text
string | predicted_class
string | confidence
float16 |
|---|---|---|
Currently, there are three main functions in this program. (1) Predict the bitterant after loading the molecule into e-Bitter program. Batch function is also developed to screen the bitter compounds against the small-molecule database. (2) Visualize the fingerprint bit in the context of 3D structure, view the feature importance of fingerprint bit contributing to the bitterant/non-bitterant classification and look over the feature partial derivative of fingerprint bit reflecting the negative or positive influence on the bitterness. (3) Check whether the compound of users' interests is located within the applicability domain of our models.
|
other
| 99.7 |
Our dataset contains the experimentally confirmed 707 bitterants and 592 non-bitterants. The bitterless dataset is composed of the experimentally validated 132 tasteless compounds, 17 non-bitter compounds and 443 sweeteners, which is different from the fully hypothetical non-bitterants employed in the works of Rodgers et al. (2006) and is also distinct from the partially hypothetical non-bitterants used in the work of Dagan-Wiener et al. (2017) and Huang et al. (2016) More specifically, Huang et al. adopt 20 experimental non-bitterants and 519 hypothetical non-bitterants, and Wiener et al. treat the 1,360 hypothetically non-bitter flavors and the experimentally validated tasteless, non-bitter and sweet compounds (557 compounds) as the non-bitterants (1,917 compounds), while our work only conservatively considers the experimentally confirmed tasteless, sweet and non-bitter molecules as the non-bitterants (592 compounds).
|
study
| 99.94 |
In the work of Wiener et al. 1,360 hypothetically non-bitter flavor compounds are assumed as the non-bitterants for the bitter/bitterless classification (Dagan-Wiener et al., 2017). However, some of hypothetically non-bitter flavors are small and volatile odorous molecules, which are modulated by about 400 olfactory receptors to give rise to the sense of smell (Hauser et al., 2017). As we know, the olfactory receptors, belonging to the class A family of GPCRs, possess the binding pocket in the TMD domain (Sayako et al., 2008), meanwhile, 25 hTas1Rs, arguably categorized as class A family of GPCRs, have the binding site in the TMD domain as well (Di Pizio et al., 2016). Hence some of these small odorants may still have chances to promiscuously activate hTas2Rs to elicit the bitterness. Hence, caution should be taken in the use of hypothetically non-bitter flavors.
|
study
| 99.94 |
Moreover, the preliminary bitter/bitterless classification model in the work of Wiener et al., which is trained on the dataset including the non-bitter flavors and 2,000 diverse molecules selected from ChEMBL database (Gaulton et al., 2012) as the hypothetical non-bitterants, is probably not very promising, since there is no detailed performance evaluation about this preliminary model mentioned in their work, and actually these 2,000 compounds from ChEMBL database are excluded in the model-training to achieve their final best classification model (Dagan-Wiener et al., 2017). Therefore, our work does not take account of the hypothetical non-bitterants in our dataset.
|
study
| 99.94 |
To examine whether the chemical space differs obviously between the bitterants and non-bitterants, the histograms of the molecular weight (MW), logP, the number of hydrogen-bond donor (NHBD) and acceptor (NHBA) for this dataset are shown in Figures S1–S4 and analyzed as follows. For logP and MW (Figures S1, S2), the overall distributions for the bitterants and non-bitterants are very similar, except the height of peak. In the histogram of logP, the peak for the bitterants is a bit sharper, which indicates that more bitterants tends to be hydrophobic, while in the histogram of MW, the peak for bitterants is a little flat. Moreover, the scatter plot of logP vs. MW (Figure 1A) also illustrate that there is no apparent separation between the bitterants and non-bitterants from this perspective. For the bitterants, the scatter plot of logP vs. MW is also very close to the counterpart in the work of Wiener et al., where they proposed the bitter domain defined by the relations as follows: −3 = < logP ≤ 7 and MW ≤ 700 (Dagan-Wiener et al., 2017).
|
study
| 99.94 |
Moreover, NHBD for the bitterants is predominantly ranging from 0 to 2 (Figure S3), and NHBA for the non-bitterants mainly varies from 1 to 3 (Figure S4). Similarly, the distributions of NHBA for the bitterants and non-bitterants have peaks at 2 and 4 respectively. Thus, the comparisons of NHBD or NHBA imply that the bitterants are slightly less hydrophilic than the non-bitterants. In addition, the scatter plots of NHBA vs. NHBD show that the distributions for the bitterants and non-bitterants cannot be easily distinguished from this view (Figure 1B). Therefore, it seems that logP, MW, NHBD, and NHBA are not good to intuitively discriminate the bitterants and non-bitterants.
|
study
| 100.0 |
Furthermore, ECFP based similarity matrix (Figure 2) clearly shows the overall Tanimoto similarities between the bitterants and non-bitterants are quite low with the average value (0.0694) over the whole matrix. This provide an important clue that ECFP fingerprint may be a good molecular descriptor for our bitterants and non-bitterants classification. Thus in our work, various ECFP fingerprints are implemented and explored.
|
study
| 100.0 |
ECFP fingerprints are extensively adopted as the molecular descriptors for the machine-learning based QSAR or QSPR study (Ekins et al., 2010; Hu et al., 2012; Braga et al., 2015, 2017; Koutsoukas et al., 2016; Varsou et al., 2017; Wang et al., 2017; Yang et al., 2017). However, it has not been adopted to predict the bitterant in the existing literature. In most of the previous works about the classification, 1024bit-ECFP4 or 2048bit-ECFP6 are often chosen by default. In this work, 1024bit-ECFP4, 2048bit-ECFP4, 1024bit-ECFP6, and 2048bit-ECFP6 are systematically explored, since ECFP6 possesses more structural features than ECFP4, and 2,048 bits can accommodate more structural features than 1024bits to alleviate the bit collision. Thus in this section, the trained models without the prior feature selection are used for the comparison.
|
study
| 99.9 |
In order to statistically compare the overall performance of those four ECFP fingerprints in the context of each machine-learning method, two-sample T-test is conducted based on F1-scores from two sets of nineteen random data-splitting schemes, and is systematically performed for each pair (Table 5 and Tables S14–S16) except the model-training with DNN, since only three random data-splitting are done for DNN due to its demanding computational time and are not enough to perform T-test for the limited sample size. Thus in this section, only KNN, SVM, GBM, and RF are statistically compared with the p-value from two-sample T-test. For KNN, Table 5 shows that the average F1-scores for 1024bit-ECFP4 (AM01), 2048bit-ECFP4 (AM07), 1024bit-ECFP6 (AM13), and 2048bit-ECFP6 (AM19) are 0.898, 0.900, 0.898, and 0.898 respectively, and there are no significant differences among two ECFPs according to the criterion (p-value < 0.0001). Thus, different ECFPs won't statistically influence the F1-score for KNN with full features. Based on the Tables S14–S16, this conclusion is also hold for SVM, GBM, and RF from the statistical perspective.
|
study
| 100.0 |
(1) “N” means that there is no significant difference between two models according to the criterion (p-value < 0.0001); (2) the number in the element of matrix is the p-value after two-sample T-test for every two average models. (3) “–” indicates that two-sample T-test between the same average models will be ignored.
|
study
| 99.56 |
In short, different ECFPs won't statistically affect the performance of our bitterant/non-bitterant classification in the context of the same machine-learning method with the full features. From this view, the default choice of 1024bit-ECFP4 or 2048bit-ECFP6 is generally acceptable in the bitterant/non-bitterant classification. In our case, four ECFPs are still adopted, since their combination with the feature selection, multiple machine-learning methods and consensus strategy may offer a better solution, which will be discussed as follows.
|
study
| 99.94 |
Feature selection is widely employed to choose a subset of features for the model training in the machine-learning studies. In this work, the top 512, 256, and 128 important bits of ECFPs are selected to build the models for the comparison with their counterparts, which are trained with the full features.
|
other
| 84.0 |
The impact of feature selection on the model performance will be assessed in the context of specific combination of ECFPs and machine-learning methods. To compare the model performance in the statistical manner, two-sample T-test is also conducted based on the F1-scores from two sets of nineteen random data-splitting schemes and all the T-test results including the p-value will be reported in Table 6 and Tables S17–S31.
|
study
| 100.0 |
(1) “N” means that there is no significant difference between two models according to the criterion (p-value < 0.0001), while “Y” suggests that there is significant difference between two models according to the criterion (p-value < 0.0001); (2) the number in the element of matrix is the p-value after two-sample T-test for every two average models. (3) “–” indicates that two-sample T-test between the same average models will be ignored.
|
study
| 99.25 |
Feature selection consistently affects the performance of model trained with KNN and four ECFPs (Table 6 and Tables S17–S19). Here KNN method with 1024bit-ECFP4 is taken as an example (Table 6). The models trained with 512 features (AM25, F1 = 0.911) or 256 features (AM49, F1 = 0.920) manifest the significantly different performance compared to its counterpart trained with full features (AM01, F1 = 0.898) according to the criterion (p-value < 0.0001). Based on the average F1-score, in this case feature number 512 or 256 after feature selection is better than the full features. However, it is not always true that feature selection will be helpful to the KNN with 1024bit-ECFP4. The models trained with 128 features (AM73, F1 = 0.910) and full features (AM01, F1 = 0.898) have the similar performance, since two-sample T-test illustrates that there are no significant difference of performance between these two models, since the p-value is larger than 0.0001. Thus, feature selection will improve the performance of KNN method depending on the specific feature number. KNN method with the other ECFPs (Tables S17–S19) gives the same conclusion.
