlled by an additional preprocessing mechanism that would link them to the current relevant cue (the color of the fixation spot, its shape, the background illumination, etc.), a sort of intermediate switchboard between possible contexts and possible gain changes. Attention has some features that fit this description -- it can select or favor one stimulus over another, it can act across modalities, and it can produce changes in gain \[[@B21]-[@B23],[@B39]\]. No specific proposals in this direction have been outlined yet, neither theoretically nor experimentally, but this speculative idea deserves further investigation.
Is exact multiplication needed?
-------------------------------
A key ingredient of the general, two-layer model is that the GM neurons must combine sensory and modulatory dependencies, *f*(*x*) and *g*(*y*), nonlinearly \[[@B13]-[@B15]\]. Results of two manipulations elaborate on this. First, when *f*and *g*were added (Equation 15) instead of multiplied, all transformations failed completely, as expected \[[@B13]\]. Second, when the sensory- and context-dependent terms were combined using other nonlinear functions (a sigmoid function, a rectification operation or a power-law; see Equations 16--18), accuracy remained approximately the same in all tasks. Results are shown in Table [1](#T1){ref-type="table"}, which compares the performance of networks that implemented different types of stimulus-context interactions but were otherwise identical. This shows that the exact form of the nonlinearity used to combine *f*and *g*is not crucial for these models.
However, in some cases a multiplication allows the synaptic connections to be learned through simple Hebbian mechanisms \[[@B14],[@B15]\], so it may be advantageous for learning. At least under some conditions, neurons combine their inputs in a way that is very nearly multiplicative \[[@B11],[@B21]-[@B23]\]. Perhaps they do so when multiplication provides a specific computational advantage.
Mixed sensory-motor activity
----------------------------
In the model for the orientation discrimination task, the level of activity of the output neurons reflects not only the evoked movement but also the difficulty of the sensory process. This is consistent with the observation that, during sensory discrimination tasks, neuronal responses in many motor areas carry information about the stimulus \[[@B40]-[@B42]\]. This activity is often interpreted as related to a decision-making process. In the discrimination model, the responses of the neurons encoding the movement toward one of the targets increased in proportion to the strength of the sensory signal linked with that target (Figs. [9a-9f](#F9){ref-type="fig"}), as observed experimentally \[[@B40]-[@B42]\]. The model was not designed to do this. It simply could not generate single, separate peaks of activity for two nearby orientations on the basis of a single feedforward step; an additional layer or additional lateral connections would be required for that. Nevertheless, when such selection mechanism is assumed to operate, remapping proceeds accurately, even when the strength of the sensory signal varies. According to the model, sensory and motor information should be expected to be mixed together when distinct, non-overlapping responses (e.g., movement to the left or to the right) are generated on the basis of small changes in a stimulus feature that varies continuously, as orientation did in this task.
Responses that depend on multiple cues
--------------------------------------
In the present framework, if sensory responses were modulated by multiple environmental cues, the responses of downstream neurons could be made conditional on highly specific contextual situations (see ref. 16). Therefore, this mechanism may also explain the capability of some neurons to drastically change their response properties in a context-dependent way. Two prominent examples are hippocampal place cells, whose place fields can be fully reconfigured depending on multiple cues \[[@B43],[@B44]\], and parietal visual neurons, which become selective for color only when behavioral context dictates that color is relevant \[[@B45]\].
In many tasks, two or more inputs are combined into conditional statements -- \'if *X*and *Y*then *Z*\'. The switching property of GM networks is useful in these situations as well. The study of abstract rule representation by Wallis and Miller \[[@B29]\] is a good example. In their paradigm, the decision to hold or release a lever depends on an initial cue and on two pictures. The cue indicates which of two rules, \'same\' or \'different\', is applied to the pictures. If the rule is \'same\', the lever is released when the two pictures are identical but not when they are different; if the rule is \'different\' the situation reverses, the lever is released when the two pictures are different but not when they are identical. To execute the proper motor action, two conditions must be checked. With the framework presented here, it is straightforward to build a model for that task; all it requires is a neural population that encodes the similarity of the pictures (i.e., is selective for matching vs non-matching pairs) and is gain modulated by the rule. Although the exact form of the modulation, for instance, whether it is close to multiplicative, is hard to infer from their data, the findings of Wallis and Miller \[[@B29]\] are generally consistent with the types of responses predicted by the model.
Experimental predictions
------------------------
Other experimental studies also include results that are consistent with gain interactions between multiple sensory cues \[[@B2],[@B27],[@B28],[@B30]\] or with gain changes due to expected reward \[[@B46]\]. Interpreting these data is problematic, however, because those experiments were not designed to test whether changes in context generate changes in gain. The tasks described here, or similar paradigms, may be simplified to eight or so stimuli and two or three conditions, generating stimulus sets that would be within the range of current neurophysiological techniques with awake monkeys. The key is to be able to construct full response curves (Figs. [2](#F2){ref-type="fig"},[5](#F5){ref-type="fig"}), so that neuronal activity across contexts can be compared for several stimuli -- not only for two, as is often done. This is because the models make three basic predictions that can only be tested with multiple stimuli and conditions: the responses should have mixed dependencies on stimulus and context, the mixing should be nonlinear, and the neurons should behave approximately as a basis-function set, in the sense that a weighted sum of their responses should approximate an arbitrary function of stimulus and context extremely well \[[@B31],[@B47]\].
Ideally, the nonlinear mixture will show up through multiplicative changes in gain, as in Figs. [2](#F2){ref-type="fig"} and [5](#F5){ref-type="fig"}, where the context-dependent variations in firing intensity respect stimulus selectivity. This could certainly happen \[[@B11],[@B21]-[@B23]\], especially for some individual neurons, but other nonlinearities are possible \[[@B19],[@B47],[@B48]\] and might work equally well. A key observation is that context can include widely different types of circumstancial information, such as expected reward, motivation, fear or social environment \[[@B1],[@B5]-[@B7]\]. Therefore, given the versatility of the models discussed here, a broader implication of the present work is the possibility that, as a basis for adaptive behavior, the brain systematically creates sensory responses that are nonlinearly mixed with numerous types of contextual signals.
Conclusions
===========
The framework discussed here demonstrates how to make a neural network adaptable to various environmental contingencies, labeled here simply as context. To achieve this flexibility, context must influence the ongoing sensory activity in a nonlinear way. This strategy was illustrated with tasks akin to those used in neurophysiological experiments with awake monkeys, but is generally applicable to the problem of executing a sensory-evoked action only when a specific set of conditions are satisfied. The mechanism works because changing the gain of multiple neurons is, in a sense, equivalent to flipping a switch that turns on and off different groups of neurons. Its main disadvantage is that all relevant combinations of stimulus and context must be covered, which may require a large number of units. On the upside, however, the switching functionality is insensitive to the quality or content of the sensory signals, is robust to changes in connectivity, and places minimal restrictions on how context is encoded. Future experiments should better characterize how cortical neurons integrate sensory and contextual information.
Methods
=======
Neuronal responses
------------------
The GM responses depend on a sensory feature *x*, which may represent stimulus location or stimulus orientation, and on a context signal *y*. For each model GM cell, these quantities are combined through a product of two factors, *f*and *g*. The former determines the sensory tuning curve of the neuron and the latter its gain or amplitude as a function of context *y*. The mean firing rate *r*~*j*~of GM unit *j*is thus written as
*r*~*j*~= *r*~*max*~*f*~*j*~(*x*) *g*~*j*~(*y*) + *B*, (1)
where *f*~*j*~and *g*~*j*~vary between 0 and 1, *B*is a baseline firing rate equal to 4 spikes/s and *r*~*max*~= 35 spikes/s. The specific functions used for *f*~*j*~and *g*~*j*~depend on the task and are described below. However, note that because these two terms are combined through a multiplication, changes in context alter the overall responsiveness of a cell, but not its selectivity, which is the defining feature of gain modulation \[[@B8],[@B9]\]. To include neuronal variability, Gaussian noise is added to all GM responses in each trial of a task. The noise is multiplicative; that is, the variance of the noise for unit *j*is equal to *αr*~*j*~, where *α*is a constant. Qualitatively similar results are obtained with additive instead of multiplicative noise.
The output neurons form a population code that represents the location of an impending eye movement. Each firing rate is determined by a weighted sum of the GM rates, where the weights correspond to synaptic connections. The mean rate of output neuron *i*is
where *w*~*ij*~is the synaptic weight from GM neuron *j*to output neuron *i*. The output rates should encode the location of the movement to be made in each trial, so their profile of activity should have a single peak indicating the location that should be reached. The synaptic connections that achieve the correct sensory-motor alignment are found through an optimal algorithm, which is described further below.
Equation 2 is used when the GM neurons drive the output neurons. But for each task, there is also an intended or desired response for each output unit. This is denoted as *F*~*i*~, and is a function of the stimulus and the context. Thus, *R*~*i*~and *F*~*i*~refer to the same postsynaptic neuron, but one is the actual driven response and the other is the intended response. The functions *f*, *g*and *F*vary across tasks, as described next.
Parameters for the antisaccade task
-----------------------------------
In this task (Fig. [1](#F1){ref-type="fig"}), stimuli appear at a location *x*in two possible contexts, labeled *y*= 1 and *y*= -1. The response should be a movement toward the location equal to *xy*. The firing rates depend on the following functions. The tuning curves are Gaussians,
with the preferred stimulus location *a*~*j*~between -25 and +25 and *σ*~*f*~= 4 (in Figs. [2](#F2){ref-type="fig"},[3](#F3){ref-type="fig"}) or *σ*~*f*~= 6 (everywhere else). Because there are two conditions, the modulatory functions *g*~*j*~take only 2 values, 1 and *γ*, where *γ*is the minimum gain. Crucially, one half of the GM neurons have *g*~*j*~(*y*= 1) = 1 and *g*~*j*~(*y*= -1) = *γ*, whereas the other half have the opposite context preference, *g*~*j*~(*y*= 1) = *γ*and *g*~*j*~(*y*= -1) = 1. The only exceptions are Figs. [3c,3f](#F3){ref-type="fig"}, in which the gain factors *g*~*j*~for each neuron were chosen randomly from uniform distributions. In this task, the desired response of output neuron *i*is
where *c*~*i*~is the preferred movement location of unit *i*and *σ*~*F*~= 4. Therefore, the output profile of activity (obtained by plotting *F*~*i*~vs *c*~*i*~) should be a Gaussian centered at the intended target location, *x*or -*x*, depending on the context.
Parameters for the scaling task
-------------------------------
The scaling task is identical to the antisaccade task, except that the context *y*can take many values. The tuning functions *f*~*j*~, are the same (Equation 3) and the output responses again encode the location given by *xy*(Equation 4). The gain factors depend on which of two possible representations is used. When context is encoded discontinuously, five scales are used, *y*= - 1, -0.5, 0, 0.5 or 1, so the modulatory functions *g*~*j*~must take five values; these are
*g*~*j*~(*y*) = {1, 0.9, 0.75, 0.65, 0.5}. (5)
Crucially, they are assigned randomly to each of the 5 conditions, with a new random permutation for each GM unit. As a final step, the *g*~*j*~values are jittered by small, random amounts (see Figs [5a,5b](#F5){ref-type="fig"}). On the other hand, when context is encoded continuously, each neuron is assigned a preferred scale *b*~*j*~between -1.4 and +1.4, and its gain is a Gaussian function of *y*,
with *σ*~*g*~= 0.3. Note that the minimum gain is 0.5 in both cases.
Parameters for the orientation discrimination task
--------------------------------------------------
In this task (Fig. [8](#F8){ref-type="fig"}), *x*is the orientation of a bar and varies between -8° and +8°, where *x*= 0° corresponds to vertical. The discrimination can occur in two ways: either left and right targets correspond to left- and right-tilted bars, respectively (*y*= 1), or viceversa (*y*= 2). In addition, there is a no-go condition (*y*= 3), for a total of three contexts. The orientation tuning curves are given by cosine functions,
where *a*~*j*~is now a preferred orientation between -90° and +90°. The modulation functions *g*~*j*~are generated as in the discontinuous version of the scaling task, except with three values,
*g*~*j*~(*y*) = {1, 0.75, 0.5}. (8)
The order in which each GM neuron prefers the three contexts is random. The responses of the motor neurons are given by
with *y*= 3 being the no-go condition and *σ*~*F*~= 4. Here, *M*(*x*, *y*) is the correct movement location, either -10 or +10, when orientation *x*is presented in condition *y*. Specifically, for *y*= 1, *M*= -10 if *x*\< 0 and *M*= +10 if *x*\> 0; and for *y*= 2, *M*= -10 if *x*\> 0 and *M*= +10 if *x*\< 0. In no-go trials, all output responses should stay at the baseline level, *B*.
Simulation results in the orientation discrimination task are presented in terms of the probability of generating a movement toward the target on the right, *P*~*R*~(*x*), which is a function of orientation. Those results are fitted to the curve
where erf is the error function. This expression has two parameters: *a*~*e*~, which is the center point, or bias, and *b*~*e*~, which is inversely proportional to the maximum slope. The discrimination threshold is defined as one half of the difference between the values of *x*that give *P*~*R*~= 0.75 and *P*~*R*~= 0.25; for Equation 10 it is equal to *b*~*e*~erfinv(1/2), where erfinv is the inverse of the error function.
Calculation of synaptic weights
-------------------------------
The synaptic weights are chosen so that, on average, the driven and desired responses of the output neurons are as close as possible. This means that
must be minimized. As in Equation 2, *w*~*ij*~is the connection from GM neuron *j*to output neuron *i*. The angle brackets indicate an average over all values of *x*and *y*and over multiple trials. The optimal connections are found by taking the derivative of the above expression with respect to *w*~*pq*~, setting the result equal to zero, and solving for the connections. The result is
where
*C*~*kj*~≡ \<*r*~*k*~*r*~*j*~\> (13)
*L*~*kj*~≡ \<*F*~*k*~(*x*, *y*)*r*~*j*~\>. (14)
Equations 12--14 are the recipe for setting the connections. Here **C**^-1^is the inverse of the correlation matrix **C**defined above. This inverse (or the pseudo-inverse) is found numerically. To calculate the averages defined above, the GM rates for all values of *x*and *y*are needed. These are found by evaluating Equation 1 plus the noise term for each GM neuron. When the noise is uncorrelated across neurons, as in the simulations, it only contributes to the diagonal of **C**. Because the variance of the noise is equal to *α*times the mean response, noise adds an amount *α*\<*r*~*i*~\> to element *C*~*ii*~of the correlation matrix. Except for this, all averages *C*~*kj*~and *L*~*kj*~are obtained from the mean input responses given by Equation 1 and the corresponding *F*~*i*~functions of the output neurons.
Having specified the parameters of the network (number of GM and output units, tuning and gain functions, stimulus-movement association), the procedure for setting the synaptic weights is run only once. Afterward, the connections are not adjusted any further.
Other response functions
------------------------
Equation 1 is based on an exact multiplication between *f*~*j*~(*x*) and *g*~*j*~(*y*). The effects of other possible interactions between stimulus and context are investigated using four alternative expressions in place of Equation 1. First, a linear combination of sensory and context signals,
Then, three nonlinear interactions. The first one is based on rectification,
*r*~*j*~= *r*~*max*~\[*f*~*j*~(*x*) + *g*~*j*~(*y*) - 1\]~+~+ *B*, (16)
where \[*x*\]~+~= max{0, *x*}. The second one uses a sigmoid function,
The sigmoid is widely used in artificial neural networks \[[@B12]\] and has two parameters, *a*~*s*~and *b*~*s*~. The third nonlinear interaction is based on a power law,
and has two parameters too. This type of expression approximates some of the gain effects observed experimentally \[[@B49]\]. The free parameters in these expressions are adjusted so that the resulting firing rates are as close as possible to those given by Equation 1. All else is as in the original simulations.
Outline of the simulations
--------------------------
Having specified a task, the tuning and gain curves of the GM neurons, and the network connections, the model is tested in a series of trials of the task. Each trial consists of the following steps: (1) specifying the stimulus and context, *x*and *y*, (2) generating all GM responses (Equation 1), (3) calculating the driven, output responses (Equation 2), and (4) determining the encoded movement *M*~*out*~by using the center of mass of the motor activity profile (Equation 19). Finally, the encoded movement is compared to the movement *M*~*desired*~that should have been performed given *x*and *y*-- their difference is the error in that particular trial.
The encoded movement *M*~*out*~is equated with the center of mass of the output population,
where *c*~*i*~is the preferred target location of output unit *i*. The root-mean-square average of the motor error is used to quantify performance over multiple trials,
where only go trials are included in the calculation. On average, the encoded movement is very near the desired one, \<*M*~*out*~- *M*~*desired*~\> ≈ 0. Thus, *σ*~*CM*~is the standard deviation of the motor error, and measures the accuracy of the output population as a whole.
In all tasks, 25 output units are used, with *c*~*i*~uniformly spaced between -25 and 25. Preferred stimulus values *a*~*j*~and preferred context values *b*~*j*~are first distributed uniformly and then jittered by small, random amounts.
All simulations were performed using Matlab (The Mathworks, Natick, MA). The source code is available on request.
Appendix A
==========
This section shows that, with a finite number of contexts, gain modulation is functionally equivalent to a switch. More precisely, for a discrete number of contexts and everything else being equal, a network of partially modulated neurons can generate the same mean downstream responses as a network of switching neurons.
Consider *M*populations or groups of sensory neurons with identical sets of tuning functions *f*~*j*~(*x*). There are *N*neurons in each population, so index *j*runs from 1 to *N*. These populations project to a postsynaptic neuron through synaptic connections , where the superscript indicates the presynaptic population of origin. Thus,  is the synaptic weight from neuron *j*in group *p*to the postsynaptic unit. The sensory neurons are gain modulated, so the mean response of unit *j*in population *p*is given by
where *x*and *y*label the stimulus and the context, as before. Next, assume that there are *M*possible contexts, so *y*can take integer values from 1 to *M*. Therefore, the gain factors can be expressed as three-dimensional arrays, and the presynaptic firing rates can be rewritten as
Here,  corresponds to the gain of unit *j*in population *p*during context *k*. With this notation, the response of the downstream neuron during context *k*becomes
where the sums are over all populations and all units in each population. Note that, for each index *j*, the coefficient in front of the tuning function is given by the product of an *M*-dimensional vector of weights times an *M*× *M*matrix of gain factors.
The essential idea is to compare the response of the postsynaptic unit under two conditions: when only one input population is active in any particular context (and all others are fully suppressed), and when the populations are only partially suppressed, with different combinations of gain factors for each context. For this, the hat symbol \^ is used to label all quantities obtained in the former case, with switching neurons; that is, the hat means \'obtained with full modulation\'.
Full modulation occurs when the matrix of gain factors for all the units with index *j*is equal to the identity matrix,
According to this expression, for population 1, the gain is 1 when *y*= 1 and is 0 otherwise; for population 2, the gain is 1 when *y*= 2 and is 0 otherwise, and so forth. Substituting this expression in equation 23 gives the firing rate of the downstream unit when driven by switching neurons,
Again, the hat simply indicates that the quantity was obtained with maximally modulated input neurons. In this case, (*x*, *y*= *k*) implements a different function of *x*for each context value, such that the function expressed in context 1 depends only on the weights from the first population of switching neurons, , the function expressed in context 2 depends only on the weights from the second population, , and so on. This is the situation depicted in Figs. [3a,3d](#F3){ref-type="fig"}.
On the other hand, the postsynaptic response driven by partially modulated units is simply as in Equation 23, where the absence of a hat means \'obtained with partial modulation\'. Under what conditions is the output response driven by partially modulated neurons, *R*(*x*, *y*= *k*), equal to the response obtained with full modulation, (*x*, *y*= *k*)? Compare the right hand sides of Equations 23 and 25; for them to be the same, the coefficients in front of the tuning functions must be equal; that is,
This condition is satisfied if the weights with partial modulation are set equal to
where **h**~*j*~is the inverse of the matrix of gain factors **g**~*j*~; that is, . Therefore, the key constraint here is that the gain factors in the partial modulation case must have linearly independent values across contexts, so that the inverses exist; in other words, the matrices **g**~*j*~must have full rank. An important consequence of this is that for *M*\> 2, the gain of each neuron as a function of context ( in Equation 21) must be nonlinear.
Equation 27 is the key result. It provides a recipe for going from a network of switching neurons to a network of partially modulated neurons (given equal numbers and types of tuning functions *f*~*j*~(*x*)). For the recipe to apply, the gain factors in the latter must have the appropriate inverses, but otherwise they are arbitrary. Because each possible function that a network can generate corresponds to a different matrix of synaptic connections, this implies that all the possible functions of *x*that the output can implement with fully switching neurons can be replicated with partial gain modulation.
This statement is exact when there is no noise; with noise it applies to the average downstream responses. Notice that this result is independent of the tuning functions *f*~*j*~(*x*). These determine the possible functions of *x*that can be generated downstream -- that is, the available sensory-motor maps -- but have no effect on how these are switched or selected. Finally, the result is also valid if the postsynaptic response is equal, not simply to the weighted sum of GM responses, but to an arbitrary function of that sum.
Appendix B
==========
To illustrate the result in Appendix A, consider a simple case with two populations and two contexts, as in Figs. [1](#F1){ref-type="fig"}, [2](#F2){ref-type="fig"}, [3](#F3){ref-type="fig"}, [4](#F4){ref-type="fig"}. With full modulation, the response of the downstream neuron is
in context 1, and
in context 2. This is simply Equation 25 for *M*= 2. Context turns one sensory population on and another off. Now, how can we obtain the same downstream responses, as functions of *x*, when the GM neurons are partially modulated? First, suppose that the modulation matrices are
The gain factors can only take two values, 1 for the preferred context, and 1 \>*γ*≥ 0 for the non-preferred one; the full-modulation case is recovered when *γ*= 0. For simplicity, these factors are the same for all units in each population, so there is no variation across index *j*. This matrix was used in Fig. [3](#F3){ref-type="fig"}, with *γ*= 0 and *γ*= 0.5 for the left and middle columns, respectively, and in Figs. [4a,4b](#F4){ref-type="fig"}. Its inverse is
Next, substitute into the transformation rule found earlier, Equation 27; the result is
With these synaptic weights, the downstream responses driven by the partially modulated neurons (Equation 23 with *p*= 1, 2 and the gain factors in Equation 30) become identical to the rates driven by the switching neurons (Equations 28,29). This is the linear transformation used in Figs. [4a,4b](#F4){ref-type="fig"}.
Appendix C
==========
Using partial instead of full modulation to switch between maps does come at a price: the variability of the postsynaptic response typically increases. This can be seen as follows.
If additive noise is included in the input firing rates, the response of neuron *j*in population *p*becomes
where  is a random fluctuation for unit *j*in population *p*during context *k*. The variance across trials of this random variable is denoted as  and is the same for all GM neurons. The downstream neuron has the same mean response as before (Equation 23), but now it has a variance, which is equal to
Here, the angle brackets indicate an average over trials, which affects the noise terms only. To go from the second to the third line above, the key is to assume that the fluctuations are independent across neurons, such that .
The next step is to compare the variance of the postsynaptic unit when driven by the switching neurons and by the regular, partially modulated GM neurons. For simplicity, consider the same 2 × 2 case as in Appendix B, where the modulation is parameterized by *γ*. The variance  of the postsynaptic response driven by switching neurons is exactly as in Equation 35, but with  and *p*= 1,2. This must be compared to the variance obtained with partial modulation for the same mean postsynaptic responses. The synaptic weights that achieve this are given by Equations 32 and 33; substituting those into Equation 35 gives
This is the variance of the postsynaptic response driven by an array of partially modulated GM neurons as a function of the variance  obtained when the response is driven by fully modulated, switching neurons. Here, *a*depends on the weights , but is not a function of *γ*,
where *b*is a constant. Note that *a* is a measure of the overlap between the sets of connections from the two populations. The dependence of  on the weights is such that *a*≤ 2.
Equation 36 shows that, although the average postsynaptic response is the same function of *x*and *y*for all *γ*, its variability changes with *γ*. A similar result is obtained when the variance of the input firing rates is proportional to their mean. In that case,
with  = 1. A calculation analogous to the one just described leads to
where *a*is the same as in Equation 37, except with a different proportionality constant. In this case, it is still true that *a*≤ 2. This expression was used to generate the continuous lines in Fig. [4a](#F4){ref-type="fig"}. For this,  was simply the variance in the postsynaptic firing rate found from the simulations with *γ*= 0, and *b*was chosen to generate the best fit to the rest of the simulation data points.
Equations 36 and 39 do not always increase monotonically with *γ*. This depends on *a*, which is a measure of the similarity between the sensory-motor maps established in the two contexts. For instance, when the two maps are the same,  for all *j*, and *a*attains its maximum value, 2. In that case, the variance with partially modulated neurons is always smaller than with switching neurons. This makes sense: if the maps in the two contexts are the same, it is always better to have the two populations active at the same time, as this reduces the noise. According to the analysis, when *a*= 2 and *γ*= 1, Equation 36 gives . The variance is divided by 2 because noise is additive and there are two active populations doing the very same thing. In contrast, when the two maps are different, their respective synaptic weights are also different, and *a* is either positive but much smaller than 2, or negative. Then,  might have a minimum for some intermediate value of *γ*, or may increase monotonically, which is what happens in Figs. [3](#F3){ref-type="fig"} and [4](#F4){ref-type="fig"}, with saccades vs antisaccades.
List of abbreviations
=====================
GM, gain-modulated.
Acknowledgements
================
I thank Terry Stanford and Nick Bentley for useful discussions, Sacha Nelson for suggesting the calculation in Appendix C, and two anonymous reviewers for their comments. Research was supported by NINDS grant NS044894.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Antisaccade task.**In each trial, a stimulus (black dot) is presented at a distance *x*from the fixation point (colored dot); the stimulus disappears; two targets appear (gray dots) and the subject responds by making an eye movement (arrow) to one of them. The color of the fixation spot indicates whether the movement should be a saccade or an antisaccade. **a**: In context 1 the fixation spot is red and the movement is to the target at *x*. **b**: In context 2 the fixation spot is green and the movement is to the opposite target, at -*x*. In the model, *x*is between -15 and +15, with distance in arbitrary units.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Responses of GM neurons in the antisaccade task.**Each graph plots the mean firing rate of a model neuron as a function of stimulus location. Red and green traces correspond to sensory responses evoked during contexts 1 and 2, respectively. The gain in the preferred condition is 1. **a**: A unit that prefers context 1 and is 100% suppressed in the non-preferred condition; its minimum gain is *γ*= 0. **b**: A unit that prefers context 2 and is 62% suppressed in the non-preferred condition; its minimum gain is *γ*= 0.38. **c**: Another unit that prefers context 1; its minimum gain is *γ*= 0.61. **d**: Another unit that prefers context 2; its minimum gain is *γ*= 0.87. Firing rates are in spikes/s. Model responses are based on Equations 1 and 3.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Network performance in the antisaccade task. a**: Firing rates of all model cells when the GM units are fully modulated (*γ*= 0). The box marks a single trial. Colored traces are the firing rates of the 60 GM neurons in the network; 30 of them (red) prefer the direct saccade condition and 30 (green) prefer the antisaccade condition. Black dots are the 25 motor responses driven by the GM neurons. For GM responses, x-axis is preferred stimulus location; for output responses, x-axis is preferred movement location. Context in each of the four trials is indicated on the left. Trials with *x*= -15 and *x*= 10 alternate. The profile of output activity always peaks at the correct location. **b**: As in **a**, but when all GM neurons are partially modulated by the same amount (*γ*= 0.5). **c**: As in **a**, but when the maximum and minimum gains of the GM units are chosen randomly from uniform distributions. **d-f**: Connection matrices for the networks in the respective columns. Each point shows the synaptic weight, coded by color, from one GM neuron to one output neuron. GM units 1--30 (red points in upper panels) prefer context 1, whereas GM units 31--60 (green points in upper panels) prefer context 2. No noise was included in the simulations (*α*= 0).
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Sensitivity to noise as a function of modulation strength in the antisaccade task.**All results are for networks of 60 GM and 30 output neurons. The x-axes indicate *γ*, which is the minimum gain of the GM neurons; the maximum is always 1. In all panels, the three curves are for three levels of noise: *α*= 0.04 (thin lines), *α*= 0.36 (medium lines), or *α*= 2.25 (thick lines). **a**: Standard deviation of single output firing rates, averaged over stimulus locations and contexts, as a function of *γ*. Data points are from simulations; continuous lines are analytic results from Equation 39, with *a*= 1.42. For each data point, the average output responses, as functions of *x*and *y*, were the same. To achieve this, the synaptic weights for *γ*\> 0 were obtained by a linear transformation of the weights for *γ*= 0 (Appendices B, C). **b**: Error between correct and encoded movement locations as a function of *γ*. Results are from the same simulations as in **a**. **c, d**: As in **a, b**, respectively, but for simulations in which the synaptic weights were computed using the standard, optimal algorithm (see Methods). Note that *σ*~*CM*~always increases with *γ*.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Responses of four model GM neurons in the scaling task.**Same format as in Fig. 2, except that there are five possible contexts, corresponding to *y*= -1, -0.5, 0, 0.5 and 1. **a**: Tuning curves for a model neuron that responds maximally to stimuli at *x*= -15 and prefers the green condition (*y*= 0). The order of effectiveness for the five scales was set randomly, so context is encoded discontinuously. **b**: As in **a**, but for another neuron that prefers *x*= 2 and *y*= 1. **c**: Tuning curves for a model neuron that encodes context in a smooth, continuous way. The unit responds maximally to stimuli at *x*= -1 and prefers the cyan condition (*y*= -0.5). The gain of the cell decreases progressively as *y*deviates from the preferred scale -- note the order of the colors. **d**: As in **c**, but for a neuron that prefers *x*= 9 and *y*= 1. All units have a maximum gain of 1 and a minimum gain near 0.5. Model responses were based on Equations 1, 3, 5 and 6.
:::
:::
::: {#F6 .fig}
Figure 6
::: {.caption}
######
**Network performance in the scaling task.**Results are from two networks, one that encodes context discontinuously (first two columns) and another that encodes it continuously (third and fourth columns). **a**: The box encloses all model responses in a single trial with *x*= -5 and *y*= 1. The color plot shows all 900 GM responses, color coded. Neurons are arranged by preferred stimulus location along the x-axis and by preferred context along the y-axis. Black traces are the firing rates of the 25 driven output neurons. The black line indicates intended target location (*xy*= -5); the red line indicates encoded target location (center of mass). Their difference (error) is -0.73. **b**: A trial with *x*= -5, *y*= -1 and error = 0.02. **c**: A trial with *x*= 15, *y*= -1 and error = -1.54. **d**: A trial with *x*= 15, *y*= 0.5 and error = 0.63. **e-h**: Same combinations of stimulus and context as in **a-d**, but using a smooth representation for context. Errors are -0.07, 0.34, -1.84, and 0.8, respectively. The variance of each GM rate is equal to its mean (*α*= 1).
:::
:::
::: {#F7 .fig}
Figure 7
::: {.caption}
######
**Robustness and generalization in the scaling task.**Left and right columns are for networks in which scale is encoded discontinuously and continuously, respectively. Each panel shows results for three noise levels: *α*= 0.09 (squares), *α*= 1 (circles) and *α*= 9 (triangles). **a**: Error in encoded movement location as a function of the number of GM neurons. Each point represents an average over stimulus locations, scales and trials; 31 stimulus locations and 5 scales were used both to set the connections and test the networks. The filled symbol indicates the network in Figs. 6a-d. **b**: As in **a**but for networks in which scale is encoded continuously. The filled symbol indicates the network in Figs. 6e-h. **c, d**: *σ*~*CM*~vs network size in networks with corrupted synaptic weights. These simulations proceed as in **a, b**, except that performance is tested after deleting 25% of the synaptic connections, chosen randomly. **e**: *σ*~*CM*~vs network size when only 8 stimulus locations (combined with the 5 scales) are used to set the connections and the network is tested with all combinations of 31 stimulus locations and 5 scales. **f**: *σ*~*CM*~vs network size when combinations of only 8 stimulus locations and 8 scales are used to set the connections and performance is evaluated with all combinations of 31 stimulus locations and 31 scales. Continuous lines are linear fits to the data points above 250 units. Note logarithmic axes.
