text
string | predicted_class
string | confidence
float16 |
|---|---|---|
Effect of hUCBSC-MO treatment on cavity formation after SCI. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean of the cavity area ± SEM (n = 5–7) and analyzed by one-way ANOVA followed by post hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. ##p < .01, and ###p < .001 versus spinal cord injury. Φ shows significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $ shows significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance.
|
study
| 100.0 |
Statistical evaluations showed the significant differences between SCI-MO, SCI-hUCBSC, and SCI-MO-hUCBSC groups when compared with SCI group based on the number of ventral horn lower motor neurons, F(5, 30) = 30.86, p < .001. Application of post hoc Bonferroni’s multiple comparisons test as well as Bartlett’s test for equal variances revealed significant increase in the number of ventral horn motor neurons in SCI-MO (p < .05), SCI-hUCBSC, and SCI-MO-hUCBSC (p < .001) treatment groups when compared with SCI group. Moreover, application of one-way ANOVA showed significant difference in SCI-MO-hUCBSC group when compared with SCI-MO (p < .001) and SCI-hUCBSC (p < .05) groups (Figures 7 and 8). Figure 7.Effect of hUCBSC-MO treatment on cell loss in ventral horn of spinal cord after injury. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean number of ventral horn motor neurons ± SEM (n = 5–7) and analyzed by one-way ANOVA followed by post hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. #p < .05 and ###p < .001 versus spinal cord injury. Φ show significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $$$ shows significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance. Figure 8.Transverse section of spinal cord showing the lower motor neurons in ventral horn of spinal cord at the level of T12-L1 of all groups which were evaluated in this study on Day 56. H&E staining showing shrinkage and reduction in ventral horn motor neurons in SCI group in comparison with SCI-MO, SCI-hUCBSC, and SCI-hUCBSC-MO groups. Black arrows illustrate the ventral horn motor neurons (400×). a = SCI, b = SCI-hUCBSC, c = SCI-MO, d = SCI-MO-hUCBSC. Note. H&E = hematoxylin and eosin; hUCBSC = human umbilical cord blood stem cell; MO = Melissa officinalis; SCI = spinal cord injury.
|
study
| 100.0 |
Effect of hUCBSC-MO treatment on cell loss in ventral horn of spinal cord after injury. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean number of ventral horn motor neurons ± SEM (n = 5–7) and analyzed by one-way ANOVA followed by post hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. #p < .05 and ###p < .001 versus spinal cord injury. Φ show significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $$$ shows significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance.
|
study
| 100.0 |
Transverse section of spinal cord showing the lower motor neurons in ventral horn of spinal cord at the level of T12-L1 of all groups which were evaluated in this study on Day 56. H&E staining showing shrinkage and reduction in ventral horn motor neurons in SCI group in comparison with SCI-MO, SCI-hUCBSC, and SCI-hUCBSC-MO groups. Black arrows illustrate the ventral horn motor neurons (400×). a = SCI, b = SCI-hUCBSC, c = SCI-MO, d = SCI-MO-hUCBSC. Note. H&E = hematoxylin and eosin; hUCBSC = human umbilical cord blood stem cell; MO = Melissa officinalis; SCI = spinal cord injury.
|
study
| 100.0 |
Statistical evaluations revealed that the number of GFAP+ astrocytes was significantly increased in SCI group. However, this activation was significantly attenuated in the treatment groups, F(5, 30) = 45.49, p < .001. Application of post hoc Bonferroni’s multiple comparisons test as well as Bartlett’s test for equal variances showed a significant reduction in the GFAP expression in SCI-MO, SCI-hUCBSC (p < .01), and SCI-MO-hUCBSC (p < .001) treatment groups when compared with SCI group. Statistically significant difference was found in GFAP expression between SCI-MO-hUCBSC group in comparison with SCI-MO (p < .01) and SCI-hUCBSC (p < .05) groups (Figures 9 and 10). Figure 9.Effect of hUCBSC-MO treatment on astrogliosis formation in ventral horn of spinal cord after injury. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean of GFAP-positive astrocytes ± SEM (n = 5–7) and analyzed by one-way ANOVA followed by post hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. ##p < .01, and ###p < .001 versus spinal cord injury. Φ shows significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $$ shows significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; GFAP = glial fibrillary acidic protein; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance. Figure 10.Transverse section of spinal cord showing the ventral horn gray matter of spinal cord at the level of T12-L1 of all groups which were evaluated in this study on Day 56. Black arrows illustrate the GFAP astrocytes. Reduced GFAP astrocytes are evident. a = Intact, b = SCI, c = SCI-MO, d = SCI-hUCBSC, e = SCI-MO-hUCBSC. Bar = 50 µm. (ECLIPSE 5Oi microscope). Note. GFAP = glial fibrillary acidic protein; hUCBSC = human umbilical cord blood stem cell; MO = Melissa officinalis; SCI = spinal cord injury.
|
study
| 100.0 |
Effect of hUCBSC-MO treatment on astrogliosis formation in ventral horn of spinal cord after injury. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean of GFAP-positive astrocytes ± SEM (n = 5–7) and analyzed by one-way ANOVA followed by post hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. ##p < .01, and ###p < .001 versus spinal cord injury. Φ shows significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $$ shows significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; GFAP = glial fibrillary acidic protein; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance.
|
study
| 100.0 |
Transverse section of spinal cord showing the ventral horn gray matter of spinal cord at the level of T12-L1 of all groups which were evaluated in this study on Day 56. Black arrows illustrate the GFAP astrocytes. Reduced GFAP astrocytes are evident. a = Intact, b = SCI, c = SCI-MO, d = SCI-hUCBSC, e = SCI-MO-hUCBSC. Bar = 50 µm. (ECLIPSE 5Oi microscope). Note. GFAP = glial fibrillary acidic protein; hUCBSC = human umbilical cord blood stem cell; MO = Melissa officinalis; SCI = spinal cord injury.
|
study
| 100.0 |
Data obtained revealed that the density of astrogliosis in the ventral horn of spinal cord was significantly reduced in treatment groups in comparison with SCI group, F(5, 30) = 17.66, p < .001. Moreover, post hoc Bonferroni’s multiple comparison test showed that the density of gliosis was significantly reduced in SCI-hUCBSC (p < .01) and SCI-MO-hUCBSC (p < .001) groups when compared with SCI group. Statistical evaluations showed significant differences between SCI-MO-hUCBSC, SCI-MO (p < .001), and SCI-hUCBSC (p < .05) groups (Figure 11). Figure 11.Effect of hUCBSC-MO treatment on density of astrogliosis in ventral horn of spinal cord after injury. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean of gliosis density ± SEM (n = 5–7) and analyzed by one-way ANOVA followed by post hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. #p < .05, ##p < .01, and ###p < .001 versus spinal cord injury. Φ shows significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $$ shows significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance.
|
study
| 100.0 |
Effect of hUCBSC-MO treatment on density of astrogliosis in ventral horn of spinal cord after injury. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean of gliosis density ± SEM (n = 5–7) and analyzed by one-way ANOVA followed by post hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. #p < .05, ##p < .01, and ###p < .001 versus spinal cord injury. Φ shows significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $$ shows significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance.
|
study
| 100.0 |
Application of one-way ANOVA demonstrated that density of myelin in the dorsal white matter of spinal cord was significantly increased in treatment groups when compared with SCI group, F(5, 30) = 62.11, p < .001. Furthermore, post hoc Bonferroni’s multiple comparison test revealed that the density of myelin was significantly increased in SCI-MO (p < .05), SCI-hUCBSC, and SCI-MO-hUCBSC (p < .001) groups compared with SCI group. Furthermore, statistical evaluations showed significant differences between SCI-MO-hUCBSC, SCI-MO (p < .001), and SCI-hUCBSC (p < .05) groups (Figures 12 and 13). Figure 12.Effect of hUCBSC-MO treatment on density of myelin in dorsal white matter of spinal cord after injury. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean of myelin density ± SEM (n = 5–7) and analyzed by one-way ANOVA followed by post hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. #p < .05 and ###p < .001 versus spinal cord injury. Φ shows significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $$ shows significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance. Figure 13.Ultra structural characteristics of myelination in dorsal white matter of spinal cord at the level of T12-L1 of all groups which were evaluated in this study on Day 56. (b) Low power view reveals the distribution of myelinated axons. Three representative high power photographs show the typical appearance of myelinated axons with extensive myelin sheath wrapped around an axon (a, d, and e). Densitometry of MBP in dorsal white matter of spinal cord at the level of T12-L1 is shown in the left part of any electron microscopy pictures. a = Intact, b = SCI, c = SCI-MO, d = SCI-hUCBSC, e = SCI-MO-hUCBSC. Note. MBP = myelin basic protein; hUCBSC = human umbilical cord blood stem cell; MO = Melissa officinalis; SCI = spinal cord injury.
|
study
| 100.0 |
Effect of hUCBSC-MO treatment on density of myelin in dorsal white matter of spinal cord after injury. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean of myelin density ± SEM (n = 5–7) and analyzed by one-way ANOVA followed by post hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. #p < .05 and ###p < .001 versus spinal cord injury. Φ shows significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $$ shows significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance.
|
study
| 100.0 |
Ultra structural characteristics of myelination in dorsal white matter of spinal cord at the level of T12-L1 of all groups which were evaluated in this study on Day 56. (b) Low power view reveals the distribution of myelinated axons. Three representative high power photographs show the typical appearance of myelinated axons with extensive myelin sheath wrapped around an axon (a, d, and e). Densitometry of MBP in dorsal white matter of spinal cord at the level of T12-L1 is shown in the left part of any electron microscopy pictures. a = Intact, b = SCI, c = SCI-MO, d = SCI-hUCBSC, e = SCI-MO-hUCBSC. Note. MBP = myelin basic protein; hUCBSC = human umbilical cord blood stem cell; MO = Melissa officinalis; SCI = spinal cord injury.
|
study
| 100.0 |
Conversely, evaluation of electron microscopic pictures from all groups by application of one-way ANOVA revealed that MI was reduced in treatment groups, F(5, 6) = 102.1, p < .001. Moreover, post hoc Bonferroni’s multiple comparison test showed that MI was significantly reduced in SCI-MO, SCI-hUCBSC (p < .01), and SCI-MO-hUCBSC (p < .001) groups than in SCI group. Statistical evaluations showed significant differences between SCI-MO-hUCBSC, SCI-MO, and SCI-hUCBSC (p < .05) groups (Figures 13 and 14). Figure 14.The effect of hUCBSC-MO treatment in reducing of myelin index in dorsal white matter of the spinal cord after injury. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean of myelin index ± SEM, (n = 2) and analyzed by one-way ANOVA followed by post-hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. ##p < .01, and ###p < .001 versus spinal cord injury. Φ shows significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $ shows significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance.
|
study
| 100.0 |
The effect of hUCBSC-MO treatment in reducing of myelin index in dorsal white matter of the spinal cord after injury. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean of myelin index ± SEM, (n = 2) and analyzed by one-way ANOVA followed by post-hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. ##p < .01, and ###p < .001 versus spinal cord injury. Φ shows significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $ shows significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance.