|
study
| 100.0 |
However, for the models trained the SVM method and ECFPs (Tables S20–S24), feature selection won't influence the performance relative to the full features, since there is no significant difference among these models according to the criterion (p-value < 0.0001). However, it is worth mentioning that the model trained with 128 features (AM92, F1 = 0.905) has significant difference of F1-score compared with the counterpart trained with 512 features (AM44, F1 = 0.921) based on the criterion (p-value < 0.0001).
|
study
| 100.0 |
Similarly, feature selection has no effect on the RF and GBM methods combined with any ECFP (Tables S25–S31), since no significant differences of F1-score are observed based on the criterion (p-value < 0.0001). This is expected for RF and GBM methods, since they have their own intrinsic capabilities to select the important features, hence RF and GBM methods are not very sensitive to the prior feature selection, which is consistent with the finding in the work of Zang et al. (2017).
|
study
| 100.0 |
Therefore, relative to the full features, 256 or 512 features will be good for the KNN method with ECFPs, while feature selection has no impact on SVM, GBM and RF methods in our cases from the statistical analysis. It should be noted that DNN2 and DNN3 are excluded in our two-sample T-test due to very limited samples, thus the effect of the feature selection on DNN2 and DNN3 is not discussed.
|
study
| 100.0 |
In this work, 1,312 models have been harvested considering the combination of different ECFPs, machine-learning methods, feature selection, and random data-splitting schemes. To reduce the bias from different random data-splitting schemes, 96 average models (AM) are obtained from 1,312 models by averaging over different data-splitting schemes.
|
other
| 97.9 |
To evaluate the performance of all 1,312 models and 96 average models, both F1-score and ΔF1-score are calculated from Tables S1, S2, which are displayed in Figure 7. F1-score is the key criterion used to select the best model during the modeling training with cross-validation, while ΔF1-score used to inspect the possible overfitting or underfitting of model is obtained with Equation (7). The scatter plot of ΔF1-score vs. F1-score (test set) demonstrates that all the average models and most of individual models have the ΔF1-score lower than 0.04, indicating that the performance on the test set and in the cross-validation is quite similar. Hence all these average models and most of individual models are probably robust without the apparent overfitting or underfitting from this perspective.
|
study
| 100.0 |
In addition, majorities of the average models and individual models are located in the bottom-right corner of the scatter plot (Figure 7), suggesting that most of them have high F1-score larger than 0.90. Moreover, MCC (test set) vs. F1-score (test set) for all the 1,312 individual models and 96 average models (Figure S12) indicates that most of these models are quite good from the perspective of MCC and F1-score. Therefore, most of our average models and individual models exhibit the promising performance.
|
study
| 100.0 |
To examine the reliability of all the 1,312 models, Y-randomization test is conducted by the random shuffling of the experimentally-observed labels, which generates a noisy dataset (Table S3). Before the model-training, the accuracy, precision, specificity, sensitivity, F1-score, and MCC of this noisy dataset are 47.5~53.3%, 51.8~57.1%, 42.4~48.7%, 51.8~57.1% 0.518~0.571, and −0.058~0.058 respectively (Table S3), which indicates that almost half of samples are still correct in this noisy dataset. Taken the randomized exp01 (Table S3) as an example, the respective numbers for the true bitterants (TP) and true non-bitterants (TN) are 315 and 223, while the respective numbers for the false bitterants (FP) and false non-bitterants (FN) are 251 and 251 due to the random shuffling (Table S3). However, the ratio between bitterants (TP and FP) and non-bitterants (TN and FN) is still 566:474, which is identical with its counterpart in the original Dataset-CV before the shuffling.
|
study
| 100.0 |
After training on these noisy datasets via the cross-validation, the models with the best performance (highest F1-score) on the internal validation dataset are also assessed on the test set, which is not transformed by any random shuffling for the labels. According to Figure S13, the F1-score and MCC of the models (blue spheres) in the Y-randomization test is strikingly decreased relative to the original models (red spheres), which suggests that our original models are quite reliable. Meanwhile, the MCC of all these models in the Y-randomization test are close to 0, indicating that these models in the Y-randomization test have no better than the random prediction. Hence the Y-randomization test substantiates the reliability of the original models.
|
study
| 100.0 |
Applicability domain of models are defined quantitatively by the average-similarity between the given test compound and its five nearest neighbors in Dataset-CV. If the given test compound is close to its neighbors with a larger average-similarity, it means that the chemical space of this given compound is covered by Dataset-CV, thus the prediction is probably interpolated from Dataset-CV, which gives more reliable estimation. According to the average-similarity histograms (Figure S11) of Dataset-CV and Dataset-Test, the compounds in the test set (Dataset-Test) are fully covered by the dataset for the cross-validation (Dataset-CV), thus the given compound with average-similarity higher than 0.1 is assumed to be within the applicability domain of our models according to Figure S11. If the compound is not located within the applicability domain of our models, the prediction for this compound is probably extrapolated from Dataset-CV and consequently is not confident.
|
study
| 100.0 |
BitterX and BitterPredict adopt SVM and Adaboost methods respectively, while e-Bitter implicitly integrates nine consensus models (CM01-CM09) and optionally includes 96 average models (AM01-AM96). To evaluate the performance of all the models, two types of comparisons are conducted. One is the direct comparison on the test set reported in the original works. The other is the more fair comparison on the three external test sets in the recent work of Wiener et al.
|
study
| 99.25 |
For the direct comparison, F1-score (test set) and MCC (test set) of the model from BitterX are 0.918 and 0.842 respectively (Table 1), and the counterparts of the model from BitterPredict are 0.708 and 0.595 respectively (Table 1). It worth mentioning that F1-score and MCC are all calculated based on the data reported in their works. From this perspective, BitterX is much better than BitterPredict. Additionally, the model from BitterX and our models including consensus models and average models are markedly better than the counterpart from BitterPredict (Red sphere in Figure 3). Moreover, one of our consensus models (CM1, Blue sphere in Figure 3) is also a little better than the model from bitterX (Black sphere in Figure 3), the consensus models (CM03, CM05, and CM07) are comparable to the model from BitterX and the consensus models (CM02, CM04, CM06, CM08, and CM09) are slightly inferior to the model from BitterX. It should be noted that the model from BitterPredict is derived from only one random data-splitting scheme, and the model from BitterX is averaged over three random data-splitting schemes. Our consensus models are built by averaging over all the constituent models (19 individual models for CM01 and 5 average models for CM02-CM09), while our average models (AM) in Table S2 are derived by averaging over 19 random data-splitting schemes for KNN, SVM, GBM, and RF or three random data-splitting schemes for DNN2 and DNN3. Therefore, our models in the e-Bitter may offer more robust results in this respect.
|
study
| 100.0 |
To seek the further objective evaluation, three external test sets in the recent work of Dagan-Wiener et al. (2017) are employed accordingly. For the “Bitter New” dataset with 23 bitterants, the prediction accuracies are 74, 74, and 100% for bitterX, BitterPredict, and e-Bitter respectively, while F1-scores are 0.85, 0.85, and 1.00 for bitterX, BitterPredict, and e-Bitter respectively (Table 2). Thus, all the consensus models and average models in the e-Bitter afford the best performance for this dataset (Figure 4 and Table 2).
|
study
| 100.0 |
As concerning the “UNIMI set” dataset (23 bitterants and 33 non-bitterants), bitterX obtains the accuracy, precision, specificity, sensitivity, F1-score and MCC of 60, 52, 56, 65%, 0.58 and 0.21 respectively (Table 3), which are lower than their counterparts reported in the work of Huang et al. (2016) BitterPredict offers the accuracy, precision, specificity, sensitivity, F1-score and MCC of 82, 78, 85, 78%, 0.78 and 0.63 respectively reported in the work of Dagan-Wiener et al. (2017) Meanwhile, e-Bitter gives the accuracy, precision, specificity, sensitivity, F1-score and MCC of 68~73%, 57~61%, 55~58%, 87~100%, 0.69~0.75, and 0.47~0.58 respectively considering the different consensus models (Table 3), which are also lower than their counterparts evaluated on our test set (Dataset-Test) in Table S2. Obviously, BitterPredict and e-Bitter (9 consensus models and 96 average models) unanimously outperform BitterX (Figure 5), while BitterPredict exhibits the slightly better performance than e-Bitter in light of F1-score and MCC (Figure 5).
|
study
| 99.94 |
Regarding to the “Phytochemical Dictionary” dataset (49 bitterants and 26 non-bitterants), BitterPredict provides the accuracy, precision, specificity, sensitivity, F1-score and MCC of 88, 86, 69, 98%, 0.91 and 0.735 respectively (Table 4, Dagan-Wiener et al., 2017) while e-Bitter predicts that the accuracy, precision, specificity, sensitivity, F1-score and MCC are 85~92%, 85~91%, 69~81%, 94~98%, 0.89~0.94, and 0.67~0.82, which is comparable with the counterparts evaluated on our test set (Dataset-Test) in Table S2. BitterX achieves the accuracy, precision, specificity, sensitivity, F1-score and MCC of 72, 72, 31, 94%, 0.814 and 0.332 respectively (Table 4). Thus for this dataset, all 9 consensus models and 96 average models in the e-Bitter are consistently better than BitterX (Figure 6), while most of the consensus models (CM01, CM02, CM03, CM04, CM05, CM08, and CM09) and some average models show more promising results than BitterPredict based on the F1-score and MCC (Figure 6).
|
study
| 99.94 |
Therefore, in light of the performance indicators F1-score and MCC, e-Bitter obtains the best performance for the “Bitter New” dataset and “Phytochemical Dictionary” dataset. e-Bitter consistently outperforms bitterX for all these three test sets. BitterPredict achieves the slightly better result than e-Bitter for the “UNIMI set” dataset. It is worth noting that in this section the performance of e-Bitter specifically refers to the performance of the consensus models, albeit the performance of 96 average models are optionally available in the e-Bitter and also discussed above.