:::
:::
::: {#F8 .fig}
Figure 8
::: {.caption}
######
**Orientation discrimination task.**In each trial, a bar oriented at an angle *x*is presented while the subject fixates; the stimulus disappears; two targets appear (gray dots), and the subject indicates whether the bar was tilted to the left or to the right by making an eye movement (horizontal arrow). Vertical bars correspond to *x*= 0°. **a**: The bar is tilted to the left (*x*\< 0). With a red fixation spot (*y*= 1), responses to left- and right-tilted bars should be to the left and right targets, respectively. The correct response is thus to the left. **b**: With a green fixation spot (*y*= 2), left- and right-tilted bars correspond to right and left targets, respectively. The correct response is now to the right. A no-go condition (*y*= 3; not shown) is included in the simulations in addition to the two go conditions. Orientation is in degrees, with *x*between -8° and +8°.
:::
:::
::: {#F9 .fig}
Figure 9
::: {.caption}
######
**Network performance during orientation discrimination.**Panels **a-h**show all 25 output responses, as driven by the GM neurons (not shown), in single trials. Continuous lines indicate intended target location (-10 or +10); dashed lines indicate the center of mass of the output activity. Right and left columns have identical stimuli and conditions, but with (*α*= 1) and without (*α*= 0) noise, respectively. A trial is deemed correct if the higher peak of activity is situated at the intended target location. **a**: Single trial with *x*= 5°, *y*= 1 and error = 0.15. **b**: As in **a**, but error = 0.001. **c**: Single trial with *x*= 1°, *y*= 1 and error= 6.7. **d**: As in **c**, but error = 12.2. The response is scored as incorrect because the tallest hill of activity is not at the intended target. **e**: Single trial with *x*= 1°, *y*= 2 and error = 6.7. **f**: As in **e**, but error = -2.9. **g, h**: No-go trials. **i**: Probability of making a movement to the right target as a function of stimulus orientation, in condition 1 (*y*= 1) and with noise. Gray lines are fits to the simulation data. The center point or bias of the fit is indicated by the dashed line and is equal to -0.06°. Discrimination threshold is 1.5°. **j**: As in **i**, but with the association between orientation and targets reversed (*y*= 2). The center point is -0.04°; the discrimination threshold is 1.4°.
:::
:::
::: {#F10 .fig}
Figure 10
::: {.caption}
######
**Robustness and generalization in the orientation discrimination task.**Left and right columns show the bias and discrimination threshold, respectively of the neurometric fits (as in Figs. 9i,j) as functions of network size. **a, b**: Bias and discrimination threshold under standard conditions, which include 64 orientations used to set the connections and test the model. For each network size, results are absolute values averaged over the two go conditions and multiple networks. Filled symbols indicate the network used in Fig. 9. **c, d**: As in **a, b**, except that performance was tested after deleting 25% of the synaptic connections, chosen randomly. **e, f**: As in **a, b**, but when only 2 stimulus orientations (-8° and +8°) are used to set the connections, in combination with the 3 possible contexts, and performance is tested with all 64 orientations and 3 contexts. Straight lines are fits to the data points above 250 units.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Errors in performance for various possible interactions between stimulus and context.
:::
**Error/Task** *fg* *f*+ *g* \[*f*+ *g*- 1\]~+~ sig (*f*+ *g*)
------------------------- ------- ---------- -------------------- ---------------- -------
(*σ*~*CM*~/Scaling (DE) 0.60 6.3 0.50 0.62 0.66
(*σ*~*CM*~/Scaling (CE) 0.60 5.5 0.51 0.61 0.69
Bias/Orientation 0.03° 36° 0.05° 0.04° 0.06°
DT/Orientation 1.39° 37° 1.09° 1.44° 1.88°
Functions *f*and *g*are the sensory- and context-dependent terms used to generate the GM responses; *fg*is the standard condition in which these functions are multiplied (Equation 1). Other combinations: *f*+ *g*, linear interaction (Equation 15); \[*f*+ *g*- 1\]~+~, rectification (Equation 16); sig(*f*+ *g*), sigmoidal function (Equation 17); and , power function (Equation 18). Network parameters were as in Figs. 6 and 9a-f, for the corresponding tasks. For each row, all numbers were generated using exactly the same model parameters, except for the specific combination of *f*and *g*terms. DE, discontinuous encoding; CE, continuous encoding; DT, discrimination threshold.
:::
|
PubMed Central
|
2024-06-05T03:55:52.049560
|
2004-11-25
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546005/",
"journal": "BMC Neurosci. 2004 Nov 25; 5:47",
"authors": [
{
"first": "Emilio",
"last": "Salinas"
}
]
}
|
PMC546006
|
Background
==========
Chronic heart failure (CHF) is a major cause of morbidity and mortality and it has a considerable impact on the health care system \[[@B1]\]. In a recent study, the prevalence in Sweden was estimated at 1.3--2.5% \[[@B2]\]. Early detection of CHF has become increasingly important, as modern drug treatment has the potential to improve symptoms and quality of life, slow down the rate of disease progression, and improve survival. However, diagnosing CHF is known to be difficult, especially in mild cases, as many features of the condition are not organ specific, and there may be few clinical features in the early stages of the disease \[[@B3]-[@B5]\]. Most of the patients are old, which also makes the diagnosis difficult. Older patients may have atypical symptoms, they may suffer from other diseases, and they may be on treatment that modifies their symptoms \[[@B3]\]. Diagnosing CHF has been found to be especially difficult in women and in obese patients \[[@B4]\]. A large proportion of patients with CHF are managed by general practitioners (GPs), especially older patients and patients early in the course of disease, i.e. those patients for whom the diagnostic process is characterized by the greatest uncertainty \[[@B6]\].
The European Society of Cardiology adopted guidelines for diagnosing CHF in 1995, and these were revised in 2001 \[[@B7]-[@B9]\]. Swedish guidelines, based on the 1995 European guidelines, were published in 1996 by the Swedish Medical Products Agency \[[@B10]\]. However, guidelines are often not easily or accurately integrated into daily practice \[[@B11],[@B12]\].
The full versions of the above-mentioned guidelines are comprehensive documents, covering epidemiology, aetiology, pathophysiology and diagnostic methods, but may be difficult to apply to specific diagnostic situations \[[@B13]\]. However, the recommendations are summarized in 1) a definition, 2) an algorithm for the diagnosis of CHF, and 3) a table of assessments to be performed routinely to establish the presence of CHF. The definition includes three criteria: *a*) one or more typical symptoms (at rest or during exercise), *b*) objective evidence of cardiac dysfunction (at rest), and *c*) response to treatment directed towards CHF (in cases where diagnosis is in doubt). Criteria *a*and *b*should be fulfilled in all cases. Echocardiography (ECHO) is mentioned as the single most effective tool in widespread clinical use for objective assessment of cardiac dysfunction. In the algorithm for the diagnosis of CHF, a sequence of investigations is recommended: suspect CHF because of symptoms and signs; assess presence of cardiac disease by electrocardiography, X-ray or Natriuretic peptides (where available); imaging by echocardiography; assess aetiology, degree, precipitating factors and type of cardiac dysfunction; additional diagnostic tests where appropriate; choose therapy. Table [1](#T1){ref-type="table"} shows the assessments to be performed routinely \[[@B9]\]. In the present study, the list of assessments recommended in Table [1](#T1){ref-type="table"} was used for evaluation of the GPs\' diagnostic reasoning. For most Swedish GPs, the main source of knowledge regarding CHF diagnostics is probably locally adapted protocols developed by cardiologists, or by cardiologists and GP representatives in collaboration.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Diagnostic assessments according to guidelines Assessments to be performed routinely to establish the presence and likely cause of heart failure (Eur Soc Cardiol 2001).
:::
Assessments The diagnosis of heart failure Suggests alterantive or additional diagnosis
---------------------------------------------------------------------------------------- -------------------------------- ----------------------------------------------- ----------------- -----------------------------------
Appropriate symptoms +++ +++ (if absent)
Appropriate signs +++ \+ (if absent)
Cardiac dysfunctioning on imaging (usually echocardiography) +++ +++ (if absent)
Response of symptoms or signs to therapy +++ +++ (if absent)
Electrocardiography +++ (if normal)
Chest X-ray \+ (if pulmonary congestion *or*cardiomegaly) \+ (if normal) Pulmonary disease
Full blood count\* Anemia/secondary polycythemia
Biochemistry and urinalysis\* Renal or hepatic disease/diabetes
Plasma concentration of natriuretic peptides in untreated patients (where available)\* \+ (if elevated) +++ (if normal)
\+ = of some importance; +++ = of great importance
\*) Recommended assessments which are *not*included in the corresponding table in the Medical Products Agency version of guidelines from 1996.
:::
Relatively few studies on how patients suspected of having CHF are diagnosed have been performed in primary health care settings, and most of them report over-diagnosis \[[@B3],[@B4],[@B14]-[@B16]\]. In the present study we used written case vignettes (case descriptions) and think-aloud technique to study how GPs\' diagnostic reasoning and diagnostic judgements about patients with suspected CHF are related to the recommendations in the European guidelines \[[@B9]\]. What clinical information is considered important by the GPs in the sense that it is used as an argument for or against the diagnosis of CHF? What information that is considered important for diagnosing CHF in the guidelines is also considered important by the GPs?
Methods
=======
Think-aloud method
------------------
Process-tracing techniques are used to study the cognitive processes involved in decision-making such as, for example, how judgements change over time as new information is presented, and which decision rules are used \[[@B17]\]. A method that is often used to describe the sequence of thoughts behind decision-making is the think-aloud technique \[[@B18]\]. Subjects are instructed to say their thoughts aloud while performing a task, and the verbal reports are usually audio-taped, transcribed to a written form, and then analysed. The think-aloud technique has been used in a number of studies in the field of medical decision-making \[[@B13],[@B19]\].
The value of conclusions reached in such studies depends on the validity of the think-aloud method, and on the reliability of the coding process. Thinking aloud while performing a task often lengthens the time for completing the task, but does not seem to change the accuracy of task fulfilment or the cognitive processes \[[@B18]\]. In a recent study we found that think-aloud data were at least as valid as ratings in describing a clinical decision process \[[@B20]\].
Participants
------------
All health care centres in northern Stockholm within a distance of 20--30 km from the city centre (*n*= 61) were listed and contacted in a random order. The distance from central Stockholm was chosen for practical reasons. In each health care centre the GPs were contacted in a random order by one of the authors (YS). Only one GP at each centre was included in the study, and this person had to be a specialist in family medicine. We contacted the GPs during their regular telephone hour, during the period October 2001 to October 2002. Our goal was to include 15 GPs in the study.
A total of 30 GPs were reached, and 15 agreed to participate. Those who declined to participate were not asked why they did so, but the majority of those who spontaneously gave a reason mentioned a heavy workload. The participants had been specialists in family medicine for an average of 14.8 (range 3--25) years, they were on average 52.7 (range 42--62) years of age, and six of them were men. The non-participating GPs were on average 52.7 (range 35--62) years of age, and seven of them were men.
Case vignettes
--------------
Six case vignettes (CV), based on authentic patients, were presented to the participants. The information presented in the case vignettes was obtained from the patient records and included information about relevant diseases (e.g. coronary heart disease, stroke, diabetes), lifestyle factors (e.g. smoking, alcohol consumption), symptoms, signs, electrocardiography (ECG), chest X-ray findings, and ECHO. Chest X-ray and ECHO results were presented in the same format as in the patient records. ECHO results could contain information about ejection fraction (EF), valvular disorders and ventricular wall motility. The diagnoses made by the attending cardiologists (based on all available clinical information, including ECHO) were used as a reference standard when assessing the participants\' diagnostic accomplishments.
The six cases were selected to represent patients with various types of potential diagnostic problems: A \"prototypical\" CHF patient (CV2), a patient with both CHF and chronic obstructive pulmonary disease (COPD) (CV6), a patient with CHF, tachycardia and mitral valve insufficiency (CV3), an obese non-CHF patient with normal ECG and EF (CV5), a non-CHF patient with COPD (CV4) and a non-CHF patient with alcohol abuse and a metabolic syndrome (CV1). [Additional file 1](#S1){ref-type="supplementary-material"} shows some of the characteristics of the six cases.
For one of the cases (CV3) there was a disagreement between the diagnosis according to the cardiologists and the diagnosis that could be deduced from a simplistic interpretation of the guidelines. This patient had typical clinical findings including gallop rhythm, cardiomegaly, and pulmonary congestion, but normal left ventricular function according to ECHO. It could therefore be categorized as not CHF according to the definition given in the guidelines. However, this patient also had a mitral valve insufficiency, which can give a \"false normal\" ejection fraction value: the left ventricle is emptied both forward (cardiac output) and backward (leakage through the mitral valve).
Procedure
---------
Before the sessions the GPs had received written information about the aim of the study (to study clinical judgements) and about the method (think aloud), but not about the kind of medical problems that would be presented to them. The study was conducted at the GPs\' offices. All visits and recordings were made by one of the authors (YS). The participants were instructed that six authentic patients, suspected by GPs to have CHF, would be presented, and that their task was to say aloud their thoughts about the case, and to try to decide whether the patient had CHF or not.
The order of the cases was the same for all participants. The order in which the information was presented was arranged to be as realistic as possible in relation to clinical practice (first history and symptoms, then findings, and then results of investigations). Each vignette was presented on a computer screen in five successive steps using QA software \[[@B21]\]. All previously shown information about a case was repeated at the top of the later screens in a different colour to reduce and control for memory effects. The participants could control the shift to a new screen by clicking with the mouse on a continue button at the bottom of the screen. After all the information had been presented, the participants were asked, on the sixth screen, to summarize their judgements about the case and to try to decide about the diagnosis. The doctors could express their diagnostic judgements freely, with their own words.
The doctors first got a test case (not recorded) in order to get acquainted with the think-aloud method, and then continued with the six study cases. The only intervention from the researcher during the think-aloud session was that a participant who was silent for more than about 15 seconds was reminded to say his or her thoughts aloud about the information presented \[[@B18]\]. All sessions were recorded and transcribed by a secretary.
Response measures and coding of data
------------------------------------
### Coding of variables in the case vignettes
All information in the case vignettes that was of relevance for the diagnosis and that could take on different values was considered to be variables. Fifty variables were defined: 19 of them were included in all six vignettes (e.g. symptoms, signs and investigations mentioned in the guidelines), six in five, one in four, six in three, one in two, and 17 in one vignette (e.g. alcohol abuse, history of a bypass operation, and panic disorder). For each case vignette, the presented variables were coded for content and value (Table [2](#T2){ref-type="table"}).
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Variabel codings
:::
Information as presented in the case vignette Content Value
----------------------------------------------- -------------- -----------------------------------------------
\"Shortness of breath on level\" Dyspnoea Positive (presence of finding)
\"Pathological R-progression on ECG\" ECG Positive (pathology)
\"He has not had swollen legs\" Oedema Negative (absence of finding)
\"Regular rhythm\" Rhythm Negative (normality)
\"Relative heart volume 630 ml/m^2^\" Heart volume 630 (numeric values as presented in the text)
Examples of coding of variables in the case vignettes. Positive value = precense of finding, or pathological finding, negative value = absence of finding, or normal finding.
:::
### Coding of think-aloud protocols
For each participant, every mention of a variable was coded for how the GP seemed to use it: as an argument for the diagnosis of CHF, as an argument against CHF, or as not being of any explicit use for the diagnosis (mentioned only). \"He has basal rales. This guy has CHF!\" is an example of a participant using the variable \"rales\" (positive value) as an argument for CHF. \"So I\'m not really sure that he has got CHF. Just a moderate cardiac enlargement, no, I wouldn\'t think so\" is an example of a participant using the variable \"relative heart volume\" (value 630 ml/m^2^) as an argument against CHF. For each participant, a specific evaluation of each variable value was only counted once for each case vignette in order not to give more weight to thoughtful repetitions of an argument than to a single, firm statement. However, if a participant used the same variable value as an argument both for and against the diagnosis of CHF, both evaluations were coded. Ten percent of the 90 case vignette protocols were selected at random and coded independently by two of the authors (YS, LB) to estimate the interrater agreement of the coding process. The rest of the protocols were coded by one of the authors (YS).
### Comparing think-aloud protocols with guidelines
The list of diagnostic assessments recommended in the guidelines (Table [1](#T1){ref-type="table"}) was used for comparing GPs\' diagnostic reasoning with the guidelines. Breathlessness, ankle swelling, and fatigue are mentioned in the guidelines as appropriate symptoms, and leg oedema, tachycardia, gallop rhythm, and pulmonary crepitations (rales) as appropriate signs. (Neck vein distension and liver enlargement are also mentioned, but these signs were not present in the case vignettes.) Use of the variables in relation to recommendations was analyzed for frequency among GPs and case vignettes.
### Classification of diagnostic judgements
The participants were not forced to express their diagnostic judgements in a specific format, and their free verbal statements therefore had to be interpreted and coded. Two of the authors (YS, LB) independently classified all the diagnostic judgements (*n*= 90) in three categories: CHF or probably CHF; uncertainty about diagnosis; probably not CHF or not CHF.
Analyses
--------
Stata 8.0 was used for the statistics. Cohen\'s kappa test (κ) was used to determine interrater agreement regarding the coding of the think-aloud protocols and the classification of the diagnostic judgements. Kappa values are classified as follows: \<0, worse than chance; 0 to 0.2, poor; 0.21 to 0.4, fair; 0.41 to 0.6, moderate; 0.61 to 0.8, good; and \>0.8, very good \[[@B22]\].
The research ethics committee of Huddinge University Hospital approved the study.
Results
=======
Reliability -- interrater coding agreement
------------------------------------------
### Think-aloud protocols
The randomly selected test protocols contained 322 segments of propositions, 36 of which were excluded since they contained variables that were not going to be investigated in this study (e.g. treatment suggestions). There was disagreement between the two coders about the content of variables in 14 of the remaining 286 segments (4.8%). The remaining 272 segments were then tested for interrater agreement on argument values (for CHF; against CHF; just a mention), which was 95% (κ 0.85).
### Diagnostic judgements
The interrater agreement was 92% (κ 0.85). For the few diagnostic judgements where there was initial disagreement, it was possible to agree upon an interpretation.
Diagnostic reasoning
--------------------
### Assessments to be performed routinely according to guidelines
The information that was used most frequently in diagnostic arguments was the ejection fraction value on ECHO, pulmonary congestion, and cardiac volume (Figure [1](#F1){ref-type="fig"}). The most frequent argument for CHF was pulmonary congestion on chest X-ray, and the most frequent argument against CHF was the ejection fraction value.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**The most frequently used diagnostic arguments**The ten most frequently used arguments making use of different categories of clinical information. Number of arguments favouring the diagnosis CHF and number of arguments against CHF are given for each category. One variable (indicated by \*) was only presented in five of the vignettes.
:::
:::
Symptoms and signs were not often used as arguments in the GPs\' diagnostic reasoning (Figure [1](#F1){ref-type="fig"}). Symptoms were most frequently used as diagnostic arguments when reasoning about CV2 ([Additional file 1](#S1){ref-type="supplementary-material"}), which represented the prototypical CHF case, with dyspnoea when walking on level ground and orthopnea (\"in need of three pillows to be able to sleep\"), and about CV5, which represented the prototypical non-CHF case (absence of dyspnoea). Signs were most frequently used for two of the case vignettes. In CV1, the presence of rales was used by nine of the GPs as an arguments for CHF (eight of them incorrectly ending up with this as the final diagnosis), and in CV3, tachycardia was used by nine of the GPs as an argument for CHF (eight of them correctly ending up with this as the final diagnosis). CV3 also had a gallop rhythm, which is reported to be fairly specific for CHF. However, only one GP used this as an argument for CHF.
According to the guidelines, a normal ECG opposes the diagnosis of CHF (Table [1](#T1){ref-type="table"}). Only one of the patients (CV5) had a normal ECG. Three of the GPs used this information as an argument against CHF when reasoning about this patient, and two of them diagnosed the patient as not CHF. Four GPs used a pathological ECG as an argument for CHF (CV1, CV6), and all four diagnosed those patients as CHF.
According to the guidelines, information about cardiac enlargement or pulmonary congestion on chest X-ray gives some support for CHF if there are any pathological findings, and is of some importance as an argument against CHF if the findings are normal (Table [1](#T1){ref-type="table"}). Chest X-ray findings were frequently used by the GPs as arguments in their diagnostic reasoning (Figure [1](#F1){ref-type="fig"}). When considered separately, information about cardiac volume was used as an argument 36 times (26 for, and 10 against CHF), and information about pulmonary congestion 38 times (32 for, and 6 against CHF).
### ECHO findings as arguments for or against CHF
Each of the 15 GPs judged 6 case vignettes, which resulted in 90 judgement situations. In 48 of them, the EF value (or the information about left ventricular function in CV3) was used as an argument for or against CHF (Table [3](#T3){ref-type="table"}). In some of the judgement situations in which the EF value was not used as an argument, other ECHO information was utilized, such as left ventricular hypertrophy or restricted motility of the ventricular wall. However, in 33 judgment situations, no ECHO information was used as an argument in the diagnostic reasoning. Table [3](#T3){ref-type="table"} shows the use of ECHO in all the judgement situations: Five GPs used information about EF in their diagnostic reasoning for five of the case vignettes, four GPs used it for four of the vignettes, one GP used it for three of the vignettes, one GP for two of the vignettes and two GPs for one of the vignettes (and in one case in the wrong direction). Two GPs never used information about EF in their diagnostic reasoning. Some of the GPs expressed uncertainty about the EF values (e.g. GP14, CV1: \"I think\... think I am not certain about the meaning of ejection fraction\").
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
GPs\' use of ECHO information
:::
CV2 CV6 CV3 CV5 CV4 CV1
------ ------- ------- ------- ------- ------- -------
GP1 **+** other other 0 0 0
GP2 \+ \+ 0 0 0 other
GP3 \+ \+ 0 \- \- \-
GP4 \+ \+ 0 \- \- 0
GP5 \+ \+ other \- \- \-
GP6 \+ \+ 0 \- \- \-
GP7 \+ \+ 0 \- \- other
GP8 \+ \+ 0 \- \- 0
GP9 \+ \+ 0 \- 0 \-
GP10 \+ \+ \- **-** \- 0
GP11 \+ \+ 0 \- \- \-
GP12 0 0 other 0 other other
GP13 \+ \+ 0 0 0 \-
GP14 0 0 other 0 0 0
GP15 0 0 0 0 \+ 0
Use of information about ejection fraction (EF) value, or, in the case of CV3, about left ventricular function, as arguments for (+) or against (-) the diagnosis CHF. \"Other\" indicates that other ECHO information than the ejection fraction was used in the diagnostic reasoning and (0) that no ECHO informtion was used. (CHF = chronic heart failure, CV = case vignette, GP = general practitioner)
:::
In 17 judgement situations, there was a conflict between the GPs\' evaluations of the chest X-ray information and their evaluations of the ECHO information. In seven of those situations, the final diagnosis was in the same direction as the ejection fraction argument (four CHF, three not CHF). In three judgement situations, the final diagnosis was in the same direction as the X-ray argument (three CHF). In seven judgement situations, the GP was uncertain about the diagnosis.
### Other relevant diseases
In our study, the GPs used other diseases as an argument in a total of 70 judgement situations, mostly as arguments for CHF (91%). Atrial fibrillation, emphysema, history of myocardial infarction, and hypertension were the diagnoses most commonly used in this way.
### Information that GPs disagree about
For certain variables, the same information value was used by some GPs as an argument for and by others as an argument against CHF. The presence of emphysema was sometimes seen as increasing the risk of CHF (e.g. GP 7, CV3: \"\...and then she has emphysema \... chronic obstructive pulmonary disease, which can also contribute to CHF.\"), or as an alternative explanation for symptoms (e.g. GP 13, CV3: \"And then she also has emphysema, which could give her this severe breathlessness.\").
Diabetes could also be seen as increasing the risk of CHF (e.g. GP 9, CV5: \"\...if I think the patient has CHF? Well, there are some facts in particular, she\'s diabetic, and she has uncontrolled hypertension, well, too high, and stasis -- so I couldn\'t rule out the idea of CHF after all.\"), or as an alternative explanation for symptoms (e.g. GP 1, CV5: \"\... we have to improve her diabetes, since her fatigue may be due to that.\").
Advanced age could be seen as increasing the probability of CHF (e.g. GP 2, CV6: \"He\'s the age for it!\"), or as an alternative explanation for symptoms (e.g. GP 1, CV6: \"I\'m not so sure that CHF alone can explain his symptoms. After all, he\'s 84 years old.\"). Age was used as a diagnostic argument only for the two patients over 80 years of age.
For relative cardiac volume, the reasoning could be compatible with GPs using different threshold values in their reasoning. The two lowest values were only used as arguments against CHF, the two highest values only as arguments for it, and the two intermediate values were used as arguments in both directions.
Diagnostic judgements
---------------------
There was total agreement among the GPs only for the prototypical CHF case; otherwise there was a large variation among GPs regarding diagnoses. Case vignettes representing CHF patients were more likely to be correctly diagnosed than those representing non-CHF patients (Table [4](#T4){ref-type="table"}).
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
GPs\' classifications of case vignettes
:::
CV2 CV6 CV3 CV5 CV4 CV1
------------------------------------------------------------------------------------ ------------- ------------- ---------- ---------- ---------- ----------
Total number of arguments used by the GPs (proportion of arguments for CHF) 60 (98%) 52 (75%) 54 (85%) 57 (44%) 63 (70%) 63 (75%)
Number of GPs classifying the patient as CHF **15** **11** **11** 3 6 11
Uncertain, no classification 0 2 2 3 3 1
Correct diagnoses of CHF cases and not CHF cases (proportion of correct diagnoses) 37/45 (82%) 18/45 (40%)
Correct diagnoses, all judgements (proportion of correct diagnoses) 55/90 (61%)
The GPs\' classification of the case vignettes and the number of arguments used by the group of GPs. Cells with bold numbers indicate correct judgements. (CV = case vignette, CHF = chronic heart failure)
:::
Discussion
==========
GPs\' diagnostic reasoning compared with guidelines
---------------------------------------------------
When comparing the GPs\' diagnostic reasoning with guidelines, we found that the clinical information in the case vignettes was not used to the extent recommended in the guidelines. It is true that information about the ejection fraction value on ECHO was the single most frequent diagnostic argument, and it was the most common argument against CHF. This is in line with the guidelines, which emphasize the need for objective evidence of cardiac dysfunction. However, in more than one third of the judgement situations, the information about ECHO that was presented was not used as an argument. Over-diagnosis of CHF in primary health care has been demonstrated in a number of studies, with ECHO findings as the gold standard \[[@B3],[@B4],[@B15]\]. Limited access to ECHO has been suggested as an explanation for this finding. However, our data indicate that simply providing access to ECHO might not be enough.
In the diagnostic algorithm, symptoms and signs are the entry criteria. However, the GPs did not seem to use them consistently in this way, except when diagnosing the prototypical CHF and non-CHF cases. One reason for this might be that most symptoms and signs considered typical for CHF are fairly non-specific as regards the diagnosis CHF.
Information about other relevant diseases, which was important in the GPs\' diagnostic reasoning, is not included in the list of assessments to be performed routinely (Table [1](#T1){ref-type="table"}) \[[@B9]\]. However, information about a history of myocardial infarction, for example, increases the probability of CHF. In a study of CHF diagnostics in primary health care, it was shown that the combination of cardiac enlargement and a history of myocardial infarction had the best positive predictive value for CHF when systolic dysfunction measured by ECHO was used as gold standard \[[@B23]\]. This finding is compatible with the notion that experienced physicians structure their knowledge more according to enabling conditions than according to biomedical reasoning \[[@B24]-[@B26]\]. Enabling conditions are patient contextual factors such as sex, age, medical history, and occupation. In most routine diagnostic situations, biomedical details of a disease and its cause are not so important, and the physician\'s images of the diseases (\'illness scripts\') are rather characterized by these enabling conditions, which form a characteristic pattern. The GPs\' frequent use of this kind of information may thus indicate that they are experienced physicians, with illness scripts for CHF which include other diseases. It might be valuable to include this kind of information in a clearer way in the guidelines, because it would reflect the higher probability of CHF in patients with those characteristics.
Some methodological considerations
----------------------------------
The case vignettes represented authentic patients referred by GPs to a cardiology department for problems related to heart failure. This may have led to a selection of more complicated patients than the \"typical\" heart failure patients in primary health care. The reason we chose this group of patients was that we wanted to include patients who were thoroughly investigated, with a well-founded clinical diagnosis, and for whom information about all variables of interest could be found in the patient records. Selecting GPs only from health care centres in, or relatively close to, the city centre may have biased the results, since differences in catchment areas, working conditions, and access to echocardiography may influence GPs\' diagnostic habits. This could make it difficult to generalize the results to other GPs. Only 50% of the GPs who were contacted agreed to take part in the study, which could bias the results. However, since the age distribution was the same in the two groups, it seems unlikely that the drop-out group would differ from the study group regarding clinical experience.
Guidelines as decision support when diagnosing CHF
--------------------------------------------------
The full version of the guidelines is difficult to apply to individual diagnostic situations and it is also difficult to use it for assessment of diagnostic behaviour \[[@B13]\]. In this study, we have used the table of routine assessments as a reference for evaluating the GPs\' diagnostic reasoning (Table [1](#T1){ref-type="table"}). This table includes a rough weighting of the importance of different types of information, which could serve as a guide for diagnostic judgements, even if it is not obvious how it should be used in individual cases. The two compulsory criteria in the definition are included in this table as necessary conditions. However, in some situations these judgment tools will not be satisfactory. One example is case vignette CV3, where the clinical picture was strongly indicative of CHF, with dyspnoea, rales, tachycardia, gallop rhythm, cardiomegaly and pulmonary congestion, while according to ECHO findings there was normal left ventricular function ([Additional file 1](#S1){ref-type="supplementary-material"}). The patient could therefore be classified as a non-heart failure patient according to the definition, while the clinical diagnosis, based on the attending cardiologist\'s judgement of all accessible information, was in fact heart failure. However, the ECHO in this case also included information about atrial dilatation, mitral insufficiency and pulmonary hypertension, i.e. a rather complex situation. A patient with clinical findings suggestive of CHF, but with a normal ejection fraction value, could be considered not to have CHF, i.e. not to have a systolic CHF, but could alternatively have a diastolic CHF \[[@B27],[@B28]\]. This situation is not dealt with in the guidelines.
Some implications of this study
-------------------------------
GPs\' tendency to over-diagnose CHF has been explained by their relying on symptoms, signs and less specific investigations such as chest X-ray, and by limited access to ECHO in the primary health care. However, this study indicates that a substantial minority of GPs seem to be less familiar with the use of ECHO and EF. Thus, access to ECHO ought to be accompanied by education about how to integrate this information better in the diagnostic reasoning.
Guidelines ought to include search of information about other cardio-vascular diseases in the list of assessments to be performed routinely (Table [1](#T1){ref-type="table"}) and in the algorithm for diagnosis of heart failure. This would reflect the increased probability of CHF in presence of those diseases. The problem of diastolic heart failure should also be addressed in a clearer way in guidelines.
Conclusions
===========
The information in the case vignettes was underused as arguments for and against the possibility of CHF as compared with the guidelines. Information about the EF value was the single most frequently used argument for or against CHF; nevertheless, in one third of the diagnostic judgements the GPs did not consider any information about the ECHO in their diagnostic reasoning. Information about symptoms and signs were not used to to the extent suggested in the guidelines. Information about other relevant diseases was frequently used in the GPs\' diagnostic reasoning, indicating that they often relied on illness scripts. Some implications of our study are that 1) GPs should be taught how to use ECHO information better in their diagnostic reasoning, 2) guidelines ought to give more importance to information about other cardio-vascular diseases in the diagnostic reasoning, and 3) guidelines ought to treat the topic of diastolic heart failure in a clearer way.
List of abbreviations used
==========================
CHF Chronic heart failure
COPD Chronic obstructive pulmonary disease
ECHO Echocardiography
ECG Electrocardiography
EF Ejection fraction
GP General practitioner
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
All authors conceived of the study and participated in the design. YS carried out the data collection, performed the statistical analyses and drafted the manuscript. All authors participated in the interpretation of the results and the discussions of the drafts. All authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2296/6/4/prepub>
Supplementary Material
======================
::: {.caption}
###### Additional File 1
**Case vignette characteristics**The additional file shows some important characteristics of the case vignettes. Format: Word-table.