|
study
| 100.0 |
For further confirmation of myelination process and the synthesis of MBP by MO-hUCBSC treatment after SCI, RT-PCR analysis was used. There was a change in the mRNA levels after SCI was determined utilizing standardized RT-PCR analysis. Qualitative analysis of RT-PCR findings in all groups revealed considerable upregulation of mRNA gene of MBP in SCI-MO-hUCBSC treated group when compared with SCI group. Application of one-way ANOVA revealed that the density of RT-PCR bands was increased in treatment groups, F(5, 6) = 1077, p < .001. Moreover, post hoc Bonferroni’s multiple comparison test revealed that the density of RT-PCR bands was significantly increased in SCI-MO (p < .01), SCI-hUCBSC. and SCI-MO-hUCBSC (p < .001) groups than in SCI group. Application of one-way ANOVA showed a significant difference between SCI-MO-hUCBSC, SCI-MO, and SCI-hUCBSC groups (p < .05; Figures 15 and 16). Figure 15.The effect of hUCBSC-MO treatment in upregulation of myelin basic protein in the spinal cord after injury. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean of RT-PCR bands density ± SEM (n = 2) and analyzed by one-way ANOVA followed by post hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. #p < .05, ##p < .01, and ###p < .001 versus spinal cord injury. Φ shows significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $, $$, and $$$ show significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance. Figure 16.Expression of myelin basic protein in injured and treated spinal cords of rats. RT-PCR analysis of myelin basic proteins depicting SCI, SCI-MO, SCI-hUCBSC, and SCI-MO-hUCBSC groups. Housekeeping gene GAPDH was utilized as loading control. Note. RT-PCR = reverse transcription-polymerase chain reaction; hUCBSC = human umbilical cord blood stem cell; MO = Melissa officinalis; SCI = spinal cord injury; GADPH = glyceraldehyde 3-phosphate dehydrogenase.
|
study
| 100.0 |
The effect of hUCBSC-MO treatment in upregulation of myelin basic protein in the spinal cord after injury. Intraperitoneal injection of MO (150 mg/kg) was started one day after injury and continued once a day for 14 days after injury. Intraspinal grafting of hUCBSCs was started 24 hr after injury. Data are represented as mean of RT-PCR bands density ± SEM (n = 2) and analyzed by one-way ANOVA followed by post hoc Bonferroni’s multiple comparison test. ***p < .001 shows significant difference between SCI versus intact. #p < .05, ##p < .01, and ###p < .001 versus spinal cord injury. Φ shows significant difference between SCI-MO-hUCBSC and SCI-hUCBSC (p < .05). $, $$, and $$$ show significant difference between SCI-MO-hUCBSC and SCI-MO (p < .05, p < .01, and p < .001, respectively). Note. hUCBSCs = human umbilical cord blood stem cells; MO = Melissa officinalis; SCI = spinal cord injury; SEM = standard error of the mean; ANOVA = analysis of variance.
|
study
| 100.0 |
Expression of myelin basic protein in injured and treated spinal cords of rats. RT-PCR analysis of myelin basic proteins depicting SCI, SCI-MO, SCI-hUCBSC, and SCI-MO-hUCBSC groups. Housekeeping gene GAPDH was utilized as loading control. Note. RT-PCR = reverse transcription-polymerase chain reaction; hUCBSC = human umbilical cord blood stem cell; MO = Melissa officinalis; SCI = spinal cord injury; GADPH = glyceraldehyde 3-phosphate dehydrogenase.
|
study
| 99.94 |
Although some research work have indicated that stem cell transplantation for treatment of SCI is unsuccessful (Růžička et al., 2013), many investigations have demonstrated that stem cells are effective in SCI (Veeravalli et al., 2009; Caron et al., 2016; Satti et al., 2016). As a result, there has been controversy about this issue.
|
review
| 99.9 |
This research was hinged on the promotion of the therapeutic properties of hUCBSCs in combination with a neuroprotective agent to induce curative effects to SCI in rats. The present study demonstrated that combination of MO extract and hUCBSCs transplantation has neuroprotective properties in treatment of SCI. Although this combination promoted the motor, sensory, and EMG functions, there were no significant differences between SCI-MO, SCI-hUCBSCs, and SCI-MO-hUCBSC groups. A number of previous investigations have revealed functional improvement after transplantation of hUCBSCs in SCI (Saporta et al., 2003; Kuh et al., 2005; Dasari, Spomar, et al., 2008; Rodrigues et al., 2012). Previous research has shown that administration of MO extract itself improved neurological and cellular outcomes in rat SCI (Hosseini et al., 2015). MO has neuroprotective and neurotrophic effects, including promotion of functional recovery, thereby suggesting that it has therapeutic effect on neurodegenerative diseases (Bayat et al., 2012; Sepand et al., 2013). MO has acetylcholinesterase inhibitory properties (Dastmalchi et al., 2009). Anticholinesterases increase the residence time of acetylcholine in the synapse. This allows rebinding of the transmitter to nicotinic receptors. It thus gives acetylcholine the competitive advantage over the neuromuscular blocking agent (Nair and Hunter, 2004). Our results revealed that the combination of hUCBSCs and MO prevented cell loss, formation of cavity, and astrogliosis in ventral horn and also enhances the myelination in the dorsal white matter of spinal cord after injury. The effect of hUCBSCs and MO on enhancing functional recovery, myelination, and reducing astrogliosis and cavity formation is likely attributable to the inhibition of pro-inflammatory cytokines. Inflammatory processes have fundamental roles in the pathophysiology of SCI. Pro-inflammatory cytokines, including IL-1 and TNF-α, are released by activation of neurons, astrocytes, microglia, and endothelial cells after injury. Thereafter, the secondary inflammatory response and activation of IL-6 and IL-8 are induced by those cytokines (Zhu et al., 2014).
|
study
| 99.94 |
It has been shown that, transplantation of human umbilical cord blood mesenchymal stem cells (hUCB-MSCs) decreases the number of activated microglia and inhibits the permeation of immune cells and cellular apoptosis in the brain after ischemic brain injury (Yang et al., 2013; Zhu et al., 2014; Wang et al., 2016). Researchers have shown that the injection of hUCB-MSCs throughout the early stage of ischemic brain injury decreased the IL-1β, IL-6, and TNF-α expression levels in the serum and increased IL-10 expression levels. The significant increase of pro-apoptotic genes such as Bad, Bax, p53, AFAP1, caspase 3, and caspase 9 has been observed after the SCI (Sabapathy et al., 2015). IL-10 can decrease the expression of those genes, oxygen free radicals, and cytokines. Upregulation of IL-1β, IL-6, and TNF-α can initiate neuronal death and improve the synthesis of nitric oxide after injury. IL-1β can improve the intracellular calcium concentration and release neurotropic factors, which can induce neuronal apoptosis. Moreover, TNF-α can induce arachidonic acid metabolite release, which enhance extracellular accumulation of glutamate and generate neurodegenerative toxicity. Glutamate-induced cell death in the CNS contains upregulation of caspase 3 and its activation via a caspase-dependent pathway involves mitochondrial signaling. hUCBSCs decline caspase-3 and -7 activities and are responsible for activation of the Akt pathway and regulation of N-methyl-d-aspartic acid receptors, thereby giving neuroprotection to cortical neurons (Dasari, Veeravalli, et al., 2008). Therefore, hUCBSCs transplantation could provide neuroprotection by regulating the balance of pro- and anti-inflammatory cytokines.
|
review
| 90.5 |
Conversely, it has been shown that MO extracts have anti-inflammatory properties (Bounihi et al., 2013; Müzell et al., 2013). Its anti-inflammatory effects are due to rosmarinic acid, flavonoids, and terpenoids present in the extract. Probably, flavonoids have a more effective role by facilitating the synthesis of prostaglandin. MO administration can suppress the pro-inflammatory cytokines such as IL-1β, IL-6, and TNF-α (Bounihi et al., 2013).
|
study
| 97.9 |
Our results have demonstrated that the combination of MO extract administration and hUCBSCs transplantation prevented cell loss and enhanced myelination. A possible explanation to this is that, the transplantation of hUCBSCs can increase the length of neurofilament-positive fibers and increase the numbers of growth cone-like structures at the lesion site. Grafted hUCBSCs can survive, move over short distances, and produce large amounts of glial cell line-derived neurotrophic factors and neurotrophin-3 (NT3) in the host spinal cord (Rodrigues et al., 2012). It has also been confirmed that the hUCBSCs can form morphologically myelin sheaths in the spinal cord. It has been revealed that oligodendrocytes derived from human umbilical cord blood secrete NT3 and brain-derived neurotrophic factor. Cord blood stem cells promote the synthesis of MBP and proteolipid protein in the injured areas, thereby facilitating the process of remyelination (Dasari et al., 2009). Conversely, a large number of experimental evidences support oxidative stress as important mediators of secondary cell death after SCI (Cuzzocrea and Genovese, 2008; Jia et al., 2012; Fatima et al., 2015). It has been shown that administration of MO extract has powerful antioxidant effects which are probably exerted through the rosmarinic acid and the benzodioxole present in the extract. Moreover, compounds such as linoleic acid, carnosic acid, and ursolic acid are also present in the extracts, all of which have antioxidant properties.
|
study
| 99.94 |
The present study demonstrated that, the combination of MO and hUCBSCs inhibited astrogliosis. This issue has shown that the matrix metalloproteinase (MMP) as a proteolytic enzyme advances functional recovery after SCI by directing the development of a glial scar. Treatment with hUCBSCs after SCI altered the expression of different MMPs in rats. hUCBSCs transplantation in SCI causes upregulation of MMP-2 and therefore reduced the development of the glial scar at the lesion site (Veeravalli et al., 2009). In addition, as mentioned earlier, MO extract can inhibit the pro-inflammatory cytokines and reactive oxygen species. These two factors are key mediators of reactive astrogliosis in SCI (Bharne et al., 2013).
|
study
| 100.0 |
The limitation of this study is the nonmeasurement of the inflammatory factors and immunostaining for macrophage markers (F4/80 or Iba-1). Further investigations on differentiation of hUCBSCs along with the use of MO extracts will provide further evidence regarding the therapeutic effectiveness of hUCBSCs after SCI.
|
study
| 100.0 |
SCI causes motor and sensory dysfunction, tissue deformity, and cell death, the formation of astrogliosis, and degeneration of axons. In conclusion, hUCBSCs enhanced motor and sensory dysfunction as well as promoting morphological improvement in SCI contusion model in comparison with SCI. Our results showed that MO extract can promote the neuroprotective effects of hUCBSCs. Further studies are needed to clarify the underlying mechanisms of these results.
|
study
| 99.94 |
Breast cancer is the second leading cause of cancer deaths for women in the United States and a significant cause of mortality worldwide. While there are many known risk factors for breast cancer, infectious disease has emerged as one likely contributor to carcinogenesis [1, 2]. Recent studies have implicated a number of different viral infections in breast cancer, including bovine leukemia virus [3, 4], human mammary tumor virus , human papillomavirus , Epstein–Barr virus (EBV) [7–9], and human cytomegalovirus (HCMV) [10, 11]. Although there is no clear causal role for any of these viruses, a combination of molecular and epidemiological evidence suggests an association between HCMV and breast cancer.
|
review
| 99.9 |
HCMV is a β-herpesvirus that infects 70–90% of the general population, causing acute, persistent, or lifelong latent infection . HCMV infections are typically subclinical and serious disease occurs mainly in immune-compromised individuals . Overall, HCMV serostatus has not been positively correlated with breast cancer; however, women with breast cancer were found to have higher mean HCMV IgG levels in an Australian case–control study , suggesting that they might have experienced a recent infection. Analysis of a Norwegian cohort by the same group also revealed that elevation of HCMV IgG, but not EBV IgG levels, preceded the development of breast cancer in some women .