|
study
| 99.94 |
ECFP fingerprint is prevalently used in the chemoinformatics, however, this “0/1” bit string is quite obscure for the food scientist to correlate the specific bit “1” to its corresponding structural feature in the context of 3D structure. For this purpose, our implementation is designed to record all the structural information including atoms and bonds for each bit during the ECFP generation, and is intended to highlight the structural feature in the context of the whole 3D structure via clicking the bit “1” of interest. This unique feature of our implementation will provide an appealing advantage. In our current version, ECFP diameter and bit length can be customized by the users, albeit the commonly-used ECFP diameter is 4 or 6, and frequently-used ECFP bit length is 1,024 or 2,048. Once setup the diameter and bit length, e-Bitter program can generate the ECFP fingerprint in the fingerprint windows via simply clicking on the menu, and can automatically associate all the bit “1” with the corresponding structural features in the 3D viewer, and is ready for the users to select the bit of interest (Figure 8).
|
other
| 97.94 |
The interactive visualization of the fingerprint bit and its corresponding structural feature in the 3D viewer. The title of fingerprint window shows the type of ECFP and also displays whether there is bit-collision occurring in the fingerprint. All the structural features can be visualized by browsing the list in the combo box (bottom) even if there is bit-collision.
|
other
| 99.94 |
Besides the structural visualization for the fingerprint bit, the associated feature importance and feature partial derivative are also very useful information. The feature importance emphasizes the importance of each bit contributing to the bitter/bitterless classification, and the feature partial derivative of each bit stresses the positive or negative influence of each bit on the bitterness of the compound. Hence, the fingerprint bit “1,” its corresponding structural feature, feature importance, and feature partial derivative are closely interconnected and can be interactively visualized in our e-Bitter program. This interactive and synchronization function can be well depicted by Figure S14.
|
other
| 99.75 |
e-Bitter is a stand-alone package, which can be freely downloaded from Dropbox shared folder (https://www.dropbox.com/sh/3sebvza3qzmazda/AADgpCRXJtHAJzS8DK_P-q0ka?dl=0). This program is well tested on the 64bit windows operating system such as Win7, Win8, and Win10, while the external Scikit-learn, Keras and Tensorflow python libraries can be easily set up via simply installing the Winpython-64bit v3.5.4.0 that integrates the complete python environment and rich python libraries on the windows operating systems. In this sense, the whole installation process on user's computer is quite handy. Once installed and configured, e-Bitter can predict whether the molecule loaded in the 3D viewer is bitter or not, or can perform an automated virtual screening against a small-molecule database to obtain the possible bitterants candidates. Additionally, e-Bitter program also implements the job management system (Figure S15), since it would be time-consuming to conduct the prediction with the consensus models, particularly for those models including the deep neuron network methods. Figure 9 briefly recapitulates all the functions in the e-Bitter program.
|
other
| 99.6 |
Besides this graphic mode, e-Bitter program can also work in the console mode, since some users prefer to use the command line to predict the bitterants. The only input for e-Bitter program is the molecular file with the Tripos Mol2 format. The detailed example and usage are described in the tutorial and manual, which are integrated in the e-Bitter program and can also be accessed from the Dropbox shared folder.
|
other
| 99.94 |
Furthermore, all the consensus models and their constituent models integrated in our program can be easily upgraded. As we know, the dataset about the bitterants and non-bitterants will grow more and more rapidly, thus we will continue to spend the effort to constantly upgrade our models with the larger experimental dataset, and upload the new models to the folder called “model” in the Dropbox shared folder. Users can download the latest models to replace the previous ones for their prediction.
|
other
| 99.9 |
Relative to the online tool bitterX (Huang et al., 2016), e-Bitter works on the local machines, which ensures that users can well exploit their own computational resource to test their propriety compounds without turning to the external web server. Moreover, e-Bitter can screen the small-molecule database in batch. More importantly, e-Bitter program adopts the consensus voting strategy based on the multiple machine-learning algorithms to enhance the reliability of prediction result.
|
other
| 99.9 |
Compared to the MATLAB tool BitterPredict (Dagan-Wiener et al., 2017), e-Bitter adopts the ECFP fingerprint as the molecular descriptors, which is natively implemented in our program. Thus, e-Bitter does not depend on any other commercial softwares such as Schrödinger package required in BitterPredict. In addition, this free e-Bitter program employs the versatile machine-learning algorithms and works compatibly with the free python environment, while BitterPredict is developed and runs in the commercial MATLAB environment, which will restrict its extensive tests or applications for most of the users.
|
other
| 99.9 |
Admittedly, our work has some limitations. (1) From the perspective of curated data, the experimentally confirmed dataset probably still has some intrinsic noise, since in the experimental taste assessment, the trained panelists have some objective factors such as the individual gene-polymorphism of Tas2Rs, and also have some subjective factors such as the mixed tastes, which are unlikely to be clearly discerned. (2) As for the parameter optimization in the model training process, it is a daunting task to explore the complete parameters combinations, since the parameter spaces for machine-learning methods such as DNN are very large. Thus, in this work only the key parameters are tuned with the grid method. (3) Regarding to the feature selection, feature importance from the random forest method is used as the criterion to select the important features. In this work, only full features, 512, 256, and 128 features are attempted, while the other feature numbers and feature selection methods, which may provide more promising results, are not tried because of the tremendous combination of options and parameters. (4) From the view of consensus models, this strategy will introduce some extra computational-burden. First, different types of ECFP fingerprints should be generated for each compound, and then each constituent model coupled with its corresponding fingerprint affords the prediction result also for each compound, finally the results from all the constituent models will be averaged to make the final prediction. Therefore, this procedure is not extremely fast, especially for the consensus model containing the deep-neuron network. (5) From the function of e-Bitter, this program is only focused on the bitterant classification for the small molecule due to our current research priority.
|
study
| 99.94 |
In the future, we will devote to further collect the high-quality dataset and constantly update the models. Moreover, we will pay more attention to the other feature selection methods and parameter optimization schemes and will implement in the future version of e-Bitter program. Furthermore, e-Bitter program will be extended to qualitatively classify the bitter/bitterless peptide, quantitatively predict the bitterness of the bitterants, and explore the possible target information of the bitterants.
|
other
| 99.9 |
In this work, it is the first time that the fully experimental bitterants/non-bitterants dataset, consensus voting based on the multiple machine-learning algorithms, and ECFP fingerprints are adopted to build the bitter/bitterless classification models. Through the exhaustive parameter exploration with the five-fold cross-validation, all the models are carefully scrutinized by the Y-randomization test to ensure their reliability, and subsequently nine consensus models are constructed based on the individual or average models, which differ in term of accuracy, speed and diversity of models. Evaluation on the three external test set from Wiener et al. demonstrates that e-Bitter outperforms bitterX on these three test sets; e-Bitter harvests the better results than BitterPredict for the “Bitter New” and “Phytochemical Dictionary” dataset; BitterPredict demonstrates slightly better performance than e-Bitter for the “UNIMI set” dataset. To automate the whole process, we develop a graphic e-Bitter program for the bitterant prediction or screening against the small-molecule database in batch. Additionally, e-Bitter program natively implements ECFP fingerprint, and more importantly, e-Bitter can vividly visualize the structural feature, feature importance, and feature partial derivative for any specific bit “1” in the ECFP fingerprint. We hope our work can provide a useful tool for the experimental scientist to rationally design and screen the bitterants.
|
study
| 99.94 |
Misfolding and aggregation of amyloidogenic proteins is associated with a vast number of neurodegenerative diseases including Alzheimer’s disease (AD), Parkinson’s disease (PD), and Huntington’s disease (HD), termed protein aggregation diseases12. Numerous studies have shown that amyloidogenic proteins are capable of spontaneous assembly into aggregates and eventually form fibrillar structures found in amyloid or amyloid-like deposits34. These studies have led to the amyloid cascade hypothesis, which posits that the disease onset involves a spontaneous assembly of amyloidogenic polypeptide, and that the accumulation of aggregates defines the disease state567. Numerous studies support this hypothesis that currently is considered the main model for the onset of AD, PD, HD, and other diseases. Although the amyloid cascade hypothesis cannot explain all facts related to the development of these diseases, it remains the major underlying hypothesis of in vitro and in vivo studies related to the molecular mechanisms of amyloid aggregation causing neurodegenerative diseases8. A strong support for the amyloid cascade hypothesis comes from recent studies that demonstrated that antibody-based immunotherapy against Aβ improved cognition in a dose-dependent manner9. Evidence of similarities of structural features of aggregates extracted from amyloid plaques with those of Aβ aggregates assembled in vitro provide additional support for the use of in vitro Aβ aggregation studies for understanding Aβ structural dynamics in vivo10. Amyloid aggregates can be visualized by electron microscopy and atomic force microscope (AFM) (e.g. refs 11 and 12 and references therein), and structural techniques including X-ray and NMR applied to fibrils have revealed a highly ordered arrangement of protein monomers13. However, recent studies have led to the discovery that oligomers rather than fibrils are neurotoxic31415. Studies of these transient oligomeric species have been enabled by the development of novel approaches1617 and have shown that at the very early aggregation stage, unstructured monomers form stable dimers due to structural transitions of monomers. In turn, these results suggest that the lack of structure in monomers can facilitate the aggregation. However, there is a serious complication with translating current knowledge on amyloid aggregation in vitro to understand the aggregation process in vivo —namely, the concentrations of amyloidogenic polypeptides are dramatically different in vivo versus in vitro. For example, whereas the critical concentration for the spontaneous aggregation of Aβ peptide in vitro is in the micromolar range, physiological concentrations of Aβ are in the low nanomolar range1819. At such low concentrations, Aβ aggregation cannot occur in vitro. Intracellular crowding effects18 can facilitate the aggregation process, but these do not solve the problem. Recently, we have found that dimers of α-synuclein (α-Syn) could be assembled at nanomolar concentrations, if the target monomer is tethered to a surface20. These data led us to hypothesize that binding to the surface can be a factor dramatically facilitating the aggregation process. This hypothesis is supported by recent studies in which the assembly of large α-Syn aggregates on a glass surface was observed with the protein concentration in the nanomolar range21.