:::
::: {.caption}
######
Click here for file
:::
Acknowledgements
================
We thank all the general practitioners who participated in the study. We also thank Bengt Ullman, Department of Cardiology, Stockholm Söder Hospital, for valuable comments on the manuscript. The study was supported by grants from the Stockholm County Council and the Swedish Heart Lung Foundation.
|
PubMed Central
|
2024-06-05T03:55:52.057408
|
2005-1-15
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546006/",
"journal": "BMC Fam Pract. 2005 Jan 15; 6:4",
"authors": [
{
"first": "Ylva",
"last": "Skånér"
},
{
"first": "Lars",
"last": "Backlund"
},
{
"first": "Henry",
"last": "Montgomery"
},
{
"first": "Johan",
"last": "Bring"
},
{
"first": "Lars-Erik",
"last": "Strender"
}
]
}
|
PMC546007
|
Background
==========
Several cross-sectional \[[@B1],[@B2]\], case-control \[[@B3]-[@B6]\], and prospective studies \[[@B7]-[@B9]\] have reported an inverse association between nonsteroidal anti-inflammatory drug (NSAID) use and the risk of Alzheimer\'s disease (AD), whereas others \[[@B10]-[@B12]\] have not. In this report, we present results of analyses of data from the Multi-Institutional Research in Alzheimer\'s Genetic Epidemiology (MIRAGE) Study in which we examined potential effect modification by APOE-ε4 carrier status and ethnicity on this association.
Methods
=======
Subjects and data collection
----------------------------
The MIRAGE Study is a multi-center family study of genetic and environmental risk factors for AD sponsored by the National Institute on Aging since 1991. The details of MIRAGE Study data collection procedures, protocols for obtaining family histories, and reports of validity studies of the MIRAGE questionnaires have been published elsewhere. \[[@B13]-[@B15]\] Briefly, families were recruited through probands meeting NINCDS-ADRDA criteria \[[@B16]\] for probable or definite AD who were ascertained through research registries and memory clinics. After obtaining informed consent from non-demented family members, and a combination of consent or assent -- along with informed consent by proxy -- on living demented subjects, questions eliciting demographic data and information about presumptive risk factors for AD were obtained using standardized MIRAGE questionnaires.
Questions pertaining to NSAID use were added to the questionnaire in 1996, and the data presented in this report were collected from May, 1996 through May, 2002. Questions about the proband were answered by a surrogate source within the family, typically the spouse or adult offspring. The same information was sought on non-demented first-degree family members of these probands over 50 years of age, usually a sibling or spouse (less commonly parents or children).
1020 family members in this analysis claimed to be cognitively normal, or were reported by family informants to be dementia-free. Of these, 982 were evaluated using the modified Telephone Interview of Cognitive Status (mTICS) \[[@B17],[@B18]\], and normal cognitive status was confirmed in 973 (99.1%). Information on both patients and first-degree family members was supplemented where available by multiple informants, and medical and nursing home records.
To elicit information on prior NSAID use, the following question was asked: \"Have you ever taken a [nonsteroidal]{.underline} anti-inflammatory medication (e.g. Advil, Motrin, etc.) *on a daily basis for more than 6 months*?\" No distinction was made between aspirin and other classes of NSAIDs. For proxy reporting about a relative with AD, the question substituted \"your relative\" for \"you\". For any affirmative answer, a follow-up question asked for the dates at which the medications were first used and the names of all NSAIDs that had been used.
A discrete \"index date\" was established within each family corresponding to the earliest date that the family or medical records reported AD symptoms to have begun in the proband. Subjects from each family (whether AD cases or non-demented family members) were considered to have been exposed to NSAIDs only if the starting date for NSAID use preceded this index date by at least one year. Age represented the age of cases and of non-demented relatives at the index date, and was treated as a continuous variable.
As shown in Figure [1](#F1){ref-type="fig"}, there were 756 probands and 1020 relatives over the age of 50 with APOE genotype who were queried about prior NSAID use. After exclusions for those subjects who had missing or unsure responses for the name of their medication, did not include a medication start date, or had missing data for the variables age, sex, education or ethnicity, there remained for analysis 682 probands and 982 relatives. Of the 982 relatives, nine were reported to be demented with the onset of their dementia prior to the index date for that family, and their diagnoses were verified by review of medical records as having probable or definite AD by research criteria, so these were classified with the probands as having AD.
::: {#F1 .fig}
Figure 1
::: {.caption}
######
MIRAGE subjects ≥ 50 years (adjusted for family index date) with APOE genotype who completed the personal history questionnaire.
:::
:::
Statistical analysis
--------------------
Analyses were performed using SAS version 8.2. The primary independent and dependent variables were prior NSAID use and AD case status, respectively. Crude odds ratios were computed in the first instance, followed by adjusted estimates using generalized estimating equation (GEE) models \[[@B19]\] to account for the possibility that variables of interest (e.g., medication use, APOE status) could be correlated among individuals within families. Adjustments were made for the following covariates: age, sex, ethnicity (categorized as White, African-American, or other), education (less than versus equal to or greater than high school level), and APOE-ε4 carrier status (one or two ε4 alleles vs. none).
We stratified the dataset in order to evaluate whether the association between NSAID use and AD was similar in APOE-ε4 carriers and non-carriers. In addition, we formally evaluated these associations by adding an interaction term (ε4 \* NSAID use) to the GEE model. Ethnicity was similarly examined as an effect modifier.
Results
=======
Characteristics of the 1664 subjects are listed in Table [1](#T1){ref-type="table"}. AD patients were more likely to be older and to be APOE-ε4 carriers compared to controls. The distributions of sex and education were not different between cases and controls.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Characteristics of AD patients and non-demented family
:::
CHARACTERISTIC AD (N = 691) NON DEMENTED (N = 973) AGE ADJUSTED PERCENT AD AGE ADJUSTED PERCENT NON-DEMENTED P-VALUE\*
------------------------ -------------- ------------------------ ------------------------- ----------------------------------- -----------
Mean Age (SD) 70.0 (8.2) 65.0 (8.8) \<0.0001
Sex (%male) 242 (35.0) 381 (39.2) 36.5 39.9 0.39
Greater than HS Ed (%) 403 (58.3) 643 (66.1) 59.3 65.1 0.10
African American (%) 215 (31.1) 204 (21.0) 28.4 21.3 0.01
Use of NSAIDs† (%) 24 (3.5) 66 (6.8) 3.5 6.7 0.08
APOE-ε4 carrier (%) 448 (64.8) 370 (38.0) 65.3 38.0 \<0.0001
\* Reported p-values use General Estimating Equations to account for correlation among observations.
†Exposure to non-steroidal anti-inflammatory drugs had to precede index date by one year or more.
:::
Sixty-six out of 973 non-demented relatives (6.8%) and 24 of 691 cases (3.5%) reported previous NSAID use (odds ratio = 0.49; 95% CI = 0.31--0.80). After adjustment for age, sex, educational level, and ethnicity, the odds ratio (OR) of NSAID use among AD cases compared to non-users was 0.57 (95% CI = 0.35--0.93); it was 0.64 (95% CI = 0.38--1.05) when APOE carrier status was added to the GEE model (see Table [2](#T2){ref-type="table"}).
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Risk of AD with and without prior use of NSAIDs, stratified by APOE-ε4 carrier status
:::
EXPOSURE AD (N = 691) NON-DEMENTED FAMILY MEMBERS (N = 973) CRUDE ODDS RATIO (95% CI) AGE-ADJUSTED ODDS RATIO (95% CI) ADJUSTED ODDS RATIO (95% CI)
----------------------------------- -------------- --------------------------------------- --------------------------- ---------------------------------- ------------------------------
**Overall**
No NSAIDs 667 907 1.0 1.0 1.0
Use of NSAIDs 24 66 0.49 (0.31, 0.80) 0.55 (0.34, 0.88) 0.64 (0.38, 1.05)\*
**Having no ε4 alleles**
No NSAIDs 231 556 1.0 1.0 1.0
Use of NSAIDs 12 47 0.61 (0.32, 1.18) 0.75 (0.38, 1.46) 0.78 (0.39, 1.52)\*\*
**Having at least one ε4 allele**
No NSAIDs 436 351 1.0 1.0 1.0
Use of NSAIDs 12 19 0.51 (0.24, 1.06) 0.47 (0.24, 0.96) 0.49 (0.24, 0.98)\*\*
\*Adjusted for age, sex, ethnicity, education, and APOE-ε4 status.
\*\*Adjusted for age, sex, ethnicity, and education.
\*\*\*NSAID\*APOE-ε4 interaction p-value = 0.04.
:::
The magnitude of inverse association was greater among APOE ε4 carriers (OR = 0.49; 95% CI = 0.24--0.98) than non-carriers (OR = 0.78; 95% CI = 0.39--1.52). However, formal evaluation of the interaction between NSAID use and APOE-ε4 carrier status did not reveal a significant difference (p = 0.40). The association between NSAID use and AD risk was similar among Caucasian and African Americans (data not shown).
Discussion
==========
This study supports the findings of previous reports \[[@B1]-[@B6]\] suggesting that use of NSAIDs for at least six months is associated with a reduced risk of AD. This association appears to be more robust among APOE-ε4 carriers than non-carriers, although the difference in associations between these two groups was not statistically significant.
The MIRAGE Study includes the largest number of well-characterized AD cases and family controls to date, and this large sample size permits adjustment for important potential confounders, as well as the power to examine effect modification by APOE genotype and ethnicity. The subjects without dementia were first-degree family members of AD cases, providing some degree of informal matching on age, socioeconomic status, and health-seeking behavior.
However, these results must be interpreted in light of some methodological limitations. Data on NSAID use was collected with a single retrospective question that did not distinguish between aspirin use and non-aspirin NSAID use. Moreover, while non-demented participants reported on themselves, a proxy historian reported on most of the demented individuals.
Differential reporting is a potential source of bias in a study that uses self-report on most of the non-demented subjects, yet relies on surrogate respondents for all of the subjects with AD. Asymmetric data collection is difficult to avoid when cases are cognitively impaired, but may be more accurate than expected in AD patients where the surrogate historian has a long association with the subject. We addressed this potential bias by performing an independent validation study to determine the accuracy of surrogate information on a number of questions, including the same questions used in this report about NSAID use. \[[@B15]\] This study found substantial reliability on the NSAID item (kappa = 0.70). While a validation study comparing proxy historians for non-demented persons does not perfectly mirror the situation in which proxy historians report on demented individuals, our study revealed excellent concordance for surrogate responses from most categories of relatives.
This result is consistent with those of prior studies which found a similar association despite differences in study design (cross-sectional \[[@B1],[@B2]\] vs. case-control \[[@B3]-[@B6],[@B10],[@B11]\] vs. prospective \[[@B7]-[@B9],[@B12]\]), sampling frame (family members \[[@B10]\] vs. registry-based \[[@B11]\] vs. general population \[[@B1]-[@B9],[@B12]\]), ascertainment of exposure, type of medication considered (aspirin \[[@B2]-[@B4],[@B6],[@B8]-[@B10],[@B12]\] vs. non-aspirin NSAIDs \[[@B1]-[@B11]\] vs. \'any\' NSAID \[[@B4],[@B8]\]), duration of exposure (current \[[@B1]-[@B3]\] vs. any history of use, duration ranging from a week to at least six months \[[@B3],[@B5]-[@B12]\]), and degree of matching or adjustment (usually adjusted for age, sex, and education, less frequently APOE genotype \[[@B7]-[@B9],[@B12]\]).
While many studies have examined the association between NSAID use and risk of AD, few have examined the impact of APOE genotype on this association. The Cache County Study \[[@B9],[@B20]\], the Canadian Study for Health and Aging (CSHA) \[[@B8]\], and the Rotterdam Study \[[@B7]\] adjusted for APOE and tested for effect modification and found none. But they had fewer AD cases, and, in the CSHA had a smaller proportion of genotyped subjects. The Rotterdam Study \[[@B7]\] reported separate odds ratios for APOE-ε4 carriers and non-carriers, but this sample did not have any subjects who were both APOE-ε4 carriers and who reported long-term use of NSAIDs. They found no difference in risk between those with at least one ε4 allele compared to ε4 non-carriers among subjects who used NSAIDs between one month and two years.
Our data suggest an enhanced protective benefit of NSAID use among those with ε4. A smaller protective effect was also evident among those lacking ε4. In our sample there were relatively few AD cases who were not ε4 carriers, thus the appearance of different patterns of association between NSAID use and AD risk among APOE genotype subgroups may be spurious. This difference could also have arisen as a result of bias and confounding. The genotype-specific association could be explained by differential inclusion of subjects into the study on the basis of APOE-ε4 carrier status and NSAID use. This might occur if there were differential mortality, according to APOE genotype, among those with AD who had a history of NSAID use; or if for any reason among NSAID users APOE-ε4 carriers were less likely than non-carriers to be diagnosed with AD (or conversely, if among non-users of NSAIDs AD was more likely diagnosed in APOE-ε4 carriers compared to non-carriers). Differential recall could also give rise to this observation. However, these explanations are unlikely because subjects were not selected on the basis of APOE genotype. It is also possible that APOE-ε4 carrier status is a proxy for differentially distributed unmeasured confounders related to NSAID use such as inflammatory disease processes.
Alternatively, our results imply that NSAID use affects AD risk differently between APOE-ε4 carriers and non-carriers. For example, because ε4 carriers are inherently more vulnerable to AD, there is a greater opportunity for attributable risk reduction. This explanation does not imply biological interaction between NSAIDs and ε4. On the other hand, the ε4 isoform may have greater pro-inflammatory properties \[[@B21]\] and ε4 individuals may be more responsive to the benefits of NSAID use than those lacking ε4.
Examination of this finding in prospective studies and clinical trials of sufficient power (such as the ADAPT Study \[[@B22]\], a prospective trial of anti-inflammatory use in the prevention of AD) to detect effect modification by APOE-ε4 carrier status is needed. Such confirmation would provide critical insights into the mechanisms by which APOE isoforms modulate AD risk and into novel therapeutic strategies.
Conclusions
===========
NSAID use is inversely associated with AD and may be modified by APOE genotype. Prospective studies and clinical trials of sufficient power to detect effect modification by APOE-ε4 carrier status are needed.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
Study concept and design (LAF, RCG, LAC); acquisition of data (LAF, RCG, MIRAGE investigators); analysis and interpretation of data (AGY, MH); drafting of the manuscript (AGY, RCG, LAF); critical revision of the manuscript for important intellectual content (AGY, LAF, RCG, LAC); statistical expertise (LAC); obtained funding (LAF, RCG, LAC, MIRAGE investigators). All authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2318/5/2/prepub>
Acknowledgements
================
Other participating investigators from the MIRAGE Study Group include Alexander Kurz, MD, Department of Psychiatry, Technische Universitat Munchen; Sanford Auerbach, MD, Department of Neurology, Boston University School of Medicine; Marsha Wilcox, EdD ScD, Department of Medicine, Boston University School of Medicine; Rodney Go, PhD, Department of Epidemiology, University of Alabama School of Public Health; Dessa Sadovnick, PhD, Department of Medical Genetics and Medicine (Neurology), University of British Columbia; Ranjan Duara, MD, The Wein Center, Mt. Sinai Medical Center and the University of Miami School of Medicine; Charles DeCarli, MD, Department of Neurology, UC Davis; Walter A Kukull, MD PhD, Department of Epidemiology, School of Public Health, University of Washington; Helena Chui, MD, Rancho Los Amigos Rehabilitation Center, Department of Neurology, University of Southern California; Timi Edeki, MD PhD, Departments of Medicine and Clinical Pharmacology, Morehouse School of Medicine; Abimbola Akomolafe, MD, Department of Medicine, Morehouse School of Medicine; Patrick A Griffith, MD, Department of Medicine, Section of Neurology, Morehouse School of Medicine; Robert P Friedland, MD, Department of Neurology, Case Western Reserve University; David Bachman, MD, Department of Psychiatry, Medical University of South Carolina.
|
PubMed Central
|
2024-06-05T03:55:52.061044
|
2005-1-12
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546007/",
"journal": "BMC Geriatr. 2005 Jan 12; 5:2",
"authors": [
{
"first": "Agustín G",
"last": "Yip"
},
{
"first": "Robert C",
"last": "Green"
},
{
"first": "Matthew",
"last": "Huyck"
},
{
"first": "L Adrienne",
"last": "Cupples"
},
{
"first": "Lindsay A",
"last": "Farrer"
}
]
}
|
PMC546008
|
Background
==========
Allogeneic hematopoietic stem cell transplantation (allo-HSCT) is performed mainly in patients with high-risk or advanced hematologic malignancies and aplastic anemias, and for some of them it is the only curative treatment. After allo-HSCT, incomplete engraftment and appearance of recipient\'s hematopoietic cells can lead to a coexistence of donor and host hematopoiesis -- a situation known as mixed chimerism. Complete recovery of hematopoiesis of the donor origin is referred to as complete chimerism.
Widely accepted molecular methods for analysis of chimerism after allo-HSCT, aimed at distinguishing precisely between donor\'s and recipient\'s cells, are PCR-based analyses of polymorphic DNA markers, such as variable number of tandem repeats (VNTR) or short tandem repeats (STR) \[[@B1]\]. Such analyses can be performed regardless of donor\'s and recipient\'s sex \[[@B2]\]. Additionally, in patients after sex-mismatched allo-HSCT, fluorescent *in situ*hybridization (FISH) can be applied. This technique is based on identification of Y-chromosome-specific sequences in the posttransplant sample examined. Variants of these techniques can be used for precise, quantitative assessment of the amount of donor\'s cells in recipient\'s peripheral blood and/or bone marrow after transplantation, in the long run giving a picture of the dynamics of changes in chimeric status within a hematopoietic compartment. Chimerism also reflects response to treatment \[[@B3]\], since it correlates with the risk of malignancy relapse. Relapse is the most frequent cause of treatment failure in recipients transplanted for hematologic malignancies, but it is still controversial if patients with mixed chimerism have an increased risk of developing relapse or graft failure. Most probably only a progressive mixed chimerism (a dynamic rise in the number of recipient\'s cells over time) seems to reflect a relapse or rejection. Successful outcome has been associated with a state of stable complete chimerism \[[@B4]\].
Here we report results of a comparison of different techniques for analysis of posttransplant chimerism: fluorescent *in situ*hybridization (FISH) and PCR-based molecular methods with fluorescent products detected in an ALF Express DNA Sequencer (Pharmacia) or ABI 310 Genetic Analyzer (PE).
Methods
=======
Patients
--------
Investigation of hematopoietic chimerism was performed in ten children (eight girls, two boys) aged 6--16 years. They were diagnosed with acute myelogenous leukemia (n = 4), acute lymphoblastic leukemia (n = 2), chronic myelogenous leukemia (n = 1), myelodysplastic syndrome (n = 1), or Fanconi Anemia (n = 2). All recipients received hematopoietic stem cells from HLA-matched sibling donors, and sex-mismatch between donor and recipient was present in all cases. The material (peripheral blood) for hematopoietic chimerism quantification was collected in different periods after transplantation. In all 10 children reconstitution of hematopoiesis was observed. Out of 8 children transplanted for hematologic malignancies, 4 are well and alive, in complete continuous remission (CCR), while in the other 4, leukemia relapse occurred 4--23 months after transplantation.
DNA isolation
-------------
High-molecular-weight DNA was extracted from frozen whole blood (approximately 5 ml) or bone marrow (approximately 3--5 ml) by the standard treatment with sodium dodecyl sulfate (SDS) and proteinase K, and the salting-out method. DNA was isolated from the donors\' and patients\' blood samples collected before and after transplantation at various intervals in order to determine the chimeric status.
Analysis of PCR products by an ABI 310 Genetic Analyzer
-------------------------------------------------------
The PCR protocol optimized for the Qiagen polymerase and the PE 9700 thermocycler was performed as described previously \[[@B5]\]. For fragment analysis (after capillary electrophoresis), an ABI 310 Genetic Analyzer (PE) was used \[[@B5]\]. All analyses were performed in the University Children\'s Hospital, Tuebingen.
Analysis of PCR products by an ALF Express DNA Sequencer
--------------------------------------------------------
The PCR protocol optimized for the Thermal Controller MJ Research (Watertown, MA) model PTC-100™ was applied. PCR was performed in a volume of 10 μl; the PCR reaction mixture contained: 2.5 pM of each forward and reverse primer, 200 μl of each dNTP, 0.4 U Taq polymerase (Qiagen, Chatsworth, CA), 1x PCR buffer (Qiagen, Chatsworth, CA), and 40 ng of genomic DNA. Conditions for PCR were as follows: 5 min at 94°C for the first denaturation; 26 cycles of amplification with a temperature profile of 45 sec at 94°C, 1 min at 55°C, 1 min at 72°C; with additional 5 min at 72°C in the last cycle. STR loci were amplified with fluorescent PCR primers described previously \[[@B6]\]. Primers for microsatellite markers were labeled with Cy5 dye (TIB MOLBIOL). A 1.5-μl aliquot of PCR reaction was resuspended in 7 μl of loading solution (formamide, bromophenol blue) containing 100 bp and 300 bp internal markers. All samples, after denaturation at 95°C for 5 min, were analyzed on 6% denaturing polyacrylamide gel with 7 M urea in the sequencer. A 50--500 sizer labeled with Cy5 dye was used as an external marker (for calculation of allele sizes). Electrophoresis was carried out in 0.6xTBE buffer at 1500 V/min. The helium-neon laser was operated at a wavelength of approximately 700 nm and laser power value of 2.5 mW. Allele sizes and peak areas of fluorescent products were analyzed and calculated with the use of Fragment Manager software (Pharmacia). PCR and analysis of PCR products by the ALF Express DNA Sequencer was performed at the Institute of Human Genetics, Poznan.
FISH
----
The experiments were performed on interphase nuclei obtained by standard short culturing of fresh whole blood samples, with probes specific for chromosomes X (locus *DXZ1*) and Y (locus *DYZ1*). The FISH procedure according to Cytocell Ltd. was used \[[@B7]\]. The number of scored nuclei was 250 to 550, with a median of 300. FISH experiments were performed at the Institute of Human Genetics, Poznan. Since the material for FISH analysis was not collected in all designated periods after transplantation in some patients, FISH experiments were not performed then.
Quantification of chimerism
---------------------------
After electrophoresis in the ABI 310 Genetic Analyzer (PE), all obtained data were analyzed by GeneScan 3.1 software and then transferred to Genotyper 2.5 software \[[@B5]\]. All data obtained after electrophoresis of fluorescent products in the ALFExpress DNA Sequencer were transferred to Fragment Manager™ software (Pharmacia). For both, calculation of the amount of recipient\'s DNA was performed using the formula:
\% of recipient\'s DNA = (R1 + R2)/(D1+ D2 + R1 + R2) × 100,
where: R1, R2 = peak areas of recipient\'s alleles; and D1, D2 = peak areas of donor\'s alleles.
Only informative markers were used for the analysis. If donor and recipient were heterozygous but shared one allele, only the area of the non-shared alleles was considered for the analysis \[[@B8]\].
To make sure that quantification is accurate, we performed serial dilution experiments, where standardized mixed chimeric samples were created by mixing donor\'s and pretransplant recipient\'s DNA in a range between 0 and 100 percent. The sensitivity strongly depends on the size of alleles, the detection level was around 3--5% of patient cells.
The results of chimerism detection by different methods were compared by the Spearman correlation test.
Results
=======
The results of chimerism quantification with the use of an ALF Express DNA Sequencer, ABI 310 Genetic Analyzer, and FISH are compared in Table [1](#T1){ref-type="table"}. In three patients (no. 5, 6, 7) only donor\'s cells were detected in all post-HSCT samples and in one patient (no. 1) only recipient\'s cells were present in all samples examined. These results were confirmed with the use of all three methods. Three other patients (no. 2, 3, 4) exhibited mixed posttransplant chimerism according to PCR and/or FISH. In the last three patients (no.8, 9, 10), complete chimerism was detected by PCR and automated DNA sizing, but in some of their samples, low numbers of recipients\' cells were detected by FISH.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Comparison of results using different DNA sizing technologies and FISH
:::
**No.** **UPN** **Sex F/M** **Diagnosis** **Material examined** **HSCT date** **Days after HSCT** **PCR/ALF Express\*** **PCR/ ABI 310 Genetic Analyser\*** **FISH with probes specific to X, Y chro-mosomes\***
--------- --------- ------------- --------------- ----------------------- --------------- --------------------- ----------------------- ------------------------------------- ------------------------------------------------------
1. 67 F AML PB 04.09.1998 542 100% 100% 100%
556 100% 100% 100%
574 100% 100% 100%
590 100% 100% 100%
639 100% 100% 100%
675 100% 100% 100%
697 100% 100% 100%
721 100% 100% 100%
750 100% 100% 100%
2. 102 M AML PB 12.10.2000 14 84% 86% 89%
19 49% 47% 51%
28 40% 40% 35%
35 47% 42% 12%
57 46% 35% 6%
3. 106 M FA PB 21.12.2000 7 100% 100% 92%
48 26% 27% 32%
91 13% 14% 9%
4. 87 F FA PB 27.12.1999 21 10% 28% ND
38 93% 94% ND
285 15% 18% ND
313 17% 19% ND
357 0% 0% ND
5. 45 F CML PB 15.11.1996 953 0% 0% ND
1030 0% 0% ND
1216 0% 0% ND
1284 0% 0% 0%
1374 0% 0% 0%
1459 0% 0% 0%
1492 0% 0% ND
6. 93 F AML PB 20.04.2000 28 0% 0% 0%
158 0% 0% 0%
277 0% 0% 0%
7. 109 F MDS PB 26.01.2001 14 0% 0% 0%
21 0% 0% ND
34 0% 0% ND
42 0% 0% ND
8. 108 F ALL PB 19.01.2001 17 0% 0% ND
21 0% 0% ND
26 0% 0% 2%
39 0% 0% ND
52 0% 0% ND
60 0% 0% ND
9. 88 F AML PB 21.01.2000 29 0% 0% ND
87 0% 0% 0%
178 0% 0% ND
267 0% 0% ND
365 0% 0% 2%
10. 95 F ALL PB 16.06.2000 27 0% 0% ND
83 0% 0% ND
200 0% 0% ND
218 0% 0% 1%
ALL = acute lymphoblastic leukemia; AML = acute myelogenous leukemia; CML = chronic myelogenous leukemia; FA = Fanconi Anemia; FISH = fluorescent *in situ*hybridization; HSCT = hematopoietic stem cell transplantation; MDS = myelodysplastic syndrome; ND = no data; PB = peripheral blood; UPN = unique patient\'s number
\* results expressed as % of recipient\'s cells
:::
Coefficients of Spearman rank correlation between results of chimerism quantification by the three different methods are shown in Table [2](#T2){ref-type="table"}. All coefficients were statistically significant (p \< 0.001). The correlation between PCR/ALF Express and PCR/ABI 310 was stronger than between FISH and both PCR methods.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Coefficients of Spearman rank correlation between results of chimerism quantification by different methods (all coefficients significant, P \< 0.001).
:::
**PCR/ABI 310 Genetic Analyzer** **FISH**
------------------------------ ---------------------------------- ----------
PCR/ALF Express 0.987 0.801
PCR/ABI 310 Genetic Analyzer 0.825
:::
Comparison of methods used
--------------------------
One of the advantages of application of automated DNA sizing techniques for detection of posttransplant chimerism is that the use of radioactivity is not necessary. It is arelatively simple and rapid method, consisting of two steps: polymerase chain reaction with fluorescent primers and automated detection of fluorescently labeled PCR products, separated by electrophoresis. The analysis of fluorescently labeled PCR products provides better accuracy and precision of measurement than traditional electrophoretic methods. The most time-consuming step, which might prolong the examination process, is the search for informative markers.
Analysis of the chimerism status by amplification of STR loci can be performed regardless of donor\'s and recipient\'s sex. The most advantageous is the high sensitivity of the detection system used, so that only small amounts of DNA are needed.
Fragment analysis in the ABI 310 Genetic Analyzer (PE) is performed after capillary electrophoresis. The examination of one sample requires about 6 hrs; 144 samples can be analyzed per day if three colors are simultaneously analyzed \[[@B5]\]. The advantage of this method is that thanks to the combination of three-color detection, multiple sets of samples or multiple loci for a single sample can be analyzed on one gel. It is possible to analyze more than one marker at a time after multiplex amplification of informative loci.
In the ALFExpress DNA Analyzer, electrophoresis is carried out in an off-vertical gel cassette specially designed for easy and safe gel casting. The number of samples to be loaded on a gel is limited to 40 per run. The ALFwin Fragment Analyzer is used afterwards for fragment analysis. It is provided with versatile application software for the control of DNA fragment separation runs and subsequent analysis of the data. Collected data are used to accurately size PCR product peaks on the basis of external and internal standards.
One analysis of 40 samples takes 7 hrs, including PCR, electrophoresis in the gel cassette, and paperwork. The appropriate assembling and cleaning of the gel cassette is critical and time consuming.
It is well known that FISH is a good quantitative method of fluorescent signal detection, but requires lots of technical experience and expertise. Fluorescent *in situ*hybridization for one patient sample lasts at least 5 hrs, including preparation of interphase nuclei, hybridization with specific probes (X, Y dual-color FISH), and analysis. The high cost of the procedure is definitely a disadvantage.
Discussion
==========
PCR-based techniques allow the relative proportions of recipient\'s and donor\'s cells in the post-HSCT period to be identified and quantified and is not only limited to sex-mismatched transplants. Although when using chimerism analysis one cannot assess whether or not the population of recipient\'s nucleated cells contains leukemic cells, samples taken at various intervals can show if the expansion rate of the particular population is consistent with hematologic and clinical symptoms of the disease. When it is not possible to find an informative marker for PCR amplification, only FISH analysis enables assessing the chimerism status. However, cytogenetic Y chromosome probing by FISH is limited exclusively to sex-mismatched transplantations. Results obtained with the use of ALFExpress DNA Sequencer and ABI 310 Genetic Analyzer are identical or very similar. We showed that appropriate quantitative assessment of chimerism after HSCT by using microsatellite genotyping and automated DNA sizing does not depend on the sequencer model used. The high correlation between results from the PCR/ALFExpress and PCR/ABI 310 Genetic Analyzer indicate that these two methods can be used interchangeably. The superiority of the ABI 310 Genetic Analyzer is limited to the possibility of analysis of three samples at the same time in one reaction tube and the technical ease of capillary electrophoresis with no need for the time-consuming and cumbersome use of glass plates. Lower, but still positive correlations were found between results of FISH analysis and these two methodological variants of PCR. However, in some samples analyzed with PCR, no recipient\'s signals were found, attesting to full donor chimerism, while at the same time residual host cells turned out to be detectable by FISH. We suggest that these results are within the range of error of the method applied.
Conclusions
===========
Finally, we conclude that all the methods applied enable a rapid and accurate detection of post-HSCT chimerism and with due caution can be used interchangeably.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
J.J. carried out the molecular chimerism studies and drafted the manuscript.
T.S. performed the statistical analysis.
A.P and D.B. supplied clinical data.
P.B. initiated quantitative analysis.
J.W. supervised clinical part and final writing
M.W. supervised laboratory part and final writing
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2326/5/1/prepub>
Acknowledgments
===============
This work was partly supported by grants from the State Committee for Scientific Research (KBN): No. 4 PO5E 045 14 (M.W.), No. 6 PO5E 035 20 (M.W.), and No. 4PO5E 108 18 (J.W.). J.J. is the recipient of a scholarship from the Postgraduate School of Molecular Medicine, funded by the L. Kronenberg Foundation. Additionally, J.J. received a short-term fellowship of the European Molecular Biology Organization (EMBO) and a UNESCO-L\'OREAL fellowship (program for women in science). Technical expertise of D. Ładoñ in FISH analysis is acknowledged.
|
PubMed Central
|
2024-06-05T03:55:52.063134
|
2005-1-10
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546008/",
"journal": "BMC Blood Disord. 2005 Jan 10; 5:1",
"authors": [
{
"first": "Justyna",
"last": "Jółkowska"
},
{
"first": "Anna",
"last": "Pieczonka"
},
{
"first": "Tomasz",
"last": "Strabel"
},
{
"first": "Dariusz",
"last": "Boruczkowski"
},
{
"first": "Jacek",
"last": "Wachowiak"
},
{
"first": "Peter",
"last": "Bader"
},
{
"first": "Michał",
"last": "Witt"
}
]
}
|
PMC546009
|
Background
==========
Essential hypertension is a common, polygenic, complex disorder resulting from interaction of several genes with each other and with environmental factors such as obesity, dietary salt intake, and alcohol consumption. Since the underlying genetic pathways remain elusive\[[@B1]\], currently most studies focus on the genes coding for proteins that regulate blood pressure as their physiological role makes them prime suspects.