|
study
| 99.7 |
While serological evidence for HCMV in breast cancer may be limited, a stronger case is made by studies that have detected viral DNA and proteins by PCR and immunohistochemistry (IHC) in tumor biopsy specimens. In Taiwan, analysis of 62 breast cancer patients found that detection of both HHV-8 and HCMV in tumor samples by PCR was associated with lower overall survival and a decrease in relapse-free time . Harkins et al. detected HCMV immediate early (IE) proteins by IHC in breast glandular epithelial cells in 31 of 32 specimens from patients with ductal carcinoma in situ (DCIS) or infiltrating ductal carcinoma (IDC) . Another study found both HCMV IE and late proteins expressed in metastatic tumor cells in 100% of breast cancer specimens analyzed (73 total), and viral DNA was detected in 12/12 samples tested . Detection of virus in breast epithelial cells is consistent with the notion that epithelial cells are a site of HCMV persistence , and further supported by the fact that transmission of infectious virus through breast milk is well-documented [16–19].
|
review
| 50.38 |
Despite the evidence indicating the presence of viral proteins and DNA in breast tumor tissue, HCMV is not typically considered an oncogenic virus . Virus infection can, however, promote many of the classic hallmarks of cancer [21, 22], such as cell cycle dysregulation, inhibition of apoptosis, increased migration and invasion, and immune evasion [20, 23]. HCMV has been linked not only to breast cancer, but to an array of other malignancies, including glioblastoma [24–27], medulloblastoma , colon cancer , and prostate cancer . Individual HCMV gene products can have profound effects on cell growth, such as immediate early proteins IE1 and IE2, which are known to stimulate entry into S phase [31, 32]. IE1 expression was found to increase the growth rate of glioblastoma cells in culture, suppress p53 and Rb tumor suppressor activity, and stimulate PI3K/Akt signaling . IE1 was detected in breast tumor tissue [10, 11] as well as in CD133+ glioma stem cells isolated from glioblastoma multiforme (GBM) patients , suggesting that IE1 may promote tumorigenesis enhancing the growth and self-renewal of tumor stem cells.
|
review
| 99.56 |
Another HCMV gene implicated in tumor development is US28, which encodes a functional chemokine receptor that binds several human chemokines, including CCL2/MCP-1, CCL5/Rantes, and CX3CL1/Fractalkine [14, 35]. US28 also exhibits constitutive signaling activity, and cells expressing US28 are highly invasive and form tumors in nude mice [36, 37]. US28 was found to induce vascular endothelial growth factor (VEGF), cyclooxygenase-2 (COX2), and Stat3 activation through upregulation of IL-6 [36–38]. Analysis of glioblastoma tumor specimens revealed the presence of both US28 and phosphorylated Stat3 [27, 38], demonstrating that US28 may play role in tumor development in vivo.
|
study
| 99.94 |
Whereas the US28 and IE1 gene products are expressed in infected cells, the UL111A gene encodes cmvIL-10, a viral cytokine that is secreted from infected cells. Although cmvIL-10 has only 27% sequence identity to human interleukin-10 (hIL-10) , the viral cytokine binds to the cellular IL-10 receptor with greater affinity than hIL-10 itself . Extensive immunosuppressive properties of cmvIL-10 have been documented, including inhibition of inflammatory cytokine synthesis, downregulation of class I and II MHC, and inhibition of dendritic cell maturation [41–48]. Engagement of the IL-10 receptor by cmvIL-10 leads to activation of Stat3 [49–53], which is commonly constitutively activated in breast cancer cells , associated with poor prognosis in ovarian cancer, and considered a key factor in metastasis formation . CmvIL-10 was found to activate Stat3 and play a pivotal role in the progression of malignant glioma by enhancing the invasiveness and migration of glioma cancer stem cells . Because cmvIL-10 is secreted from the infected cell, it has the potential to act on any cell type, infected or not, that expresses the IL-10 receptor.
|
study
| 99.94 |
We have previously shown that the breast cancer cell lines MDA-MB-231 and MCF-7 express the IL-10R and that exposure to cmvIL-10 results in enhanced cell proliferation and migration [57, 58]. In the present study, we examined the impact of cmvIL-10 on MDA-MB-231 cell invasion through a simulated basement membrane and investigated the effect of cmvIL-10 on a panel of metastasis-related genes. We found that cmvIL-10 was a potent enhancer of invasion and influenced expression of genes strongly linked to the metastatic spread of breast cancer.
|
study
| 100.0 |
MDA-MB-231 human breast adenocarcinoma cells (American Type Culture Collection, Manassas, VA) were cultured in L-15 Leibovitz’s Medium (Mediatech, Manassas, VA) supplemented with 10% fetal bovine serum (FBS) (Atlanta Biologicals, Flowery Branch, GA) at 37 °C with atmospheric CO2. The human foreskin fibroblast (HFF) cell line (ATCC) was cultured in Dulbecco’s Modified Eagle Medium (DMEM) with 10% FBS at 37 °C in a humidified chamber with 5% CO2. Human cytomegalovirus strain AD169 (ATCC) was used to infect confluent monolayers of HFFs at the indicated multiplicities of infection. Purified recombinant cmvIL-10, human IL-10, and epidermal growth factor (EGF) were purchased from R&D Systems (Minneapolis, MN). IL-10R neutralizing antibody and S3I-201 Stat3 inhibitor were from Santa Cruz Biotechnology (Santa Cruz, CA).
|
study
| 100.0 |
Transwell migration was monitored using 96-well BD Fluoroblock plates with 8 µm filters (Corning, Inc., Corning, NY). Cells were harvested and suspended at density of 2 × 106 cells/ml in migration media (L-15 + 1% FBS), and a volume of 75 μl cell suspension was placed on top of the filter inserts. Where indicated, IL-10R neutralizing antibody was added at a concentration of 30 μg/ml. The bottom wells were loaded with the indicated concentrations of EGF in the presence of conditioned medium from mock or HCMV-infected fibroblasts (96 h post infection) in a total volume of 235 μl. After 5 h at 37 °C, cells that traversed the filter and entered the lower chamber were quantified by the addition of Cell Titer Glo (Promega, Madison, WI) using a Turner Veritas luminometer. For invasion, 96-well matrigel-coated BD Fluoroblock transwell invasion plates (Corning) were used. Invasion plates were re-hydrated with warm media at 37 °C for 3 h and then 75 μl cell suspension loaded onto the hydrated filters as described above. Where indicated, 10 μM Stat3 inhibitor was included with cells in the top chamber; cmvIL-10, hIL-10 or conditioned medium was present in both chambers. The bottom plates received the indicated EGF concentrations, and then transwell system was incubated for 22 h at 37 °C with atmospheric CO2. At harvest, cells that had degraded the matrigel and entered the lower chamber were quantified by the addition of Cell Titer Glo as above.
|
study
| 100.0 |
RNA was harvested from 10 × 106 MDA-MB-231 cells that were mock treated or treated with 100 ng/ml cmvIL-10 or hIL-10 for 5 h using the RNeasy Midi Kit and RNAse-Free DNase set (Qiagen, Valencia, CA). From the isolated RNA, cDNA was prepared using the RT2 First Strand Kit (SA Biosciences, Frederick, MD) and subsequently loaded into a 96-well breast cancer metastasis profiler PCR array (PAHS-028ZD) with system RT2 SYBR Green Mastermix (SA Biosciences). The plates were run using the CFX96 Real-Time system cycler (BioRad, Hercules, CA) with the following amplification program: 95 °C for 10 min, 95 °C for 15 min with a slow ramp rate for 1.0 c/s and 60 °C for 1 min. The data from three biological replicates for each treatment was analyzed by the ΔΔCT method according to manufacturer’s instructions using the RT2 profiler PCR array data analysis program located on the SABiosciences web portal and is reported as fold change relative to control.
|
study
| 100.0 |
DuoSet ELISA kits (R&D Systems) were used to quantify uPAR, PAI-1, and MMP-3. For uPAR and PAI-1 measurement, MDA cells were seeded in triplicate in 96-well plate at 5.0 × 104 cell/ml density with complete L-15 media and treated with 10 ng/ml of either cmvIL-10 or hIL-10 for the indicated times and supernatants were collected daily. The ELISA was carried out on supernatants according to manufacturer’s instructions using and following the addition of substrate and stop solution, absorbance of the plate was measured at 450 nm using a Dynex Opsys MR microplate reader. Sample concentrations were interpolated from a standard curve using linear regression analysis. For cell-associated MMP-3, MDA cells were seeded in 96-well plates and treated with cmvIL-10 as above. Cells were treated with cell lysis buffer (150 mM NaCl, 20 mM HEPES, 0.5% Triton-X-100, 1.0 mM NaOV4, 1.0 mM EDTA, 0.1% NaN3) supplemented with 1× protease inhibitors (Calbiochem, EMD Chemicals, San Diego CA) and were collected daily for the indicated time points. The lysates were evaluated for MMP-3 according to the manufacturer’s instructions (R&D Systems).
|
study
| 100.0 |
Confluent T-75 flasks of MDA-MB-231 cells were treated with 10 ng/ml cmvIL-10 (R&D systems) for the indicated times, then scraped and harvested into cell lysis buffer (150 mM NaCl, 20 mM HEPES, 0.5% Triton-X-100, 1.0 mM NaOV4, 1.0 mM EDTA, 0.1% NaN3) containing 1× protease inhibitors (Calbiochem). Cell lysates were clarified via centrifugation, heated at 70 °C for 10 min in reducing buffer, and the proteins separated on a 4–12% Tris-Base SDS-PAGE gel (Life Technologies, Grand Island, NY). After transfer to nitrocellulose, the membrane was incubated in blocking solution (5% milk + TBS) for 1 h at room and then probed with primary antibody: 1:1000 dilution for MMP-3 or MTSS-1 antibodies (Santa Cruz), or MAPK antiserum (Cell Signaling Tech, Danvers, MA), in blocking solution overnight, oscillating on a platform rocker at 4.0 °C. After three washes, the membranes were incubated with a 1:2000 dilution of appropriate AP-conjugated secondary antibody on a platform rocker at room temperature for 1 h. Protein bands were detected using western blue stabilized AP substrate (Promega, Madison, WI). For zymography, cell lysates were denatured in SDS buffer under non-reducing conditions without heat, and run on a 4–16% Zymogram gel using Tris–Glycine SDS running buffer according to manufacturer’s instructions. After electrophoresis, the enzyme was renatured by incubating the gel in Zymogram Renaturing Buffer containing a non-ionic detergent, then equilibrated in Zymogram Developing Buffer (to add divalent metal cations required for enzymatic activity), and then stained and destained to reveal digested (clear) areas corresponding to active enzyme.
|
study
| 100.0 |
MDA-MB-231 cells were seeded onto FBS-coated glass coverslips at a density of 2.0 × 105 cells/well and cultured for 48 h at 37 °C. Cells were treated with 100 ng/ml of purified recombinant cmvIL-10 for 96 h, then fixed with 2% paraformaldehyde in DPBS for 20 min, washed, permeabilized with 0.2% Triton-X-100 in PBS for 15 min. The cells were washed and blocked with 10% FBS for 1 h at 37 °C, then incubated with anti-MTSS-1 antibody at a 1:100 dilution for 1 h at 37 °C followed by the addition of goat anti-mouse TRITC secondary antibody at a 1:150 dilution for 30 min (Life Technologies) and Alexa Fluor 488 phalloidin (Molecular Probes, Eugene, OR). Coverslips were washed and excess fluid was removed before inverting the coverslip onto a glass slide containing 20 μl of DAPI-containing Prolong Gold mounting medium (Life Technologies, Grand Island, NY). Images were taken on a Zeiss AX10 Imager.A1 microscope (Carl Zeiss Inc., Oberkochen, Germany) using AxioVision 4.7.2 imaging software.