|
study
| 99.9 |
In the present study, we developed an approach enabling us to directly test our hypothesis. We utilized AFM to image assembly of aggregates on surfaces and compare this effect with their assembly in the bulk solution. We used full-length Aβ protein (Aβ-42), its aggregation-prone segment Aβ(14-23) peptide and the full-size α-Syn. The experiments demonstrate that all these proteins on mica surface assemble into aggregates at nanomolar concentration with essentially no aggregation propensity in the bulk solution. Time-lapse AFM imaging experiments in solution demonstrate that assembled aggregates can be released to the solution to act as seeds for the aggregation in bulk solution. Computational modeling allowed us to reveal the mechanism of the accelerated on-surface aggregation process. Given that the on-surface aggregates are oligomeric in nature, which are known to be the most neurotoxic species, we hypothesize that prevention of the on-surface aggregation should block the disease-prone process and can be considered a means for the development of future preventions and treatments for Alzheimer’s and similar neurodegenerative protein aggregation diseases.
|
study
| 99.94 |
To directly test the hypothesis that surface interactions facilitate the self-assembly of amyloidogenic polypeptides, we performed systematic AFM studies of the on-surface aggregation of full size Aβ-42, Aβ(14-23) peptide and α-Syn protein at the nanomolar range. The on-surface aggregation schematically is shown in Fig. 1A. The mica sheets functionalized with 1-(3-aminopropyl) silatrane (APS) were immersed in a 100 nM solution of Aβ-42 in 10 mM sodium phosphate buffer (pH 7.4). After a certain incubation time, a mica sheet was removed from the solution, rinsed with water, dried and imaged with AFM. The samples were prepared for incubation times 0 h, 24 h, 48 h and 72 h, and typical images are shown in Fig. 1B, plates i-iv, respectively. The surface for the initial sample (0 h, plate (i)) is clean, but globular aggregates appear after a 24 h incubation time (plate (ii)), and the surface is quite densely coated with aggregates after 48 h (plate (iii)). The aggregate sizes are larger for the 72 h sample (plate (iv)). In parallel with the on-surface aggregation, aggregation of the same solution of the protein in the absence of a mica surface (aggregation in the bulk solution) was performed. A set of typical images is shown in Fig. 1C. No aggregation was detected for these samples. Only after 72 h of incubation in the bulk solution, a few globular features are noticeable (Fig. 1C, plate viii).
|
study
| 100.0 |
To quantitatively characterize the aggregation process, we measured the aggregate volume and number at all incubation times, and the data are shown in Fig. 1D and Supplementary Fig. 1. The distribution of aggregate volumes is approximated with a single Gaussian for the 24 h incubation with a maximum of 107.2 nm3 (Supplementary Fig. 1A), but the distributions are bimodal for two longer aggregation times (48 h (Supplementary Fig. 1B) and 72 h (Supplementary Fig. 1C)), indicating the formation of two sets of aggregates. The sizes of aggregates that appear after a 72 h incubation in the bulk solution were also measured, and the distribution is shown in Supplementary Fig. 1D. The overall number of counts is low due to a low yield of globular aggregates, and their sizes are more than 50 times smaller than those found for the aggregates formed on the surface. Thus, interaction with the surface dramatically enhances the Aβ-42 aggregation process.
|
study
| 100.0 |
Next, to characterize the aggregation process further, we performed time-lapse AFM imaging experiments in which images were taken over the same surface area continuously after injecting 100 nM Aβ-42 protein in 10 mM phosphate buffer (pH 7.4) into the fluid cell. AFM images for selected observation times are shown in Supplementary Fig. 2A. Frame (i) shows the image of the surface prior to adding the protein, illustrating that the surface is clean. The next frame (ii) shows the image taken after 10 min, and here again no protein aggregates are seen. As the incubation time increases, new aggregates appear on the surface. For example, frame (iii) illustrates that quite a large number of aggregates appear after 5 h incubation. Comparison with the images taken after 6 h demonstrates the dynamics of the aggregation process in which aggregates are formed and dissociate and grow. For example, an aggregate circled in yellow in frame (iii) is no longer present in (iv), and the aggregate circled in yellow in (iv) disappears in (v), suggesting that these aggregates have dissociated from the surface. At the same time, a new aggregate circled in green appears in (iv). The aggregate indicated by a dotted black circle remains throughout the time-period of imaging (Supplementary Fig. 2iii–vii) and serves as a marker to monitor the changes occurring in other aggregates. The aggregates marked with a yellow dotted circle dissociate from the surface, whereas the new aggregates which appear on the surface are indicated with a green dotted circle. The results are summarized graphically in Supplementary Fig. 2B. In this graph, the time-dependent changes of overall volume and number of aggregates are plotted. The graphs demonstrate that the particle volumes increase over time monotonically, although the number of aggregates is not monotonic, indicating that aggregation occurs as a dynamic process in which aggregates can dissociate from the surface. To illustrate these dynamics for individual aggregates, we numbered a few particles as shown in Supplementary Fig. 2 iii-vii. For example, the aggregate marked ‘1’ stays in plates iii and iv (Supplementary Fig. 2A) but disappears in plate v. The aggregate marked ‘2’ shows up in plate iv and disappears in plate vii (Supplementary Fig. 2A). Measurement of the aggregate sizes (Supplementary Table 1) demonstrated that some of the aggregates grow in size leading to an increase in volume (aggregates ‘3’, ‘4’, ‘5’ and ‘6’), whereas some dissociate from the surface (aggregates ‘1’ and ‘2’). To estimate the subunit numbers of the oligomers, we compared the volumes determined here with volumes of isolated Aβ-42 oligomers of specific sizes obtained by photoinduced cross-linking22. These measurements (data not shown) show that the heptamer and decamer have volumes of 60.3 ± 9.7 nm3 and 73.4 ± 12.8 nm3 respectively, suggesting that the oligomers formed after 8 h of the time-lapse experiment correspond to the oligomer order of heptamer to decamer.
|
study
| 100.0 |
The stretch of Aβ peptide spanning residues 14 through 23 is the segment of the Aβ-42 peptide (designated as Aβ(14-23) from hereon) that adopts β-sheet structure when Aβ-42 assembles into fibrils, and Aβ(14-23) itself forms fibrils in solution2324. The on-surface aggregation experiments were performed under the same conditions as in the above described experiments (100 nM Aβ(14-23), 10 mM sodium phosphate buffer (pH 7.4) at room temperature), and the results are shown in Fig. 2. A few aggregates appear as globular features after 24 h (Fig. 2A-plate i) and their sizes have increased at the 72 h incubation time point (Fig. 2A, plate iii). Fibrillar features were also observed and are indicated with arrows. Zoomed images of the fibrils are shown as insets (Fig. 2Ai–iii), and a 3D view is shown in Fig. 2A(iii). Statistical analysis for the globular aggregates sizes was performed, and the data are shown in Fig. 2B. There is a considerable growth of the aggregates between 48 h (Fig. 2B(ii)) and 72 h (Fig. 2B(iii)), with a bimodal distribution reflecting the presence of large aggregates evident in the histogram corresponding to the later time point (Fig. 2B(iii)). A graph in which these values as well as the number of aggregates were plotted against incubation time (Fig. 2C) illustrates a monotonic growth of aggregates in size and number. As a control, aggregation in the bulk solution was monitored in parallel. Typical AFM images for these samples are shown in Supplementary Fig. 3. Only a few aggregates were formed after 72 h, and no fibrillar features were found.
|
study
| 100.0 |
Time lapse experiments, similar to those carried out with the full-length Aβ42 peptide as described above, were also performed for the Aβ(14-23) peptide. The same surface area was imaged continuously after injecting 100 nM of Aβ(14-23) peptide in 10 mM sodium phosphate buffer in a fluid cell. Supplementary Fig. 4A shows typical images of aggregates at different time intervals: frame (i) 10 minutes after protein addition; (ii) 4 h; (iii) 5 h, and (iv) 6 h respectively. Frame (i) illustrates the presence of a very small number of globular aggregates after 10 min of peptide injection. As time progresses, the number and size of aggregates increases as shown in frames (ii) to (iv). Dynamics of association/dissociation of amyloid aggregates with the surface were also observed in time lapse experiments for this small fragment Aβ(14-23). The black circles in frames (ii), (iii) and (iv) indicate a subset of aggregates and serve as markers to monitor the changes occurring in other aggregates. A few examples of association (green dotted circle) and dissociation (yellow dotted circles) of Aβ(14-23) aggregates are shown in frames (iii) and (iv). The statistics on number and size (volume) of globular aggregates over time are shown in Supplementary Fig. 4B indicating a monotonic increase in both variables with time.