The Renin-Angiotensin System (RAS) has a central role in regulating blood pressure and sodium homeaostasis. Genes encoding components of RAS, including angiotensinogen (AGT), angiotensin-converting enzyme (ACE), angiotensinogen II type-1 receptor (AGTR1), and renin, have been extensively investigated as genetic determinants of essential hypertension \[[@B2]\]. Polymorphisms of RAS\[[@B3]\] genes seem also to play a role in the development of diseases that cause secondary hypertension\[[@B4],[@B5]\]. Subjects carrying the ACE D allele have unanimously been shown to have increased ACE serum activity\[[@B3],[@B6]\] while the T235 AGT variant has been associated with elevated angiotensinogen levels \[[@B7]\]. However, so far there are no consistent findings.
In 1992, the M235T AGT TT polymorphism was first reported to be associated with hypertension \[[@B8]\]. This finding has not been confirmed by all investigators\[[@B9],[@B10]\]. Although no relationship between the ACE gene and hypertension was observed in one early linkage study \[[@B11]\] and most recent studies including one meta-analysis \[[@B12]-[@B14]\], several studies have suggested a role: hypertensive individuals have a high prevalence of the D allele or DD genotype \[[@B3],[@B15],[@B16]\].
The inconsistent results might be explained in part by the genetic and environmental heterogeneity among different ethnic groups \[[@B13]\]. On the other hand, one recent study \[[@B17]\] reported that the MM, AA, CC, DD/ID genotype combination was associated with a substantially higher prevalence of hypertension in the participants to the Olivetti Heart Study, even though no individual effect of each isolated genotype was detected.
The present study investigates the relationship between variants of the I/D ACE gene and M235T AGT gene, and the presence and severity of essential hypertension in a large homogeneous German population. The effect of a combination of ACE and AGT gene polymorphisms on hypertension was also examined.
Methods
=======
Study design
------------
The design of the study followed the guidelines proposed by Cooper et al \[[@B18]\], and the study was carried out in accordance with the Declaration of Helsinki \[[@B19]\].
Study population
----------------
This cross-sectional study comprised a total of 1358 individuals from Weisswasser, a county town of 25,000 in Saxony, Germany. After giving informed consent, 720 normotensive subjects were selected from local blood donors and 638 hypertensive patients from the local renal care center. All hypertensive patients included in the study had been diagnosed as suffering from primary hypertension by the attending consultants on first contact with the clinic. Hypertensives were defined as those who received at least one antihypertensive medication. At the time of blood sampling, 34.2% were diabetic (6 type 1, 212 type 2), and 65.1% were suffering from kidney disease. Of the 37 patients with K/DOQI stage 5 (Kidney failure: GFR, 15 ml/min/1.73 m^2^or dialysis), 15 were due to diabetic nephropathy, 5 to chronic pyelonephritis, 1 to chronic glomerulonephritis, 1 light chain deposits, 1 polycystic kidney disease, and 14 of unknown cause (no biopsy obtained). The severity of hypertension was estimated based on the number of antihypertensive medications used, a surrogate marker for the severity of hypertension\[[@B8]\]. Age and gender distribution is described in table [1](#T1){ref-type="table"}.
Genotyping
----------
RFLP (restriction fragment length polymorphism) and restriction analysis were used to determine the frequencies of the I/D polymorphisms of the ACE gene, and homo-/ heterozygoty of the M235T AGT gene\[[@B20],[@B21]\]. ACE I/D polymorphism was studied by PCR based amplification of a 597 bp long gene fragment of the ACE gene, which lacks 287 bp in case of the deletion (D) variant. The primers used were: sense- 5\'GATGTGGCCATCACATTCGTCAGAT3\', and antisense- 5\'CTGGAGACCACTCCCATCCTTTCT3\'.
AGT M235T polymorphism was studied by first amplifying a 104 bp long fragment of the AGT gene using the following primer sequences: sense- 5\'CCGTTTGTGCAGGGCCTGGCTCTCT3\', and antisense: 5\'CAGGGTGCTGTCCACACTGGACCCC3\'.
The M -\> T point mutation at position 235 creates a detection site for the restriction enzyme *Tth 111I*.
Statistical analysis
--------------------
Statistical analysis was carried out using SPSS personal computer statistical package (version 11.5, SSPS Inc, Chicago, IL). Demographic characteristics were compared by *t*test for continuous data and Pearson\'s χ^2^test for categorical data. Allele frequencies were calculated with the gene-counting method. χ^2^test was used for assessment of the Hardy-Weinberg equilibrium for the distribution of genotypes. Odds ratios were calculated with a 95% confidence interval. A *P*\< .05 was considered significant.
Meta-analysis
-------------
A meta-analysis was performed using Review Manager 4.2 (The Cochrane Collaboration) to further examine the association of the AGT M235T gene polymorphisms with essential hypertension in Caucasians. A systematic literature search in PubMed Medline for articles published between April 2002 and June 2004 was carried out using the following MESH-headings: \"angiotensinogen/genetics\", \"hypertension/genetics\", \"blood pressure/genetics\", and \"adult\". The search was limited to articles published in English and studies on Caucasian human subjects. Only 2 studies were left after strict examination according to the exclusion criteria listed in \[[@B22]\]. A total of 25 studies were finally included: 22 were from the most recent meta-analysis \[[@B22]\], which covered articles from January 1992 to March 2002; 2 were selected from the query described above, and the last one was the present study. Homogeneity among studies was assessed on the basis of χ^2^test using P-value \< 0.05. The Mantel-Haenszel odds ratios were calculated by applying both fixed effect model and random effect model in case of heterogeneity.
Results
=======
Demographic data are summarized in Table [1](#T1){ref-type="table"}. Hypertensives were older (58.80 ± 13.22 years) than controls (41.24 ± 12.66 years, p \< .0001), and more often male (53.2% vs. 44.4%, p = .001).
AGT M/T genotyping was successful in 637 hypertensives and 720 normotensives, and ACE I/D genotype was analyzed in 636 hypertensive and 719 control subjects. The roles of age and gender in the association between hypertension and ACE and AGT gene polymorphisms were examined by comparing the effects of ACE and AGT genotypes for hypertension in men and women, young and elderly subjects respectively. \"Young\" was defined as \"age \< 50 years old\", and \"elderly\" was characterized as \"age ≥ 50 years old\".
ACE polymorphism
----------------
No differences were observed in ACE allele and genotype frequency distribution between hypertensives and controls with respect to gender and age, and no deviations from Hardy-Weinberg equilibrium were observed in any of subgroups (P \> 0.1). ACE genotypes I/I, I/D and D/D were of almost identical frequency within both groups (P = 1.0, Figure [1](#F1){ref-type="fig"}). Risk assessment showed that there were no significant risk changes for hypertension in the subjects either with the ACE DD genotype (odds ratio: 1.00, 95% CI: 0.74 to 1.36, P = .98, Table [2](#T2){ref-type="table"}) or D allele (odds ratio: D vs. I: 1.00, 95% CI: 0.86 to 1.17, P = .98).
AGT polymorphism
----------------
AGT T/T homozygotes tended to be more frequent in controls than in hypertensives (4.6% vs. 2.7%, Figure [2](#F2){ref-type="fig"}). In women, this finding became significant (5.3% vs. 1.7%, Figure [3](#F3){ref-type="fig"}), but no difference in AGT genotype frequency was found in men. The distribution of the AGT genotypes in all subgroups of the sample population was not in Hardy-Weinberg equilibrium (P \< .001). AGT TT genotype was associated with a significant 48% decrease in the risk of being hypertensive (Table [2](#T2){ref-type="table"}, odds ratio: 0.52; 95% CI: 0.28 to 0.96; P = .034), and this risk decreased even more to 72% in women (odds ratio: 0.28; 95% CI: 0.1 to 0.78; P = .01). However, no difference was observed in the AGT allele frequency distribution with respect to age and gender (Table [3](#T3){ref-type="table"}, P \> 0.05), and the effect of the AGT T allele did not reach a significant level in the decrease of hypertension risk (odds ratio: T vs. M: 0.88; 95% CI: 0.75 to 1.03; P = .12).
Meta-analysis of studies on AGT polymorphisms in Caucasians
-----------------------------------------------------------
When all studies were pooled, Caucasian individuals with TT genotype had an odds ratio for hypertension of 1.21(95% CI: 1.11 to 1.32) compared with those with MM genotype (Figure [4](#F4){ref-type="fig"}). The studies included in the meta-analysis are \[[@B8],[@B20],[@B23]-[@B43]\]. The pooled odds ratio (odds ratio: TT vs. MM: 1.23; 95% CI: 1.13 to 1.34) increased by 2.5% when the presented study was excluded. Tests for heterogeneity were significant (P \< 0.001) in the above two cases, and the odds ratios (TT vs. MM) rose to 1.30 (95% CI: 1.10 to 1.54), and 1.35 (95% CI: 1.15 to 1.59) respectively when applying the random effects model.
Combination of ACE and AGT polymorphisms
----------------------------------------
The effect of eight combinations (TT, DD/ID; TT, II/ID; MM, DD/ID; MM, II/ID; DD MM/MT; DD, TT/MT; II, MM/MT; II, TT/MT) on hypertension was examined. No statistically significant association was observed between any combination above and hypertension in all subjects combined. Nonetheless, in women, both genotypes of TT, DD/ID and TT, II/ID were significantly associated with lower prevalence of hypertension (Table [2](#T2){ref-type="table"}, 20% vs. 43.3%, odds ratio: 0.33, P = 0.038; 19% vs.43.6%, odds ratio: 0.31, P = 0.026), while MM, DD/ID genotype significantly increased the risk for hypertension (Table [2](#T2){ref-type="table"}, 49.2% vs. 40.1%, odds ratio: 1.45, P = 0.028).
No association could be identified between severity of hypertension and a specific ACE or AGT genotype (Table [4](#T4){ref-type="table"}).
Discussion
==========
Although some recent studies \[[@B15],[@B16],[@B44],[@B45]\] suggested a unique sex-specific role of ACE in essential hypertension, no significant association of essential hypertension with the ACE gene I/D polymorphism was observed in this German population of 1,358 for either gender. This finding confirms earlier observations in another German population \[[@B31]\], in other Caucasian populations \[[@B11],[@B12],[@B14],[@B46]\], and in one meta-analysis \[[@B13]\]. The distribution of the ACE genotypes was in Hardy-Weinberg equilibrium in this German population while that was not the case in Pereira et al\'s study \[[@B1]\]. One possible explanation is the ethnic difference. Pereira et al. \[[@B1]\] showed that there were statistically different ACE I/D polymorphism genotypic frequencies in different ethnic groups.
Surprisingly, the cross-sectional study presented here shows a higher prevalence of the T/T M235T AGT gene in the control group compared to the hypertensive group. AGT TT genotype was associated with a *decrease*in the risk for hypertension (odds ratio-TT vs. MM: 0.52; 95% CI: 0.28 to 0.96) and a more significant association was found in women (odds ratio-TT vs. MM: 0.28; 95% CI: 0.1 to 0.78), compared to men. This is in stark contrast to findings from previous studies, including two German datasets \[[@B31],[@B32]\], and three meta-analyses \[[@B10],[@B22],[@B47]\], which reported that the *AGT*235 T-allele and/or TT genotype significantly *increased*the risk for essential hypertension in Caucasians: odds ratio T vs. M was 1.20 (95% CI: 1.11 to 1.29) \[[@B10]\], odds ratio TT vs. MM was 1.31 \[[@B47]\], and odds ratio TT vs. MM was 1.19 (95% CI: 1.10 to 1.30)\[[@B22]\]. In agreement with the previous meta-analyses, the meta-analysis presented here showed increased odds for hypertension (odds ratio: TT vs. MM: 1.21) in Caucasians conferred by TT, and the odds ratio rose by 2.5% when the present study was excluded. Of the 25 studies included in the present meta-analysis, the present study was the only one in which the AGT T235T genotype decreased odds for hypertension. Nevertheless, the quality of meta-analysis results depends on the quality of the individual studies included, and unusual sample sizes might bias the finding. For example, one single study \[[@B43]\] included in the previously largest meta-analysis \[[@B22]\] was exceptionally large, giving it enormous weight. The highly variable study quality implies that all interpretations must be made with great caution, as was explicitly pointed out by Kunz et al. \[[@B10]\]
Although no significant difference was observed in AGT T allele frequency distribution between hypertensives and controls with respect to age and gender, the frequency of the *AGT*T allele among normotensives was higher than that among hypertensives (0.36 vs. 0.33). This was inconsistent with one previous meta-analysis in Caucasians, which showed that among controls, the mean allele frequency for the *AGT*T allele was 0.41 (95% CI: 0.34 to 0.48), and among cases, increased to 0.45 (95% CI: 0.38 to 0.52) \[[@B10]\]. In the present study, the frequency of the *AGT*T allele among hypertensives (0.33) was outside the lower border of the 95% confidence interval (0.38 to 0.52) reported in \[[@B10]\]. This may reveal the specific genetic background of this particular German population.
The AGT genotype distribution is not in Hardy-Weinberg equilibrium (P \< 0.001) while no deviations from Hardy-Weinberg equilibrium are observed on the ACE genotype distribution in the same population. This may be explained by a shift toward a higher frequency of MT individuals (62.4%) instead of TT individuals (3.7%) in this specific population.
The study population presented here contains a large proportion (65%) of patients with renal disease. While selection of participants based on patient records excluded those patients that had symptoms suggesting the diagnosis of secondary hypertension at first contact, the possibility remains that at least a part of the study population suffers from renal rather than essential hypertension. It should be noted, however, that the majority of studies included in the presented meta-analysis does not give specific information regarding renal function of hypertensives, and the largest study \[[@B43]\] is population based and does not name any specific, kidney related exclusion criteria.
Another possible explanation is that the AGT T allele frequency may decrease with age, which was reported from the United Arab Emirates \[[@B48]\]. The prevalence of hypertension increases with advancing age. According to the National Health and Nutrition Examination Survey III (NHANES III) prevalence estimates for the years 1988--1994, American Whites aged 55 to 64 years have a more than threefold higher prevalence of hypertension (42.1%) than those aged 35 to 44 years (11.3%) \[[@B49]\]. Frossard et al detected that AGT T allele frequency decreased with age in the United Arab Emirates\[[@B48]\]. The ACE DD genotype was found associated with human longevity \[[@B50]\]. In the present study, normotensives are significantly younger than hypertensives (41 ± 12 yrs vs. 59 ± 13 yrs). It is conceivable that many of the young individuals are at hypertension risk because of their ACE or AGT genotype, but have not yet shown hypertension at the time of genotyping, and may develop hypertension in their older age. This might lead to some misclassifications and hence reduce the power of this study. On the other hand, two studies of German populations \[[@B32],[@B51]\] reported that the AGT T allele was a risk factor for hypertension in individuals younger than 50 years of age. In the present study, young hypertensives had a higher frequency of the AGT T allele than elderly hypertensives (0.36 vs. 0.32), but the difference didn\'t reach statistical significance (p = .22). It is possible that the small percentage (24.9%) of the studied hypertensives under 50 years of age biased the finding.
A number of studies \[[@B51]-[@B54]\] examined the relationship between RAS genotype and the severity of hypertension, but their results were contradictory. In accordance with two \[[@B52],[@B54]\] of them including one German dataset, the present study fails to find an effect of the AGT or ACE genotype on the severity of hypertension. Nevertheless, in the present study, hypertensives that carried at least one copy of the AGT T allele (TT or MT: n = 397) were less likely to take two or more antihypertensive medications than those with MM genotype (n = 225) (odds ratio -TT or MT vs. MM: 0.797; 95% CI: 0.57 to 1.12; P = .185), and their average number of antihypertensive drugs was lower (2.09 vs. 2.20; P = .276). Despite not reaching statistical significance, this observation was in contrast to Schunkert et al\'s study \[[@B51]\] on another German population (subjects initially participated in the MONICA Augsburg cohort baseline survey), which found that the carriers of the AGT T allele (n = 418) had a 2.1-fold higher probability of taking two or more antihypertensive drugs than individuals with the MM genotype (n = 216). It is worth pointing out that in the present study, the number of subjects taking antihypertensive medications is much larger than in Schunkert et al\'s study (622 vs. 143). While the data is far from statistical significance, the trend is in line with those findings that associate rather the M allele with hypertension in the present study.
The effect of a combination of RAAS genes\' polymorphisms on blood pressure has been investigated in the participants to the Olivetti Heart Study \[[@B17]\], in which the MM, AA, CC, DD/ID genotype was detected to be associated with a substantially higher prevalence of hypertension in the absence of detectable effects of each individual polymorphism at any single locus. The present study showed very similar findings in women: MM, DD/ID genotype significantly increased the odds for hypertension (odds ratio: 1.45; 95% CI: 1.04 to 2.02), while TT, DD/ID and TT, II/ID were significantly associated with lower prevalence of hypertension (odds ratio: 0.33, 95% CI: 0.11 to 0.99; odds ratio: 0.31, 95% CI: 0.10 to 0.92). The risk of hypertension in the women with TT, DD/ID and TT, II/ID, however, didn\'t change much compared with those with TT alone (odds ratio: 0.31, 95% CI: 0.12 to 0.83). This suggests that there is only a slight synergistic effect between the AGT and ACE genes.
Although in most surveys the prevalence of hypertension appears to be equal in women and men \[[@B55]\], sex-specific effects of ACE or AGT genes on hypertension have been reported recently \[[@B16],[@B43]\]. For instance, Sethi et al. \[[@B43]\] found the AGT TT associated with an increase in risk for hypertension in women but not in men from the Copenhagen City Heart Study with a population of 9100 subjects, and an association of the ACE DD genotype with increased diastolic blood pressure was detected in men, but not in women from the Framingham Heart Study \[[@B16]\]. In the present study, the AGT TT genotype was negatively correlated to hypertension in women only while no sex-specific effect of the ACE gene was shown. It is possible that the fact is covered by the different gender distributions: in this population, there are more women in normotensives than in hypertensives (55.6% vs. 46.8%, P = .001).
In addition to the factor declared above, the result may be influenced by the study design and the composition of the sample population. The study design itself may influence the results. As mentioned above, this study was a cross-sectional study where subjects were assessed at a single time, and at that time most of the controls were younger than 50 years old. Animal studies have shown that hypertension genes may be activated for only certain periods during the life history of an organism \[[@B2]\]. Hence, some of them might develop hypertension at an older age, resulting from the activated hypertension genes. Longitudinal studies are needed to further examine the relationship between hypertension and genes at different ages.
The population can be described as static population with a mixed Germanic-Slavonic background. Due to the location (a provincial town on the German-Polish border) and political and historical setting (little population fluctuation during the over 140 years of imperial, fascist and socialist rule, no influx due to lack of economic attractivity after reunification), the population may be assumed as homogeneous. Several recent studies \[[@B56]-[@B61]\] have reported significant differences in prevalence of hypertension between Germany and Poland, which, however, are assumed to be largely dependent on life-style differences, mostly salt intake. These differences, however, are unlikely to play a role in the study population due to its homogeneity with regards to life style preferences following successful assimilation over many generations. In the case of ACE polymorphism with all allele frequencies greater than 15%, there is no need to examine population stratification \[[@B62]\]. This becomes more urgent for AGT where there is no Hardy- Weinberg equilibrium. As the sample information, however, does not include data on the ethnic background of the probands, additional haplotyping carried out on the samples would not have allowed to rectify for the historical ethnic background (Germanic versus Slavonic).\[[@B62],[@B63]\]
In the meantime, it should be noted that the present study was carried out on an unusually large population (second only to one in 63 studies included in Sethi\'s meta-analysis\[[@B22]\]). Given the large homogeneous population-based sample, the findings cannot be attributed to simple selection bias. Therefore, the finding that the AGT TT genotype associated with a decreased risk for essential hypertension is likely to be true for this particular German population.
Conclusions
===========
Despite the limitations mentioned, this cross-sectional study does not support the notion that the ACE I/D polymorphism contributes to the prevalence and severity of essential hypertension. However, the M235T TT genotype of AGT gene was detected to confer a significantly decreased risk for the prevalence of hypertension in women from this particular population. Despite the large sample size, the present study fails to revise the odds ratio in a meta-analysis of a total of 25 studies on the association between the AGT M235T polymorphisms and hypertension in caucasians. This observation may reflect a very specific local inheritance pattern of the AGT genotypes. If this holds true, studies aiming at drug development based on genomic traits must be scrutinized rigorously as therapeutic recommendations may be valid for selected subpopulations only\[[@B64]\].
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
M.N. designed and initiated the study, analyzed the samples and wrote part of the manuscript. A-L.Z. and M.L. did the statistical analysis. A.M. devised the concept for meta-analysis and wrote the manuscript. L.P. created the figures. M.N. and A.M. should be considered as joint co- authors.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2369/6/1/prepub>
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
ACE genotypes and allele frequencies, both genders combined. No significant differences. F(D): frequency of D- allele.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
AGT genotypes and allele frequencies, both genders combined. No significant differences. F(T): frequency of T- allele.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
AGT genotypes and allele frequencies in women (significant at p = 0.035). F(T): frequency of T- allele.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
Risk of hypertension associated with the AGT M235T genotypes in Caucasians.
:::
:::
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Demographic features. M: male. F: female.
:::
**Hypertensives**(n = 638) **Normotensives**(n = 720) P- value
---------------------- ---------------------------- ---------------------------- ----------
Gender (M/F) 339/298 319/400 0.001
Age (yrs, + SD 58.80 ± 13.22 41.24 ± 12.66 \< 0.001
Age \< 50 yrs: n (%) 159 (24.9%) 547 (76.0%) \< 0.001
Age \> 50 yrs: n (%) 479 (75.1%) 173 (24.0%) \< 0.001
:::
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Hypertension risks estimates for AGT and ACE genotypes by gender.
:::
**Odds ratio**(95%CI)
----------------------- ---------------- -------------- -------------- --------------------- --------------------- --------------------- ---------------------
MT vs. MM TT vs. MM ID vs. II DD vs. II MM DD/ID vs. Others MM II/ID vs. Others TT DD/ID vs. Others TT II/ID vs. Others
0.87 0.52 1 1 1.2 1.04 0.56 0.56
(0.69--1.09) (0.28--0.96)\* (0.77--1.31) (0.74--1.36) (0.94--1.53) (0.81--1.34) (0.28--1.09) (0.28--1.12)
0.82 0.83 0.83 0.79 1 1.16 0.85 1.08
(0.59--1.15) (0.36--1.93) (0.57--1.21) (0.52--1.22) (0.76--1.43) (0.81--1.66) (0.34--2.11) (0.39--3.01)
0.88 0.28 1.26 1.32 1.45 0.97 0.33 0.31
(0.64--1.21) (0.10--0.78)\* (0.85--1.85) (0.86--2.03) (1.04--2.02)\* (0.69--1.37) (0.11--0.99)\* (0.10--0.92)\*
\*P \< 0.05; CI, confidence interval.
:::
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
AGT T allele frequencies by age and gender.
:::
f(T)
--------------- ------------ ------------ ----
Age \< 50 yrs 0.36 (159) 0.36 (547) NS
Age ≥ 50 yrs 0.32 (479) 0.37 (173) NS
Male 0.35 (339) 0.37 (319) NS
Female 0.32 (298) 0.36 (400) NS
Total 0.33 (638) 0.36 (720) NS
f(T), frequency of AGT T allele; NS, not significant; yrs, years.
:::
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Prevalence of ACE and AGT genotypes by severity of hypertension.
:::
**ACE Genotype**n (%) **AGT Genotype**n (%)
---------------------------- ----------------------- ----------------------- --------- ------- ---------- ---------- -------- -------
Hypertensive, one drug 56 (22) 120 (48) 74 (30) 0.458 83 (33) 163 (65) 5 (2) 0.225
Hypertensive, \> two drugs 89 (24) 189 (51) 93 (25) 0.458 142 (38) 217 (59) 12 (3) 0.225
:::
|
PubMed Central
|
2024-06-05T03:55:52.065809
|
2005-1-11
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546009/",
"journal": "BMC Nephrol. 2005 Jan 11; 6:1",
"authors": [
{
"first": "Adrian",
"last": "Mondry"
},
{
"first": "Marie",
"last": "Loh"
},
{
"first": "Pengbo",
"last": "Liu"
},
{
"first": "Ai- Ling",
"last": "Zhu"
},
{
"first": "Mato",
"last": "Nagel"
}
]
}
|
PMC546010
|
Background
==========
A family of vertebrate proteins containing a six transmembrane domain (6TM) has been recently implicated in apoptosis and cancer. Up to date, this family contains four vertebrate members: STEAP, STAMP1/STEAP2, TIARP and TSAP6/pHyde. STEAP (six-transmembrane epithelial antigen of the prostate) was the first described member of this family and identified as a prostate-specific cell-surface antigen overexpressed in cancer, located at the cell-cell junction of the secretory epithelium of prostate, and found as well in both colon and bladder cancer cell lines \[[@B1],[@B2]\]. STAMP1 (six transmembrane protein of prostate 1), also known as STEAP2, also overexpressed in prostate cancer, has been located in the trans-Golgi network and shuttles to plasma membranes, which suggest a role in the secretory/endocytic pathways \[[@B3],[@B4]\]. TIARP (Tumor necrosis factor-alpha-induced adipose-related protein) is a cell surface protein induced by TNF-α and IL-6, probably implicated in resistance to insulin \[[@B5],[@B6]\]. TSAP6 (tumor suppressor activated pathway-6), also known as pHyde \[[@B7]-[@B10]\], is a p53 inducible protein which regulates apoptosis and the cell cycle via direct interaction with Nix (a pro-apoptotic Bcl-2 related protein) and Myt1 kinase (a negative regulator of the G2/M transition) \[[@B9]\]. TSAP6 has been shown to be interacting with TCTP (Translationally controlled tumor protein) and could be implicated in its secretion \[[@B10]\].
In this work we present evidence of remote homology of this family to other two families: the mainly eukaryotic Nox and the bacterial YedZ (Fig. [1](#F1){ref-type="fig"}), both involved in redox functions \[[@B11]-[@B17]\]. The Nox family is involved in the production of reactive oxygen species (ROS) \[[@B11],[@B12]\]. The first member of the family (gp91phox) was discovered in phagocytes and contains an N-terminal transmembrane heme binding domain and two C-terminal domains with binding sites for both flavin adenine dinucleotide (FAD) and NADPH (Fig. [2](#F2){ref-type="fig"}) \[[@B11]-[@B13]\]. Originally, ROS were thought to be used just as a mechanism of host defence. The discovery of gp91phox homologues in several other tissues has suggested their implication in many other functions, such us signal transduction, cancer, mitogenic signalling, cellular growth, angiogenesis, and modification of extracellular matrix proteins \[[@B11]-[@B14]\]. Regarding the bacterial YedZ family, the only experimentally characterized member so far is the *Escherichia coli*YedZ protein, which binds a single heme and is involved in electron transfer to the molybdopterin cofactor in YedY, its operon neighbour gene \[[@B16]-[@B18]\]. The exact function of the operon YedZ/YedY remains unknown.
Methods
=======
To do the sequence analysis of the new domain we took advantage of the possibility of connecting distant protein families via intermediate sequences \[[@B19]\] and methods of iterative profile sequence similarity search: HMMer \[[@B20],[@B21]\] and PSI-BLAST \[[@B22]\] over the Uniprot 90% non redundant sequence database \[[@B23]\]. We used NAIL to view and analyse the HMMer results \[[@B24]\]. The alignment of transmembrane regions using standard substitution matrices might be inaccurate because of the different roles played by amino acids in globular proteins and in transmembrane media \[[@B25]\]. In the case of the ACRATA domain, the regions to be aligned mostly consist of amino acids located in transmembrane regions (Fig. [1](#F1){ref-type="fig"}). For this reason, we used a method based on a hidden Markov model (HMMer hmmalign), which does not rely on a general substitution matrix \[[@B20],[@B21]\], using as a guide both the transmembrane predictions from TMHMM \[[@B26],[@B27]\] and the results of multiple sequence alignment using T-Coffee \[[@B28],[@B29]\]. The genomic neighborhood of the bacterial sequences (YedZ family) was analyzed to find potentially related genes in operons using STRING \[[@B17],[@B18]\].
Results
=======
The global hidden Markov profile \[[@B20],[@B21]\] generated for STEAP and related vertebrate proteins (STEAP family, henceforth) localized the first bacterial sequence (SpTrembl Q7VKI9 from *Haemophilus ducreyi*) with an E-value of 0.63. This protein belongs to the large YedZ family of bacterial oxidoreductases. The corresponding YedZ global profile detected the STEAP family (most similar member: SpTrembl Q8IUE7, human STAMP1) with an E-value of 0.00087. The global profile of STEAP and YedZ detected the Nox family with an E-value of 0.032 and the corresponding Nox global profile localized the YedZ family with an E-value of 0.007 (Fig. [3](#F3){ref-type="fig"}). Only the regions from transmembranes 3 to 5 were considered to build the profiles because the transmembranes 1, 2 and 6 are highly variable among families. We have named this 6TM domain the ACRATA domain after **[A]{.underline}**poptosis, **[C]{.underline}**ancer and **[R]{.underline}**edox **[A]{.underline}**ssociated **[T]{.underline}**ransmembr**[A]{.underline}**ne domain.
To investigate the consistency of our results we performed iterative database searches using the PSI-BLAST program \[[@B22]\]. We used as query the most conserved region of the ACRATA domain in *E. coli*YedZ protein (residues 71--166). These searches detected all of the ACRATA domain-containing families after 15 iterations (using a cut-off of E = 0.005 for the inclusion of retrieved sequences in the profile). None of these profile searches retrieved new unrelated sequences, and reciprocal searches produced convergent results. Therefore, we have concluded that ACRATA is a previously undetected, conserved domain that is commonly found in members of the STEAP, YedZ and Nox protein families. The similarity between these proteins was suspected before \[[@B30]\]. However, no alignment, domain definition or substantial statistical evidence was provided to demonstrate the evolutionary relationships of these proteins.
The complete conservation of two histidines in all ACRATA domain containing proteins (Fig. [1](#F1){ref-type="fig"}) indicates that the STEAP protein family could bind at least an heme group, as was previously experimentally characterized for Nox and YedZ families \[[@B13],[@B16]\], or for other analogous proteins such as cytochromes \[[@B31]\].
Discussion
==========
Experimental evidence show that Nox and YedZ families share heme binding capabilities and also involvement in electron transfer chains \[[@B11]-[@B13],[@B16]\]. For the STEAP family, the electron transfer capability is consistent with the presence of an N-terminal cytoplasmic NADP oxidoreductase coenzyme F420 dependent domain, being the only exception the STEAP protein itself (Fig. [2](#F2){ref-type="fig"}). Therefore we conclude that ACRATA domain is a heme binding 6TM domain that originated before the onset of eukaryotes (ancestral, YedZ), transmitted by vertical descent in a conventional manner (Nox family), and further expanded in vertebrates (STEAP family) (Fig. [4](#F4){ref-type="fig"}).
Although the mechanism of action of the ACRATA domain could be the same in all the proteins containing it, its variable cellular role is made conspicuous by the different effects produced by the modifications in the expression of the corresponding genes. For example, the knock-out of the whole YedYZ operon seems not to affect *E. coli*in a number of conditions tested (Brokx, S.J. and Weiner, J.H. personal communication). Very differently, changes in the expression patterns of human genes containing the ACRATA domain could be related with apoptosis or cancer \[[@B1]-[@B14]\]. The well known functional flexibility of oxidoreductases is exemplified in bacterial proteins such as cupredoxins and cytochromes, normally involved in electron transfer during respiration but that can enter in eukaryotic cells to induce apoptosis or inhibition of cell growth \[[@B32]\], or in the Nox family, with functions as different as host defence in phagocytes or extracellular matrix modification \[[@B11],[@B12]\].
Conclusions
===========
We have described for the first time a 6TM domain present in three protein families (STEAP, YedZ and Nox): the ACRATA domain. The common functions of the proteins of those families suggest that this domain is involved in electron transfer, mediated by its heme binding capability. We hypothesize that STEAP, STAMP1, TSAP6, and TIARP have this function, and that they form part of electron transfer systems involved in cellular regulation, apoptosis, and cancer. Additional experimental approaches using different members of the ACRATA domain-containing families are required to confirm these hypotheses.
Abbreviations
=============
STEAP, six-transmembrane epithelial antigen of the prostate;
STAMP1, six transmembrane protein of prostate 1;
TSAP6, tumor suppressor activated pathway-6;
TIARP, Tumor necrosis factor-alpha-induced adipose-related protein;
TCTP, Translationally controlled tumor protein;
Nox, NADPH oxidase;
ROS, reactive oxygen species;
TM, Trans-membrane;
FAD, flavin adenine dinucleotide;
HMM, Hidden Markov Models;
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
LSP, AMR, and MAA carried out the sequence analysis of the domain.