|
study
| 99.94 |
The tumor microenvironment is a complex milieu that includes not only malignant cells, but immune cells, fibroblasts, signaling molecules, the extracellular matrix (ECM), and blood vessels. We have previously found that cmvIL-10 enhances migration of MDA-MB-231 breast cancer cells in vitro toward epidermal growth factor (EGF) . In order to more faithfully replicate conditions under which cmvIL-10 might be found in the tumor microenvironment, we examined the ability of cmvIL-10 secreted from virus-infected cells to stimulate movement of MDA cells. Monolayer cultures of human foreskin fibroblasts were mock-infected or infected with HCMV strain AD169 at a range of multiplicities of infection (MOI). After 96 h, supernatants were harvested and placed in the lower chamber of a transwell migration plate in the presence or absence of EGF. MDA cells were placed in the upper chamber, separated from the EGF and conditioned medium by a porous filter. After five hours, cells that traversed the filter were quantified. MDA cells did not exhibit any significant movement toward conditioned medium from mock or infected cells, which is consistent with our previous finding that cmvIL-10 is not a chemoattractant for tumor cells . However, when conditioned medium from mock infected cells was supplemented with EGF, cell migration was observed (Fig. 1a). When EGF was added to conditioned medium from HCMV-infected cells, the amount of cell migration increased, suggesting that substances released from virus-infected cells amplified chemotaxis to EGF. Moreover, the enhanced MDA cell movement was greater when EGF was provided in supernatants from higher MOI infections, and thus greater concentrations of cmvIL-10, indicating a dose-dependent effect. To confirm that cmvIL-10 was the virally produced substance mediating this increase in cell movement, MDA cells were pre-incubated for 30 min with a neutralizing antibody (NAb) directed at the cellular IL-10R. The NAb was also included in the top chamber with MDA cells during the 5 h incubation, and resulting migration was reduced to levels seen when only EGF was present in medium from mock infected cells. These results demonstrate that cmvIL-10 secreted from virally infected cells has the ability to interact with the cellular IL-10R on tumor cells to enhance directed movement.Fig. 1MDA-MB-231 cells exhibit increased migration and matrigel invasion in the presence of cmvIL-10. a Transwell migration toward EGF after 5 h in the presence of conditioned medium from mock or HCMV-infected fibroblasts at the indicated MOIs. IL-10R neutralizing antibody (NAb) was included at 30 μg/ml. b Matrigel invasion toward EGF in the presence of 100 ng/ml purified recombinant cmvIL-10 or hIL-10 after 22 h. c Matrigel invasion toward EGF in the presence or absence of 100 ng/ml purified recombinant cmvIL-10 or conditioned medium from mock or infected fibroblasts (MOI = 1). Where indicated, 10 μM Stat3 inhibitor was included. Error bars SEM. *p < 0.05, **p < 0.001. These results are representative of three independent experiments
|
study
| 100.0 |
MDA-MB-231 cells exhibit increased migration and matrigel invasion in the presence of cmvIL-10. a Transwell migration toward EGF after 5 h in the presence of conditioned medium from mock or HCMV-infected fibroblasts at the indicated MOIs. IL-10R neutralizing antibody (NAb) was included at 30 μg/ml. b Matrigel invasion toward EGF in the presence of 100 ng/ml purified recombinant cmvIL-10 or hIL-10 after 22 h. c Matrigel invasion toward EGF in the presence or absence of 100 ng/ml purified recombinant cmvIL-10 or conditioned medium from mock or infected fibroblasts (MOI = 1). Where indicated, 10 μM Stat3 inhibitor was included. Error bars SEM. *p < 0.05, **p < 0.001. These results are representative of three independent experiments
|
study
| 100.0 |
To further recapitulate the tumor microenvironment, we examined whether cmvIL-10 could also promote invasion through matrigel, a gelatinous protein mixture derived from mouse sarcoma cells widely used to simulate the ECM in vitro . MDA cells were place atop a matrigel-coated transwell system with EGF placed in the lower chambers. Purified recombinant cmvIL-10 or hIL-10 was added to both chambers. After incubation for 22 h, invasion was assessed by counting cells in the lower chamber, which should contain only the cells that were able to degrade the matrigel coating to access the porous filter. As shown in Fig. 1b, cmvIL-10 was found to be a strong enhancer of cell invasion. Surprisingly, cmvIL-10 was able to increase invasion of MDA breast cancer cells to a significantly greater extent than hIL-10, suggesting that the viral cytokine may trigger signaling events that are distinct from the cellular cytokine. Since activation of the transcription factor Stat3 by cmvIL-10 is well-documented [39, 49–51, 53, 60, 61], we next examined the need for Stat3 in cmvIL-10-enhanced invasion. Treatment with a Stat3 inhibitor reduced the cmvIL-10-induced increase in invasion through matrigel toward EGF seen when either recombinant purified protein or cytokine produced during virus infection were present (Fig. 1c). Taken together, these results demonstrate the novel finding that not only does cmvIL-10 produced during virus infection stimulate enhanced migration and invasion of breast cancer cells, but it does so more effectively than hIL-10.
|
study
| 100.0 |
Given the impact of cmvIL-10 on MDA cell invasion, we wanted to investigate whether the viral cytokine brought about changes in the expression of genes associated with tumor metastasis. Transcriptional profiling was performed using a tumor metastasis qPCR array designed to analyze 84 genes known to be involved in breast cancer metastasis. MDA cells were mock-treated or incubated with either cmvIL-10 or hIL-10 for 5 h, then RNA was extracted, cDNA synthesized, and qPCR performed. Additional file 1: Table S1 contains a complete list of genes analyzed with fold changes for cmvIL-10 or hIL-10 treated cells compared to mock treated control cells indicated. Select genes encoding proteins associated with either the ECM (Fig. 2a) or cell adhesion (Fig. 2b) are shown graphically. Overall, plasminogen activator inhibitor (PAI-1) was the most highly upregulated gene for both cytokines, with expression increased by 2.68-fold by cmvIL-10 and 3.12-fold by hIL-10. Interestingly, increased expression of urokinase plasminogen receptor (uPAR) was also common to both cmvIL-10 and hIL-10 (1.59- and 1.87-fold increases, respectively). Matrix metalloproteinase-3 (MMP3) was specifically upregulated by cmvIL-10 only (2.75-fold increase), while collagen type 4 (COL4A) expression was increased by hIL-10 only (1.51-fold increase). Changes in cell adhesion genes were more modest, with only one gene, metastasis suppressor 1 (MTSS1) exhibiting a statistically significant change of more than twofold, and this was observed for cmvIL-10 treatment only (0.305-fold change, or −3.28). Slight decreases in integrin alpha 7 (ITGA7, 0.561- or −1.78-fold change), melanoma cell adhesion factor (MCAM, 0.768- or −1.32-fold change), and cadherin 6 (CDH6, 0.811- or −1.23-fold change) were also found with cmvIL-10 treatment. Both cmvIL-10 and hIL-10 induced a slight decrease in expression of vascular endothelial growth factor (VEGFA, 0.554- or −1.80-fold for cmvIL-10; 0.5987- or 1.67-fold change for hIL-10). Chemokine receptor CXCR2 expression was also strongly decreased by cmvIL-10 and hIL-10, but those changes were not statistically significant. Overall, these transcriptional profiling results indicate that cmvIL-10, as well as human IL-10, can affect expression of genes that are likely to promote metastatic spread of tumor cells.Fig. 2CmvIL-10 induces changes in metastasis-related gene expression in MDA-MB-231 cells. RNA was extracted from cells treated with 100 ng/ml cmvIL-10 or hIL-10 for 5 h, then RNA was purified and expression of 84 genes was analyzed using the Human Tumor Metastasis RT2 Profiler PCR Array. a Select genes encoding extracellular matrix proteins or b cell adhesion proteins is shown. Fold changes represent comparison to untreated MDA-MB-231 and are the average of three biological replicates. Error bars SEM. *p < 0.05. A complete list of genes analyzed is found in Additional file 1: Table S1
|
study
| 100.0 |
CmvIL-10 induces changes in metastasis-related gene expression in MDA-MB-231 cells. RNA was extracted from cells treated with 100 ng/ml cmvIL-10 or hIL-10 for 5 h, then RNA was purified and expression of 84 genes was analyzed using the Human Tumor Metastasis RT2 Profiler PCR Array. a Select genes encoding extracellular matrix proteins or b cell adhesion proteins is shown. Fold changes represent comparison to untreated MDA-MB-231 and are the average of three biological replicates. Error bars SEM. *p < 0.05. A complete list of genes analyzed is found in Additional file 1: Table S1
|
study
| 100.0 |
The most significantly upregulated gene by both cmvIL-10 and hIL-10 was PAI-1, or plasminogen activator inhibitor 1. PAI-1 is a 43 kDa glycoprotein that inhibits the function of urokinase plasminogen activator (uPA), a serine protease that catalyzes the conversion of inactive plasminogen to plasmin and has been implicated in many aspects of tumor progression . The activity of uPA system is regulated by the receptor uPAR and two endogenous inhibitors, PAI-1 and PAI-2 . PAI-1 is constitutively secreted by many cell types and high levels have been found to inhibit cell adhesion and promote migration [63, 64]. In order to confirm that changes in gene expression identified by the qPCR array correlated with protein expression, MDA cells were treated with cmvIL-10 or hIL-10 and PAI-1 levels measured by ELISA. As expected, PAI-1 was produced by untreated cells, however, the amount of protein secreted was significantly increased by cmvIL-10 after 12 h of exposure (Fig. 3a). After 24 h, both cmvIL-10 and hIL-10 stimulated a significant increase in PAI-1 production, and this was maintained over 72 h. In addition, we examined uPAR protein levels and found that they were also elevated upon exposure to cmvIL-10 or hIL-10 (Fig. 3b). These results demonstrate that expression of two elements of the uPA serine protease system, its receptor uPAR and its serpin inhibitor PAI-1, are significantly increased by cmvIL-10 and hIL-10 in human breast cancer cells.Fig. 3PAI-1 and uPAR levels are elevated upon exposure to cmvIL-10 or hIL-10. a MDA-MB-231 cells were cultivated in the presence of 10 ng/ml purified recombinant cmvIL-10 or hIL-10. At the indicated time points, culture supernatants were collected and levels of PAI-1 measured by ELISA. b MDA cells were cultured as above and levels of uPAR in the supernatant were determined by ELISA. Error bars represent standard error among three replicates for each data point. *p < 0.05. These results are representative of three independent experiments
|
study
| 100.0 |
PAI-1 and uPAR levels are elevated upon exposure to cmvIL-10 or hIL-10. a MDA-MB-231 cells were cultivated in the presence of 10 ng/ml purified recombinant cmvIL-10 or hIL-10. At the indicated time points, culture supernatants were collected and levels of PAI-1 measured by ELISA. b MDA cells were cultured as above and levels of uPAR in the supernatant were determined by ELISA. Error bars represent standard error among three replicates for each data point. *p < 0.05. These results are representative of three independent experiments
|
study
| 100.0 |