|
study
| 100.0 |
To test that the observed enhancement effect of the surface is not limited to Aβ peptides, we performed surface-aggregation experiments with α-synuclein (α-Syn) protein. Given that α-Syn is more than three times larger than Aβ-42 and the fact that we were previously able to image α-Syn monomers20, we anticipated being able to visualize the formation of oligomers starting with dimers. The surface-mediated aggregation experiment was carried out similarly to the analyses described above. Similar to the experiments with Aβ peptides, mica sheets were incubated in 10 nM α-Syn solution for different time-periods and imaged with AFM. Typical AFM images for samples corresponding to 0 h, 24 h and 48 h incubation times are shown in Fig. 3A. A few globular features were seen in the initial sample (plate (i)). To characterize their sizes, volume measurements as described in ref. 20 were performed. These measurements are shown in Fig. 3B(i) and they indicate that these features are monomers. After 24 hours (Fig. 3A, plate ii), the particles become brighter and larger in size; the volume measurements (Fig. 3B(ii)) are consistent with the formation of oligomers larger than dimers. Longer incubation (48 hours) leads to the formation of aggregates with a broad range of sizes as evidenced by images in plate (iii) and the corresponding volume distribution (Fig. 3B(iii)). Control experiments for α-Syn aggregation in the bulk solution (Fig. 3C) did not reveal such a pronounced aggregation. Even after a 24 h incubation, the bulk solution did not show aggregates (Fig. 3C, plate ii). The volume measurements for the 48 h sample reveal a major peak at 39 nm3 which is very close to the monomer volume (Fig. 3D)20. There is a secondary minor peak with a volume value of 66 nm3 that can be assigned to dimers-trimers.
|
study
| 100.0 |
Time-lapse measurements were also performed with a 100 nM α-Syn solution. AFM images for 2, 3 and 5 hours of continuous observation, shown in Supplementary Fig. 5, clearly illustrate the time-dependent aggregation of α-Syn. Similar to the previous experiments, the surface was imaged first in 10 mM sodium phosphate buffer (pH 7.4) to confirm that the surface was clean (Supplementary Fig. 5A(i)). Then 100 nM α-Syn was injected onto the surface, and imaging was continued. A considerable number of aggregates were observed after 2 h (Supplementary Fig. 5A(ii)). The number of aggregates was further increased at the 3 and 5 h time points. (Supplementary Fig. 5A(iii, iv)). The volume distribution of the aggregates indicates that the size of the aggregates increases with time (Supplementary Fig. 5B). Similar time-lapse experiments with 10 nM α-Syn demonstrating the surface induced aggregation has also been performed and the data are shown in Supplementary Fig. 6. Apart from globular aggregates, the formation and growth of fibrillar features was also visualized (Supplementary Fig. 6B). In this dataset, an initially small fibrillar complex was observed in plate (i) and found to grow in size over time (plate (ii) and (iii)), but became smaller later on (plate (iv)). This set indicates that the assembled complex is not stable and can dissociate. A control experiment in connection with all of the time-lapse studies mentioned above was performed by imaging a mica surface in 10 mM sodium phosphate buffer (pH 7.4) without any protein for an extended period of time (up to 6 h) (Supplementary Fig. 7). This experiment did not show the appearance of any aggregate like features even after 6 h.
|
study
| 100.0 |
In order to understand the effect of the surface on aggregation, we performed all-atom molecular dynamics (MD) simulations of interactions of Aβ(14-23) monomers with the mica surface, Fig. 4A. Two systems were simulated, Mica1 and Mica2, with initial monomer structure being adopted from23; for detailed description of the simulation parameters please see the Methods section.
|
study
| 100.0 |
The interaction of two peptide monomers with each mica surface was simulated, the secondary structure of the monomers is characterized using the defined secondary structure of proteins (DSSP) method25 (Fig. 4B). In the Mica1 system, Aβ(14-23) monomer A rapidly interacts with the mica surface; within the first 50 ns of the simulation the monomer approaches the surface. Binding of the monomer is accompanied by its structural transformation, going from having a small helical segment to assuming a bend structure, as seen with the change in secondary structure in Fig. 4B. The interaction between the monomers was characterized by the distance between them as a function of time as shown in Fig. 4C and D. A few snapshots illustrating the peptide structure are indicated along the time trajectory. The surface induces a conformation that is favorable for dimer formation, as is evident from the rapid recruitment of the free monomer and the formation of a dimer bound to the surface, Fig. 4D. Recruitment of the free monomer, monomer B, happens within the first 100 ns of the simulation; the dimerization causes a structural change in the previously free monomer, Fig 4B. However, the newly formed dimer is only formed transiently and for the next ~200 ns forms and dissociates multiple times as is demonstrated by the fluctuation of the dimer-surface plot in Fig 4C. The dimers interact with the surface primarily staying in contact with the surface through interactions involving a few residues and rarely lie fully on the surface. The behavior of the Aβ(14-23) monomers in the Mica2 system is very similar to the Mica1 system, with the exception that the interaction of peptides with the surface is not as strong, and the conformation of the dimers formed is less compact, Supplementary Fig. 8.
|
study
| 100.0 |
Next we performed similar MD simulations using the DLPE lipid bilayer system that mimics membrane surfaces2627. Two monomers were placed above and below the bilayer (Fig. 5A inset). The initial orientations of the monomers on the outside and inside facing leaflets are opposite of each other to determine if the initial approach orientation has an effect on the surface interaction. Similar to the results for the mica surface, the monomers assemble into dimers, but the process occurs more rapidly. As shown in Fig. 5A, in both dimers, the monomers interact with the surface within 20 ns followed by a rapid dimer formation process. Structural changes occur in monomers upon the interaction with the surface. Fig. 5B demonstrates the structural dynamics of the monomers over the entire simulation time. The initial structure containing small helical segments becomes largely unstructured with β-bridge, turn, or bends. The dimer interacts with the surface and re-organizes on the surface. Through this re-organization, monomers are able to dissociate from and re-associate with the surface – but still remain as dimers. This dynamic behavior is seen as fluctuations in the number and duration of contacts during the simulation, Fig. 5C.
|
study
| 100.0 |
The simulations suggest the following model for the surface effect on dimer formation. Interaction of a monomer with the surface is accompanied by the structural transition of the monomer. Another monomer binds to the immobilized monomer, forms a dimer in which both monomers undergo a structural transition. As a result, the interaction with the surface accelerates the formation of dimers. Compared to our previous simulations for dimer formation by free Aβ(14-23) monomers in which no major structural changes were observed during the first ~200 ns23, we have an almost five-fold faster structural transition when the peptides interact with the surface.
|
study
| 100.0 |
We have shown that the interaction of amyloid proteins with surfaces presents a unique opportunity to allow the protein to assemble into aggregates at concentrations that are non-permissive for aggregation in solution. This finding eliminates a major problem with understanding the spontaneous appearance of plaques in the AD brain due to the apparent gap between known extracellular Aβ concentrations in vivo (in the low nM range) and the concentration required for spontaneous aggregation in vitro (which is in the μM range). Importantly, the on-surface aggregation is a dynamic process, so the assembled aggregate can dissociate from the surface to the bulk solution. This mechanism is confirmed by direct measurement of the concentration of Aβ-42 aggregates in solution depending on the presence of the mica surface. The data shown in Supplementary Fig. 9 demonstrate that in the presence of mica, the concentration of Aβ-42 aggregates in solution is considerably higher than in the control, and this number grows with time. As a result, these dissociated oligomers can play roles as seeds for aggregation in the bulk solution, or start a neurotoxic effect such as phosphorylation of the tau protein to initiate its misfolding and aggregation followed by neurodegeneration8. Additionally, in the vast majority of cases, we found that aggregates formed on the surface are oligomers, which are considered to be the most neurotoxic amyloid aggregates. Although the experimental data presented was for APS-mica, a very similar effect was observed for the bare mica. Additionally, we are currently performing experiments with the use of a lipid bilayer surface. Preliminary results show an even stronger on-surface aggregation effect for the lipid bilayers compared with APS and bare mica. Therefore, we posit that on-surface aggregation is a general mechanism by which neurotoxic amyloid aggregates are produced under physiological conditions.
|
study
| 100.0 |
With regard to applications of this work to AD development, we propose the following model in the framework of the amyloid cascade hypothesis in which interaction of amyloidogenic polypeptides with cellular membranes plays an important role for the disease-related aggregation process. Under normal conditions, the interaction of intracellular or extracellular amyloid proteins with intracellular or extracellular membranes is weak, so small aggregates assemble. These are unstable and dissociate into monomers either on the surface or after dissociation from the membrane. A change in membrane properties leading to an increase in affinity of amyloid proteins to the membrane surface will shift the process to the formation of more stable oligomers that remain intact after dissociation from the surface, and this assembly triggers the disease-related aggregation process. This mechanism does not require elevation of the amyloid peptide concentration, and indeed the concentration of amyloid beta peptide in blood fluctuates weakly regardless of the disease state and does not differ from the controls28. Also, the Aβ clearance in late-onset AD patients drops by about one quarter29, and only a fraction of the Aβ produced is trapped in amyloid plaques30. Our model is in line with recent findings31323334 that demonstrate that the aggregation rate of amyloidogenic proteins measured in the presence of membranes of various types depends on the membrane composition and mechanical properties. Indirect support for the concept of membrane-induced aggregation comes from findings on the elevated yield of Aβ dimers in membrane-containing fractions of blood from AD patients compared with controls35. Note as well our direct observation of α-Syn on-surface assembly at nanomolar concentrations when the initial monomer was covalently bound to the surface20. Recent NMR studies on the intracellular structure of α-Syn showed that it primarily exists in cells as monomers in an essentially unstructured, compact conformation, but transient interaction of the protein with the membrane was considered9. Indeed, the on-surface assembled oligomers comprise a very small fraction of the protein mass in bulk solution, and only a fraction of the aggregates dissociate into the solution, so their detection by NMR is a challenge.