LSP and MAA provided with the initial input of the research.
LSP, AMR, AV, CMA, and MAA authored the manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2407/4/98/prepub>
Acknowledgements
================
We are grateful to S.J. Brokx and J.H. Weiner (University of Alberta, Edmonton, Canada) for useful comments on early versions of the manuscript. Thanks also to G.R. Reina and J. Menéndez (Hospital Puerta de Hierro, Madrid, Spain) for their encouragement and moral support.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Representative multiple alignment of the ACRATA domain.**It is viewed with the Belvu program \[33\]. The colouring scheme indicates average BLOSUM62 score (correlated to amino acid conservation) in each alignment column: cyan (greater than 3), light red (between 3 and 1.5) and light green (between 1.5 and 0.5). The limits of the domains are indicated by the residue positions on each side. The TMHMM helix transmembrane \[29, 30\] consensus prediction is shown below the alignment. The asterisks above the alignment mark the conserved histidine residues mentioned in the text. Different groups of the ACRATA sequences are shown by coloured lines to the left of the alignment: red, YedZ family; yellow, STEAP family; violet, Nox family. The sequences are named with their swissprot or sptrembl identifiers, and also, if necessary, with their gene name. Species abbreviations: Homsa, *Homo sapiens*; Glovi, *Gloeobacter violaceus*; Anasp, *Anabaena sp*.; Cloac, *Clostridium acetobutylicum*; Musmu, *Mus musculus*; Xenla, *Xenopus laevis*; Sacce, *Saccharomyces cerevisiae*; Emeni, *Emericella nidulans*; Dicdi, *Dictyostelium discoideum*; Caeel, *Caenorhabditis elegans*; Drome, *Drosophila melanogaster*; Arath, *Arabidopsis thaliana*. Complementary information is accessible at: <http://www.pdg.cnb.uam.es/STEAP>.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**Schematic representation of the domain architecture and common features in representative proteins containing the ACRATA domain.**The representative sequences selected correspond to: YEDZ, SW:YEDZ\_ECOLI; STEAP, SW:STEA\_HUMAN; TSAP6, SP:Q80ZF3; NOX1, SW:NOX1\_HUMAN; NOX5, SP:Q96PH2; DUOX1, SP:Q9NRD9. NOX1, NOX5, and DUOX1, belong to the Nox family. The proteins are drawn approximately to scale. The domains are named and located according to the Pfam and SMART protein domain databases \[34-37\]. Abbreviations: SP, SPTREMBL; SW, SWISSPROT.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**HMMer E-Values between the ACRATA domain containing families.**The numbers correspond to HMMer E-values from global profile search results \[20, 21\] that connect independently each family with the others. The arrows indicate the profile search direction, for example: the YedZ family profile search finds sequences of the STEAP family with a 0.00087 E-value and the global profile of the YedZ/STEAP families (dotted line) detected the Nox family with an E-value of 0.032.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Phylogenetic tree of selected ACRATA domain containing proteins.**The scale bar represents the number of inferred substitutions per 100 sites (amino acid residues). The tree branches for YedZ proteins are in red, STEAP branches are in yellow, and Nox branches are in violet. Organism name abbreviations are as in Fig. 1. We generated trees using a Bayesian, Neighbor Joining, and Minimum Evolution methods, and found that all three gave the same topology suggesting that the overall structure of the tree is correct. For illustrative purposes only the Neighbor Joining tree is shown \[38-40\].
:::
:::
|
PubMed Central
|
2024-06-05T03:55:52.068017
|
2004-12-29
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546010/",
"journal": "BMC Cancer. 2004 Dec 29; 4:98",
"authors": [
{
"first": "Luis",
"last": "Sanchez-Pulido"
},
{
"first": "Ana M",
"last": "Rojas"
},
{
"first": "Alfonso",
"last": "Valencia"
},
{
"first": "Carlos",
"last": "Martinez-A"
},
{
"first": "Miguel A",
"last": "Andrade"
}
]
}
|
PMC546011
|
Background
==========
Prostate cancer is a significant health problem among American men. This year about 230,110 men will be diagnosed with the disease and about 30,000 will likely die in this country alone \[[@B1]\]. Treatment options are limited and are associated with significant morbidity and mortality \[[@B2]\]. Localized cancer is treated with radical prostatectomy, brachy- or cryotherapy, and external beam radiation while cancer that has escaped the prostatic capsule is generally treated by androgen ablation. However, the eventual development of an androgen-independent phenotype leads to incurable disease, indicating the need for better treatment strategies. Experimental approaches include delivery of oncolytic viruses, immunomodulatory molecules, p53 and p21, enzymes that metabolize prodrugs and agents that can induce apoptosis \[[@B3]\]. One agent that has received considerable attention as a novel apoptosis-inducing agent is tumor necrosis factor-related apoptosis inducing ligand (TRAIL/Apo2L).
TRAIL is a type II membrane protein that can induce apoptosis by binding to death domain containing receptors DR4 and DR5 \[[@B4]\]. Unlike other death receptor ligands such as TNF and FasL, which cause septic shock and hepatotoxicity, respectively, TRAIL is tolerated well in mice and non-human primates \[[@B5]-[@B7]\]. TRAIL induces apoptosis in a variety of cell lines *in vitro*and *in vivo*. Apoptosis is initiated by binding to receptors DR4/DR5 (TRAIL-R1, 2), which is followed by assembly of the death inducing signaling complex and activation of caspase-8. Subsequent activation of caspases-3/7 (in a mitochondria-dependent or independent fashion) leads to execution of apoptosis \[[@B4]\].
Active caspase-8 tetramers are generated by cis- and transcatalytic cleavage from pro-caspase-8 homodimers \[[@B8]\]. These cleavage steps are inhibited by heterodimer formation of caspase-8 with c-FLIP. The apoptosis inhibitor c-FLIP is structurally similar to caspase-8 but lacks the cysteine residue essential for catalytic activity. A strong correlation between c-FLIP expression and malignant potential has been observed in carcinomas of the colon and liver as well as melanomas \[[@B9]-[@B11]\]. In addition, a high c-FLIP/caspase-8 ratio has been associated with resistance to death receptor-mediated apoptosis \[[@B11]-[@B13]\]. Thus downregulation of c-FLIP is a desirable strategy to enhance the apoptotic response to death receptor ligands.
We have previously shown that prostate cancer cells are relatively resistant to recombinant TRAIL but can be sensitized by pretreatment with the chemotherapeutic agent doxorubicin \[[@B7],[@B14]\]. The enhanced susceptibility of prostate cancer cells was independent of androgen-phenotype or p53 status but correlated with doxorubicin-mediated downregulation of c-FLIP expression. In this study we extended those observations to an *in vivo*model using xenografts of PC3 prostate carcinoma cells.
Our results show that doxorubicin reduces c-FLIP expression in xenografts and that combination of doxorubicin with TRAIL is more effective in tumor growth inhibition than either agent alone.
Methods
=======
Cells and reagents
------------------
PC3 cells were purchased from the ATCC and were maintained in RPMI1640 supplemented with 10% FBS at 37°C with 5% CO~2~. Apo2L/TRAIL was generously provided by Genentech Inc., San Francisco CA. KillerTRAIL was purchased from Alexis, San Diego, CA. Doxorubicin was obtained from the MUSC pharmacy. The CellTiter Aqueous One Solution Cell Proliferation Assay and DeadEndTUNEL kits were purchased from Promega, Madison WI. The c-FLIP antibody NF-6 was kindly provided by Dr. Marcus Peter, University of Chicago. The Dave-2 (c-FLIP) antibody and anti-actin were purchased from Alexis, San Diego, CA and from Sigma, St. Louis, MO, respectively. Supersignal DuraWest was obtained from Pierce Biotechnology Inc., Rockford, IL.
MTS viability assay
-------------------
Cells were seeded into 96-well plates at 1 × 10^4^cells/well, incubated overnight and subsequently treated with doxorubicin and/or Killer-TRAIL/Apo2L/TRAIL. The MTS \[3-(4,5-dimethylthiazol-2-yl)-5-(3-carboxymethoxyphenyl)-2-(4-sulfophenyl)-2H-tetrazolium\] reagent was added 24 hours after initiation of treatment and plates read at an absorbance of 490 nm one to two hours later using a Vmax kinetic microplate reader (Molecular Devices, Sunnyvale, CA). All treatments were performed in triplicate. Background absorbance was determined by incubating media with substrate alone and subtracting the values from wells containing cells. Percent cytotoxicity was calculated as follows: % cytotoxicity = 1 - \[(OD of experimental/OD of control) × 100\]. Experiments were repeated three times with similar results. There were no discrepancies between results obtained in the MTS assay and visual assessment of the cells prior to adding the MTS reagent.
Animal experiments
------------------
Athymic male nude mice (3--4 weeks old) were purchased from Harlan, Indianapolis IN and were housed under pathogen-free conditions according to Medical University of South Carolina animal care guidelines. All animal experiments were reviewed and approved by the Institutional Animal Care and Use Committee at MUSC. PC3 cells (4 × 10^6^) were injected subcutaneously into the flanks of mice. Animals bearing tumors were randomly assigned to treatment groups (five or six mice per group) and treatment initated when xenografts reached volumes of about 100 mm^3^. Tumors were measured using digital calipers and volume calculated using the formula: Volume = Width^2^× Length × 0.52, where width represents the shorter dimension of the tumor. Treatments were administered as indicated using vehicle (PBS containing 0.1% BSA), doxorubicin (2--8 mg/kg), Apo2L/TRAIL (500 μg/animal), or a combination of 4 mg/kg doxorubicin followed by 500 μg Apo2L/TRAIL. Doxorubicin was administered systemically whereas Apo2L/TRAIL was given either intra-tumorally (Fig [4](#F4){ref-type="fig"}) or systemically (Fig [5](#F5){ref-type="fig"}). All treatments were given once. Mice were monitored daily for signs of adverse effects (listlessness and scruffy apparance). Treatments seemed to be well tolerated. The mean ± SEM was calculated for each data point. Differences between treatment groups were analyzed by the student t-test. Differences were considered significant when P \< 0.05.
Western blotting
----------------
Following doxorubicin treatment, mice were sacrificed, tissue removed and immediately frozen in liquid nitrogen. Protein was prepared in RIPA buffer containing freshly added mammalian protease inhibitor cocktail. Lysates were stored at -70°C and were centrifuged (20,000 × g) prior to performing protein assays on the supernatant. Protein (50 μg) was separated on 4--12% Bis/Tris NuPage gels in MES buffer and transferred to nitrocellulose for 60--90 minutes at 30 V. After transfer and blocking (5% milk), blots were probed with Dave-2 (1:500 in 0.5 % milk) or NF-6 (1:5 in TBS-Tween) overnight at room temperature. Following three washes with TBS-Tween, membranes were incubated with the anti-mouse (1:5,000) or anti-rat (1:15,000) HRP-conjugated secondary antibodies for 1 hour at room temperature in 5% milk in TBS-Tween. Membranes were washed three times in TBS-Tween followed by chemiluminescent detection of the secondary conjugates with DuraWest Supersignal. Membranes were reprobed with anti-actin (1:2000) to ensure equal loading.
TUNEL staining
--------------
Tumors were excised and fixed in 10% formalin and processed by standard procedures. Sections were analysed for apoptotic cells using the DeadEND Tunel kit according to the manufacturers instructions. Stained slides were examined for TUNEL positive cells using a Zeiss Axiovert 200 microscope (20×). TUNEL positive cells from 4 fields/slide were counted by two investigators to calculate the mean ± SEM. P-values were calculated using the student\'s t-test.
Results
=======
Comparison of KillerTRAIL and Apo2L/TRAIL *in vitro*
----------------------------------------------------
We have previously shown that the TRAIL apoptotic response in prostate cancer cells can be enhanced by doxorubicin \[[@B14],[@B15]\]. These studies were conducted using KillerTRAIL, a commercially available form of the protein that is crosslinked for maximal activity and contains a histidine tag. Concerns about hepatotoxicity were raised when a polyhistidine tagged recombinant version of TRAIL induced apoptosis in human hepatocytes \[[@B16]\]. A subsequent study revealed that different recombinant versions of TRAIL vary widely in their biochemical properties and their potential to cause toxicity. TRAIL containing a histidine tag (Apo2L/TRAIL.His) contained less Zinc, had a less ordered conformation and was more heterogeneous than TRAIL without a tag (Apo2L/TRAIL.0) \[[@B17]\]. In contrast to Apo2L/TRAIL.His, Apo2L/TRAIL.0 was non-toxic to human hepatocytes. Thus non-tagged Apo2L/TRAIL.0 (referred to as Apo2L/TRAIL) is the preferable form to use in preclinical studies.
Initially we compared the susceptibility of PC3 prostate carcinoma cells to KillerTRAIL and Apo2L/TRAIL in parallel assays *in vitro*. As shown in Figure [1A](#F1){ref-type="fig"}, at 1000 ng/ml KillerTRAIL resulted in about 50% cytotoxicity whereas less than 20% cytotoxicity was obtained with Apo2L/TRAIL. This difference may stem from the histidine tag of KillerTRAIL. In the presence of doxorubicin, KillerTRAIL and Apo2L/TRAIL were about equally effective (Fig [1B--D](#F1){ref-type="fig"}). Doxorubicin lowered the concentration requirement for TRAIL with near maximal killing achieved at 10 ng/ml ligand.
Effect of doxorubicin on growth and c-FLIP expression *in vivo*
---------------------------------------------------------------
To determine an appropriate dose of doxorubicin for the *in vivo*experiments, mice bearing PC3 xenografts were injected with 2, 4 or 8 mg/kg doxorubicin and tumor volume was measured over time (Figure [2](#F2){ref-type="fig"}). A dose of 2 mg/kg did not affect tumor growth while higher dosages delayed tumor growth initially (p \< 0.05 at days 18 and 22). However, no statictically significant differences were detected between untreated and doxorubicin treated groups at later time points.
Previously, we have examined possible targets of doxorubicin that may be responsible for increasing the susceptibility of prostate cancer cells to TRAIL *in vitro*. These included TRAIL receptors, Bax, Bcl-2, Bcl-xl, and c-FLIP \[[@B15]\]. In PC3 cells, the only change observed in response to doxorubicin was a decrease in c-FLIP that correlated with onset and magnitude of caspase-8 activation and apoptosis. To investigate whether doxorubicin would have a similar effect on c-FLIP *in vivo*, protein from PC3 xenografts was isolated for western blotting 24 hours after systemic delivery of the drug. As shown in Figure [3A](#F3){ref-type="fig"}, we found that either 4 mg/kg or 8 mg/kg doxorubicin significantly reduced levels of c-FLIP in PC3 xenografts. The antibody used for the detection of c-FLIP (NF-6) is human-specific and thus detects only c-FLIP of PC3 origin. We also investigated the effect of doxorubicin on c-FLIP in the heart from the same mice. The heart was chosen because it expresses high levels of c-FLIP \[[@B18]\]. In addition, c-FLIP-deficient mice do not survive past day 10.5 of embryogenesis due to impaired heart development \[[@B19]\], indicating that c-FLIP plays a critical role in this organ. Protein isolated from the heart was analyzed with the rat monoclonal antibody Dave-2 that detects both mouse and human c-FLIP, although all c-FLIP detected should be exclusively of mouse origin, since PC3 cells are grown subcutaneously. Two major proteins, each migrating somewhat slower than the human c-FLIP isoforms, were detected with the Dave-2 antibody (Fig. [3B](#F3){ref-type="fig"}). The signal of the lower band (c-FLIP~S~) was not diminished following doxorubicin treatment, whereas the signal of the upper band (c-FLIP~L~) was slightly reduced following administration of 8 mg/kg doxorubicin. Since 4 mg/kg doxorubicin reduced c-FLIP in PC3 xenografts as effectively as 8 mg/kg without affecting endogenous mouse c-FLIP, this dose was chosen for combination therapy.
Effectiveness of Apo2L/TRAIL doxorubicin combination therapy *in vivo*
----------------------------------------------------------------------
To determine apoptosis *in vivo*following treatments, mice bearing bilateral PC3 xenografts were injected systemically with vehicle or 4 mg/kg doxorubicin, followed by intratumoral injection of vehicle (left tumor) or Apo2L/TRAIL (right tumor). Twenty-four hours after Apo2L/TRAIL injection, tumors were harvested and analysed for apoptosis by TUNEL staining (Fig. [4](#F4){ref-type="fig"}). There was no significant difference in TUNEL-positive cells in control and doxorubicin treated tumors. Treatment with Apo2L/TRAIL resulted in 70% more apoptotic cells than in vehicle treated tumors. However, combination treatment of doxorubicin and Apo2L/TRAIL yielded the most TUNEL positive cells (278% compared to the control) and was statistically different from all other treatment groups.
Next, we determined whether this pattern would be reflected in tumor growth of PC3 xenografts. Animals received systemic admininstration of doxorubicin followed by systemic injection of TRAIL after 24 hours, when c-FLIP levels were expected to have decreased. Tumors in animals injected with doxorubicin reached volumes of approximately 1000 mm^3^within 26 days after treatment initiation (1025 ± 138 mm^3^), which was not significantly different from the control group (1025 ± 118 mm^3^, p = 0.41). Tumors in animals treated with Apo2L/TRAIL were reduced compared to either untreated or doxorubicin treated groups (833.5 ± 150 mm^3^, p = 0.03). However, tumors that had been exposed to both doxorubicin and Apo2L/TRAIL were significantly smaller (224 ± 145 mm^3^, p \< 0.001) and remained smaller at day 33 (490 ± 197 mm^3^) when the experiment was terminated. Analysis of tumor volume over time suggests that a single round of combination treatment delays growth until day 26 (Fig [5C](#F5){ref-type="fig"}). After tumors escape this delay they resume growth at rates similar to the other groups.
Discussion
==========
In this study we have shown that combination of Apo2L/TRAIL and doxorubicin is more effective in retarding tumor growth of PC3 prostate carcinoma xenografts than either agent alone. Doxorubicin is an anthracycline, which intercalates into DNA thereby activating DNA repair pathways and elevating levels of p53. PC3 cells are p53^-/-^, which may explain why doxorubicin as a single agent was relatively uneffective against tumor growth inhibition \[[@B20]\]. However doxorubicin was able to reduce expression of the anti-apoptotic protein c-FLIP *in vivo*. We have previously shown that sequential treatment of doxorubicin followed by TRAIL resulted in cell death *in vitro*\[[@B15]\]. Our data suggest that downregulation of c-FLIP is an important step *in vivo*that enhances TRAIL-induced apoptosis as evidenced by increased TUNEL staining in tumors treated with combination therapy. Other agents that reduce c-FLIP and increase death ligand mediated apoptosis include cycloheximide, 9-nitrocamptothecin, cisplatin, and the proteasome inhibitor PS-341 \[[@B21]-[@B24]\]. These agents affect multiple cellular responses suggesting that downregulation of c-FLIP may not be the sole factor by which death ligand-induced apoptosis is enhanced. However, selective reduction of c-FLIP using RNA interference or anti-sense technology is sufficient to sensitize human cell lines, including Du145 prostate carcinoma cells to death receptor ligands, indicating that c-FLIP is a major provider of resistance in this apoptotic pathway \[[@B25],[@B26]\].
Tumors in mice that received Apo2L/TRAIL alone were about 20% smaller than those of the control group. Thus the growth inhibitory effect of Apo2L/TRAIL on PC3 xenografts *in vivo*was reflective of the results obtained *in vitro*. In contrast to single agent therapy, combination treatment with doxorubicin and TRAIL resulted in tumors that had 80% less volume than those in the control group. Tumors that received combination therapy continued to grow slowly, indicating that complete regression of PC3 xenografts was not achieved using a single treatment. In a recent study, PC3 xenografts were treated with irradiation followed by TRAIL, which resulted in complete growth inhibiton following three weekly rounds of therapy \[[@B27]\]. This indicates that multiple treatments may be neccessary to achieve a complete response.
We observed that a single administration of 4 mg/kg doxorubicin reduced cFLIP protein in PC3 xenografts but not mouse heart. One possibility is that cFLIP expression is differentially affected in normal versus malignant cells. If so, then doxorubicin would preferentially lower the apoptotic threshold in cancer cells and facilitate the selective elimination of these cells. Alternatively, the difference may be species specific. When grown in mice, human xenografts of various origins have successfully been treated by TRAIL in combination with other agents \[[@B5],[@B24],[@B27]-[@B29]\]. In contrast, combination of doxorubicin and TRAIL *in vitro*can induce cell death in normal human breast, mesothelial, or prostate epithelial cells, \[[@B14],[@B30],[@B31]\], which suggests that combination therapy may lack specificity when applied to humans. Could combination therapy in mice be non-toxic because endogenous mouse c-FLIP remains unaffected by chemotherapy? To avoid the possible complication of toxicity the effect of chemotherapy on c-FLIP expression in human patients should be carefully evaluated before considering systemic combination of chemotherapy and Apo2L/TRAIL.
Conclusions
===========
We found that combination of doxorubicin chemotherapy and Apo2L/TRAIL is more effective in tumor growth inhibition than either agent alone, indicating that this may represent a novel treatment strategy against prostate cancer. One of the mechanisms by which doxorubicin may enhance Apo2L/TRAIL apoptosis in PC3 xenografts is reduced expression of the anti-apoptotic protein c-FLIP. Future studies are needed to further investigate the effect of chemotherapy on c-FLIP in human patients and to optimize the treatment schedule to achieve complete tumor regression.
Competing interests
===================
The author(s) declare that they have no competing interests
Authors\' contribution
======================
AZ carried out the animal experiments using doxorubicin and Apo2L/TRAIL and TUNEL staining. JM carried out the animal experiments using doxorubicin. CVJ carried out viability assays and western blot analysis, conceived of the study, participated in its design and coordination, and prepared the manuscript. All authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2407/5/2/prepub>
Acknowledgements
================
The authors would like to thank Genentech for its generous supply of Apo2L/TRAIL, Dr. Marcus Peter for the NF-6 hybridoma supernan and Margaret Romano from the Hollings Cancer Center histopathology core laboratory for excellent technical assistance. This research was supported by grants from the Department of Defense (N6311600MDM0601) and National Institutes of Health (NIH RO1 CA102218) to CVJ.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**Cytotoxic effects of KillerTRAIL and Apo2L/TRAIL *in vitro*.**PC3 cells were incubated with ligand at the indicated concentrations without (A), or with 0.25 μg/ml (B), 0.5 μg/ml (C) or 1 μg/ml (D) doxorubicin for 24 hours. Data shown are the mean ± SEM from a representative experiment. Similar results have been obtained three times.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**The effect of doxorubicin on growth of PC3 xenografts.**Athymic nude mice were injected with PC3 cells (day 0) and treated with doxorubicin (i.p.) on day 10. Tumor size in animals treated with 4 or 8 mg/kg doxorubicin were significantly different from the untreated group on day 18 and 22, P \< 0.05.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Effect of doxorubicin on c-FLIP *in vivo***Mice bearing PC3 xenografts were untreated or treated with 4 mg/kg or 8 mg/kg doxorubicin (i.p.). After 24 hours, protein was isolated from xenografts (A) or mouse heart (B) and probed for c-FLIP expression by western blotting. Actin was included as a loading control.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**TUNEL staining of PC3 xenografts**Mice bearing bilateral PC3 xenografts were untreated or treated with doxorubicin (i.p.) before receiving intratumoral injections of 500 μg Apo2L/TRAIL (left) or vehicle (right). After 24 hours, tumors were harvested, fixed, processed and analysed for TUNEL staining. Data shown are the mean ± SEM calculated from TUNEL positive cells counted in four fields.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Effect of combination therapy on the growth of PC3 xenografts**Mice bearing PC3 xenografts were untreated or treated with 4 mg/kg doxorubicin (i.p.), followed by either vehicle or 500 μg Apo2L/TRAIL (i.p.). Data shown represent tumor size twenty-six days after treatment was initiated (A). Tumor size in animals treated with combination therapy was monitored an additional week (33 days after treatment initiation) (B). Tumor growth over time (C) demonstrates that tumor growth is significantly different on and after day 21 (\* indicates p \< 0.05).
:::
:::
|
PubMed Central
|
2024-06-05T03:55:52.069227
|
2005-1-7
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546011/",
"journal": "BMC Cancer. 2005 Jan 7; 5:2",
"authors": [
{
"first": "Ahmed",
"last": "El-Zawahry"
},
{
"first": "John",
"last": "McKillop"
},
{
"first": "Christina",
"last": "Voelkel-Johnson"
}
]
}
|
PMC546012
|
Background
==========
Telomerase is a ribonucleoprotein complex that consists of an essential RNA molecule, hTR, with a template domain for telomeric DNA synthesis and of a catalytic protein, hTERT, with reverse transcriptase activity. Functional telomerase is minimally composed of both hTR and hTERT \[[@B1],[@B2]\]. The transcriptional control of these two genes is a major step in the regulation of telomerase expression in human cells, with high expression of both genes detected in cancer cells relative to normal cells \[[@B3]-[@B6]\]. Several groups have recently reported transcriptionally targeted cancer gene therapy strategies based on the differential activities of hTR and hTERT promoters between normal and cancer cells \[[@B7]-[@B11]\]. Thus, investigation of the activating and repressive mechanisms of telomerase gene transcription has become an area of intense interest in cancer research.
The molecular regulation of hTR gene transcription in cancer cells remains poorly understood. The previously identified core promoter region in the hTR gene has several features utilised by the basal RNA PolII transcription machinery, including one CCAAT-box and four Sp1 sites termed Sp1.1-Sp1.4. The activity of the hTR promoter is controlled by NF-Y, Sp1 and Sp3 in bladder cancer cells in vitro and we have recently shown that an Sp1 site mutation in the hTR promoter detected in a blood sample taken from a paroxysmal nocturnal haemoglobinuria (PNH) patient can alter core promoter activity in vitro, raising the possibility that mutation might affect hTR gene transcription in hematopoietic cells *in vivo*\[[@B12]-[@B14]\]. Several other known transcriptional regulators, including the retinoblastoma protein pRB, are able to affect hTR transcription in the experimental setting of over-expression \[[@B12],[@B13]\].
The mechanism whereby pRB activates hTR remains unknown though pRB is not known to interact directly in a specific fashion with DNA, relying instead on recruitment to genes through interaction with other transcriptional regulators including Sp1 and mdm2. The recent finding that pRB induces Sp1 activity by binding to mdm2 resulting in the physical release of Sp1 from mdm2 and enhancement of its binding to consensus sequence implies that mdm2 might inhibit promoters such as hTR that are positively regulated by pRb and Sp1 \[[@B15]\]. In this study, we investigated regulation of hTR reporter constructs by Sp1, pRb, NF-Y and mdm2 and performed chromatin immunoprecipitation (ChIP) assays to determine whether mdm2 plays a role in hTR regulation in the p53 and pRb negative bladder cancer cell line 5637 which also expresses relatively low levels of mdm2 \[[@B16]\]. We found that mdm2 interacts with the hTR promoter in vivo and that mdm2 expression can down-regulate hTR promoter activity and suppress pRb, Sp1 and NF-Y-mediated transactivation. Sp1 sites within the hTR core promoter were not absolutely required for this negative effect. These studies demonstrate that hTR transcription is dominantly repressed by mdm2 through functional, and possibly physical, interactions with the hTR promoter complex.
Methods
=======
Materials and cell culture
--------------------------
Antibodies to Sp1, TFIIB and mdm2 were purchased from Santa Cruz Biotechnology Inc. (Santa Cruz, CA). Rabbit polyclonal antibodies directed against NF-YA, B and C were obtained from R. Mantovani (University of Milan, Milan, Italy). The 5637 cell line, originally established from the primary bladder carcinoma of a 68-year-old man in 1974, was purchased from DSMZ (No: ACC 35). 5637 cells were maintained at 37°C in 5% CO~2~in 1640 medium supplemented with 10% foetal bovine serum, penicillin, and streptomycin.
Plasmid construction and site-directed mutagenesis
--------------------------------------------------
Construction of the promoter fragment hProm867 and subcloning as an Xho I/Hind III fragment in the luciferase reporter pGL3-Basic (Promega, Madison, WI) was previously reported \[[@B17]\]. The reporter contains an 867 bp fragment of the hTR promoter. For generation of the Sp1 site mutation construct a two step cloning strategy was used to prevent unexpected mutations in luciferase reporter vectors; (i) an hTR 176 bp fragment (2923 wt, spanning from -107 to +69 bp) was cloned into the Xho I/Hind III sites in pCR-Script™ plasmid vector (Stratagene, La, Jolla CA), which was used as template for PCR mutagenesis using a QuikChange™ site-directed mutagenesis kit (Stratagene, La, Jolla CA) following the manufacturer\'s instructions. (ii) All mutation fragments were reconstructed into the Xho I/Hind III sites of pGL3-basic vectors and verified by DNA sequencing. The multiple-site mutation construct was generated in several separate PCR reactions as previously described \[[@B12]\].
Transfection and dual-luciferase reporter assay
-----------------------------------------------
The hTR promoter plasmids containing firefly luciferase reporters were cotransfected into tumour cells with an internal *Renilla*luciferase control, pRL-SV40 (Promega) using Superfect Transfection Reagent (Qiagen) as previously described \[[@B12],[@B13]\]. 5637 cells were cotransfected with 0.5 μg of expression vectors encoding wild-type NF-YA, B and C (kindly donated by Dr R. Mantovani \[[@B18]\]), titrations of Sp1, pRb and mdm2, 3 μg of the plasmids containing the luciferase reporter gene and 0.5 μg of pRL-SV40 plasmid for control of transfection efficiency. The total amount of DNA was kept constant at 10 μg with Salmon sperm DNA. The activity of both firefly and *Renilla*luciferase was determined 48 h later using the Dual Luciferase Assay kit (Promega). A minimum of three independent transfections were performed in duplicate and specific hTR promoter activity was normalized to protein as described elsewhere \[[@B12],[@B13]\].
Chromatin immunoprecipitation assays
------------------------------------
Formaldehyde cross-linking and chromatin immunoprecipitation were performed as described previously \[[@B19]\]. In brief, 5637 cell cultures were treated with formaldehyde for 10 min followed by the addition of glycine to a final concentration of 0.125 M. Cells were then washed twice with cold PBS and were resuspended in lysis buffer (1% SDS, 10 mM EDTA, 50 mM Tris-HCI, pH 8.1) with a proteinase inhibitor. After brief sonication to fragment the DNA with an average fragment size of 500 bp, the DNA fragments crosslinked to the proteins were enriched by immunoprecipitation with specific antibodies. A \"No-Ab\" sample was included as a negative control for the immunoprecipitation step. After reversal of the crosslinks and DNA purification, the extent of enrichment was monitored by PCR amplification of promoters using forward and reverse primers to the hTR (5\'-TACGCCCTTCTCAGTTAGGGTTAG-3\' and 5\'-AGCCCGCCCGAGAGAGTGAC-3\') gene promoter fragments \[Zhao, 2003 \#9\] and to the GAPDH coding region as a negative control (5\'-TGAAGGTCGGAGTCAACGGATTTGGT-3\' and 5\'-CATGTGGGCCATGAGGTCCACCAC-3\'). The PCR product was separated by agarose gel electrophoresis. The input sample was processed with the rest of the samples from the point at which the cross-links were reversed. All chromatin immunoprecipitation experiments were repeated three times and PCR analysis of individual experiments was also performed at least twice.
Results
=======
Mdm2 interacts with hTR core promoter in vivo
---------------------------------------------
Regulation of hTR promoter activity by pRb and Sp1 suggests that mdm2 might also play a role in hTR promoter regulation. To investigate whether mdm2 protein targets the hTR promoter *in vivo*, we performed chromatin immunoprecipitation experiments using antibodies to Sp1, TFIIB and mdm2. The presence of the hTR promoter in chromatin immunoprecipitates was detected by semi-quantitative PCR. Antibodies to Sp1 and TFIIB, both of which have previously been detected at the hTR promoter in chromatin immunoprecipitation, and to mdm2 all precipitated the hTR core promoter DNA sequence but failed to precipitate the negative control GAPDH coding sequence from 5637 cells (Fig. [1](#F1){ref-type="fig"}). These results establish that mdm2 binds the hTR core promoter *in vivo*.
Mdm2 represses the hTR promoter and pRb and Sp1 mediated transactivation
------------------------------------------------------------------------
To address whether mdm2 affects the activity of the hTR promoter *in vitro*, a luciferase reporter carrying an 867 bp fragment of the hTR promoter (-798/+69) was used in transient co-transfections with a full-length mdm2 expression vector. As shown in Fig. [2](#F2){ref-type="fig"}, the hTR promoter was repressed by mdm2 in a dose dependent manner. Promoter activity under maximal repression by mdm2 was 42% of basal activity using 3 μg of the mdm2 expression vector.