Next we examined MMP-3, a member of the matrix metalloproteinase family that has the ability degrade many components of the extracellular matrix, such as collagen III-V, and IX-XI, as well as laminins, elastins, fibronectin, vitronectins and proteoglycans . Mouse epithelial mammary cells cultured with MMP-3 had decreased expression of cytokeratin markers and increased expression of vimentin, a clear sign of the epithelial-to-mesenchymal transition (EMT), in which epithelial cells morph into a mesenchymal-type cell to eliminate their connection to the basement membrane and initiate migration towards subsequent intravasation into blood vessels . MMP-3 can also activate other MMPs, and high levels of MMP-3 correlate with poor prognosis in breast cancer patients . MDA cells were treated with cmvIL-10 and then total MMP-3 levels were measured by ELISA. We were unable to detect any MMP-3 in cell supernatants, but cell-associated MMP-3 was detected by analysis of cell lysates. Relatively low levels of MMP-3 were produced by untreated MDA cells, but this amount increased significantly after 48 h of treatment with cmvIL-10 (Fig. 4a). Since MMPs are generally secreted as inactive pro-enzymes that require cleavage to become activated, we further examined MMP-3 by western blotting and zymography. Consistent with the ELISA results, an increase in total MMP-3 protein was observed over time with exposure to cmvIL-10 (Fig. 4b). The amount of active MMP-3 enzyme was also increased by cmvIL-10 treatment, as evidence by increased digestion of casein in the zymogen gel. Taken together, these results indicate that cmvIL-10 promotes increased expression and activation of MMP-3 by breast cancer cells, which is likely to contribute to increased degradation of the ECM and greater risk of metastasis.Fig. 4MMP-3 expression and activity are increased by cmvIL-10. a MDA-MB-231 cells were cultured in the presence or absence of 10 ng/ml cmvIL-10 for the indicated times and then cell lysates were analyzed by ELISA. Error bars SEM, *p < 0.05. b MDA cells cultured with 100 ng/ml cmvIL-10 for the indicated times were harvested and lysates examined by western blotting with anti-MMP3 or anti-MAPK as a protein loading control. Lysates from cells receiving the same treatment were analyzed under non-reducing conditions on a 4–16% Zymogram gel (lower panel). Arrow indicates active MMP-3. These results are representative of three independent experiments
|
study
| 100.0 |
MMP-3 expression and activity are increased by cmvIL-10. a MDA-MB-231 cells were cultured in the presence or absence of 10 ng/ml cmvIL-10 for the indicated times and then cell lysates were analyzed by ELISA. Error bars SEM, *p < 0.05. b MDA cells cultured with 100 ng/ml cmvIL-10 for the indicated times were harvested and lysates examined by western blotting with anti-MMP3 or anti-MAPK as a protein loading control. Lysates from cells receiving the same treatment were analyzed under non-reducing conditions on a 4–16% Zymogram gel (lower panel). Arrow indicates active MMP-3. These results are representative of three independent experiments
|
study
| 100.0 |
MTSS1 was notable as the gene most strongly downregulated by cmvIL-10 treatment. Also known as missing-in-metastasis (MIM), MTSS1 was originally identified as a tumor suppressor gene whose expression was lost in metastatic bladder and prostate cancers . The tumor suppressor works as a scaffold to inhibit the dissociation of cell junctions and to increase adherens junction formation, so when MTSS1 is lost, recruitment of F-actin to the cytoskeleton is reduced, enabling tumor cells to detach from the basement membrane and from neighboring cells. MTSS1 has been found to be inversely correlated to the aggressive invasive potential in several breast cell lines and with overall survival in breast cancer patients . To confirm that the reduced gene expression observed on the PCR array correlated with a decrease in MTSS1 protein levels, immunoblotting was performed on lysates from MDA-MB-231 cells treated with 10 ng/ml cmvIL-10. The expected 82 kD band was detected for MTSS1 in untreated cells and was still visible after 24 h of incubation with cmvIL-10 (Fig. 5a). However, as time progressed, the cmvIL-10-treated samples showed a significant decrease in MTSS1 expression. In contrast, the β-actin bands that serve as a loading control remained constant. We further examined MTSS1 expression via immunofluorescence microscopy and found the protein to be widely distributed throughout the cytoplasm in untreated MDA cells (Fig. 5b), which is consistent with its role in regulating cytoskeletal rearrangement. After exposure to cmvIL-10, dramatic reduction in the amount of MTSS1 protein was observed. This reduction in MTSS1 corresponded to a noticeable change in cellular architecture, as cmvIL-10-treated cells appeared to be thinner and have fewer substrate attachment points. These results demonstrate that treatment with cmvIL-10 reduced the expression of MTSS1 in MDA cells, which could contribute to the increased migration and invasion observed in the presence of cmvIL-10.Fig. 5MTSS1 expression is significantly decreased upon exposure to cmvIL-10. a MDA-MB-231 cells were cultured with 100 ng/ml cmvIL-10 for the indicated times and lysates examined by western blotting with anti-MTSS1 or anti-β-actin as a protein loading control. b MDA cells were seeded onto glass coverslips in the presence of absence of 100 ng/ml cmvIL-10 for 96 h, then fixed, permeabilized, and stained as indicated. Scale bar 10 μm. These results are representative of three independent experiments
|
study
| 100.0 |
MTSS1 expression is significantly decreased upon exposure to cmvIL-10. a MDA-MB-231 cells were cultured with 100 ng/ml cmvIL-10 for the indicated times and lysates examined by western blotting with anti-MTSS1 or anti-β-actin as a protein loading control. b MDA cells were seeded onto glass coverslips in the presence of absence of 100 ng/ml cmvIL-10 for 96 h, then fixed, permeabilized, and stained as indicated. Scale bar 10 μm. These results are representative of three independent experiments
|
study
| 100.0 |
The results that we present here characterize a new role for cmvIL-10 beyond its well-known function as an immune modulator during HCMV infection . The viral cytokine conferred MDA-MB-231 cells with heightened migration and invasion abilities and is likely to promote the development of breast cancer metastasis. While our studies have been conducted in vitro using cell lines, we propose an in vivo scenario in which latently infected monocyte/macrophages infiltrate the developing tumor and release cmvIL-10 (Fig. 6). Our observations that cmvIL-10 reduces expression of MTSS1, while increasing expression of uPAR, PAI-1, and MMP-3 suggest that the secreted viral cytokine can act directly on breast epithelial cells expressing IL-10R to promote reduced cell–cell adhesion and increased movement, ultimately leading to invasion into the surrounding stromal tissue and entry into bloodstream.Fig. 6Model depicting possible role of cmvIL-10 in promoting tumor metastasis. A monocyte that is latently infected with HCMV infiltrates a localized tumor, releasing cmvIL-10 that acts on tumor cells expressing the IL-10 receptor. This leads to changes in levels of MTSS1, uPAR and PAI-1, which reduce cell adhesion. Increased levels of uPAR and PAI-1 are strongly associated with increased migration and can also help activate MMP-3. Active MMP-3 degrades proteins in the extracellular matrix, facilitating access for tumor cells to invade surrounding stromal tissue and enter the bloodstream
|
study
| 100.0 |
Model depicting possible role of cmvIL-10 in promoting tumor metastasis. A monocyte that is latently infected with HCMV infiltrates a localized tumor, releasing cmvIL-10 that acts on tumor cells expressing the IL-10 receptor. This leads to changes in levels of MTSS1, uPAR and PAI-1, which reduce cell adhesion. Increased levels of uPAR and PAI-1 are strongly associated with increased migration and can also help activate MMP-3. Active MMP-3 degrades proteins in the extracellular matrix, facilitating access for tumor cells to invade surrounding stromal tissue and enter the bloodstream
|
study
| 100.0 |
The evidence linking HCMV to breast cancer continues to grow, yet there is still considerable controversy in field. While some groups have found traces of viral DNA or proteins in tumor samples [6, 10, 11], other labs have not been able to detect evidence of HCMV [70, 71]. This may be due to differences in detection technique, sample preparation, and even tumor type. El-Shinawi et al. have found evidence for HCMV in inflammatory breast cancer (IBC), a highly metastatic and aggressive form of breast cancer that is often associated with pregnancy and occurs in at higher frequency in women of Northern African and Egyptian descent . They reported detection of more HCMV DNA in IBC tissues compared to IDC, and higher HCMV IgG titers in IBC patients compared to IDC patients . In a follow-up study, they assessed viral genotypes and found a correlation between mixed phenotypes and disease progression, notably lymphovascular invasion and formation of lymphatic emboli in IBC patients, but not in women with other forms of breast cancer . The findings suggest that HCMV may be more closely associated with specific subtypes of breast cancer.
|
review
| 99.9 |
HCMV can infect a wide range of cell types that may be present within the tumor microenvironment, such as monocyte/macrophages, fibroblasts, epithelial and endothelial cells. While several studies have found evidence for direct infection of tumor cells via IHC staining [10, 11], detection of viral DNA by PCR from fixed tissue samples precludes direct identification of the cell type harboring the virus. It may be that the cell type infected varies from case to case, just like the combination of specific mutations that lead to tumor initiation in a given individual. We favor the notion that immune cells harboring latent HCMV infiltrate the tumor, because monocytes are a well-documented reservoir for HCMV [74–76] and this scenario could lead to significant variability from tumor to tumor depending on the number of infected infiltrating cells and whether they reactivate virus that goes on to infect other cells in the tumor microenvironment. Transcriptional analysis of HCMV-infected monocytes revealed a unique M1/M2 polarization signature that included induction of both M1 type inflammatory cytokines like IL-1, IL-6, and TNFα, as well as upregulation of M2 type cytokines like IL-10 and IL-18 . The presence of these conflicting signals in the tumor environment has been associated with neoplastic progression, suggesting that HCMV could tip the balance in favor of this process .
|
study
| 100.0 |
Analysis of the secretome, or proteins produced by HCMV-infected cells, has revealed high levels of both MMP-3 and PAI-1 , which is consistent with our observations that cmvIL-10 induces higher expression of both of these proteins. The secretome was found to promote angiogenesis and wound healing, and contained many growth factors, cytokines, chemokines, and enzymes associated with metastasis, including MMP-1, MMP-2, MMP-9, and MMP-10 . Somiari and colleagues were able to detect elevated levels and activity of MMP-2 and MMP-9 in plasma from breast cancer patients, as well as in women determined to be high risk based on Gail Model predictions . This suggests that it may be possible to develop a plasma protein profile with a characteristic signature that identifies individuals likely to develop breast cancer.
|
study
| 100.0 |
Although cmvIL-10 has not yet been quantified in patient serum, measurement of cmvIL-10 and hIL-10 may also have prognostic value in breast cancer. Elevated levels of hIL-10 (27-2134 pg/ml) have already been detected in the serum of some cancer patients and correlate with poor prognosis [81–86], suggesting that hIL-10 may contribute to immune suppression and tumor progression. In vitro, hIL-10 has been found to promote resistance to apoptosis in human breast and lung cancer cell lines [87, 88]. Importantly, recent evidence suggests that cmvIL-10 induces increased expression of hIL-10, potentially amplifying the immune suppressive environment and enabling the invasive spread of tumor cells . Our results show that cmvIL-10 increased the migration and invasive ability of MDA-MB-231 breast cancer cells and affected expression of several metastasis-related genes. Taken together, these findings suggest a new mechanism for HCMV oncomodulation, as secretion of cmvIL-10 is expected to manipulate the tumor microenvironment, enhancing the potential of a developing breast tumor to invade surrounding tissue, and ultimately establish metastatic tumors. Ultimately, it may be that a signature profile of factors like cmvIL-10, hIL-10, MMPs, and PAI-1 could have predictive or prognostic value for breast cancer. If HCMV is truly involved in promoting tumor progression, chemotherapy treatment regimens that include anti-cmvIL-10 specific antibodies or even anti-viral drugs may help improve the overall survival of cancer patients.