|
study
| 100.0 |
Our surface mediated amyloid aggregation model is a significant departure from the current model in which amyloid aggregation is linked with elevated synthesis of amyloid proteins and their accumulation in the cell to initiate the aggregation process. The lack of evidence for the change of amyloid protein level during the disease progression is a weak point of this model. Our model does not require an elevation of amyloid synthesis as the on-surface assembly can occur at the nanomolar concentration range of the protein. The approaches described in this paper can be extended to various types of surface, so future studies will provide detailed characterization of the on-surface aggregation process and pinpoint possible changes in cellular membrane defining the disease prone state of the organism.
|
study
| 100.0 |
Aβ-42 protein (AnaSpec, Fremont, CA) and Aβ(14-23) having the sequence HQKLVFFAED (Peptide 2.0 company, VA, USA) were dissolved and sonicated for 5 min in 100 μL of 1,1,1,3,3,3 Hexafluoroisopropanol (HFIP) to destroy pre-aggregated oligomers. Then the solvent was evaporated in vacufuge (1 hr) for complete removal of HFIP. The stock solution of the protein/peptide was prepared in DMSO and kept at −20 °C. Wild-type A140C α-Syn in which the C-terminal alanine was replaced with a cysteine was prepared as described previously36. α-Syn solutions were freshly prepared by dissolving 0.4 to 0.8 mg of the lyophilized powder in 200 μL water (the pH has been adjusted to 11 with NaOH), with the addition of 1 μL of 1 M dithiothreitol (DTT) to break disulfide bonds, followed by the addition of 300 μL of 10 mM sodium phosphate buffer (pH 7.0). The obtained solution was filtered through an Amicon filter with a molecular weight cutoff of 3 kDa at 14,000 rpm for 15 min. The filtration was repeated 3 times to completely remove free DTT. The concentration of the stock solutions was determined by spectrophotometry (Nanodrop® ND-1000, DE) using the molar extinction coefficients 1280 cm−1·M−1 and 120 cm−1·M−1 for tyrosine and cysteine at 280 nm, respectively. In general, freshly prepared stock solutions were used for all the experiments.
|
study
| 100.0 |
To monitor the effect of surface on aggregation of Aβ42, Aβ(14-23) and α-Syn, 1-(3-aminopropyl) silatrane (APS)-functionalized mica surfaces were used. APS-functionalized mica surfaces were prepared by incubating freshly cleaved mica into 167 μM of APS solution for 30 min, then rinsed with deionized water and dried properly with Argon stream37. The small pieces of surfaces were put inside the eppendorf tube (low protein binding tubes) containing 100 nM of protein solution [Aβ42 and Aβ(14-23)] in 10 mM sodium phosphate buffer, pH 7.0. The substrates were taken out at desired time periods and rinsed thoroughly with deionized water and dried under Argon flow for imaging in ambient condition. To compare the on-surface aggregation propensity with that of in bulk solution, a tube containing the same stock of 100 nM protein was incubated along with the tubes with surfaces. 5 μl solution was taken out from this tube and deposited onto the APS-mica at the same time periods to that of surfaces to monitor the aggregation in bulk solution. For α-Syn, a similar method as described above was adopted. The protein concentration used was 10 nM.
|
study
| 100.0 |
AFM images were acquired in tapping mode by Multimode Nanoscope III system (Bruker, Santa Barbara, CA) and MFP-3D AFM (Asylum Research, CA) using TESPA and MSNL (Bruker, CA) probes in air and in buffer medium respectively. The nominal spring constant for TESPA and MSNL probes were ~42 N/m and ~0.1 N/m, respectively. AFM topographic images were analyzed using Femtoscan Online software (Advanced Technologies Center, Moscow, Russia). The volume of the aggregates was obtained from the ‘Enum feature’ and ‘grain analysis’ tool in the software. The data obtained have been plotted in Origin software (OriginPro 2016) to generate the histograms and then they were fitted with Gaussian distribution to measure the mean value and errors38.
|
study
| 99.94 |
Time-lapse imaging was carried out in Asylum MFP3D instrument using the MSNL probe (Bruker Corporation). First the functionalized mica surface was imaged in 10 mM sodium phosphate buffer (pH 7.0) to obtain a clean surface, then the desired amount of protein solution was added (100 nM) and imaging was carried out. On-surface aggregation was monitored by scanning the surface at one hour time intervals. The results shown in Supplementary Fig. 6B have been obtained using Nanoscope Multimode VIII system (Bruker, Santa Barbara, CA) in Peakforce mode. A small droplet of 10 nM alpha-synuclein was created under the fluid cell, then the cell was inserted into scanner head and imaging was being continued after proper alignment of the laser. The imaging was continued in the same area to monitor the changes in morphology of the aggregates.
|
study
| 100.0 |
Molecular dynamics (MD) simulations were conducted using NAMD v 2.10 and employing CHARMM27 force field39, extended with INTERFACE FF v1.5 (INTFF) parameters for mica40, and the TIP3P water model41. A single layer of mica, spanning 52 × 54 Å, was constructed using the INTFF provided structures. Two monomers of Aβ(14-23) were then placed at Center-of-Mass (COM) distance of 2 nm above the mica surface. The initial monomer structure (Figure S9) was adopted from ref. 23. To mimic the experimental design, a Cys residue was added to the N-terminus of the peptide. The index of this Cys residue was set to 0 to keep the context of the other residues as the actual Aβ42 protein. Because we do not know the behavior of the cations on the mica surface, we performed simulations of two different mica surfaces: one, which allowed the K cations to freely move during the simulation, called Mica1, and another system where the K cations were restrained to their crystal positions as obtained from42, called Mica2. Both systems were then solvated with TIP3P water molecules. Na+ and Cl− ions were then added to neutralize the charges and maintain an ionic concentration of 150 mM. Other details of the simulations setup were adopted from our previous work23. 20 ns NVT simulation was then performed. After which 520 ns NPT production simulation, at 1 bar and 300 K, were carried out for each system using Crane at the Holland Computing Center (HCC) and Comet at San Diego Supercomputer Center (SDSC).
|
study
| 100.0 |
Similarly, an equilibrated bilayer containing 128 DLPE (1,2- didodecanoyl-sn-glycero-3-phosphoethanolamine) molecules was obtained from http://www.fos.su.se/~sasha/SLipids and used together with Slipids force field parameters43 and four Aβ(14-23) monomers, two on each side of the lipid bilayer. The Aβ(14-23) molecules were placed 2 nm COM above the lipid head groups. The rest of the simulation parameters, steps, and duration were the same as the mica simulations.
|
study
| 100.0 |
For analysis, the first 20 ns of the NPT simulation was discarded. The interaction between peptides was examined using the COM distances between each of the peptides. Likewise, the minimum distance between the peptide and the mica layer was also calculated using the COM of each peptide and the Si atoms of the mica surface, this was done using g_distance. Similarly for the DLPE system, the distances were calculated with respect to the PO4 groups. Additionally, the backbone interactions of each of the monomers were also monitored using g_mindist. To follow the frequency of interaction of each of the peptides as they interact with the surface, the number of contacts between peptide backbone and the surface were monitored; contact being defined as distances less than 1 nm.
|
study
| 100.0 |
Over time, farms have evolved towards factory production units. This has led to a decline of the welfare of animals that becomes an important concern for consumers . Moreover, this type of farming has led to the selection of animals with high production traits such as rapid growth, lean meat, or large litters. However, the strong selection focus on these characteristics is suspected to reduce functional traits, such as viability of the newborns or disease resistance. Consequently, the genetic potential of animals is usually not fully expressed in commercial conditions, due to the limiting influence of the environment. Robustness is a specific quality of an individual to express a high production potential in a wide variety of environmental conditions and is now a major specific breeding goal in the context of sustainable farm animal breeding. Various strategies are available to increase robustness, and we have suggested that the reinforcement of the neuroendocrine stress responses may favour the processes of adaptation and dampen the negative consequences of the environment . The hypothalamic-pituitary-adrenocortical (HPA) axis is the main neuroendocrine system involved in adaptation to stress and is strongly influenced by genetic factors . It is therefore a primary candidate for the selection of more robust animals .
|
study
| 99.56 |
In modern intensive livestock production, pigs are easily threatened by different types of inflammation. Immunological stress is a comprehensive process involving immunological, neurological, and endocrinological responses . The reciprocal “subjugation” of the brain and the immune system via cytokines and stress hormones is now well demonstrated [5, 6]. The resulting balance has more recently been demonstrated at the level of blood cell transcriptome , with chronic stress increasing the expression of genes regulated by inflammatory mediators and decreasing those regulated by glucocorticoid hormones . This approach has been used to evaluate the negative consequences of adverse environmental conditions, mostly in humans but also in farm animals (horses ). More recently, individual differences have also been described as related to personality dimensions in humans . However the relationships with individual variations of HPA axis activity, including genetic factors, is still unexplored.
|
review
| 99.7 |
We have previously shown large variations in biological and transcriptomic responses to an ACTH stimulation test . Indeed, the adrenal response to ACTH is a major source of variability of HPA axis function . The present study aims at describing blood transcriptomic, hormonal, and metabolic responses of pigs to a systemic challenge using lipopolysaccharide (LPS), a major component of the outer membrane in gram-negative bacteria . LPS provokes an acute inflammatory syndrome resulting eventually in all kinds of pathophysiological damages . The objective is to analyze the individual variation of the biological responses in relation to the activity of the HPA axis. This was assessed through the level of cortisol released by LPS (this experiment) and also, in the same animals, through the level of cortisol released after an ACTH stimulation test (in an experiment previously published ).
|
study
| 100.0 |
The same 120 piglets (63 females and 57 males) as described in were used for this study. In addition to the ACTH stimulation test, previously described, each animal was injected in the neck muscles with LPS at 8 weeks (E. coli serotype 055:B5, Sigma-Aldrich, Saint Quentin Fallavier, FR) at a dose of 15 μg/kg body weight. Injections occurred from 10:00-11:00 AM to avoid nycthemeral variations. Blood samples were collected before the injection (t=0) and 1 h (t=+1), 4 h (t=+4) and 24 h (t=+24) after injection with the same protocol as described in .