Mdm2 has been shown to interact with the transcription factor Sp1 *in vitro*and *in vivo*and to inhibit transactivation of Sp1-activated promoters \[[@B15],[@B20]\]. We next tested whether mdm2 could interfere with Sp1 and pRb mediated transactivation of the hTR promoter. The 5637 cell line expresses low levels of mdm2 and does not express functional p53 or pRb \[[@B16]\]. As shown in Fig. [3A](#F3){ref-type="fig"}, and as previously reported, co-transfection of 5637 cells with Sp1 increased hTR promoter activity 2-fold relative to the vector control. Mdm2 overexpression was able to repress Sp1 mediated activation at all mdm2 concentrations, with Sp1 mediated induction completely abolished in the presence of 1 μg mdm2. Even in the continued presence of Sp1 overexpression, at high mdm2 concentrations promoter activity was reduced to sub-basal levels similar to those observed with mdm2 alone. Similarly, pRb overexpression led to a 3.75-fold induction of promoter activity that could be repressed in a dose dependent manner by titration of mdm2 (figure [3B](#F3){ref-type="fig"}). Interestingly though, at the plasmid concentrations used, even the highest concentration of mdm2 did not completely inhibit induction by pRb. Promoter activity was still induced by 1.7-fold in the presence of both pRb and 3 μg mdm2, rather than repressed to the sub-basal levels observed with Sp1/mdm2 co-expression.
To investigate whether this repression of the hTR promoter by mdm2 was entirely dependent on Sp1, we transfected a core promoter construct harboring functional mutations in all Sp1 binding sites. It should be noted the construct used here contains a shorter promoter fragment than that used throughout. Our previous study showed that mutation of all four Sp1 binding sites does not impair the activity of this hTR core promoter or its ability to be transactivated by NF-Y, but disrupts Sp1 binding and activation \[[@B13]\]. As shown in figure [3C](#F3){ref-type="fig"}, mdm2 was also able to repress this mutant, indicating that Sp1 is not essential for mdm2 mediated repression of the hTR promoter *in vitro*. Furthermore, gel shift analysis using a probe corresponding to the Sp1.3 site did not produce an mdm2 specific supershift, suggesting that mdm2 may not directly bind the hTR promoter DNA (data not shown).
Mdm2 interferes with NF-Y-dependent activation of hTR promoter
--------------------------------------------------------------
Since mdm2 can influence the basal activity of the hTR reporter without targeting its Sp1 binding sites, it is likely that at least part of the repressive effect is directed through another pathway. One alternative mechanism by which mdm2 could regulate the construct is through interfering with the NF-Y-CCAAT-box complex. Therefore, we next tested whether mdm2 could repress NF-Y function. As shown in Fig. [4](#F4){ref-type="fig"}, the hTR promoter is strongly stimulated (more than 5-fold induction) by co-transfection of expression vectors encoding the three NF-Y subunits, NF-YA, B and C. Titration of mdm2 again resulted in dose-dependent inhibition of promoter induction, reducing NF-Y mediated activation to 1.4-fold relative to basal levels at the highest mdm2 concentration. Thus, forced expression of mdm2 also attenuates NF-Y-activated transcription. This finding suggests that in addition to Sp1 dependent effects, down-regulation of hTR promoter activity by mdm2 may also be mediated partly by a mechanism involving inhibition of NF-Y-activated transcription.
Discussion
==========
The retinoblastoma gene product pRb acts as a positive regulator of hTR promoter activity by an unknown mechanism while Sp1 and NF-Y activate the hTR promoter by directly binding to DNA. Recent studies have suggested that pRb can mediate stimulatory effects at Sp1 stimulated promoters by liberating Sp1 from negative regulation by mdm2 \[[@B15],[@B21]-[@B23]\]. Mdm2 regulates the activities of both p53 and pRB, and physically interacts with Sp1 to repress transcription \[[@B24]-[@B26]\]. Conversely, pRB interaction with mdm2 displaces Sp1 and restores Sp1 DNA binding and transactivation activity. Additionally, mdm2 protein activates the p53 pathway under stress conditions and also represses transcription directly by interaction with elements of the basal transcription machinery and their binding sites \[[@B27],[@B28]\].
In this study, we used the p53 and pRb negative bladder cancer cell line 5637 to provide evidence that mdm2 interacts with the hTR core promoter *in vivo*(figure [1](#F1){ref-type="fig"}) and serves as a negative regulator of the hTR gene promoter in vitro (figure [2](#F2){ref-type="fig"}). Mdm2 opposed Sp1 and pRb directed transactivation of the hTR promoter, suggesting a plausible mechanism whereby transfected pRb may elicit its effects by opposing the action of mdm2 (figure [3](#F3){ref-type="fig"}). The reciprocal scenario in which mdm2 might oppose pRB cannot explain the negative effect of transfected mdm2 alone since the cells used here lack functional pRB. Rather, other mechanisms, probably including direct inhibition of Sp1 must be involved.
Mdm2 can bind directly to Sp1 and inhibit its DNA binding and can also bind to Sp1 sites at some promoters such as that of p65 \[[@B27]\]. Mdm2 suppressed Sp1 mediated transactivation in this study, but also had a more general repressive effect that cannot have been entirely dependent on inhibition of Sp1 since an hTR core promoter construct carrying functional mutations in all Sp1 sites was still repressed by mdm2 (figure [3C](#F3){ref-type="fig"}). This clearly argues against direct binding of mdm2 to those sites and implies the existence of additional repressive mechanisms.
Here we demonstrated that activation of the hTR promoter by over-expression of all three sub-units of NF-Y could also be inhibited by mdm2 (figure [4](#F4){ref-type="fig"}). The CCAAT box binding protein NF-Y is a ubiquitous factor with central roles in PolII mediated transcription at numerous promoters. In cooperation with accessory proteins, NF-Y binds promoters *in vivo*before gene activation and \"pre-sets\" the promoter architecture allowing access by other regulatory proteins. There is also increasing evidence that NF-Y recruits multiple components of the basal PolII machinery \[[@B29]-[@B33]\]. Mdm2 and pRb have also been reported to interact with the basal PolII transcriptional apparatus \[[@B28],[@B34]-[@B36]\]. Therefore, the possibility exists that mdm2 also inhibits this promoter through one of these other interactions although this remains to be tested (figure [5](#F5){ref-type="fig"}). Thus either an interaction with the basal transcriptional machinery or specific transcription factors may regulate hTR promoter activity.
Conclusions
===========
In conclusion, the hTR promoter is dominantly suppressed by mdm2. Mdm2 may utilise more than one mechanism to attenuate hTR promoter activity (Fig. [5](#F5){ref-type="fig"}). Mdm2 may directly repress activation by both pRB and Sp1, or activation by NF-Y. Furthermore, the ability of mdm2 to interact and interfere with components of the general transcription machinery might partly explain the general repressive effect seen here. Elucidation of new regulators affecting hTR basal promoter activity in cancer cells provides a basis for future studies aimed at improving our understanding of the differential hTR expression between normal and cancer cells. This will be essential for a thorough understanding of the regulation of telomerase and for advancement of telomerase directed gene therapies.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
JZ, AB and KJ carried out the transfection and genetic analysis. AB helped draft the manuscript. WNK conceived the study, participated in design and coordination and helped draft the manuscript. All authors read and approved the final manuscript.
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-2407/5/6/prepub>
Acknowledgements
================
This work was supported by the Cancer Research UK and Glasgow University.
Figures and Tables
==================
::: {#F1 .fig}
Figure 1
::: {.caption}
######
**mdm2 interacts with the hTR core promoter in vivo.**Formaldehyde cross-linked chromatin was prepared from 5637 cells and immunoprecipitated with antibodies to Sp1 (lane 4), TFIIB (lane 5) and mdm2 (lane 6), or in the absence of antibody (lane 3). PCR detection of DNA sequences immunoprecipitated with each antibody is shown in these lanes. PCR was performed with specific primers for the hTR promoter and for the GAPDH coding region as a negative control. A sample representative of the total input chromatin (input DNA lanes 1 and 2) was included in the PCR analysis. Lanes 7 and 8 show PCR positive and negative controls. Chromatin immunoprecipitation experiments were repeated three times and PCR analysis of individual experiments was also performed at least twice.
:::
:::
::: {#F2 .fig}
Figure 2
::: {.caption}
######
**mdm2 represses hTR promoter activity.**The pRB/p53 negative cell line 5637 were co-transfected with 1.5 μg of the pLh2023(-796/+69) construct and increasing amounts of mdm2 expression vector (from 0.5 to 4.0 μg). Total input DNA amount for transfection was adjusted with Salmon Sperm DNA to ensure a constant amount in all transfections. After 48 hour of culture, cells were harvested, and cell lysates were assayed for firefly and Renilla luciferase activity. The data is expressed as fold induction of luciferase activity relative to the promoter alone and data presented are the mean of three independent experiments performed in duplicate.
:::
:::
::: {#F3 .fig}
Figure 3
::: {.caption}
######
**Mdm2 represses pRb and Sp1 mediated hTR promoter transactivation a) mdm2 suppresses Sp1 activation of hTR promoter activity.**5637 cells were co-transfected with 1.5 μg of the pLh2023(-796/+69) construct, 2.0 μg of the expression vector for either Sp1 and a titration of mdm2. Total input DNA amount for transfection was adjusted with Salmon Sperm DNA to ensure a constant amount in all transfections. After 48 hour of culture, cells were harvested, and the cell lysates assayed for firefly and Renilla luciferase activity. The data is expressed as fold induction of luciferase activity relative to the promoter alone and data are the mean of three independent experiments performed in duplicate. **b) mdm2 suppresses pRB activation of hTR promoter activity**5637 cells were co-transfected with 1.5 μg of the pLh2023(-796/+69) construct, 2.0 μg of the expression vector for either pRb and a titration of mdm2. Total input DNA amount for transfection was adjusted with Salmon Sperm DNA to ensure a constant amount in all transfections. After 48 hour of culture, cells were harvested, and the cell lysates assayed for firefly and Renilla luciferase activity. The data is expressed as fold induction of luciferase activity relative to the promoter alone and data are the mean of three independent experiments performed in duplicate. **c) Sp1 sites in the hTR promoter are not essential for repression by mdm2**5637 cells were co-transfected with 3.0 μg of the 2923 wild type or 2923 mSp1(4) mutant constructs \[13\], together with mdm2. After 48 hour of culture, cells were harvested, and the cell lysate were assayed for firefly and Renilla luciferase activity. The data is expressed as fold induction of luciferase activity relative to the promoter alone and the means of the three independent experiments performed in duplicate.
:::
:::
::: {#F4 .fig}
Figure 4
::: {.caption}
######
**Mdm2 inhibits NF-Y mediated hTR transactivation.**5637 cells were co-transfected with 1.5 μg of the pLh2023(-796/+69) construct, 0.5 μg of expression vectors encoding wild-type NF-YA, B and C together with increasing amounts of mdm2 expression vector (from 0 to 2.5 μg). Total input DNA amount for transfection was adjusted with Salmon Sperm DNA to ensure a constant amount in all transfections. After 48 hour of culture, cells were harvested, and the cell lysate were assayed for firefly and Renilla luciferase activity. The data is expressed as fold induction of luciferase activity relative to the promoter alone and the means of the three independent experiments performed in duplicate.
:::
:::
::: {#F5 .fig}
Figure 5
::: {.caption}
######
**Multiple mechanisms of the transcriptional repression of the hTR gene by mdm2.**Schematically shown are various mechanisms by which mdm2 might act on the hTR core promoter. The hTR core promoter elements include Sp1 sites (recognised by Sp1 family factors) and a CCAAT-box (recognized by NF-YA, B and C). NF-Y may facilitate the access of upstream activators (such as Sp1) to their cognate enhancer/promoter sequences, recruit TAFs components of TFIID \[33\] and serve to instigate the formation of a pre-initiation complex of general transcription factors and Pol II. Co-regulators such as pRB/mdm2 for modulation of signals between transcription factors and the core transcriptional machinery might act through the hTR core promoter directly or indirectly.
:::
:::
|
PubMed Central
|
2024-06-05T03:55:52.070978
|
2005-1-18
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546012/",
"journal": "BMC Cancer. 2005 Jan 18; 5:6",
"authors": [
{
"first": "Jiangqin",
"last": "Zhao"
},
{
"first": "Alan",
"last": "Bilsland"
},
{
"first": "Katrina",
"last": "Jackson"
},
{
"first": "W Nicol",
"last": "Keith"
}
]
}
|
PMC546013
|
Background
==========
Trichotillomania (TTM) is characterized by repetitive stereotypical hair-pulling from different sites resulting in noticeable hair loss \[[@B1]\]. Phenomenological observations have suggested that symptoms of repetitive hair-pulling are reminiscent of the compulsions seen in obsessive-compulsive disorder (OCD) \[[@B2],[@B3]\]. For example, both TTM and OCD patients describe compulsive urges and ritualistic behaviours \[[@B2],[@B4]\]. Comorbidity data also suggest some overlap between TTM and OCD \[[@B2]\]. Thus, a number of authors have suggested that TTM might be classified with OCD in a spectrum of disorders having similar phenomenology \[[@B4]-[@B8]\].
However, in addition to overlapping phenomenology between OCD and TTM, there are also significant differences. For example, in contrast to compulsions in OCD, hair-pulling in TTM is not in response to obsessive thoughts (such as worry about harm to self or others) but rather because of an irresistible urge and the promise of gratification when pulling out hair \[[@B2],[@B6]\]. Also, unlike patients with OCD whose symptoms change over time in terms of focus and severity (e.g. from washing of hands to checking locks, stoves, appliances, etc) \[[@B9]\], TTM patients usually only present with hair-pulling without evolution to non-self-injurious compulsive rituals.
Examination of demographic variables in OCD and TTM supports the argument that these are two distinctive disorders. TTM is much more prevalent in females (10:1 female to male ratio) \[[@B10]\] whereas OCD is equally common in males and females \[[@B11]\]. Age of onset also differs somewhat: TTM typically presents in early adolescence, with the mean age of onset of hair-pulling in males later than that in females \[[@B10],[@B12],[@B13]\] whereas OCD has its onset from childhood through to early adulthood \[[@B14]\], but with males reporting an earlier onset compared to females \[[@B15]\].
Additional clinical observations further support a distinction between OCD and TTM. Patients with TTM tend to have fewer comorbid obsessive-compulsive symptoms, as well as less depression and anxiety compared to OCD patients \[[@B16]\]. Response prevention in OCD patients eventually leads to anxiety reduction, whereas in people with TTM it may lead to an increase in anxiety \[[@B17]\]. Although a selective response to serotonergic reuptake inhibitors (SRI\'s) has been suggested to characterize both OCD and TTM, there is good evidence that response to SRI\'s is sustained in OCD, whereas the evidence-base for the efficacy of these agents in TTM is much more mixed.
Relatively few empirical studies have, however, documented the phenomenological similarities and differences between OCD and TTM \[[@B3],[@B18],[@B19]\]. A large clinical database comprised of patients with OCD and TTM provided us an opportunity to investigate the relationship between these conditions in terms of demographic and clinical variables.
Methods
=======
Subjects
--------
Two hundred and seventy eight OCD patients (n = 278: 148 male; 130 female), and 54 TTM patients (n = 54; 5 male; 49 female), ranging in age between 8 and 75 years, took part in the study (Table [1](#T1){ref-type="table"}). These patients were referred to our research unit from a wide range of sources (including the OCD Association of South Africa, community based primary care practitioners, and psychiatrists). Either a clinical psychologist or a psychiatrist with expertise in the field interviewed participants. Participants met the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) criteria \[[@B1]\] for either a primary diagnosis of OCD or TTM on the Structured Clinical Interview for Axis I Disorders (SCID-I) \[[@B20]\]. Patients were included irrespective of whether they were at baseline (i.e. not receiving any form of treatment for their primary psychiatric disorder), or were receiving treatment for OCD / TTM, but those with comorbid OCD and TTM (N = 25) were excluded from subsequent analysis. A history of psychosis was also an exclusion criterion. Referring clinicians were contacted to establish, where possible, a longitudinal expert evaluation of the diagnostic status of the patient. All subjects gave informed written consent to participate after confidentiality was guaranteed and risks and benefits had been fully explained. The study was approved by the Institutional Review Board of the University of Stellenbosch.
::: {#T1 .table-wrap}
Table 1
::: {.caption}
######
Demographic information: OCD and TTM
:::
**Variables** **OCD (N = 278)** **TTM (N = 54)** **χ^2^** **P**
-------------------- ------------------------------------------------------- ------------------------------------------------------- ---------- ----------- ------ --------
Gender 148 male 130 female 5 male 49 female 40.7 \<.001
Age (SD) 33.1 (14.4) 31.3 (12.5) NS
Population group 86.6% Caucasian 79.6% Caucasian NS
Level of education 50.7% completed high school or higher level education 44.4% completed high school or higher level education NS
Employment 6.8% unemployed 3.7% unemployed NS
NS = non-significant
:::
Interview
---------
Specific demographic data, including age when interviewed, age of onset of OCD/TTM, highest level of education, current employment status, and population group were obtained from all participants. In addition to the SCID-I, and selected parts of the SCID-II (obsessive-compulsive, avoidant, schizotypal, borderline personality disorders) for adult patients (aged 18 or older) \[[@B20]\], the interview also included the Structured Clinical Interview for Obsessive-Compulsive Spectrum Disorders (SCID-OCSD) to determine the presence of other obsessive-compulsive related conditions \[[@B21]\].
The Yale-Brown Obsessive-Compulsive Severity Scale (Y-BOCS) \[[@B22]\] was implemented to assess the severity of OCD symptoms. Severity of hair-pulling symptoms was assessed with the Massachusetts General Hospital Hair-pulling Scale \[[@B23]\]. The Trichotillomania Behaviour Profile (TBP, available from the first author on request) was administered to TTM patients to assess hair-pulling phenomenology.
Patients\' level of insight into the senselessness or excessiveness of their symptoms was assessed on the relevant YBOCS item. When an adequate trial of pharmacotherapy with an SRI (i.e. for both OCD and TTM groups, at least 10 weeks on the medication with a minimum of 6 weeks on mid-range dose) had been undertaken, response to pharmacotherapy was assessed using the global improvement item of the Clinical Global Impression (CGI) scale; subjects with CGI scores of 1 (\'very much improved\') or 2 (\'much improved\') were defined as responders \[[@B24]\]. Similarly, when patients received an adequate trial of cognitive behavioural therapy (CBT) (i.e. for both OCD and TTM groups, 8 or more sessions with an expert CBT psychotherapist), response to treatment was rated using the CGI. The Disability Profile questionnaire (DP) \[[@B25]\] was included in the interview to assess current (i.e. past two weeks) and lifetime impairment in eight domains. The DP was initially developed for use in patients with social anxiety disorder; nevertheless, the scale has since been used to assess disability in patients with other anxiety disorders as well \[[@B26]\].
Questions addressing potential precipitating or exacerbating factors, including the impact of menstrual/reproductive cycle changes, brain trauma and history of autoimmune infections on OCD/TTM symptom fluctuations, were included in the interview.
Self-report questionnaires
--------------------------
Severity of comorbid depression was evaluated with the Beck Depression Inventory (BDI) \[[@B27]\]. The Childhood Trauma Questionnaire (CTQ) \[[@B28]\], a scale proven to be a valid and reliable measure of past traumatic experiences \[[@B29]\], was used as a self-report questionnaire to assess the nature and severity of childhood trauma. Sub-scales of the CTQ include measures of emotional abuse, physical abuse, sexual abuse, emotional neglect and physical neglect.
The self-report Temperament and Character Inventory (TCI) \[[@B30]\] was also used to measure behaviours associated with seven personality dimensions, namely novelty seeking, harm avoidance, reward dependence, persistence, self-directedness, cooperativeness, and self-transcendence. In addition, participants completed the self-report Young Schema Questionnaire (YSQ) \[[@B31]\] to assess the current profile of fundamental maladaptive beliefs (cognitive schemas) in OCD and TTM. For each item of the 75-item \"short form\" of the YSQ (which includes 15 schemas), the answer is required to be placed on a 6-point Likert-type scale (1= \'completely untrue of me\', 2 = \'mostly untrue of me\', 3 = \'slightly more true than untrue\', 4 = \'moderately true of me\', 5 = \'mostly true of me\', 6 = \'describes me perfectly\').
Data analysis
-------------
As there were few males with TTM (Table [2](#T2){ref-type="table"}), only clinical data from females with OCD and TTM were analyzed. Chi-square and t-tests were performed to investigate the differences in OCD/TTM phenomenology where appropriate. A one-way analysis of variance (ANOVA) was done to investigate the effects of the primary and comorbid disorders on disability. Subsequently, a two-way fixed effects ANOVA was used to assess the main interactions between primary diagnosis and comorbidity on disability and to test for the main (fixed) effects of the different diagnoses on disability. Residuals of ANOVA\'s of cognitive schema data in OCD and TTM groups suggested non-normality of the data. As a result, pair-wise comparison tests (Mann-Whitney U) were implemented to compare the two groups on cognitive schemas.
::: {#T2 .table-wrap}
Table 2
::: {.caption}
######
Comparison of symptomatology: OCD *vs*TTM
:::
**Variables** **OCD** **TTM** **P**
--------------------------------------------- ----------------------------------------- ------------------------------------------ --------
Age of onset (SD) 19.3 (12.0) 11.8 (7.6) \<.001
Symptom severity (SD) YBOCS score: 20.1 (8.0); range: 0 -- 39 MGHHPS score: 16.1 (6.5); range: 0 -- 26
Severity of depressive symptoms (BDI score) 8.9 (11.3) 5.5 (7.2) .04
Poor insight 13.6% 0% NS
Treatment response SRI: 90.7% SRI: 42.9% .003
CBT: 73.3% CBT: 33.3% .02
Tics 12.3% 6.1% NS
:::
Results
=======
Demographics
------------
Gender distribution of the sample differed significantly, with a marked predominance of female participants with TTM compared to almost equal numbers of male and female participants with OCD. TTM patients had an earlier age of onset of illness compared to patients with OCD (Table [2](#T2){ref-type="table"}).
Clinical features
-----------------
### Comorbidity
A number of disorders were more frequent in females with OCD: major depressive disorder (MDD), dysthymia, panic disorder, hypochondriasis and intermittent explosive disorder. In terms of the selected Axis II disorders, obsessive-compulsive personality disorder (OCPD) was more frequent in females with OCD (Table [3](#T3){ref-type="table"}).
::: {#T3 .table-wrap}
Table 3
::: {.caption}
######
Lifetime comorbidity: OCD *vs*TTM
:::
**Disorder** **OCD (N = 130)** **TTM (N = 49)** **χ^2^** **P**
---------------------------------- ------------------- ------------------ ---------- ----------
Major depressive disorder 66.9% 49.0% 4.8 .03\*
Dysthymia 13.8% 2.0% 6.8 .009\*\*
Bipolar disorder 0.8% 0% 0.6 0.4
Panic disorder 20.8% 6.1% 6.4 .01\*
Alcohol abuse 5.4% 2.0% 1.1 0.3
Alcohol dependence 0.8% 0% 0.6 0.4
Substance abuse 0.8% 4.1% 2.0 0.2
Substance dependence 1.5% 2.0% 0.1 0.8
Social phobia 10.0% 8.2% 0.1 0.7
Specific phobia 18.5% 18.4% 0.0 \<1.0
Posttraumatic stress disorder 3.1% 0% 2.6 0.1
Generalized anxiety disorder 13.1% 20.4% 1.4 0.2
Body dysmorphic disorder 6.2% 6.1% 0.0 \<1.0
Anorexia Nervosa 8.5% 2.0% 2.9 0.1
Bulimia Nervosa 7.7% 6.1% 0.1 0.7
Binge-eating disorder 4.6% 10.2% 1.8 0.2
Hypochondriasis 4.6% 0% 3.9 \<.05\*
Stereotypic movement disorder 3.5% (N = 85) 0% (N = 18) 1.2 0.3
Tourette\'s disorder 3.8% 2% 0.4 0.5
Tics 12.3% 6.1% 1.6 0.2
Kleptomania 4.6% 4.1% 0.0 0.9
Pyromania 0% 2.0% 2.6 0.1
Compulsive shopping 6.9% 4.1% 0.5 0.5
Hypersexual disorder 1.5% 0% 1.3 0.3
Intermittent explosive disorder 16.2% 6.1% 3.5 .06
OCPD 39.2% 13.3% 11.3 .001\*\*
Avoidant personality disorder 21.2% (N = 99) 0.03% (N = 26) 2.9 0.9
Schizotypal personality disorder 5.3% (N = 94) 0% (N = 25) 2.4 0.1
Borderline personality disorder 22.3% (N = 94) 8% (N = 49) 3.0 0.1
\* p \< .05 (2-tailed)
\*\* p \< .01 (2-tailed)
:::
### Symptom severity
The severity of OC symptoms in OCD patients, as measured by the YBOCS severity scale, was 20.1 (± 8.0). TTM patients scored 16.1 (± 6.5) on average on the MGHHPS. Compared to TTM patients, females with OCD had significantly higher depressive symptom scores on the BDI (Table [2](#T2){ref-type="table"}).
### Disability
The DP was administered to a total of 95 OCD and 30 TTM patients (Table [4](#T4){ref-type="table"}). OCD patients reported significantly more lifetime impairment due to their illness than TTM patients. More specifically, OCD patients were more impaired in terms of work-related functioning, family functioning, marriage / dating, activities of daily life, and other activities (which included religious activities, membership of clubs, having hobbies, participation in sports etc.) and had more suicidality. One-way ANOVA\'s showed that the primary diagnosis (either OCD or TTM) (F = 11.84; p = 0.001) and panic disorder (F = 6.73; p = 0.01) had a significant effect on the levels of disability. However, there was no significant interaction effect between the primary diagnosis and panic disorder, suggesting that the levels of disability were dependent on primary diagnosis (either OCD or TTM) (F = 5.79; p = 0.02) and not influenced by the absence or presence of panic disorder (F = 0.001; p = 0.98).
::: {#T4 .table-wrap}
Table 4
::: {.caption}
######
Disability profile: OCD *vs*TTM
:::
**DOMAIN** **OCD (N = 95)** **TTM (N = 30)** **Mann-Whitney U**
-------------------------- ------------------ ------------------ -------------------- ----- --- ------ ------ ---------
School 1.0 0 4.0 1.0 0 4.0 -1.2 NS
Work 2.0 0 4.0 1.0 0 3.0 3.7 \<0.001
Family 2.0 0 4.0 1.0 0 4.0 2.4 0.02
Marriage / dating 2.0 0 4.0 1.0 0 4.0 3.1 0.002
Friendships 1.0 0 4.0 1.0 0 3.0 1.1 NS
Other activities 2.0 0 4.0 0 0 3.0 2.3 0.02
Activities of daily life 2.0 0 4.0 0 0 3.0 5.5 \<0.001
Suicide 1.0 0 4.0 0 0 3.0 2.6 0.009
Total disability 12.0 1.0 26.0 6.5 0 18.0 3.3 \<0.001
:::
### Character / Temperament
Compared to OCD patients, patients with TTM scored significantly higher on novelty seeking, whereas OCD patients had significantly greater harm avoidance (Table [5](#T5){ref-type="table"}).
::: {#T5 .table-wrap}
Table 5
::: {.caption}
######
Temperament and Character Inventory: OCD *vs*TTM
:::
**TEMPERAMENT / CHARACTER TRAITS\*** **OCD (N = 68)** **TTM (N = 21)** **F** **P**
-------------------------------------- ------------------ ------------------ ------- -------
NS 17.6 (6.8) 21.6 (6.0) .5 .02
HA 22.3 (7.9) 15.8 (7.3) .4 .001
RD 22.3 (4.2) 21.8 (4.9) .4 NS
SD 25.1 (8.2) 28.2 (9.6) .7 NS
C 30 (5.1) 29.3 (6.8) 2.8 NS
ST 15.2 (6.7) 21.3 (18.3) 5.0 NS
\* NS = novelty seeking total score SD = self-directedness total score
HA = harm avoidance total score C = cooperativeness total score
RD = reward dependence total score ST = self-transcendence total score
:::
### Schemas
Fifty-nine OCD and 26 TTM patients fully completed the YSQ. Pair-wise comparison tests (Mann-Whitney U) indicated that OCD and TTM patients differed significantly on 5 schemas, i.e. mistrust / abuse, social isolation, shame / defectiveness, subjugation and emotional inhibition (Table [7](#T7){ref-type="table"}). More specifically, OCD patients had significantly higher scores on each of these schemas compared to TTM patients.
::: {#T7 .table-wrap}
Table 7
::: {.caption}
######
OCD and TTM scores on the YSQ subscales
:::
**Schemas** **OCD (n = 59)** **TTM (n = 26)** **Mann-Whitney U**
----------------------- ------------------ ------------------ -------------------- -------- ----- ----- ------ --------
Median Min Max Median Min Max Z P
Emotional deprivation 2.4 1.0 5.6 2.3 1.0 5.4 -0.6 NS
Abandonment 2.8 1.0 6.0 2.1 1.2 6.0 -0.5 NS
Mistrust / abuse 2.6 1.0 5.8 1.9 1.9 5.4 -2.3 .02
Social isolation 2.4 1.0 6.0 1.9 1.9 6.0 -2.7 .007
Shame / defectiveness 2.2 1.0 6.0 1.4 1.4 4.8 -3.0 .003
Failure to achieve 2.0 1.0 6.0 1.9 1.9 4.8 -0.9 NS
Incompetence 2.0 1.0 4.8 1.8 1.8 4.4 -1.2 NS
Vulnerability to harm 2.2 1.0 6.0 1.6 1.6 5.2 -1.8 NS
Enmeshment 1.8 1.0 6.0 1.7 1.7 4.2 -0.6 NS
Subjugation 2.2 1.0 6.0 1.7 1.7 5.6 -2.8 .005
Self-sacrifice 3.4 1.2 6.0 3.2 3.2 5.8 -1.0 NS
Emotional inhibition 2.4 1.0 5.0 1.4 1.4 4.4 -3.8 \<.001
Unrelenting standards 3.8 1.2 6.0 3.6 3.6 6.0 -0.9 NS
Entitlement 2.6 1.0 5.8 2.7 2.7 6.0 -0.7 NS
Self-discipline 3.0 1.2 6.0 2.8 2.8 4.8 -1.3 NS
:::
Precipitating factors
---------------------
### Interpersonal trauma history
OCD patients reported more childhood sexual abuse than did TTM patients (p = .04).
### Brain trauma history
OCD and TTM patients did not differ significantly in terms of a history of serious head injury associated with the onset of OCD or TTM.
### History of autoimmune infections
Compared to none in the TTM group, 9 OCD patients reported onset of their OCD with an episode of bacterial pharyngitis (p = .06). In terms of other autoimmune infections, OCD and TTM patients did not differ significantly.
### Hormonal influence
Female OCD and TTM patients did not differ significantly in terms of the impact of premenstrual/menstrual/menopausal symptoms on their illness. Compared to 42 (38.5%) of 109 OCD patients who reported OC symptom changes in the premenstrual/menstrual period, 16 of 32 TTM patients (50%) reported regular changes in their symptoms during this time. Seventeen (n = 17) OCD patients were menopausal and 35.3% (n = 6) of these women reported that their OC symptoms only started with menopause. One of the TTM patients had gone through menopause with no effect on her hair-pulling symptoms. However, OCD and TTM patients differed significantly in terms of the temporal association between pregnancy/puerperium and onset of illness: 42.6% (26 of 61) OCD patients reported OCD onset while pregnant or within a month of childbirth, compared to 7.7% (1 of 13) of TTM patients (χ^2^= 6.8; p = .009).
Treatment response
------------------
Significantly fewer TTM patients reported a clinical response to either CBT- or SRI-treatment than did OCD patients (Table [2](#T2){ref-type="table"}).
Discussion
==========
A comparison of women with TTM and with OCD found significant differences in clinical variables; OCD patients had more comorbidity, greater disability, increased childhood interpersonal trauma (specifically sexual abuse) and more maladaptive schemas. Fewer TTM patients, however, reported having responded to treatment.