|
study
| 99.94 |
Bitter taste is one of the basic taste modalities including the bitterness, sweetness, umaminess, saltiness, acidness, and fatness (Besnard et al., 2016; Roper and Chaudhari, 2017). Evolutionarily, bitter taste is pivotal to the survival by protecting organisms from the consumption of potentially poisonous substances, which often taste bitter. Perception of the bitter taste is mainly mediated by the taste receptors type 2 (Tas2Rs) family of G-protein coupled receptors (GPCRs) on the apical membrane of the taste receptor cells located in the taste buds (Jaggupilli et al., 2016; Roper and Chaudhari, 2017). Intriguingly, Tas2Rs are also expressed in the extra-oral tissues, e.g., the gastrointestinal tract and respiratory system, etc., indicating that they are also intricately involved in the other crucial biological processes (Clark et al., 2012; Shaik et al., 2016).
|
review
| 99.7 |
In humans, 25 different hTas2Rs are evolved to bind the bitterants with diverse chemical structures (Behrens and Meyerhof, 2013; Jaggupilli et al., 2016). Some Tas2Rs such as Tas2R10, Tas2R14, and Tas2R46 broadly accommodate the various bitterants, while some Tas2Rs such as Tas2R50 exquisitely select the specific bitterants (Meyerhof et al., 2010; Brockhoff et al., 2011; Ji et al., 2014). At the same time, the promiscuous bitterant can interact with multiple Tas2Rs, while the selective bitterant can only activate one or few specific Tas2Rs (Di Pizio and Niv, 2015). Therefore, the compound that can stimulate at least one Tas2R can be treated as a bitterant, however, only the compound that cannot activate any of 25 hTas2Rs can be defined as a non-bitterant, since a compound cannot stimulate one specific Tas2R, which could still elicit the bitterness via targeting another Tas2R.
|
study
| 99.94 |
Bitterness is often perceived as an unpleasant taste, albeit it is considered desirable in some products such as tea, coffee, and beer etc. In most cases, bitterness influences the palatability of the functional beverage and food containing the bitter ingredients, and also poses a major problem for the patient acceptability and compliance of the bitter-taste drugs, especially for the pediatric formulations (Drewnowski and Gomez-Carneros, 2000; Mennella et al., 2013). Therefore, the bitter-tasting assessment is imperative in the functional food/beverage development, and could be considered in advance during the drug discovery process.
|
review
| 99.1 |
Bitter-tasting assessment is often an arduous and tedious task. Basically there are two types of experimental taste evaluations: in-vivo and in-vitro approaches, which are systematically reviewed by Anand et al. (2007). One of the most direct methods is called “human taste panel studies,” which evaluates the taste of standard and test stimuli in the healthy human volunteers with the well-designed protocols (Anand et al., 2007). However, this experimental method has its major disadvantage due to the higher probability of toxicity for the bitter compounds, which will cause the safety and ethical issues and consequently limit its application in the high-throughput screening of the bitterants. In contrast to the experimental methods, in-silico method provides a cheap and rapid alternative to identify the most likely bitterants from the small-molecule database (Bahia et al., 2018). Thus, the computational prediction of the bitterant becomes more and more important prior to the laborious and time-consuming experimental taste assessment.
|
review
| 99.9 |
Current commonly-used computational methods for the bitterant prediction are categorized by Bahia et al. which are listed as follows: (Bahia et al., 2018) structure-based method (Floriano et al., 2006; Brockhoff et al., 2010; Singh et al., 2011; Tan et al., 2011; Marchiori et al., 2013; Sandal et al., 2015; Acevedo et al., 2016; Karaman et al., 2016; Suku et al., 2017), ligand-based method (Roland et al., 2013, 2015; Levit et al., 2014) and machine-learning based method (Rodgers et al., 2006; Huang et al., 2016; Dagan-Wiener et al., 2017). Structure-based method requires the 3D structures of Tas2Rs, whose crystal structures still remain unresolved. In contrast, ligand-based method approach such as the 3D-pharmacophore method still works even in the absence of 3D structures of Tas2Rs. Both methods work well for the particular bitter-taste receptor, Nevertheless, a compound that cannot activate one specific Tas2R could still trigger the bitter taste via stimulating the other 24 hTas2Rs. Thus, both methods are not suitable for the general classification of bitterant/non-bitterant. Machine-learning based approach can effectively circumvent the aforementioned problems and can directly predict the bitter or bitterless compounds (Rodgers et al., 2005, 2006; Huang et al., 2016; Dagan-Wiener et al., 2017; Bahia et al., 2018). In this emerging method, certain experimental dataset including both the bitter and bitterless compounds is employed to establish the prediction model, while the target information of the bitter compound is not necessary, which confers the unique advantage on this method.
|
review
| 99.9 |
There are three typical studies about the general bitterant prediction with the machine-learning approach based on the relatively large dataset (Rodgers et al., 2006; Huang et al., 2016; Dagan-Wiener et al., 2017), although there are several studies about the congeneric systems with the comparatively small dataset (Takahashi et al., 1982; Spillane et al., 2002; Cravotto et al., 2005; Scotti et al., 2007). In addition, all these studies focus on the prediction of small-molecule bitterant, which is our current main research interest, thus the prediction of bitter peptide explored in the other studies (Ney, 1979; Soltani et al., 2013) will not be reviewed here.
|
review
| 99.9 |
Rodgers et al. employ the Naive Bayes algorithm and circular fingerprint (MOLPRINT 2D, Willett et al., 1998) to classify the bitter/bitterless compounds (Rodgers et al., 2006). The dataset consists of 649 bitterants and 13,530 hypothetical non-bitterants. All the bitterants are from their proprietary database, while 13,530 hypothetical non-bitterants are randomly selected from the MDL Drug Data Repository (MDDR). The prediction model gives the best accuracy, precision, specificity, and sensitivity of 88, 24, 89, and 72% respectively in the five-fold cross-validation. It's the first bitterant prediction model trained with the large dataset. Nevertheless, the bitterless compounds in their study are not experimentally confirmed, and their work didn't provide a practical prediction tool for the users to have a test on their model.
|
study
| 99.0 |
Huang et al. developed the first online prediction tool called “BitterX,” which combines Support Vector Machine (SVM) approach (Vapnik, 1995) with the physicochemical descriptors (Huang et al., 2016). In their study, the dataset is composed of 539 bitterants and 539 non-bitterants. Five hundred thirty-nine bitterrants are gathered from the literature and the publicly available BitterDB (Wiener et al., 2012). For 539 non-bitterants, 20 non-bitterants are from their in-house bitterless compounds validated by the experiments, and 519 non-bitterants are the representative structures clustered from the compounds without the tag of “bitter” in the Available Chemicals Directory (ACD) database (http://accelrys.com). Their bitterant prediction model offers the impressive accuracy (91~92%), precision (91~92%), specificity (91~92%), and sensitivity (91~94%) on the test set. However, 519 compounds assumed as the non-bitterants are still not confirmed by the experiments. Thus, the limited number of experimental non-bitterants are the bottleneck for the machine-learning based approach.
|
study
| 99.9 |
Recently Wiener et al. published a prediction tool named “BitterPredict,” which adopts 12 basic physiochemical descriptors and 47 Schrödinger QikProp descriptors (Dagan-Wiener et al., 2017). In this work, the classification method is the adaptive ensemble machine-learning method “Adaptive Boosting” (AdaBoost), whose advantage is that this method is simple, fast, less susceptible to the overfitting. Meanwhile, the dataset is larger than the counterpart in Huang et al. and comprises 691 bitterants and 1,917 non-bitterants. The bitterants are mainly from their BitterDB (Wiener et al., 2012), and the work of Rojas et al. (2016) The non-bitterants are composed of 1,360 non-bitter flavors, 336 sweeteners, 186 tasteless compounds, and 35 non-bitter molecules (Eric Walters, 1996; Arnoldi et al., 1997; Ley et al., 2006; Rojas et al., 2016). The last four sets of compounds are experimentally confirmed, whereas 1,360 non-bitter flavors are gathered from the Fenroli's Handbook of Flavor Ingredients (Burdock, 2004), and are hypothetically defined as the non-bitterants if the word “bitter” is not explicitly mentioned in the description section of each compound in the book (Burdock, 2004). Their prediction model gives the accuracy (83%), precision (66%), specificity (86%), and sensitivity (77%) on the test set. Nevertheless, majorities of non-bitterants (1,360 non-bitter flavors) are still hypothetical. In addition, BitterPredict works in the environment of commercial MATLAB package and requires the commercial Schrödinger software to generate molecular descriptors, which will hamper the extensive test by the users. So far, users can only send the data to authors for the prediction.
|
study
| 98.7 |
In short, all three works adopt Naive Bayes, SVM or Adaboost as the classification method. The recently popular machine-learning methods such as deep neuron network (DNN) (LeCun et al., 2015), random forest (RF) (Breiman, 2001), and gradient boosting machine (GBM) (Friedman, 2002), frequently manifest the promising performance in the kaggle competition (www.kaggle.com/competitions), but were not used in the bitterant prediction before. In addition, the simple K-nearest neighbors (KNN) method (Itskowitz and Tropsha, 2005), which is generally used as the baseline for the comparison of machine-learning methods, was never applied in the bitterant prediction as well. Moreover, the consensus voting strategy based on the multiple machine-learning methods also was not employed to build the bitterant classification model in the past. Secondly, the previous works make use of the fully or partially hypothetical non-bittterant dataset. Therefore, there is a pressing need to make use of the fully experimental dataset with the relatively large size to develop the consensus model for the bitterant prediction that can be utilized by the food scientists in an easy-to-use and free software.
|
review
| 99.9 |
In this work, we collect only the experimentally confirmed bitterants and non-bitterants. Based on this fully experimental dataset, we adopt the popular Extended-connectivity Fingerprint (ECFP) (Rogers and Hahn, 2010) as the molecular descriptors and propose the consensus voting from the current mainstream machine-learning methods such as KNN, SVM, RF, GBM, and DNN to build the bitterant/non-bitterant classification models. All the models are carefully inspected by the Y-randomization test to ensure their reliability, and some promising models are subsequently selected to construct nine consensus models that are integrated in our program for the bitterant prediction. To aid the food scientists to automatically predict whether the compound of interest is bitter or not, we present a convenient graphic program called “e-Bitter,” which natively implements ECFPs for the automatic generation of the molecular descriptors. More importantly, e-Bitter can intuitively visualize the inter-connected 3D structural feature, feature importance and feature partial derivative for any specific bit “1” in ECFP. At last, the performance and functions of our program compared with other bitterant prediction tools are discussed.
|
study
| 99.94 |
An appropriate experimental dataset including both the bitterants and non-bitterants are critical to properly build the reasonable prediction model. Three criteria are defined for our data curation. (1) All the disconnected structures such as salts are not considered. (2) Only the compounds with the common elements C, H, O, N, S, P, Si, F, Cl, Br, or I are collected. (3) The same compound labeled with the different taste qualities will be excluded. (4) The duplicated compounds from the different sources will be removed. Based on these criteria, all the compounds are curated as the Tripos mol2 format.