|
study
| 100.0 |
Cortisol, glucose, free fatty acid (FFA), blood cell counts (including: white cells count, proportion of lymphocytes, monocytes and granulocytes, red cells count, hematocrit, concentration of hemoglobin, red cells width and volume, platelets count and platelets width and volume) were obtained using the same protocol as in the previous study. Fifteen biological variables were measured on the 120 pigs in addition to birth and weaning weights. All these variables were preprocessed for outlier and missing value correction and to ensure normality as in the previous study.
|
study
| 100.0 |
Gene expression analysis was performed at the GeT-TRiX facility (GénoToul, Génopole Toulouse Midi-Pyrénées) using Agilent SurePrint G3 porcine microarray GPL16524 (Agilent, 8 ×60 K) following the manufacturer’s instructions (Agilent Technologies, Santa Clara, California). For each of 120 samples, Cyanine-3 (Cy3) labeled cRNA was prepared from 200 ng of total RNA using the One-Color Quick Amp Labeling kit (Agilent) according to the manufacturer’s instructions, followed by Agencourt RNAClean XP (Agencourt Bioscience Corporation, Beverly, Massachusetts). 600 ng of Cy3-labelled cRNA were hybridized on the microarray slides following the manufacturer’s instructions. Immediately after washing, the slides were scanned on Agilent G2505C Microarray Scanner using Agilent Scan Control A.8.5.1 software and fluorescence signal extracted using Agilent Feature Extraction software v10.10.1.1 with default parameters (grid 037880_D_F_20120213 and protocol GE1_1010_Sep10).
|
study
| 99.94 |
Blood samples of 2 pigs at one time step each were of poor quality and thus not used. The same experimental design than the one described in was used to secure the kinetics of the response for each individual and to prevent confounding effects between batch and array. After quality control and filtering, 27,837 probes were kept and log2 transformed. Technical biases and missing data were handled similarly than in the previous study.
|
study
| 100.0 |
For validation of array data by Fluidigm technology 22 animals (88 samples) were kept to fit the technical constraints of this technique. Total RNA (1 μg) used in microarray experiments was reverse-transcribed as previously described . Primer sequences for genes were designed using Primer3plus software (http://primer3plus.com) and are given in Additional file 1. The TFRC gene (transferrin receptor) and EPRS gene (glutamyl-prolyl-tRNA synthetase) were used as internal controls. Pre-amplified samples were analyzed with a 96 ×96 Dynamic Array™ IFC (Fluidigm) following the protocol defined by , with some modifications. All measurements were performed on the same plate. Each gene was tested twice for each sample. Four dilution points containing a pool of all samples were used to determine PCR efficiency. Data were analyzed using BioMark Gene Expression Data Analysis software (Fluidigm) to obtain Ct values. The Pfaffl method was applied to compute the relative expression of each gene . Pearson correlations were computed to compare the expression values of microarray and quantitative real-time PCR. Quantitative RT-PCR data were also analyzed for time effect by ANOVA with repeated measurements for every gene.
|
study
| 100.0 |
All analyses were performed with the R software, version 3.2.2 . They were designed so as to address two main questions: the first one is the study of the evolution over time of the different variables (plasma metabolites, cortisol and gene expression) after LPS injection. The second one is the study of the relation between the different variables and one of the most relevant measure of sensitivity of the adrenals, the cortisol level one hour after ACTH injection (data from , obtained on the same animals).
|
study
| 100.0 |
Longitudinal data analysis of the evolution over time of the different variables can be performed using different statistical methods. A very common approach is to fit curves (for instance splines as in [19, 20]) a as prior processing to the statistical analysis. However, four time steps are too few number of time points to obtain an accurate fit. The analysis was thus performed using two main approaches: the first one relies on linear models with the time as a factor covariate and the second one is based on a decomposition of sources of variations, as was already proven successful for repeated measurements analysis in and for longitudinal data analysis in .
|
study
| 100.0 |
First, all variables were subjected to a one-way ANOVA with repeated measures and the time step as a factor covariate. In order to control the false discovery rate (FDR) , p-values were adjusted using a Benjamini-Hochberg (BH) approach (Table 1). Variables with an adjusted p-value (FDR <0.05) were then subjected to 3 paired t-tests to assess the difference between t=0 and t=+1, between t=0 and t=+4 and between t=0 and t=+24. The full list of p-values was adjusted using a BH approach (Fig. 1). Fig. 1Mean evolution of the biological variables over time. Vertical bars correspond to + and - SEM at each point Table 1Reference values (at t=0) for the biological variables, birth weight and weaning weight (n=120)UnitsMinMaxMeanSEM F Tympanic temperature°C36.10040.25739.1680.050258.110White cellslog10(G/l)0.4911.4721.1880.011572.970Lymphocytes%46.60091.90067.4770.555112.180Monocytes%3.90016.2008.5570.19169.500Granulocytes%2.50035.60022.6080.52778.210 L/G ratio1.35536.7603.4660.29564.650Red cellsT/l1.4907.3305.1630.05469.420Mean corpuscular volumefl39.70063.70052.0080.33362.120Hematocrit%6.80037.40026.8280.29978.990Hemoglobing/dl6.90012.8008.9470.09246.680Red blood cells distribution widthfl29.10033.80032.0290.08174.440Plateletslog10(G/l)2.3302.9982.6670.011227.400Mean platelet volumefl7.60013.0009.6820.10271.210Platelet distribution width%9.60012.00010.7710.045122.790Cortisollog10(ng/ml)1.0412.0331.4750.017370.240Free fatty acids \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$\sqrt {(\text {mmol/l})}$\end{document}(mmol/l) 0.0790.5600.1620.005111.040Glucosemmol/l5.8509.5258.0350.061123.990Bilirubin μmol/l4.66013.0008.5230.190178.610Birth weightkg0.4002.6801.4920.033Weaning weightkg5.46016.5649.4860.174Results of the ANOVA for time effect (F value): all variables varied significatively during the experiment, with an FDR <10−12, except for the weights (not tested because constant)
|
study
| 100.0 |
Cortisol levels measured one hour after ACTH injection are the most relevant measure to assess the sensitivity of the adrenals to ACTH (data from ). Hence, correlations between biological variables at t∈{0,+1,+4,+24} and the level of cortisol one hour after ACTH injection were investigated using t-tests of the linear regression on ACTH level. p-values were adjusted using a BH approach.
|
study
| 100.0 |
The whole blood is composed of different types of white cells with distinct roles which express different kinds of transcripts . It is thus likely that a modification in blood cell composition may influence the gene expression level without having cells actually express transcripts differently. As blood cell composition was found to vary over time after LPS injection, we used the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$\frac {\text {Lymphocyte}}{\text {Granulocyte}}$\end{document}LymphocyteGranulocyte (L/G) ratio as a covariate in our analyses.
|
study
| 100.0 |
Three different approaches were used to identify relevant probes. The first two are longitudinal analyses aiming, respectively, at identifying probes with an expression significantly varying from their basal levels after LPS injection and probes with a varying contribution of the L/G ratio to their expressions after LPS injection. The last analysis searched for probes correlated to the level of cortisol one hour after ACTH injection.
|
study
| 100.0 |
Firstly, we identified probes differentially expressed at each time step while taking blood cell composition into account. Blood cell composition was measured by the L/G ratio. Three models (one for each time step t ′ where t ′∈{+1,+4,+24}) were fitted to each probe using observations at t=0 and t=t ′. 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document} $$ \text{expr}_{{it}} = \mu_{0} + \tau_{t'}\mathbb{I}_{\{t = t'\}} + \beta^{t'} L/G_{{it}} + \epsilon_{{it}} $$ \end{document}exprit=μ0+τt′𝕀{t=t′}+βt′L/Git+εit
|
study
| 100.0 |
with i=1,…n is animal i. exprit is the expression of the probe being studied for animal i at time step t (t∈{0,t ′}), μ 0 is the specific contribution of time step t=0, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$\phantom {\dot {i}\!}\tau _{t^{\prime }}$\end{document}τt′ is the effect of time step t ′, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$\phantom {\dot {i}\!}\beta ^{t^{\prime }}$\end{document}βt′ is the effect of L/G ratio in this model and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$\epsilon _{it} \sim N\left (0, \sigma _{e}^{2}\right)$\end{document}εit∼N0,σe2 is an error term.
|
study
| 99.9 |
We then tested the contribution of time step t ′ against the null hypothesis H 0: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$\phantom {\dot {i}\!}\tau _{t^{\prime }} = 0$\end{document}τt′=0. The full list of p-values was globally adjusted using a Bonferroni approach. As the Bonferroni approach exerts a more stringent control than the BH approach, it was used to obtain a narrowed list of the most significant probes. Probes with at least one adjusted p-value <0.01 were probes for which the expression adjusted by the L/G ratio was significantly different from the basal level. In the sequel, this list of genes will be referred to as (List1).
|
study
| 100.0 |
Secondly, we identified probes for which the L/G ratio effect is different according to the time step. To that aim, we compared a complete model, including all time step contributions and the L/G ratio effect according to the time step (Eq. (2)): 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document} $$ \text{expr}_{{it}} = \tau_{t} + \beta_{t} L/G_{{it}} + \epsilon_{{it}} $$ \end{document}exprit=τt+βtL/Git+εit
|
study
| 100.0 |
(with t∈{0,1,4,24} and β t is the interaction effect between time step t and the L/G ratio of individual i at time step t), to a reduced model, including only the average L/G ratio and all time step contributions (Eq. (3)): 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document} $$ \text{expr}_{{it}} = \tau_{t} + \beta L/G_{{it}} + \epsilon_{{it}}. $$ \end{document}exprit=τt+βL/Git+εit.