The gender ratio findings in both OCD and TTM groups were similar to other surveys where a mean female:male ratio of 1.5:1.0 in OCD \[[@B11],[@B32],[@B33]\] and approaching 10:1 in TTM \[[@B34]\] were documented. OCD patients\' mean total score on the YBOCS (i.e. 20.1 ± 8.0) puts them in the \"moderate\" severity category \[[@B22]\]. The mean hair-pulling severity score on the MGHHPS (16.1 ± 6.5) was similar to that reported in other studies \[[@B35],[@B36]\]. Taken together, these data suggest that our patients are not dissimilar from those assessed at other sites.
Our comorbidity findings are consistent with existing data suggesting that depressive and anxiety disorders are highly prevalent in both OCD and TTM, and significantly more prevalent in OCD \[[@B3],[@B16],[@B18]\]. Indeed, compared to TTM, comorbidity in OCD is greater across a range of different diagnostic categories including mood (MDD, dysthymia), anxiety (panic disorder), OCD-related (hypochondriasis) and personality disorders (OCPD). Such comorbidity appears to extend also to impulse control disorders (intermittent explosive disorder), a finding which argues against the current classification of trichotillomania as a member of this spectrum of conditions.
Our findings of increased disability in OCD is consistent with studies on OCD suggesting it is one of the most impairing of all medical disorders \[[@B37]\]. A number of clinical studies have emphasized the burden of OCD across different domains, including higher rates of divorce and separation than in subjects without OCD \[[@B13]\] and significantly impaired instrumental functioning (work, school, home making and family life) \[[@B38]-[@B41]\]. However, the impairment and distress due to TTM should not be underestimated: TTM can be associated with serious sociological and psychological effects (e.g. strong feelings of shame and embarrassment \[[@B42]\], as well as avoidance behaviour including potentially dangerous avoidance of medical care \[[@B43]\]) resulting in a significant decline in quality of life (QOL) for patients, their family members and significant others \[[@B44],[@B45]\].
OCD patients reported significantly more sexual abuse than TTM patients (p = .04). This finding differs from our previous data suggesting similar rates of childhood interpersonal trauma (CIT) in OCD and TTM \[[@B46]\]. However, the current sample size is much increased, resulting in more power to detect smaller differences. Indeed, increased rates of OCD (and other anxiety disorders) have previously been linked with a history of physical and sexual abuse during childhood \[[@B47],[@B48]\]. Nevertheless, in both OCD and TTM, dissociative symptoms -- which are present in a minority of patients in both conditions -- are positively correlated with a history of childhood interpersonal trauma \[[@B49]\], so that a potential role for CIT in some TTM patients should not be ignored \[[@B50]\].
TTM patients had significantly more novelty seeking (NS) than OCD patients, whereas OCD patients scored significantly higher on harm avoidance (HA) compared to TTM. Our findings are consistent with previous work on temperament / character in OCD, showing increased HA and decreased NS \[[@B51]-[@B53]\]. Of note, compared with mean temperament scores obtained in a normal community sample \[[@B30]\], both TTM and OCD scored high on HA. NS scores in the TTM sample were higher than in the OCD sample, but compared to normal controls, these fell in the \"medium\" range. The higher NS in TTM may however point to greater dopaminergic involvement in this disorder, and might also be used to argue that TTM lies closer to the more impulsive risk-/novelty-seeking pole of an impulsive-compulsive (IC) spectrum of disorders \[[@B54]\].
OCD patients had more maladaptive cognitive schemas than TTM, i.e. mistrust / abuse, social isolation, shame / defectiveness, subjugation and emotional inhibition. The schemas that OCD and TTM patients differed on are included in 2 of the 4 higher order factors (i.e. \"impaired autonomy\" and \"disconnection\") described by Lee and colleagues\' YSQ factor model \[[@B55]\]. While schemas are thought to represent responses to life experience, including the experience of a disorder, they may also reflect underlying symptoms. Given that maladaptive schemas in OCD were not reminiscent of its characteristic symptoms, it is likely that they at least partly reflect life experience. An increased number of maladaptive schemas in OCD is consistent with higher rates of comorbidity, disability, and functional impairment. Nevertheless, further empirical investigation is needed to assess the relationship between schemas and illness course.
Hormonal influences have previously been investigated in OCD \[[@B56]\] and TTM \[[@B57]\]. For example, it has been noted that menarche, premenstruum, pregnancy \[[@B58]\], and menopause \[[@B59]\] may be related to onset or relapse in OCD. Similarly, in a study that investigated the relationship of the menstrual cycle and pregnancy to compulsive hair-pulling, premenstrual symptom exacerbation was reported for actual hair-pulling, urge intensity and frequency, and ability to control pulling \[[@B57]\]. In that study the impact of pregnancy on TTM was less clear. Our findings suggest that significantly more OCD patients than TTM patients report an association between pregnancy/puerperium and the onset of illness. This finding is in part consistent with previous work suggesting that the postpartum may constitute a risk for the onset of OCD in women \[[@B60]\]. Taken together our data suggest both similarities and differences in the role of sexual hormones in the mediation of OCD and TTM.
Although rare, brain injury may play a role in some cases of OCD \[[@B61],[@B62]\]; in only one OCD patient (and none of the TTM patients), head injury was associated with onset of obsessive-compulsive symptoms. No data could be found on the potential role of brain injury in the etiology of hair-pulling.
A number of patients associated the onset of their OCD onset with an infection, possibly bacterial pharyngitis. This finding is consistent with a body of data suggesting post-streptococcal disease is a cause of OCD in children and adolescents \[[@B63]\], and perhaps also adults \[[@B64],[@B65]\]. There is less work demonstrating a role for autoimmune factors in TTM \[[@B66]\]. Notably, the data on bacterial pharyngitis were based on retrospective assessment and could have been contaminated by memory bias.
In our study, more OCD patients reported a positive response to treatment (with CBT or SRI\'s) than TTM patients. These data should be interpreted cautiously given the retrospective assessments. Nevertheless, there is evidence that SRI\'s in TTM may not be as effective over the long-term as in OCD \[[@B2]\]. About 40--60% of OCD patients respond to the first trial of an SRI \[[@B67]\], with a proportion of non-responders to a single SRI responding to administration of a second SRI \[[@B68]\]. In comparison, it has been suggested that TTM patients judge their treatment (including pharmacotherapy, psychotherapy, and behaviour modification) to be relatively ineffective \[[@B69]\]. The usefulness of SRI\'s in TTM has been investigated in a number of studies with results so far being equivocal. For example, Christenson et al \[[@B70]\] were unable to document efficacy for fluoxetine in a placebo-controlled trial in which patients received 6 weeks of the active agent in doses of up to 80 mg/day. Anecdotal evidence also suggests that the effectiveness of SRI\'s in TTM may wane with time \[[@B71]\].
Although there is evidence for the usefulness of behaviour therapy in both OCD \[[@B72]\] and TTM \[[@B73]\], the focus of the treatment differs in the two disorders (exposure and response prevention in OCD versus habit reversal in TTM). Keuthen et al \[[@B74]\] have suggested that \"state-of-the-art\" behavioral and pharmacological treatments offer substantial clinical benefit to patients with TTM, but in general clinics there may be relatively little experience with highly specialized interventions.
Several limitations of the current study should be acknowledged. First, interviewers were not blind to the patients\' psychiatric diagnosis, so potentially biasing clinician\'s assessments. Nevertheless, a structured diagnostic instrument ensures a reasonable degree of reliability. Second, instruments employed in the current study are intended for use in adults rather than younger subjects. However, in the case of children and adolescents, the SCID-I was supplemented with a clinical interview of parents or guardians, and self-report data was included only when it was clear that questionnaires had been completed meaningfully. Third, source of referral, and the duration of OCD/TTM, were not controlled for in the analysis, so potentially biasing the analyses. However, given the chronicity of both conditions, this is unlikely to have materially affected the findings. Fourth, males, as well as patients with comorbid OCD and TTM were excluded from the investigation; so that the results here may not be generalizable to all OCD or TTM subjects. Given evidence that the phenotype of OCD \[[@B75]\] and of TTM \[[@B10]\] varies with gender, additional work on male subjects should be undertaken in the future.
Conclusions
===========
In conclusion, our data suggest that despite some overlap, TTM differs from OCD in terms of demographics (gender distribution), associated clinical variables (e.g. comorbidity, cognitive schemas, temperament/character profiles and disability), precipitating factors (trauma history) and treatment response. It has been suggested that although TTM is not the same as OCD, it lies on a compulsive-impulsive spectrum of disorders \[[@B54]\]. However, it is notable that impulsivity may be an important component of OCD \[[@B76]\], and rather than viewing OCD and TTM on a single dimension, compulsivity and impulsivity should arguably therefore be seen as lying on orthogonal dimensions. Although TTM patients had more novelty seeking, OCD patients were more likely to have intermittent explosive disorder; such data support a view that TTM should not be classified as an impulse control disorder. Indeed, TTM may have more in common with conditions characterized by stereotypical self-injurious symptoms, such as skin-picking \[[@B77]\]. Differences between OCD and TTM may reflect contrasts in underlying psychobiology, and may necessitate contrasting treatment approaches.
Competing interests
===================
The author(s) declare that they have no competing interests.
Authors\' contributions
=======================
CL:
• participated in the design of the study,
• was responsible for patient recruitment,
• did most of the clinical assessments,
• performed the statistical analyses,
• helped to obtain funds for the research, and
• was responsible for the final writing up of data.
SS:
• participated in the design of the study,
• assisted with recruitment of patients, and
• supervised writing and statistical analyses.
PLduT:
• participated in the design of the study,
• assisted with recruitment of patients, and
• did some of the clinical assessments.
DGN:
• was the primary statistical consultant for analyses.
DJHN:
• participated in the design and coordination of the study,
• was responsible for patient recruitment, and
• assisted with clinical assessments.
RS:
• assisted with literature review of cognitive schema data, and
• asisted statistical analysis of schema data.
DJS:
• conceived of the study
• supervised coordination, statistical analysis, and writing,
• did the final revision of paper before submission, and
• assisted with obtaining of funds.
All authors read and approved the final manuscript.
::: {#T6 .table-wrap}
Table 6
::: {.caption}
######
TCI-temperament cut-off scores: A normal community sample\*
:::
**TCI -- TEMPERAMENT TRAITS**
------------------------------- -------- ------ ----- -------- ------
**NS** **HA**
LOW MEDIUM HIGH LOW MEDIUM HIGH
16 19.5 22 8 12.5 16
\*from Cloninger et al, 1994 (reference nr. 31)
:::
Pre-publication history
=======================
The pre-publication history for this paper can be accessed here:
<http://www.biomedcentral.com/1471-244X/5/2/prepub>
Acknowledgements
================
This work is supported by the Medical Research Council of South Africa, the National Research Foundation, and by a grant from the Obsessive-Compulsive Foundation. The help of the Obsessive-Compulsive Association of South Africa is gratefully acknowledged.
|
PubMed Central
|
2024-06-05T03:55:52.073050
|
2005-1-13
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546013/",
"journal": "BMC Psychiatry. 2005 Jan 13; 5:2",
"authors": [
{
"first": "Christine",
"last": "Lochner"
},
{
"first": "Soraya",
"last": "Seedat"
},
{
"first": "Pieter L",
"last": "du Toit"
},
{
"first": "Daniel G",
"last": "Nel"
},
{
"first": "Dana JH",
"last": "Niehaus"
},
{
"first": "Robin",
"last": "Sandler"
},
{
"first": "Dan J",
"last": "Stein"
}
]
}
|
PMC546037
|
Introduction {#s1}
============
The circadian clock controls biological processes such as behavior, gene expression, and physiology in diverse organisms, ensuring that these processes to take place at appropriate times of the day. This is crucial for many organisms, such as plants, which must synchronize photosynthesis with day-night cycles. Animals also synchronize to environmental cycles because of more subtle but nevertheless important needs such as predator avoidance, food availability, and optimal temperatures for various processes. Circadian clocks in all species share the following properties: They persist even in constant environmental conditions with periods near 24 h; they can be reset by environmental stimuli such as light and temperature; and their periods are relatively constant at different temperatures.
Recently, remarkable progress has been made in elucidation of molecular mechanisms of circadian clocks in diverse organisms. The common feature of clocks is cycling gene expression due to intracellular transcriptional feedback loops \[[@pbio-0030034-b01],[@pbio-0030034-b02],[@pbio-0030034-b03]\]. Genetic analysis in higher metazoan species, for example the fruit fly *Drosophila,* has been extremely valuable in identifying important players of the clockworks and their roles \[[@pbio-0030034-b04],[@pbio-0030034-b05],[@pbio-0030034-b06]\]. Identification of clock genes in *Drosophila* lead to molecular dissection of clock mechanisms in mammals mainly by testing whether homologs of *Drosophila* clock genes are involved in mammalian clocks. While this approach is informative, it harbors the risk of missing important factors that would have been found by forward genetic searches without preconceptions. Furthermore, gaps in our understanding of the clock mechanisms include factors responsible for the expression of positive transcription factors *Clock* and *Bmal* and mechanisms for clock protein turnover. In this regard, zebrafish is in a unique position as a vertebrate species in which large-scale forward genetic screens are convenient \[[@pbio-0030034-b07],[@pbio-0030034-b08]\]. Furthermore, zebrafish circadian clocks have been shown to possess unique properties such as light entrainability of molecular rhythms in cultured organs and cells \[[@pbio-0030034-b09],[@pbio-0030034-b10],[@pbio-0030034-b11]\].
A behavioral screening for circadian mutations has been successfully carried out in zebrafish \[[@pbio-0030034-b12]\]. However, due to the limited capacity of this method, it is not suited to high-throughput screening. A method relying on bioluminescence rhythms mediated by luciferase reporting has been successfully used to screen for mutants affecting circadian gene expression in plants, cyanobacteria, and flies \[[@pbio-0030034-b13],[@pbio-0030034-b14],[@pbio-0030034-b15]\]. In vertebrates, however, luciferase reporting has been used mainly for recording circadian gene expression in cultured tissues and cells, because of technical difficulties in these species \[[@pbio-0030034-b16],[@pbio-0030034-b17],[@pbio-0030034-b18],[@pbio-0030034-b19],[@pbio-0030034-b20]\]. Bioluminescence rhythms mediated by a *zper4-luc* promoter fusion construct have been studied successfully in the zebrafish PAC-2 cell line \[[@pbio-0030034-b21]\]. While this approach was useful for promoter dissection, generation of transgenic animals is necessary for mutagenesis screening.
In this study, transgenic zebrafish were made in which cycling expression of the firefly *luc* gene is driven by the promoter of *per3* \[[@pbio-0030034-b22]\]. This promoter was chosen because *per3* mRNA has been shown to oscillate rhythmically in embryos as well as in a cell line \[[@pbio-0030034-b11],[@pbio-0030034-b22]\]. For mutagenesis screening, it is most convenient and economical to test the youngest possible animals and avoid raising the animals that give negative results. In this regard, *per3-luc* was considered ideal for mutagenesis screening, because an in situ hybridization study showed that *per3* mRNA cycles starting on day 1 postfertilization with or without any entraining signals \[[@pbio-0030034-b22]\]. It was suggested that maternal *per3* mRNA present in the oocyte can set the phase of *per3* mRNA rhythms in early embryos. This result, however, is not consistent with other studies involving development of circadian rhythms: Rhythms of melatonin production require a light-dark (LD) transition later than 20 h postfertilization \[[@pbio-0030034-b23]\]; circadian swimming rhythms in larval fish develop during the first 4 d of development and require entraining signals late in embryonic development \[[@pbio-0030034-b24]\]; and rhythms of cell proliferation in larval fish develop only after exposure to several LD cycles \[[@pbio-0030034-b25]\].
In order to determine the earliest possible developmental time when rhythmic *luc* expression can be monitored in the *per3-luc* transgenic fish, embryos from the transgenic lines were monitored for bioluminescence from day 1 of development ([Protocol S1](#sd001){ref-type="supplementary-material"}). To our surprise, very low and non-oscillating levels of bioluminescence were detected during the first 4--5 d into development. Furthermore, consistent with the rhythms of melatonin production, locomotor activity, and cell division, rhythmicity of the *per3* gene expression gradually developed during the first several days postfertilization, and was observed only if fish were exposed to LD cycles during the hatching period or later. It was also found that ambient temperature affects *per3-luc*-mediated bioluminescence in a complex way. This study defines conditions under which the *per3-luc* transgenic fish can be used for mutagenesis screening and other types of studies.
Results {#s2}
=======
Generation of *per3-luc* Transgenic Fish {#s2a}
----------------------------------------
To develop a system in which circadian gene expression in zebrafish can be monitored in vivo, transgenic fish were generated in which the expression of the firefly *luc* gene is driven by the promoter of the *per3* gene \[[@pbio-0030034-b22]\]. The construct was made by modifying a bacterial artificial chromosome (BAC) originally screened for sequences in the first coding exon of *per3* ([Figure 1](#pbio-0030034-g001){ref-type="fig"})*.* By comparison of *per3* cDNA sequence to the genomic sequence from another BAC clone (CH211--138E4) from this region, it was found that the cDNA contains another exon 5′ to the first coding exon. By comparing BAC-end sequences from the construct to genomic sequences from CH211--138E4 as well as with the Ensembl Zebrafish whole genome shotgun assembly sequence version 4 (<http://www.ensembl.org/Danio_rerio/>, the construct was found to be approximately 72 kb long, and spans from 26 kb upstream of exon 1 to intron 19 of *per3,* and contains part of another gene 5′ to *per3* ([Figure 1](#pbio-0030034-g001){ref-type="fig"}). One canonical and two noncanonical E-boxes were found within 1 kb of the *per3* promoter (unpublished data). In the modified BAC, the coding portion of the first coding exon was replaced by the *luc* and *kanamycin resistance (Km^r^)* genes ([Figure 1](#pbio-0030034-g001){ref-type="fig"}). This rather long construct was made because it has been shown in zebrafish that a reporter gene is more consistently expressed in a context of a longer BAC construct than in a conventional short construct with just a few kilobases of promoter sequences \[[@pbio-0030034-b26]\].
::: {#pbio-0030034-g001 .fig}
Figure 1
::: {.caption}
###### Schematic Map of the *per3-luc* Construct
The top graphic shows the exon-intron structures of *per3* and the flanking gene. The BAC clone 8M06 screened for the first coding exon of *per3* is approximately 72 kb long, and extends from about 26 kb upstream of exon 1 to intron 19 of *per3*. The bottom graphic shows the magnified view of the first coding exon in the BAC 8M06 and the modified BAC construct. The white and black boxes represent noncoding and coding sequences, respectively, of the first coding exon of *per3*. The coding sequence of this exon was replaced with an approximately 3-kb fragment containing *luc* and *Km^r^*. Arrows under the construct represent primers used for the screening of transgenic lines.
:::
:::
After screening 147 injected founders by PCR, five independent transgenic lines were found. Each positive founder was bred to a wild-type fish, and their progeny were individually tested for bioluminescence as 5--7-d-old larval fish. Larval fish with bioluminescence above background (more than 100 counts per second \[cps\]) were raised as transgenic F1 fish. Three of the five lines emitted bioluminescence above background. The level of bioluminescence varied depending on the line. The strongest-glowing line (\#23) was used for this study unless otherwise stated. All the animals used in this study were the progeny of crosses between a transgenic line and the \*AB wild-type strain. Therefore, these animals carried the transgene in hemizygous condition.
Light Signals Are Necessary for *per3-luc* Rhythms {#s2b}
--------------------------------------------------
One of the intended usages of the *per3-luc* transgenic zebrafish is screening for mutations that affect bioluminescence rhythms. Since it is most convenient to screen the youngest possible animals, bioluminescence from transgenic embryos was monitored first. It was also expected that *per3-luc*-mediated bioluminescence in embryos should cycle from day 1 of development even in constant conditions, because *per3* mRNA expression detected by in situ hybridization has been demonstrated to oscillate from day 1 postfertilization in constant conditions \[[@pbio-0030034-b22]\]. Therefore, embryos carrying the transgene in hemizygous condition were collected and their bioluminescence monitored for 10 d starting from day 1 postfertilization. Surprisingly, when embryos were exposed to only one 14 h light: 10 h dark (14:10 LD, lights on at 8 [A.M.]{.smallcaps}; lights off at 10 [P.M.]{.smallcaps} CST) cycle on day 1, the majority of the animals showed no circadian rhythmicity of bioluminescence ([Figure 2](#pbio-0030034-g002){ref-type="fig"}A and [2](#pbio-0030034-g002){ref-type="fig"}B; [Table 1](#pbio-0030034-t001){ref-type="table"}). Nevertheless, a characteristic developmental profile of *luc* expression was observed. Bioluminescence mediated by *per3-luc* stays rather low until day 4, when there is a small peak of bioluminescence, followed by a small dip on day 5 and a rapid increase that reaches the second peak on days 7--9. The slow decline of luminescence after that point may be due to substrate deprivation common in luciferase reporting \[[@pbio-0030034-b27]\]. Importantly, this developmental profile was also observed in two other lines of *per3-luc,* albeit with much lower overall luminescence counts (unpublished data).
::: {#pbio-0030034-g002 .fig}
Figure 2
::: {.caption}
###### Bioluminescence in Embryos That Experienced Different Numbers of LD Cycles during Development
Embryos hemizygous for *per3-luc* were collected and monitored for bioluminescence while exposed to different numbers of 14:10 LD cycles starting on day 1 postfertilization followed by DD. (A) One LD, (B) one LD, (C) two LD, (D) three LD, (E) four LD, (F) five LD, (G) six LD, and (H) six LD. In (I), embryos were exposed to two LDs, one on day 1 and the other on day 6 of development. Black and white bars on top of each plot represent the times when the lights were off and on, respectively. Since overall bioluminescence levels can vary among clutches and experiments, normalized bioluminescence was averaged and plotted in each graph. Number of animals that were averaged is given at top right corner of each plot. Error bars represent standard error of the mean (SEM) For the one-LD and six-LD groups, plots for two experiments are shown here. These experiments showed small differences in developmental profiles, possibly due to differences in room temperature (about 1 °C), therefore could not be pooled. For the two- to five-LD groups, data from two experiments were pooled. The small but abrupt increase of luminescence that occurred only during the light period of LD cycles is considered an artifact made visible by low bioluminescence counts during the first several days of development, because the same level of fluctuation was observed in empty wells under LD condition.
:::
:::
::: {#pbio-0030034-t001 .table-wrap}
Table 1
::: {.caption}
###### Rhythmicity and Periods of Larval Zebrafish: Varying Number of LD Cycles
:::
Data are included for 5.5- to 10-d-old larval fish that had experienced varying numbers of LD during development. For each group, results of two experiments were pooled. Mean ± SEM of rhythm statistic \[[@pbio-0030034-b54]\] and period were calculated for rhythmic individuals only. A series of G-tests showed that percent rhythmic values were significantly different (*p* \< 0.001) among the seven groups
^a^ Further tests showed that these percent rhythmic values are significantly higher (*p* \< 0.05) than the others. Except for the one- to three-LD groups that were mostly arrhythmic, differences in rhythm statistic among groups were not quite significant (Wilcoxon/Kruskal-Wallis test, *p* = 0.027, α = 0.017)
:::
Since the light signal on day 1 was not enough to elicit detectable rhythmicity of bioluminescence, an increasing number of LD cycles were given to embryos while they were monitored for bioluminescence (see [Protocol S1](#sd001){ref-type="supplementary-material"}). LD cycles on days 2 and 3 did not increase rhythmicity on subsequent days, although small fluctuations of bioluminescence were discernible on days 7--10 in the averaged plots ([Figure 2](#pbio-0030034-g002){ref-type="fig"}C and [2](#pbio-0030034-g002){ref-type="fig"}D; [Table 1](#pbio-0030034-t001){ref-type="table"}). The number of animals expressing significant rhythmicity during the last 4.5 d of the record increased gradually when the number of LD cycles was increased from three to six ([Figure 2](#pbio-0030034-g002){ref-type="fig"}D--[2](#pbio-0030034-g002){ref-type="fig"}H; [Table 1](#pbio-0030034-t001){ref-type="table"}). The circadian fluctuation was superimposed on the developmental profile also seen in embryos entrained by fewer numbers of LD cycles ([Figure 2](#pbio-0030034-g002){ref-type="fig"}).
To determine whether the number of LD cycles or the developmental stage at which the last LD transition occurred is more important for robust rhythmicity of luciferase reporting, embryos were monitored for bioluminescence while experiencing two LD cycles, one on day 1 and the other on day 6 ([Protocol S1](#sd001){ref-type="supplementary-material"}). Luminescence rhythms on the last 4.5 d of the record for this group of animals were as robust as those exposed to six LDs ([Figure 2](#pbio-0030034-g002){ref-type="fig"}G--[2](#pbio-0030034-g002){ref-type="fig"}I; [Table 1](#pbio-0030034-t001){ref-type="table"}). Thus, the developmental stage at which the last LD transition occurred, rather than the number of LD cycles, was important for light entrainment of *per3-luc* rhythms. There was no systematic effect of the time of the last lights-off on free-running periods ([Tables 1](#pbio-0030034-t001){ref-type="table"} and [2](#pbio-0030034-t002){ref-type="table"}).
::: {#pbio-0030034-t002 .table-wrap}
Table 2
::: {.caption}
###### Rhythmicity and Periods of Larval Zebrafish: Varying LD and Temperature
:::
Data are included for 5.5- to 10-d-old larval fish that had experienced varying numbers of LD at different temperatures during development. Results from one of two experiments with similar results are shown here. Mean ± SEM of rhythm statistic and period were calculated only for rhythmic individuals. A series of G-tests showed that percent rhythmic values among the six groups were significantly different (*p* \< 0.001)
^a^ Further tests showed that these percent rhythmic values were significantly higher (*p* \< 0.01) than the others. For the six-LD groups, the rhythm statistic was not significantly different between the two temperature groups (*t*-test, *p* = 0.61)
:::
Luciferase Reporting Reflects *per3* Expression {#s2c}
-----------------------------------------------
The lack of bioluminescence rhythms during the first few days of development was rather unexpected because of the previously reported *per3* mRNA rhythms \[[@pbio-0030034-b22]\]. Therefore, mRNA cycling of *per3* and *luc* was compared by real-time quantitative PCR (qPCR). As in the experiment shown in [Figure 2](#pbio-0030034-g002){ref-type="fig"}G and [2](#pbio-0030034-g002){ref-type="fig"}H, embryos were exposed to six LD cycles and transferred to constant darkness (DD). Embryos were collected and their mRNA was extracted on days 3 and 8. Both *per3* and *luc* mRNA levels were much lower on day 3 compared to day 8 ([Figure 3](#pbio-0030034-g003){ref-type="fig"}). What appears to be approximately 2-fold oscillations of *per3* and *luc* mRNA on day 3 (see the insets on [Figure 3](#pbio-0030034-g003){ref-type="fig"}A and [3](#pbio-0030034-g003){ref-type="fig"}C) were not statistically significant (*p* = 0.47 for *per3*, and *p* = 0.08 for *luc* by the Wilcoxon/Kruskal-Wallis test). In contrast, approximately 5-fold fluctuations of the transcripts were observed on day 8 ([Figure 3](#pbio-0030034-g003){ref-type="fig"}B and [3](#pbio-0030034-g003){ref-type="fig"}D). This increase in overall expression levels and cycling amplitudes reflect the observed bioluminescence profile, although there were qualitative differences between bioluminescence and RNA as well as between *per3* and *luc* mRNA (see Discussion). Importantly, the peak phase of mRNA cycling on day 8 was 5--7 h advanced compared to the phase of bioluminescence cycling (compare [Figure 2](#pbio-0030034-g002){ref-type="fig"}G and [2](#pbio-0030034-g002){ref-type="fig"}H to [Figure 3](#pbio-0030034-g003){ref-type="fig"}B and [3](#pbio-0030034-g003){ref-type="fig"}D).
::: {#pbio-0030034-g003 .fig}
Figure 3
::: {.caption}
###### Temporal Expression of *luc* mRNA Is Similar to That of *per3* during Development
Embryos hemizygous for *per3-luc* were collected from naturally breeding parents and kept in 14:10 LD cycles (lights on at 8 [A.M.]{.smallcaps} CST) at 22 °C for 6 d. Larval fish were shifted to DD at the end of the light phase on day 6. Total RNA was extracted from 2- to 3-d-old embryos and 7- to 8-d-old larval fish, and was subjected to real-time PCR for *per3* and *luc* mRNA levels.
\(A) Expression of *per3* mRNA per embryo on day 3 was determined every 6 h starting at 10 [A.M.]{.smallcaps} (2 h after lights-on).
\(B) The *per3* mRNA per animal on day 8 was measured every 4 h starting at 10 [A.M.]{.smallcaps}
\(C) Levels of *luc* mRNA on day 3 were determined as in (A). White and black bars at the bottom represent light and dark phases, respectively, for (A) and (C).
\(D) Cycling of *luc* mRNA on day 8 was determined as in (B). The gray and black bars represent the time when the light would have been on and off, respectively, had the LD cycles continued.
For each of *per3* and *luc*, mRNA levels were normalized to the peak level on day 8 (10 [A.M.]{.smallcaps} time point). The y-axis scales were set at 120% maximum for all plots to allow direct comparison of mRNA extracted on days 3 and 8. The x-axis scales are given in both hours and days postfertilization to facilitate the comparison with [Figure 2](#pbio-0030034-g002){ref-type="fig"}. In order to show more detailed temporal profiles of mRNAs on day 3, plots with smaller y-axis scales were shown in the insets at top right corners of (A) and (C). Each plot is the average of three identical experiments, and error bars represent SEM. An identical experiment was also done at 24 °C with essentially the same results (unpublished data).
:::
:::
Effects of Ambient Temperature on *per3-luc*-Mediated Bioluminescence {#s2d}
---------------------------------------------------------------------
The experiments presented above were done at 21--24 °C simply because fish survived better at these rather low temperatures (see [Materials and Methods](#s4){ref-type="sec"}). However, the previously documented circadian studies on zebrafish, including those involving development of rhythmicity, have been done mainly at the higher temperatures of 25--28.5 °C \[[@pbio-0030034-b22],[@pbio-0030034-b23],[@pbio-0030034-b24],[@pbio-0030034-b25]\]. Since higher temperatures accelerate development in general, it was conceivable that development of bioluminescence rhythms may be faster at higher temperatures. However, it was not possible to do the same experiment at higher temperatures, because fish do not survive well in microtiter wells at temperatures higher than 25 °C. Therefore, embryos were raised in petri dishes at two different temperatures, 22 °C and 28.5 °C, while exposed to two, four, or six LD cycles. Subsequently, they were placed in microtiter wells and bioluminescence was recorded in DD at 21--24 °C ([Protocol S1](#sd001){ref-type="supplementary-material"}).
The majority of embryos that were raised at 22 °C were arrhythmic after they were entrained by two or four LD cycles, but they were highly rhythmic after six LD cycles ([Figure 4](#pbio-0030034-g004){ref-type="fig"}A, [4](#pbio-0030034-g004){ref-type="fig"}C, and [4](#pbio-0030034-g004){ref-type="fig"}E; [Table 2](#pbio-0030034-t002){ref-type="table"}). This is largely consistent with the trends observed in [Figure 2](#pbio-0030034-g002){ref-type="fig"} and [Table 1](#pbio-0030034-t001){ref-type="table"}, although more fish were rhythmic in this experiment for the two-LD and six-LD groups, and less in the four-LD group. The increased percentage of rhythmicity in this experiment for the six-LD group may be due to the fact that embryos raised in petri dishes are generally healthier than those raised in 96-well plates. Embryos raised at 28.5 °C showed significantly higher rhythmicity for the four-LD group than the same group raised at 22 °C (*p* \< 0.05; [Figure 4](#pbio-0030034-g004){ref-type="fig"}C and [4](#pbio-0030034-g004){ref-type="fig"}D; [Table 2](#pbio-0030034-t002){ref-type="table"}). This result shows that embryos raised at higher temperatures can be entrained earlier than those raised at lower temperatures.
::: {#pbio-0030034-g004 .fig}
Figure 4
::: {.caption}
###### Effects of Temperature on Development of *per3-luc* Expression
Transgenic embryos were entrained by two, four, or six LD cycles at either 22 °C or 28.5 °C, and monitored for bioluminescence in DD at 21--24 °C. (A) Two LDs at 22 °C, (B) two LDs at 28.5 °C, (C) four LDs at 22 °C, (D) four LDs at 28.5 °C, (E) six LDs at 22 °C, and (F) six LDs at 28.5 °C. The insets in (E) and (F) show the last 3 d of the record with magnified y-axis scales. Black and white bars on top of each plot represent the times when the lights were off and on, respectively. Actual amount of bioluminescence in cps is averaged and plotted in each graph. Number of animals that were averaged is given at top right corner of each plot. Error bars represent SEM. Results of one of two identical experiments with similar results are shown here.