|
study
| 99.9 |
In our work, majorities of bitterants are downloaded from the publicly available BitterDB (Wiener et al., 2012), and the others are retrieved from the literature (Rodgers et al., 2006; Rojas et al., 2016). The total number of the bitterants is 707. However, the data source of the non-bitterants raises a tough issue, since most of the published works often did not report the non-bitterants due to the less scientific significance. Hence only 132 tasteless and 17 non-bitter compounds retrieved from the literature are treated as the non-bitterants (Huang et al., 2016; Rojas et al., 2016). In order to further extend the size of the bitterless dataset, we tentatively propose to use the sweet molecules that can be generally assumed as the non-bitterants. The sweet compounds are downloaded from the SuperSweet (Ahmed et al., 2011) and SweetenersDB (Chéron et al., 2017). Database and additionally gathered from the literature (Zhong et al., 2013; Cristian et al., 2016; Rojas et al., 2016), which results in 443 compounds. The whole dataset, containing 707 bitterants and 592 non-bitterants, is publicly available in our e-Bitter program, with which users can handily view the 3D structure of each compound and its corresponding label (Y: bitterant or N: non-bitterant).
|
study
| 99.94 |
To explore the chemical space of our dataset, molecular weight (MW), logP, and the numbers of hydrogen-bond donor and acceptor (NHBD and NHBA) for all the bitterants and non-bitterants are calculated with Openbabel v2.4 (O'Boyle et al., 2011). The histograms of logP, MW, NHBD, and NHBA are plotted in Figures S1–S4 and the scatter plots of logP vs. MW and NHBA vs. NHBD are shown in Figures 1A,B respectively. Furthermore, the Tanimoto similarity matrix (Figure 2) between bitterants and non-bitterants is calculated based on the 2048bit-ECFP6 due to its more features and less bit collisions.
|
study
| 100.0 |
In order to derive and validate the bitter/bitterless classification model, the whole dataset is randomly split into two chunks: the dataset for the cross-validation and the test set for the independent validation. The detailed data-splitting scheme is given as follows: 20% of the bitterants and 20% of the non-bitterants (141 bitterants and 118 non-bitterants) randomly selected from the whole dataset are treated as the test set (Dataset-Test), while the rest of them (566 bitter and 474 bitterless compounds) are adopted to train the model with the cross-validation (CV), which is denoted as Dataset-CV. In the five-fold cross-validation, Dataset-CV is randomly split into five chunks. One chunk is employed as the internal validation set (Dataset-Internal-Validation), and the remaining four chunks are combined to form the training set (Dataset-Training). This procedure will be repeated for five times, which is prepared for the five-fold cross-validation. Finally, to reduce the bias from the data-splitting scheme, the whole data-splitting procedure will be repeated for nineteen or three times depending on the different machine-learning methods. More specifically, 19 data-splitting schemes are adopted for the model-training with KNN, SVM, GBM, and RF, while only three data-splitting schemes are used for the model-training with DNN2 and DNN3 that are very computationally expensive.
|
study
| 99.9 |
In this work, Extended-connectivity Fingerprint (ECFP) (Rogers and Hahn, 2010) is adopted as the molecular descriptor. Thus, ECFP is implemented natively in our e-Bitter program due to the following three factors. (1) ECFP, one typical class of topological fingerprints, was manifested to be powerful in the classification (Ekins et al., 2010; Rogers and Hahn, 2010; Chen et al., 2011; Hu et al., 2012; Braga et al., 2015, 2017; Koutsoukas et al., 2016; Rodríguez-Pérez et al., 2017; Varsou et al., 2017; Wang et al., 2017; Yang et al., 2017). However, ECFP has not been applied to the general classification of diverse bitter/bitterless compounds in the literature. (2) The existing softwares with the ECFP function such as Pipeline Pilot (http://accelrys.com), JCHEM (https://www.chemaxon.com), and RDKit (http://www.rdkit.org) etc. cannot provide a facile and intuitive mean to highlight the fingerprint bit “1” in the context of the 3D structure and also cannot inform us the importance of each bit. But it is worth mentioning that Bioalerts program (Cortes-Ciriano, 2016), which is developed based on the RDKit, can offer a very useful function to generate the 2D structure image highlighting with one ECFP bit. Nevertheless, Bioalerts doesn't have a 3D graphic frontend to support the interactive visualization. (3) The native integration of ECFP in our e-Bitter program will decrease the software dependency on any other packages and will be convenient for the users to deploy this program on their own computers. The implementation of ECFP is given as follows.
|
study
| 99.94 |
The generation procedure of ECFP can be divided into the following steps (Rogers and Hahn, 2010). (1) Initial assignment of the atom identifiers. The initial integer identifiers, which are assigned to all the non-hydrogen atoms of the given molecule, encode the local information about the corresponding atom such as Sybyl atom type, atomic number, and connection count etc. (2) Iterative update of the atom identifiers. Each atom identifier is recursively updated to reflect the identifiers of each atom's neighbors till a specified diameter is reached. The commonly defined diameter is 4 or 6. (3) Duplication elimination of the atom identifiers. The multiple identifiers representing the equivalent atom-environment are removed. (4) Folding operation of the atom identifiers. All the identifiers are mapped into the bit string with the fix-sized length, in spite of the occasional bit collision. The frequently-used length of the fingerprint is 1,024 or 2,048 bits. (5) Record of the structural features. Finally all the fingerprint bits “1,” original identifiers and their corresponding structural features are recorded for the subsequent visualization. This step is purposely designed to couple with our 3D visualization platform. In this study, 1024bit-ECFP4, 2048bit-ECFP4, 1024bit-ECFP6, and 2048bit-ECFP6 will be harnessed in the following model-training, since more bits will reduce the chance of bit collision during the folding operation, while the larger diameters will provide more structural features.
|
study
| 99.94 |
In this work, both full features without the feature selection and feature subset with the feature selection are attempted to examine whether the feature selection is beneficial to our bitterrant/non-bitterant classification. Herein, feature selection is conducted based on the feature importance (Teixeira et al., 2013) derived from the model-training with the random forest (RF) method, which will be elaborated in the following section. Thus, the full features without the feature selection, and the feature subset after the feature selection are adopted as the molecular descriptors to systematically evaluate the performance of bitter/bitterless classification.
|
study
| 99.94 |
In this work, five algorithms (KNN, SVM, RF, GBM, and DNN) will be utilized to train the models via the Scikit-learn, Keras and TensorFlow python libraries, which are fully integrated in the Windows version of python package (Winpython 3.5.4.0) with the download site (https://winpython.github.io/). For the sake of the fine tuning of hyper-parameters, the five-fold cross-validation is conducted to explore the corresponding optimal parameters for each machine-learning method, which will be succinctly introduced as follows.
|
other
| 99.75 |
K-nearest neighbors (KNN) algorithm is a non-parametric method used for the classification, which is based on the closest training instances in the feature space (Itskowitz and Tropsha, 2005). The number of nearest neighbors (K) and the weighting methods will affect the performance of KNN model. In this study, K (1, 3, 5, 7, 9, 11, 13, and 15) and two weighting schemes (uniform weight or distance-dependent weight) are explored during the cross-validation.
|
study
| 99.94 |
Support Vector Machine (SVM) is a popular machine-learning technique that performs the classification by constructing the hyper-planes in the multi-dimensional space that separates the different classes (Vapnik, 1995). The radial basis function (RBF) is used as the kernel, and the grid search is harnessed to optimize the penalty parameter C (1,000, 5,000, 10,000, 50,000, and 100,000) and the kernel parameter gamma (0.0001, 0.0005, 0.001, 0.005, 0.01, and 0.1).
|
other
| 99.9 |
Random forest (RF) is an ensemble learning method by generating the multiple decision trees via the bootstrap sampling of training set and random selection of feature subset from the total descriptors (Breiman, 2001). Finally, RF predicts the class based on the consensus votes from these multiple decision trees. In addition, RF can provide the importance of each feature, which is very useful for the intuitive interpretation of the prediction model and is the key criterion for our feature selection in the following section. The number of decision trees (10, 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, and 1,000) will be probed during the cross-validation.
|
other
| 99.0 |
Gradient boosting machine (GBM) is also an ensemble machine learning technique to construct the multiple decision trees in a step-wise manner. Each decision tree is not randomly generated as in the random forest, but is consecutively built to give a better estimate of the response variable. More specifically, GBM is to stepwisely construct a new decision tree as a weak learner with the maximum correlation to the negative gradient of the loss function (Friedman, 2002). The number of decision trees (10, 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, and 1,000), and the learning rate (0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, and 0.9) will be tried during the cross-validation.
|
other
| 99.7 |
Deep neuron network (DNN) is a neural network with more than one hidden layer between the input and output layers. In DNN, thousands of neurons in each layer can be extensively applied to the dataset with thousands of features, and more advanced regularization technique such as the dropout can be used to prevent the overfitting problem (LeCun et al., 2015). Nevertheless, DNN requires the users to adjust a variety of parameters. The number of epochs, the size of mini-batches and the dropout rate are the most important parameters. The number of epochs refers to the number of times that the model is exposed to the training dataset. The size of mini-batches defines the number of training samples exposed to the model before updating of the weight. The dropout rate is the percentage of neurons that are randomly-selected and ignored during the training. In this study, the number of epochs (100, 200, 300, 400, 500, and 600), the size of mini-batches (60, 80, 100, 120, 140, and 160), and the dropout rate (0.1, 0.2, 0.3, 0.4, and 0.5) will be probed in the cross-validation. The dropout technique is exerted only after each hidden layer. Moreover, four configurations of deep neuron network with the different numbers of hidden layers (2 or 3 layers) and neurons per layer (1,024 or 2,048) are explored, which are defined in detail as follows: DNN2 (Figure S5) contains two hidden layers [input layer: X (1,024 or 2,048) neurons; hidden layer1: X neurons; hidden layer2: X neurons; output layer: 2 neurons], and DNN3 (Figure S6) includes three hidden layers [input layer: X (1,024 or 2,048) neurons; hidden layer1: X neurons; hidden layer2: X neurons; hidden layer3: X neurons; output layer: 2 neurons]. Additionally, the rectified linear unit function (ReLU) is used as the activation function. adam algorithm is adopted as the optimizer and “binary crossentropy” is employed for the loss function.
|
study
| 100.0 |
Upon completion of model-training with the five-fold cross-validation, the optimal parameters and the corresponding best models are achieved based on highest F1-score in Equation (1). Thus, the combination of four ECFP fingerprints, different random splits of the dataset, and different machine-learning methods (KNN, SVM, RF, GBM, DNN2, and DNN3) will totally offer 328 trained models with the optimal parameters in Table S1. Subsequently, all those models are evaluated on the test set with the following metrics: accuracy, precision, specificity, sensitivity, Matthews correlation coefficient (MCC) and F1-score (Equations 1–6), which are also listed in Table S1.
|
study
| 100.0 |
Where TP, TN, FP, and FN refer to the true bitterant, true non-bitterant, false bitterant, and false non-bitterant respectively. F1-score and Matthews correlation coefficient (MCC) are commonly used to measure the quality of binary classifications. F1-score (cross-validation) denotes that F1-score is evaluated on the internal validation dataset during the cross-validation, and F1-score (test test) marks that F1-score is assessed on the test set. ΔF1-score is the absolute value of the difference between F1-score (cross-validation) and F1-score (test set). ΔF1-score is calculated to monitor the potential overfitting or underfitting. If ΔF1-score is small, it means that the model performances are similar on the internal-validation dataset and test set. For the sake of the conciseness, F1-score (test set) is reduced to F1, hence the symbol “F1” specifically means that F1-score is evaluated on the test set by default if there is no additional statement in this work.