|
study
| 100.0 |
An F-test was then performed to test the null hypothesis, H 0: β 0=β 1=β 4=β 24, against the alternate hypothesis, H 1: ∃t 1,t 2 such as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$\beta _{t_{1}} \neq \beta _{t_{2}}$\end{document}βt1≠βt2. Multiple testing was handled by applying a BH approach (FDR <0.05). Probes for which the test was significant were probes for which the effect of L/G varied over time. In the sequel, this list of genes will be referred to as (List2).
|
study
| 100.0 |
Finally, we studied correlations between all probes and cortisol level when it reaches its peak in blood after LPS injection. Thus, Pearson correlations, ρ, were computed between DEP expression at each time step and cortisol level at t=+4. A correlation test was then performed to test the null hypothesis, H 0: ρ=0 against H 1: ρ≠0. Multiple testing was handled by using a BH approach (FDR <5%). This list of genes will be referred as (List3) in the sequel.
|
study
| 100.0 |
Enrichment analysis was performed using tools available at WEB-based GEne SeT AnaLysis Toolkit (WebGestalt) [24, 25]. Entrez gene IDs were used as unique gene identifiers. Target gene lists for main effects and interactions and a background gene set consisting of all 9530 genes were used to identify enrichment in GO, KEGG, Transcription Factor Target, Microarray Target, Protein Interaction Network Module, and Phenotype Analysis in WebGestalt using Fisher’s exact test and BH correction for multiple testing.
|
study
| 100.0 |
A pathway is an interconnected arrangement of processes, representing the functional roles of genes in the genome. The biological processes in which individual genes may participate were identified using the “Gene Ontology” database AmiGO (http://amigo.geneontology.org). The DEG were assembled into networks using Ingenuity Pathway Analysis (IPA Ⓒ) (http://www.ingenuity.com). This application includes algorithms that automatically identify the biological pathways and functions of selected genes. It is based on a large bibliographic database with various types of interaction already identified between pairs of genes. Every biological network extracted by IPA corresponds to the best possible arrangement of the genes, and are associated with a score derived from a p-value (right-tailed Fisher’s exact test, − log10-transformed).
|
study
| 100.0 |
In the case of time course analyses, the approach previously described (applying a univariate linear model on each variable followed by multiple test correction) is common. However, this approach disregards the dependencies between genes and does not allow for a global view of the relationships between the repeated measurements in high dimensional data. The multilevel approach, already proven successful to investigate the relationships between repeated measurements in was thus used and combined with multivariate data analysis methods.
|
study
| 100.0 |
The multilevel approach is inspired by the mixed-model framework and uses a split-up variation of the (n T)×p matrix X that contains the observations of p variables (clinical biology variables or gene expressions) on n animals with T=4 times of measurements: 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document} $$ {\begin{aligned} X = \underbrace{X_{..}}_{\textrm{offset term}} + \underbrace{X_{b}}_{\textrm{between-animal deviation}} + \underbrace{X_{w}}_{\textrm{within-animal deviation}} \end{aligned}} $$ \end{document}X=X..⏟offset term+Xb⏟between-animal deviation+Xw⏟within-animal deviation
|
study
| 99.94 |
Similarly to what was performed in , multivariate approaches were performed on X w to bring out the most relevant correlations between variables in the dataset, independently from individual variations. First a multilevel PCA was performed on the biological variables to study the overall effect of LPS on plasma metabolites and cortisol over time. Then, a multilevel multiple factor analysis (MFA) was used to investigate the overall relationships between clinical biology and transcriptomic data.
|
study
| 100.0 |
Tympanic temperature peaked at t=+4 (40.8 °C vs 39.1 °C) and returned to basal levels at t=+24. A decrease of total count of white blood cell count was observed, maximal at t=+4 (5.70 vs 15.35 G/l) and the mirror changes in the respective proportions of lymphocytes and granulocytes. This indicated that the lymphocytes/granulocytes ratio (L/G) was a good measure to use in order to take into account these changes that result mainly from the redistribution of lymphocytes into the tissues . The L/G ratio was maximal at t=+1 (9.32 vs 3.67) and back to basal levels at t=+4. The red blood cell count and associated measures (hematocrit and hemoglobin concentration) showed a biphasic change, with an initial increase, maximal at t=+4 (5.47 vs 5.16 T/l) and a subsequent long-lasting decrease (4.82 T/l at t=+24). The platelet count showed a steady decrease until at least t=+24 (284 vs 475 G/l). These measures were not influenced by sex, except the mean red cell volume and hematocrit that were slightly lower in males (FDR <0.05).
|
study
| 100.0 |
Cortisol levels peaked at t=+4 with a 3.83-fold increase (114.3 vs 29.8 ng/ml). Circulating glucose levels were reduced by 26.9% to 5.95 mmol/l at t=+4. The circulating concentration of free fatty acids increased from 0.026 to 0.146 mmol/l at t=+4. None of these biochemical measures was influenced by sex.
|
study
| 100.0 |
The overall effect of LPS over time was investigated with a multilevel PCA (Fig. 2). The first component of the multilevel PCA opposes the observations at t=0 (negative coordinates on this axis) to the observations at t=+4 (positive coordinates on this axis), this time step corresponding to the peak of LPS effect. The second component opposes the observations at t=+24 (positive coordinates on this axis) to the other observation times (negative coordinates on this axis). The representation of the variables shows that the first axis is mainly driven by an opposition between free fatty acids (FFA), bilirubin, temperature and cortisol (high measures at t=+4), and white cell count and glucose (low measures at t=+4). The second axis of the PCA is driven by L/G ratio and platelet count that are high at t=+1. Fig. 2Multilevel PCA on the biological variables responding to LPS. Colors represent the time step; Black: t=0; Red: t=+1; Green: t=+4; Blue: t=+24; a: Projection of the individuals on dimensions 1 - 2; b: Projection of the variables on dimensions 1 - 2
|
study
| 100.0 |
In our study, we used a comprehensive gene expression profiling by means of microarray analysis to identify clusters of genes differentially expressed in peripheral blood cells, taking into consideration the kinetic of the response with 4 time points (t∈{0,+1,+4,+24}). LPS induces dramatic changes in blood cell number and lymphocyte/granulocyte (L/G) ratio that introduces a confusion between time and cell type effects, and a major challenge for the interpretation of transcriptomic data. Therefore we based the interpretation of the results on three different lists of genes, (List1), (List2), and (List3).
|
study
| 100.0 |
The first list of genes (List1) consists of 9530 unique genes (22,794 transcripts, Additional files 2 and 3) for which the expression adjusted by the L/G ratio was significantly different from basal level. (List1) was submitted to gene ontology and enrichment analysis. These analyses showed 106 classes significant at FDR < 0.05. Due to the important number of DEG, generic classes were removed (such as morphogenesis, transcription, locomotion and others). Only genes that were well-known and well described in the literature were chosen to define a final selected list of 284 genes. These genes were grouped into 6 functional classes that were all found enriched for genes (List1) (Immunity and Inflammation, Chemotaxis, Apoptosis, Calcium ion transport, Metabolism, Hormonal Response).
|
study
| 100.0 |
The “immunity and inflammation” class (175 genes) is related to the inflammatory cascade after activation of leukocytes by LPS via TLR4 receptor (a receptor for bacterial lipopolysaccharide). TLR4 is a critical driver of immune responses to bacterial infections. Signals from TLR4 promote NF- κB and AP-1 activation, leading to inflammatory gene expression (DEG for TLR4, TNF, JUNB, and NF-B pathway).
|
study
| 100.0 |
The “chemotaxis” class is composed of 59 genes. Among them ABHD2, ACADS, AIF1, ANXA7, ARPC1A, ARPC2, CD97, CHL1, CLIC1, CNTFR, COQ3, DGKD, DNASE2, GP1BA, GPI, HCLS1, HPS6, IL1RN, IL8RA, KAT5, LOC100523056, LSP1, MAN2B1, PARK7, PTPN6, SMAD7, SPG21, TMEM173, TMSB10, TMSB4X, TRDMT1, and TSPO genes are related to immune cell trafficking. This observation is in agreement with the observed blood cell redistribution.
|
study
| 99.94 |
The “apoptosis” class (33 genes) includes C5AR1, CCL24, CCR1, CCR3, CXCL13, IRG1, ALDOC, C3AR1, CADM1, CAPN3, HEXA, ID3, MAEA, PLAU, PRDX5, PROC, and CXCR2 genes related to apoptosis and inflammatory response, and TNFSF13B and NFKBIA involved in cell-activating factor signalling pathway.
|
study
| 98.75 |
The 284 remaining genes from the first list (List1) were clustered into 4 clusters using HAC (Fig. 3, Additional file 3). Fig. 3Black: Average evolution the genes in each of the clusters identified by the HAC on the 284 DEG identified in list (List1). Evolution of each gene is translated so that it is equal to 0 at t=0; Red: Average evolution over all genes in the cluster (cluster 1: 8 genes, cluster 2: 12 genes, cluster 3: 159 genes, cluster 4: 77 genes). 28 genes were unclassified
|
study
| 100.0 |
Black: Average evolution the genes in each of the clusters identified by the HAC on the 284 DEG identified in list (List1). Evolution of each gene is translated so that it is equal to 0 at t=0; Red: Average evolution over all genes in the cluster (cluster 1: 8 genes, cluster 2: 12 genes, cluster 3: 159 genes, cluster 4: 77 genes). 28 genes were unclassified
|
study
| 99.94 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.