:::
:::
In contrast to the effects on rhythmicity, temperature during the first several days of development seemed to have no systematic effects on periods: The two methods used for period estimation showed opposite effects of developmental temperature on periods (see the six-LD groups in [Table 2](#pbio-0030034-t002){ref-type="table"}). It should be pointed out that this may not mean that *per3-luc* rhythm is temperature compensated, because all the fish were monitored in the same temperature condition.
Besides development of rhythmicity, the developmental profile of *luc* expression and baseline level of bioluminescence were also affected by prior ambient temperature. For the two-LD groups, the first and second developmental peaks came earlier in the 28.5 °C than in the 22 °C group ([Figure 4](#pbio-0030034-g004){ref-type="fig"}A and [4](#pbio-0030034-g004){ref-type="fig"}B). Higher temperatures also caused elevated levels of baseline bioluminescence, especially in the six-LD group ([Figure 4](#pbio-0030034-g004){ref-type="fig"}E and [4](#pbio-0030034-g004){ref-type="fig"}F). This higher bioluminescence cannot be caused by high specific activity of equivalent luciferase enzyme, because all of the fish were monitored at the same temperature, and the difference in bioluminescence level persisted through over 4 d of monitoring. Therefore, this difference in luminescence most likely reflects a difference in the level of *luc* expression. Part of this difference between the two temperature groups may be explained by the fact that animals raised at higher temperatures are more mature and therefore express more luciferase than do those raised at lower temperatures. However, 10-d-old animals raised at 22 °C, which should have reached the plateau of luminescence (see [Figure 2](#pbio-0030034-g002){ref-type="fig"}), showed much lower bioluminescence than 10-d-old fish raised at 28.5 °C (compare [Figure 4](#pbio-0030034-g004){ref-type="fig"}E and [4](#pbio-0030034-g004){ref-type="fig"}F). Therefore, maturity of animals cannot explain the difference either. It seems that animals raised at higher temperatures simply express more luciferase than do those at lower temperatures. This was confirmed by a real-time PCR experiment ([Figure 5](#pbio-0030034-g005){ref-type="fig"}). Cycling amplitudes and peak levels of *per3* and *luc* mRNA on day 6 were much higher in fish raised at 28.5 °C than those at 22 °C. The difference was especially large for *luc,* for which an approximate 40-fold difference between the two temperature groups was found at Zeitgeber Time 0 (2 h after lights-on; [Figure 5](#pbio-0030034-g005){ref-type="fig"}C).
::: {#pbio-0030034-g005 .fig}
Figure 5
::: {.caption}
###### Levels of *per3* and *luc* mRNA Are Elevated by High Temperatures during Development
\(A) A schematic diagram showing how embryos were entrained and collected for RNA extraction. Embryos were entrained in 14:10 LD cycles (lights on at 8 [A.M.]{.smallcaps}; lights off at 10 [P.M.]{.smallcaps} CST) at two different temperatures, 22 °C and 28.5 °C. On day 6, half of the animals were sacrificed for RNA at 10 [A.M.]{.smallcaps} (2 h after lights-on) and at 10 [P.M.]{.smallcaps} (at lights-off). The rest of the animals were transferred to DD at 22 °C at 10 [P.M.]{.smallcaps} on that day, and sacrificed for RNA on day 10 at 10 [A.M.]{.smallcaps} and 10 [P.M.]{.smallcaps} The white and black bars represent day and night, respectively, and the gray bars the time at which lights would have been on had the LD cycles continued. The arrowheads indicate the time at which the animals were sacrificed for RNA extraction.
\(B) Relative mRNA level per animal for *per3* on days 6 and 10 of the experiment quantified by real-time qPCR. The levels were normalized to the value of the 10 [A.M.]{.smallcaps} time point on day 6 at 28.5 °C.
\(C) Relative RNA level per animal for *luc* measured from the same samples used in (B). For both (B) and (C), averages of three experiments are shown. Error bars represent SEM.
:::
:::
Bioluminescence in the high-temperature group gradually decreased over several days after they were transferred to lower temperatures, but did not fully return to the level of the low-temperature group (see [Figure 4](#pbio-0030034-g004){ref-type="fig"}E and [4](#pbio-0030034-g004){ref-type="fig"}F). This is consistent with *per3* and *luc* mRNA levels determined by real-time PCR ([Figure 5](#pbio-0030034-g005){ref-type="fig"}). Taken together, high temperatures elevate the level and cycling amplitudes of *per3* and *luc* mRNA, at least in larval fish, and this level and amplitude can gradually decrease after the animals are shifted down to lower temperatures.
Optimal Condition for Recording Larval *per3-luc* Rhythms {#s2e}
---------------------------------------------------------
Fold amplitudes of bioluminescence rhythms were higher when embryos were entrained by LD cycles for 6 d at 22 °C rather than at 28.5 °C (see [Figure 4](#pbio-0030034-g004){ref-type="fig"}E and [4](#pbio-0030034-g004){ref-type="fig"}F). Furthermore, survival of the animals was better if they were raised at 22 °C than at 28.5 °C (96.9% and 0% survival, respectively, on day 7). Therefore, embryos were entrained by six LDs at 22 °C in a petri dish, and tested for bioluminescence rhythms in DD ([Protocol S1](#sd001){ref-type="supplementary-material"}). In this experiment, embryos survived better than they did when they were placed in 96-well plates from day 1 onward (89.6% survival on day 12 in this experiment, compared to 71.5% on day 10 for the experiments presented in [Figure 2](#pbio-0030034-g002){ref-type="fig"}). Furthermore, 88.1% (*n* = 42) of the fish were rhythmic with a 25.2 ± 0.7 h (mean ± standard deviation) period under this condition, and their rhythms persisted for 6 d, albeit with some damping ([Figure 6](#pbio-0030034-g006){ref-type="fig"}A and [6](#pbio-0030034-g006){ref-type="fig"}B). In addition, *per3-luc* rhythms were tested in LD ([Protocol S1](#sd001){ref-type="supplementary-material"}). Amplitudes of bioluminescence rhythms in LD were higher than in DD, although they also damped slightly, possibly due to substrate deprivation ([Figure 6](#pbio-0030034-g006){ref-type="fig"}C and [6](#pbio-0030034-g006){ref-type="fig"}D). A slightly higher percentage of fish was rhythmic in LD (95.0%, *n* = 179) compared to DD, although the difference was not significant (*p* \> 0.1, G-test). The waveform of the rhythm in LD was different from that in DD: The ascending part of the wave that happens during the day was steeper in LD than in DD ([Figure 6](#pbio-0030034-g006){ref-type="fig"}), suggesting that light may induce transcription of *per3*.
::: {#pbio-0030034-g006 .fig}
Figure 6
::: {.caption}
###### Bioluminescence Rhythms Mediated by *per3*-*luc* in DD and LD Measured for Six Days
\(A) Average plot of bioluminescence rhythms in DD. Animals were entrained in 14:10 LD cycles for 6 d at 22 °C and tested for approximately 6.5 d in DD.
\(B) Representative plot of bioluminescence rhythm in DD for an individual.
\(C) Average plot of bioluminescence rhythms in 14:10 LD cycles. Animals were entrained in 6 LD cycles at 22 °C prior to the monitoring.
\(D) Individual plot of bioluminescence rhythm in LD cycles. For each of DD and LD experiments, two experiments have been performed with essentially the same results. Only one of two experiments is shown for each of DD and LD. The first 12 h of data were deleted from each plot. Black and white bars on top of each plot represent the time when lights were off and on, respectively. Numbers of animals that were averaged is given at top right corner of (A) and (C). Error bars represent SEM in (A) and (C).
:::
:::
Discussion {#s3}
==========
The *per3-luc* transgenic zebrafish system presented here is unique, because it is the only vertebrate system in which circadian gene expression in the whole animal can be studied in a high-throughput manner. This property of *per3-luc* in combination with zebrafish genetics makes the transgenic fish suitable for mutagenesis screening for circadian mutants. Embryos were tested initially, because it is more efficient to screen embryos than older animals, and movements of older animals can cause noise in bioluminescence signals \[[@pbio-0030034-b28]\]. However, the current study clearly demonstrates that monitoring embryos is not an option for any circadian studies using these lines. Of the conditions tested, raising fish to 6 d of age at 22 °C under LD cycles produced the most robust free-running rhythms. Rhythms measured under LD cycles were even stronger, and this condition has been used successfully in previous screens \[[@pbio-0030034-b15]\]. In addition to mutagenesis screening, it will facilitate other studies such as circadian organization of central and peripheral oscillators, entrainment pathways by various environmental cues, and physiological effects on circadian rhythms as have been studied in other organisms \[[@pbio-0030034-b17],[@pbio-0030034-b18],[@pbio-0030034-b29],[@pbio-0030034-b30],[@pbio-0030034-b31],[@pbio-0030034-b32],[@pbio-0030034-b33]\].
Development of *per3* RNA Cycling in Larval Zebrafish {#s3a}
-----------------------------------------------------
Larval fish that experienced an LD-DD transition later during development were more rhythmic than those shifted to DD earlier. This may mean that there is a critical developmental period after which *per3-luc* rhythms can be entrained. Alternatively, *per3-luc* rhythms might damp so fast that rhythms entrained a few days earlier could not be detected. We think the latter possibility unlikely for the following reasons: *per3-luc*-mediated bioluminescence rhythms persisted reasonably well for at least 6 d in DD if entrained properly ([Figure 6](#pbio-0030034-g006){ref-type="fig"}A); at 28.5 °C, the rhythmicity of larval fish entrained for 4 d was comparable to that of fish entrained for 6 d, suggesting that bioluminescence rhythms do not damp in 2 d ([Table 2](#pbio-0030034-t002){ref-type="table"}); and *luc* mRNA cycling was almost undetectable during the first few days of development (see [Figure 3](#pbio-0030034-g003){ref-type="fig"}C), and this low-amplitude fluctuation cannot give rise to high-amplitude oscillations unless there is a separate rhythm-amplifying mechanism operating during development, such as synchronization of cellular oscillators.
The results presented here, namely the gradual development of rhythmicity and responsiveness to entraining stimuli during the first several days of development and requirement of light signal, is consistent with previous observations on rhythms of melatonin production, locomotor activity, and the cell cycle in larval zebrafish \[[@pbio-0030034-b23],[@pbio-0030034-b24],[@pbio-0030034-b25]\]. It is also consistent with several studies in other vertebrate species involving gene expression \[[@pbio-0030034-b34],[@pbio-0030034-b35],[@pbio-0030034-b36]\] and physiological rhythms \[[@pbio-0030034-b37]\]. In both insects and mammals, free-running behavioral rhythms can develop normally in the absence of entraining signals, but their phases are not synchronized \[[@pbio-0030034-b38],[@pbio-0030034-b39],[@pbio-0030034-b40]\]. DeLaunay et al. \[[@pbio-0030034-b22]\] reported that *per3* RNA detected by in situ hybridization cycled synchronously from day 1 of development \[[@pbio-0030034-b22]\]. In our hands, however, *per3* and *per3*-driven *luc* RNA in 3-d-old larval fish detected by real-time PCR were not significantly rhythmic. Low-amplitude oscillations of both mRNAs may exist at this early developmental stage, because a similar trend (high in the morning) was observed in all three independent experiments done. However, this does not mean that the low-amplitude *per3* RNA oscillations that occur earlier during development can amplify into high-amplitude ones without any entrainment by LD cycles.
The increase of cycling amplitude over the course of several days of development in LD may happen within each cell that expresses *per3*. Alternatively, cycling amplitude may increase because various cellular oscillations present become synchronized. However, it was not only the cycling amplitude that increased with age, but the overall levels of *per3* mRNA also. Therefore, more cells may start expressing *per3* with high-amplitude oscillations later during development. This rather simple scenario is in fact what happens in developing *Drosophila;* as soon as the central pacemaker Lateral Neurons start expressing the PERIOD protein, the molecular rhythm is entrainable, and so is the eventual behavioral rhythmicity \[[@pbio-0030034-b40],[@pbio-0030034-b41]\]. In any case, increasing amplitudes within cells, synchronization of oscillators, and more cells with high-amplitude oscillations are not mutually exclusive.
It is worth mentioning that rhythms of melatonin release starts as early as day 2 postfertilization, and this roughly corresponds to the time when the light-sensitive pineal gland is formed \[[@pbio-0030034-b23]\]. Another photosensitive organ, the retina, becomes photoresponsive as early as day 3 postfertilization \[[@pbio-0030034-b42]\]. Therefore, these organs become photosensitive, and/or develop rhythmicity prior to robust oscillations of *per3* RNA. Again, clocks in some tissues may develop earlier than in other tissues. It is also possible that *per3* may not be expressed in all the clock cells.
The biological significance of the developmental profile of *per3*-driven *luc* expression is not known. Most animals seemed to hatch between the first minor developmental peak and the subsequent trough (unpublished data). However, hatching itself is unlikely to induce *per3* expression in early embryos, because dechorionated *per3-luc* embryos showed developmental profiles of bioluminescence similar to those of nondechorionated siblings, albeit with an accelerated second rise of bioluminescence (unpublished data). Larval fish after the hatching period are supposed to have completed most of their morphogenesis and start swimming actively \[[@pbio-0030034-b43]\]. It may be that *per3* is important for rhythmic processes specific to hatched animals, such as behavioral rhythms \[[@pbio-0030034-b24]\].
Consistency between *per3* Expression and Bioluminescence {#s3b}
---------------------------------------------------------
The developmental and circadian profiles of *per3*-driven bioluminescence largely reflected endogenous *per3* expression. However, the amplitude of bioluminescence cycling was greatly reduced compared to that of *per3* or *luc* RNA. In general, proteins synthesized from cycling mRNAs show amplitude reduction and phase delay due to protein stability \[[@pbio-0030034-b44]\]. Consistently, luciferase-reporting studies in other organisms also showed dampening of cycling. For instance, in the *per*-*luc* transformants of *Drosophila,* approximately 6-fold cycling of *luc* RNA was reduced to 3- to 4-fold bioluminescence rhythms \[[@pbio-0030034-b27]\]. Similar reduction in cycling amplitude was observed for suprachiasmatic nuclei from the *per1-luc* mouse \[[@pbio-0030034-b18]\]. However, the reduction of amplitude was even more dramatic in *per3-luc* larval fish. This may mean that the luciferase protein is somehow more stable in larval fish than in flies or mice.
It should be noted that the luciferase protein itself is quite stable, but the enzymatic activity of this protein is unstable \[[@pbio-0030034-b28],[@pbio-0030034-b45]\]. Therefore, the apparent stability of the luciferase protein in larval fish may be in fact stability of luciferase activity.
Besides the difference between bioluminescence and mRNA, there were differences between *per3* and *luc* mRNA, measured by real-time qPCR. The difference between days 3 and 8 of development was much larger for *luc* than *per3*. Furthermore, the difference between the 22 °C and 28.5 °C groups was larger for *luc* than for *per3*. These differences may be due to positional effect of the insert in the particular transgenic line used in this study. However, differences in bioluminescence levels between the first few days and older fish were found in two other independent *per3-luc* lines (unpublished data). Therefore, it is more likely that these differences between the two mRNA species reflect a property of the transgene itself. Although the BAC transgene used in this study has approximately 26 kb of upstream sequences, there may be critical sequences missing from this transgene, such as the first coding exon. Alternatively, posttranscriptional modification could be responsible for the difference. Posttranscriptional control of clock gene mRNA expression has been documented elsewhere \[[@pbio-0030034-b18],[@pbio-0030034-b28],[@pbio-0030034-b46],[@pbio-0030034-b47]\].
Effects of Ambient Temperature on *per3* Expression {#s3c}
---------------------------------------------------
Increasing ambient temperature had three effects on development of *per3* expression: It increased peak levels and cycling amplitudes of mRNA, and accelerated the developmental profile and development of responsiveness to the entraining stimuli. The latter two effects presumably are due simply to the fact that development is accelerated by higher temperatures. These results are not consistent with the report that development of a cell-cycle rhythm in larval fish was not affected by ambient temperature \[[@pbio-0030034-b25]\]. The increase in peak level and cycling amplitudes of *per3* mRNA induced by increased temperature is mostly independent of developmental speed. The peak *per3* mRNA level in the high-temperature group on day 6 was higher than that in the low-temperature group on day 10, when bioluminescence levels should have reached the plateau (see [Figure 5](#pbio-0030034-g005){ref-type="fig"}). Ambient temperatures are known to affect levels of clock gene expression in *Drosophila* and *Neurospora* \[[@pbio-0030034-b47],[@pbio-0030034-b48],[@pbio-0030034-b49]\], and this could be the mechanism for clock resetting by temperature shifts \[[@pbio-0030034-b49]\]. Since our study was done on developing animals, it remains to be seen whether the increase in *per3* expression by elevated temperatures holds true in adults. Furthermore, it was not possible to test larval fish at 28.5 °C, and therefore it is not known whether the *per3* mRNA rhythm is temperature-compensated. Levels of *per3* and *luc* mRNA in fish raised at higher temperatures did not fully return to the levels in fish raised at lower temperatures, even after 4 d at lower temperatures. Therefore, there may be a mechanism for maintaining circadian periods despite change in *per3* expression levels.
Materials and Methods {#s4}
=====================
{#s4a}
### Animals {#s4a1}
Animals used in this study were derived from the University of Oregon \*AB strain. Adults were kept under 14:10 LD (lights on at 8 [A.M.]{.smallcaps}; lights off at 10 [P.M.]{.smallcaps} CST) cycle, and group-housed in plastic tanks in a Z-MOD holding system (Marine Biotech) with recirculating filtered water at about 28.5 °C. They were fed commercial flake food in the morning, baby brine shrimp at midday, and adult brine shrimp in the evening. Experimental protocols were approved by the Institutional Animal Care and Use Committee.
Embryos were collected from naturally breeding fish in the morning, by plastic mesh traps that prevented parents from eating their progeny \[[@pbio-0030034-b23]\]. For microinjection, one- to two-cell-stage embryos were required. Therefore, male and female breeders were separated by a divider when they were placed in a trap. By removing the divider, fish were allowed to breed just before microinjection was performed \[[@pbio-0030034-b50]\].
### Construction of *per3-luc* transgene {#s4a2}
First, a zebrafish BAC library was screened for BACs containing 5′ coding sequences of *per3* \[[@pbio-0030034-b22]\] using a PCR-based screening kit (Incyte Genomics, Wilmington, Delaware, United States). One of two such clones, 8M06, was used for the construction of the transgene. The primers used for the screening were: forward, 5′- CCAGTAAAACGTCGTCGTCA-3′; reverse, 5′- GTCTGGGCCTGGAGAAGAGT-3′. The *per3* sequence from the initiation codon to the end of the first coding exon was replaced by a gene cassette containing the *luc* and *Km^r^* genes by homologous recombination in E. coli (see [Figure 1](#pbio-0030034-g001){ref-type="fig"}) \[[@pbio-0030034-b51],[@pbio-0030034-b52]\].
The *luc*/*Km^r^* gene cassette was constructed as follows. First, the *Km^r^* gene was cloned into NotI and SacI sites of pBluescriptSK+ (Stratagene, La Jolla, California, United States); then a HindIII-BamHI fragment containing *luc* and the SV40 polyA signal from pGL3-Basic (Promega, Madison, Wisconsin, United States) was cloned upstream of *Km^r^* into HindIII and BamHI sites of pBluescriptSK+; finally, a NotI site between *Km^r^* and the polyA signal was destroyed by digesting the clone with NotI and BamHI, followed by treatment with the Klenow fragment, and ligation of these blunted ends together.
Using this gene cassette as a template, a PCR fragment flanked by approximately 50-bp homology arms was amplified. The primer sequences used for the PCR reaction were: forward, 5′- GGGTTGTGAATCAGATCTTCAGTAGAGGAGGACAGGAGATCTCACAGGGAATGGAAGACGCCAAAAACATAAAGAAAG-3′; reverse, 5′- GTGCAGATTAAGTCAAATTCCACATAAAAAAAGCCACATTTCAAGTGTAC CGTTAATAATTCAGAAGAACTCGTC-3′. The forward primer contains 25 bp of sequence from the 5′ end of the *luc* coding sequences flanked by a 53-bp overhang corresponding to sequences just upstream of the initiation codon of *per3*. The reverse primer consists of a 50-bp overhang that corresponds to the intron sequences just downstream of the first coding exon, and 25 bp of sequence from the 3′ end of *Km^r^* as the primer.
The PCR fragment was purified and electroporated into DH10B cells containing the BAC and the plasmid pBAD-αβγ \[[@pbio-0030034-b51]\] as a source of recombinase genes. Cells in which the BAC was successfully modified were selected by kanamycin. The *luc* sequences in the modified BAC clones were checked for PCR errors by sequencing. One PCR error that resulted in a Val 217 to Ala change in the luciferase protein sequence was found. However, this is a conservative change, and fish injected with this construct showed bioluminescence above background. Therefore, it was judged that luciferase encoded by this construct can still function.
### Generation of *per3-luc* transgenic lines {#s4a3}
The *per3-luc* transgene was purified, linearized by NotI digestion, and injected into one- to two-cell-stage zebrafish embryos according to \[[@pbio-0030034-b53]\] with minor modifications. Injected embryos were raised to adulthood and individually bred to a wild-type fish or pairwise bred to each other. The progeny were tested for the presence of the transgene by PCR. PCR primers used for the screening were: Per35F, 5′- GCACCAGTAAAACGTCGTCA-3′; Per33R, 5′- TCATTCTCACTGGCAGAGCA-3′; Luc5R2, 5′- GTTTTAGAATCCATGATAATA-3′; Kan3F2, 5′- CTTTTTGTCAGAAGACCGACC-3′. The approximate positions of these primers with respect to the construct are shown in [Figure 1](#pbio-0030034-g001){ref-type="fig"}. Transgenic fish were identified as those fish that gave an approximately 600-bp PCR product with the Per35F/Luc5R2 primer combination, and an approximately 800-bp product by the Kan3F2/Per33R combination, when their genomic DNA was used as templates. Nontransgenic fish gave no products with either of these primer combinations.
### In vivo measurement of bioluminescence rhythms {#s4a4}
Embryos or larval fish were placed individually in every other well of a white 96-well Optiplate (Perkin-Elmer, Wellesley, California, United States) with 200 μl of Holtfreter solution (7.0 g of NaCl, 0.4 g of sodium bicarbonate, 0.2 g of CaCl~2~, and 0.1 g of KCl \[pH 7.0\] in 2 l of ddH~2~O) aerated overnight and containing 0.5 mM D-luciferin potassium salt (Biosynth, Naperville, Illinois, United States) and 0.013% Amquel Instant Water Detoxifier (Kordon brand; Novalek, Hayward, California, United States). Once loaded with animals, four such plates were subjected to automatic monitoring of bioluminescence every 30 min by the Topcount multiplate scintillation counter (Perkin-Elmer) equipped with six detectors and plate stackers. The room temperature was set at 21--22 °C, and the machine at 24 °C. However, due to the heat created by the machine, temperature at the bottom of the stacker was 1--2 °C higher than the room temperature. In order to minimize high background counts under lighted conditions, each plate was dark-adapted for approximately 5 min before being counted for bioluminescence. Each well was counted for 4.8 s every 30 min. The plates were illuminated with two white fluorescent lamps, each facing the left or right side of the stacker. The approximate intensity of the light that reached the plates was 17--35 lux, depending on the position of the plates within the stacker.
### Experimental protocol {#s4a5}
The experimental protocol for each experiment involving in vivo bioluminescence measurement is described in [Protocol S1](#sd001){ref-type="supplementary-material"}.
### Bioluminescence data analyses {#s4a6}
Bioluminescence data from the Topcount were imported into Microsoft Excel 2000 by the Import and Analysis macro (kindly supplied by Steve Kay, Scripps Institute). In many of the experiments performed on the Topcount, some plates were placed in the machine several days earlier than others in order to monitor fish that experienced different numbers of LD cycles ([Protocol S1](#sd001){ref-type="supplementary-material"}). Therefore, at the end of the recording period that typically lasted for approximately 2 wk, many fish that were placed in the machine earlier had been dead for a few days. Since only the data up to 10-d old larval fish were analyzed for the experiments presented in [Figures 2](#pbio-0030034-g002){ref-type="fig"} and [4](#pbio-0030034-g004){ref-type="fig"}, simple observation of fish after the recording period could overestimate the number of fish that had died while the 10-d worth of data were collected. When a *per3-luc* fish dies in luciferin solution, it emits a burst of high bioluminescence counts (\> 2,000 cps). This burst of luminescence is typically followed by a low background level of luminescence (\< 50 cps). Furthermore, intermediate levels of spikes were also found in many plots just before the burst of high bioluminescence. Therefore, in order to eliminate data from dead fish, data that exceeded 5,000 cps, or those that went down below 50 cps, at any point of the analyzed portion of the data were first discarded. Then an averaged plot of the remaining data from each clutch of embryos was examined, and the highest count on days 8--9 was determined for experiments presented in [Figure 2](#pbio-0030034-g002){ref-type="fig"}. For the other experiments, highest counts from the entire averaged plots except for day 1 of the record were determined. Then data that exceeded twice that value were also discarded as those with medium-sized spikes. This second round of data elimination was done in this way because overall luminescence counts varied among different clutches, possibly due to varying sizes of eggs laid. It should be noted that this procedure also eliminated records from fish that were alive, but that showed one or more transient spikes of bioluminescence. Those data with transient spikes were eliminated anyway, because the spikes can severely affect the accuracy of the data analysis program.
Period and rhythmicity for each animal were determined by a macro \[[@pbio-0030034-b54]\] based on MatLab 6.5 (Mathworks, Natick, Massachusetts, United States). With this macro, periods were determined by the maximum entropy spectral analysis (MESA) and autocorrelation, and rhythmicity by autocorrelation. Some fish that were apparently arrhythmic by visual examination of the plot gave rhythm indexes with a confidence interval higher than 95% by autocorrelation due to spurious peaks and small confidence intervals. Autocorrelation plots of these fish, however, were almost always nonsinusoidal and/or did not have five clear peaks (one at the center and two on each side of the center peak). Therefore, each set of data was judged blindly by three people as to whether its autocorrelation plot was sinusoidal with five peaks. Fish were judged as rhythmic only if two to three people found their autocorrelation plots sinusoidal, and their rhythm statistic values exceeded 1.
### Real-time qPCR {#s4a7}
Total RNA was extracted from 9--42 embryos or larval fish raised in petri dishes using TRIzol reagent (Invitrogen, Carlsbad, California, United States). The number of animals used for each extraction was recorded. Once extracted, total nucleic acid concentration was determined by a spectrophotometer. In order to prevent genomic DNA contamination, RNA samples were treated with Turbo DNA-free (Ambion, Austin, Texas, United States), and the concentration determined again by a spectrophotometer. Total RNA (0.5--1 μg) was subjected to cDNA synthesis by Superscript II Reverse Transcriptase (Invitrogen) using Oligo (dT)~12--18~ (Invitrogen) as the primer in a 25--40 μl reaction volume. Real-time PCR was performed in a 25 μl reaction volume containing a probe, forward and reverse primers, and qPCR Mastermix according to the manufacturer\'s instruction (Eurogentec, Seraing, Belgium). Each reaction was quadrupled in order to minimize pipetting errors. The primers and TaqMan MGB probes for *per3* and *luc* were designed and synthesized by the Assays-by-Design Gene Expression service (Applied Biosystems, Foster City, California, United States): *per3* forward, 5′- GCCCTGGCAGCACCA-3′; *per3* reverse 5′- GAAAGCTGGAGGACGAGGAA-3′; probe, 5′-6-FAM- CTAAGAGCTCAAAATCC-NFQ-3′; *luc* forward, 5′- GCAGGTGTCGCAGGTCTT-3′; *luc* reverse, 5′- GCGACGTAATCCACGATCTCTTTT-3′; probe, 5′-6-FAM- TCACCGGCGTCATCG-NFQ-3′. The ABI Sequence Detection System 7000 (Applied Biosystems) was programmed to perform the following protocol: 50 °C for 2 min, 95 °C for 10 min, followed by 40 cycles of 95 °C for 15 s and 60 °C for 1 min.
In this study, relative amount of *per3* or *luc* cDNA per animal was calculated by the standard-curve method \[[@pbio-0030034-b55]\] rather than by normalizing those RNA species to a constitutive control gene, for the following reasons: Both *per3* and *luc* were compared between two different developmental stages as well as among different times of the day. It was also important to calculate the amount of each mRNA species per animal in order to compare these data to bioluminescence data. The amount of a specific control RNA, as well as the total RNA, may differ among fish of different ages, in which case RNA per animal cannot be calculated by the relative quantification method using a constitutive control. As a concentration standard, a single-stranded DNA oligonucleotide of known concentration was used for each gene. These oligonucleotides span from the 5′ end of the forward primer to the 5′ end of the reverse primer, and including 75 bp for *per3* and 110 bp for *luc* (Biosource, Camarillo, California, United States). The standard concentration was varied from 10^2^ to 10^7^ copies per reaction in 10-fold increments. For every qPCR experiment, reactions for standards were performed in four replicates along with reactions for cDNA samples.
### Statistics {#s4a8}
To test whether percentages of rhythmic fish among different experimental groups were equal, the G-test was performed using Microsoft Excel 2000 according to Sokal and Rohlf \[[@pbio-0030034-b56]\]. If multiple tests were performed for a set of data, critical value of the G-statistic was adjusted for the experimentwise error rate \[[@pbio-0030034-b56]\]. For all the other numerical data, JMP 3.1.5 (SAS Institute, Cary, North Carolina, United States) was used for the following tests: Each set of data were first subjected to the test for normality. If the data were normally distributed, the one-way analysis of variance or the *t*-test was performed. The nonparametric Wilcoxon/Kruskal-Wallis test was performed on data that were not normally distributed even after various transformations (logarithmic, square root, and inverse) were tried. Where multiple tests were performed on a set of data, the experimentwise error rate (α) was adjusted by the Dunn-Sˇidák method \[[@pbio-0030034-b56]\].
Supporting Information {#s5}
======================
Protocol S1
::: {.caption}
###### Experimental Protocol for Bioluminescence Experiments
(27 KB DOC).
:::
::: {.caption}
######
Click here for additional data file.
:::
Accession Numbers {#s5a2}
-----------------
The GenBank (<http://www.ncbi.nlm.nih.gov/>) accession numbers of the sequences discussed in this paper are *per 3* cDNA (NM\_131584) and BAC clone CH211--138E4 (AL929204). The Ensembl (<http://www.ensembl.org/Danio_rerio/>) ID of the flanking gene of *per 3* mentioned in [Figure 1](#pbio-0030034-g001){ref-type="fig"} is ENSDARG00000023492.
We are grateful to A. F. Stewart for the plasmid pBAD-αβγ, and J. D. Levine for assistance with the rhythm analysis macro. We thank J. Nassif for help with PCR screening of transgenic fish, H. Borsetti, E. Fumagalli, and N. Hernandez-Borsetti for scoring rhythmicity of luciferase data, D. A. Martinez and S. Tanoue for help with the real-time PCR technique, and R. Spotts for maintenance of our zebrafish facility. MK was an O\'Donnell Foundation fellow of the Life Sciences Research Foundation. This work was supported by National Institutes of Health grant MH60939 and Texas Advanced Research Program grant 3652--761 awarded to GMC.
**Competing interests.** The authors have declared that no competing interests exist.
**Author contributions.** MK and GMC conceived and designed the experiments. MK performed the experiments and analyzed the data. MK and GMC wrote the paper.
¤ Current address: Division of Biology, University of California-San Diego, La Jolla, California, United States of America
Citation: Kaneko M, Cahill GM (2005) Light-dependent development of circadian gene expression in transgenic zebrafish. PLoS Biol 3(2): e34.
BAC
: bacterial artificial chromosome
cps
: counts per second
DD
: constant darkness
*Km^r^*
: *kanamycin resistance* gene
LD
: light-dark
*luc*
: *luciferase* gene
MESA
: maximum entropy spectral analysis
*per*
: *period* gene
qPCR
: quantitative PCR
SEM
: standard error of the mean
|
PubMed Central
|
2024-06-05T03:55:52.077361
|
2005-2-1
|
{
"license": "Creative Commons - Attribution - https://creativecommons.org/licenses/by/4.0/",
"url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC546037/",
"journal": "PLoS Biol. 2005 Feb 1; 3(2):e34",
"authors": [
{
"first": "Maki",
"last": "Kaneko"
},
{
"first": "Gregory M",
"last": "Cahill"
}
]
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.