|
study
| 99.94 |
Feature selection is commonly adopted to eliminate the redundant features in the machine-learning study. In order to demonstrate whether there is any improvement for our bitterant/non-bitterant classification, feature selection is performed based on the feature importance derived from the random forest (RF) method.
|
study
| 99.9 |
More specifically, as described in the previous section about the model-training without the feature selection, 76 runs of random forest are conducted considering the combination of four ECFP fingerprints and different random splits of the dataset, which will lead to 76 models and the attendant 76 sets of feature importance. Then the feature importance for all the bits in the ECFP fingerprint is sorted descendingly and plotted in Figures S7–S10. Thus, the top 512, 256, and 128 important features (Figures S7–S10) are selected respectively as the typical feature subsets for the following model-training, since the exhaustive and systematic scan of feature number ranging from 1 to fingerprint length is really time-consuming especially for the training of deep neuron networks such as DNN2 and DNN3.
|
study
| 100.0 |
Subsequently, each set of important features are combined with the machine-learning algorithms (KNN, SVM, GBM, RF, DNN2, and DNN3) to train the models respectively. The training process is nearly identical to the aforementioned model-training without the feature selection. The only difference is existed in the configuration of DNN: DNN2 with two hidden layers (input layer: X (512, 256, or 128) neurons; hidden layer1: X neurons; hidden layer2: X neurons; output layer: 2 neurons) and DNN3 with three hidden layers (input layer: X (512, 256, or 128) neurons; hidden layer1: X neurons; hidden layer2: X neurons; hidden layer3: X neurons; output layer: 2 neurons). Thus, the combination of three sets of important features and six machine-learning methods (KNN, SVM, RF, GBM, DNN2, and DNN3), different random data-splitting schemes (three splits for DNN2/DNN3 and nineteen splits for the others) and four ECFPs will lead to 984 models.
|
study
| 99.94 |
After the five-fold cross-validation, the best models are harvested according to the highest F1-scores, and then all the best 984 models are assessed on the test set, which are appended to Table S1. Hence 1,312 models including 984 models with feature selection and 328 models without feature selection are obtained. To reduce the bias from the random data splitting, 96 average models (AM) are derived from 1,312 individual models by averaging over the different data splitting schemes and are tabulated in Table S2.
|
study
| 100.0 |
Y-randomization test (Rücker et al., 2007) is conducted to inspect the reliability of all the 1,312 models. In this test, the experimentally observed labels (bitter or bitterless) for Dataset-CV are randomly shuffled without changing the total number of bitterants and non-bitterants (Table S3). Worthy of notice is that some labels are still correct due to this random operation. Thus, the newly generated Dataset-CV still contains some true samples but with lot of noise, and its detail is described in Table S3. Subsequently, the five-fold cross-validation on this new dataset is performed with exactly the same molecular descriptors and protocols mentioned in the previous section about the model-training. The best models are determined based on the highest F1-scores assessed on the internal validation dataset during the cross-validation, and further evaluated on the test set (Dataset-Test) without any random shuffling. All the evaluation metrics are collected in Table S4.
|
study
| 100.0 |
Generally speaking, compounds that are highly dissimilar from all the compounds used in the model-training cannot be predicted reliably (Tropsha, 2010), thus the applicability domain of our models should be defined in accordance with the guideline of Organization for Economic Cooperation and Development (OECD). In this work, each compound in the test set (Dataset-Test) is compared with every compound in the cross-validation dataset (Dataset-CV) according to the Tanimoto similarity based on 2048bit-ECFP6 fingerprint due to its more structural features and less bit collisions. Subsequently, five most similar compounds from Dataset-CV are retrieved and treated as five nearest neighbors for the given compound in Dataset-Test, and the average of five similarities is defined as the “average-similarity” between this given compound and these five nearest neighbors. It should be noted that five nearest neighbors are selected here, because the optimal nearest neighbors is five for the best KNN model with full 2048bit-ECFP6 (M0255 in Table S1) based on the highest F1-score (0.927).
|
study
| 100.0 |
Following the definition above, each compound in Dataset-Test finds its own five nearest neighbors in Dataset-CV and compute its corresponding average-similarity. Similarly, each compound in Dataset-CV also retrieves its own five nearest neighbors in Dataset-CV and calculates its corresponding average-similarity. Finally the histograms of the average-similarity for Dataset-Test and Dataset-CV are given in Figure S11 to address the applicability domain of our models. This average-similarity is used to reflect the closeness between the given compound and its neighboring compounds in the cross-validation dataset (Dataset-CV). If the average-similarity is close to 1, it means that the given compound can find very similar compounds in the training set, and the prediction for the given compound based on our models is not extrapolated and can be considered as a reliable inference. Nevertheless, in reality it is often very difficult for us to expect that the compound of user's interest can always find very similar neighboring compounds in our dataset. Thus, an appropriate threshold for the average-similarity should be defined based on Figure S11.
|
study
| 100.0 |
In this work, 1,312 individual models (M0001-M1312 in Table S1) and 96 average models (AM01-AM96 in Table S2) are achieved. Although all the models are public available and can be used for the bitterant prediction through the flexible function of “customized model” in our e-Bitter program. However, it would be confusing for the users without any recommendation.
|
other
| 99.5 |
Thus, nine consensus models are proposed based on the balance among the accuracy, speed and diversity of machine-learning methods, and are implicitly integrated in our e-Bitter program. Consensus model 1 (CM01) selects 19 best individual models (Table S5) from Table S1 purely based on the highest F1-scores in each data-splitting scheme. Consensus model 2 (CM02) chooses the average models (AM32, AM28, AM31, AM11, and AM69 in Table S6) considering each machine-learning method with the highest F1-scores to balance the diversity and performance of machine-learning methods. Consensus model 3 (CM03) considers the top average models (AM32, AM26, AM28, AM62, and AM52 in Table S7) with the highest F1-scores. Consensus model 4 (CM04) selects the top five average models (AM31, AM49, AM55, AM67, and AM43 in Table S8) trained with KNN. Consensus model 5 (CM05) comprises the top five average models (AM32, AM26, AM62, AM50, and AM56 in Table S9) trained with SVM. Consensus model 6 (CM06) includes the top five average models (AM69, AM63, AM51, AM33, and AM57 in Table S10) trained with GBM. Consensus model 7 (CM07) combines the top five average models (AM28, AM52, AM10, AM46, and AM70 in Table S11) trained with RF. Consensus model 8 (CM08) consists of the top average models (AM11, AM23, AM35, AM05, and AM29 in Table S12) trained with DNN2. Consensus model 9 (CM09) contains the top average models (AM06, AM36, AM12, AM18, and AM30 in Table S13) trained with DNN3. All the evaluation metrics for each consensus model (Table 1) are obtained by averaging over all the constituent models.
|
study
| 100.0 |
(1) No standard deviation can be given for BitterPredict, since only one random data-splitting scheme is adopted in their work. (2) Three random data-splitting schemes are used for BitterX, thus the evaluation metrics for BitterX are shown here via averaging over three different data-splitting schemes.
|
other
| 99.9 |
Models from the BitterX, BitterPredict and e-Bitter will be compared in two manners: (1) the direct comparison of F1-score (test set) and MCC (test set), which are derived from their own works and (2) the more objective comparison on three external test sets from the recent work of Wiener et al.
|
study
| 99.56 |
For the first direct comparison, two performance indicators F1-score (test set) and MCC (test set) should be given for each model. However, both evaluation metrics are not directly reported in the works of Wiener et al. and Huang et al. thus F1-score (test set) and MCC (test set) are indirectly derived from their works. To vividly demonstrate the performance of each model, the scatter plot of MCC (test set) vs. F1-score (test set) is plotted in Figure 3 based on Table 1 and Table S2.
|
study
| 100.0 |
The scatter plot of MCC (test set) vs. F1-score (test set) for our models (9 consensus models and 96 average models) and the models from BitterPredict and BitterX. The MCC (test set) and F1-score (test set) for the models from BitterPredict and BitterX are calculated based on the data reported in the original works.
|
study
| 100.0 |
For the further fair comparison, three external independent test sets from the recent work of Dagan-Wiener et al. (2017) are used for the independent assessment and are given as follows: “Bitter New” dataset (23 bitterants), “UNIMI set” dataset (23 bitterants and 33 non-bitterants) and “Phytochemical Dictionary” dataset (49 bitterants and 26 non-bitterants). The prediction results by BitterPredict for these three test sets are reported in the work of Dagan-Wiener et al. (2017) and compiled in Tables 2–4 for the convenience of comparison. BitterX prediction is conducted by the manual uploading of each molecule to the web server one by one, which gives results in Tables 2–4. The prediction by e-Bitter is performed in batch for these three datasets and offers the results in Tables 2–4. In addition, scatter plot of MCC (test set) vs. F1-score (test set) or accuracy (test set) vs. F1-score (test set) for all the models are plotted in Figures 4–6.
|
study
| 100.0 |
(1) No standard deviation can be given in this table, since only one unambiguous predicted label should be provided by our e-Bitter and this predicted label is obtained from the predicted probability averaging over all their respective constituent models. (2) Specificity (test set) and MCC (test set) cannot be given due to the zero number of TN and FP.
|
other
| 93.0 |
The scatter plot of accuracy (test set) vs. F1-score (test set) for all the models (9 consensus models and 96 average models) evaluated on the “Bitter New” dataset with 23 bitterants. Accuracy instead of MCC is used as Y axis, since MCC cannot be calculated due to the zero number of TF and FP in this dataset without the non-bitterants.
|
study
| 100.0 |
The feature importance of ECFP bit is also of particular interest for us. In this work, feature importance of each ECFP bit is derived from the random forest (RF) after the five-fold cross-validation as mentioned above, and feature importance will be automatically linked to the corresponding fingerprint bit “1” and structural feature in our e-Bitter program. However, the feature importance from RF can only tell us whether these features are vital to the bitter/bitterless classification, but cannot inform us whether each ECFP bit “1” in a compound positively or negatively influences the bitterness, which can be described by the concept “feature partial derivative.” (Hasegawa et al., 2010; Marcou et al., 2012; Polishchuk, 2017).
|
study
| 100.0 |
The feature partial derivative, exactly defined by Equation (8) and Hasegawa et al. (2010) is firstly proposed in the work of Byvatov and Schneider (2004) and is systematically reviewed in the work of Polishchuk (2017). To derive the feature partial derivative of each ECFP bit, the backward finite difference approach is adopted and briefly described as follows (Hasegawa et al., 2010). Firstly, the fingerprint bit nullification is simply done by the replacement of the bit “1” with zero, and then the difference between the predicted probabilities, from the original prediction and the new prediction after the bit nullification, is defined as the feature partial derivative for this fingerprint bit. If the feature partial derivative for one bit is positive, it means that this bit “1” is important to the bitterness of this compound in the positive manner, otherwise, this bit negatively affect the bitterness of this compound. This procedure is repeated for each bit “1” in a compound, thus the feature partial derivative of each ECFP bit in the compound can be derived, which can be done automatically in the e-Bitter program.
|
study
| 99.94 |
In order to automate the whole process, e-Bitter is developed for the convenience of users. In the current implementation, there are two basic parts. One is the generation and visualization of ECFP fingerprint, which is natively implemented in the e-Bitter program; the other is the underlying model-prediction with the diverse machine-learning approaches via evoking the external Scikit-learn, Keras, and TensorFlow python libraries natively integrated in the Winpython v3.5.4.0. For the sake of seamless fusion between these two parts, various python scripts have been implemented and integrated in the e-Bitter program.
|
other
| 99.9 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.