Strains and Media

All experiments used the diploid S. cerevisiae strain BY4743, which has been previously described [23]. Cells were grown using standard techniques [24].

Growth rate analysis

Cells were grown in 96 well plates in 100 μL cultures (10 μL of starter culture and 90 μL of fresh media) with continuous shaking in a BioTek Synergy 4 Hybrid Microplate Reader for 10 h. Growth rates were assayed in 8 conditions: (1) Synthetic complete medium with glucose (ScD) at 30°C, (2) ScD at 37°C, (3) Synthetic complete medium with glycerol (ScG) at 30°C, (4) ScG at 37°C, (5) Yeast extract, peptone medium with glucose (YPD) at 30°C, (6) YPD at 37°C, (7) Yeast extract, peptone medium with glycerol (YPG) at 30°C, and (8) YPG at 37°C. Absorbance was measured at 660nm at 3 min intervals. Using custom R scripts, the doubling times were calculated from the linear regression curve through the log growth phase using the log of the absorbance and time of growth. A two-tailed t-test of independence with Bonferroni correction for the 11 comparisons (7 comparisons of the control, YPD at 30°C, to the test conditions, 3 comparisons of the observed concurrent double stimuli effect to the expected sum of individual stimulus effects, and 1 comparison of the observed concurrent three stimuli effect to the expected sum of the effects of the three individual stimulus) was used to calculate the statistical significance of a stimulus effect on the growth rate [25].

Preparation of yeast protein extracts

Five mL of YPD (1% yeast extract, 2% peptone, 2% glucose) was inoculated with a single yeast colony from a YPD agar plate and grown overnight. Three replicates were grown under each growth condition: YPD at 30°C and 37°C and YPG at 30°C and 37°C. Fifty mL of YPD was inoculated with 50 μL of the overnight culture and grown at 30°C and 37°C. One hundred mL of YPG (1% yeast extract, 2% peptone, and 3% glycerol v/v) was inoculated with overnight cultures and grown at 30°C and 37°C. The cultures were grown with constant shaking at 175 rpm in Innova 44 shaker incubators (New Brunswick Scientific). For all four growth conditions, cells were harvested at mid-log phase as determined by OD₆₀₀ measurements. Cells grown in YPD were harvested after 14 h, while cells grown in YPG were harvested after 24 h. All cultures were centrifuged at 2000 rpm for 5 min at 4°C using a Sorvall HLR6/H600A/HBB6 rotor in Sorvall RC-3B centrifuge and washed with ice cold deionized H₂0. The cell pellets were resuspended in 1 mL ice cold wash buffer (10 mM Tris pH 8.0, 5 mM beta-mercaptoethanol, 500 mM ammonium chloride, 100 mM magnesium acetate) and lysed at 4°C using glass beads and a Bead Beater (BioSpec, Inc) for 10 min as previously described [26]. The whole cell extracts (WCE) were clarified by centrifugation at 20,000g for 15 min at 4°C, and a 200 μL aliquot of the cleared WCE was stored at -80°C.

Isobaric tag for relative and absolute quantitation (iTRAQ) labeling

The total protein concentration was determined using a Bradford assay according to the manufacturer’s protocol (Sigma Aldrich). For each growth condition, 50 μg of total protein was mixed with 50 ng of bovine serum albumin (Thermo Scientific) as an internal standard. Each protein sample was acetone precipitated and resolublized in 25 μl iTRAQ dissolution buffer (500 mM triethylammonium bicarbonate, 0.1% sodium dodecyl sulfate). The proteins were reduced with tris(2-carboxyethyl)phosphine at 60°C for 60 min and the cysteines were derivatized with methyl methanethiosulfonate at room temperature for 10 min. All samples were digested with sequencing-grade modified trypsin (1:50; Promega Corporation) overnight at 37°C. Equal fractions of the tryptic digests from the three replicates grown in YPD at 30°C were pooled separately and used as a control for the iTRAQ experiments. Fifty μg of the pooled control and 50 μg of each of the replicates were used for iTRAQ labeling. The iTRAQ labeling reagents were resolublized in 150 μL anhydrous ethanol (Sigma Aldrich). 75 μL of iTRAQ reagent solutions were added to each 50 μg sample, incubated with shaking for 1 h at room temperature on an Eppendorf Thermomixer R, pooled, frozen, lyophilized, resolublized in 200 μL of buffer A (0.1% formic acid), and stored at -80°C.

Liquid chromatography and mass spectrometry

The iTRAQ-labeled samples were analyzed with MudPIT as previously described [27]. Precursor ions were analyzed in the Orbitrap mass analyzer followed by four CID fragment ion scans in the ion trap and four HCD fragment ion scans (normalized collision energy = 45%) in the Orbitrap.

iTRAQ data analysis: RAW files generated by the MudPIT experiments were searched using the Sequest HT database search engine running under Proteome Discoverer v1.4 (Thermo Scientific) against a forward and reverse yeast protein database (S.cererevisiae_orf_trans_all_SGD.fasta.6718) with appended common contaminant sequences [28,29]. Beta-methylthiolation and iTRAQ modifications were included as constant modifications. Oxidation of methionine and tryptophan, and deamidation of glutamine and asparagine were used as variable modifications. Precursor mass tolerance was set to 3 Da and fragment mass tolerance was set to 0.8 Da. Protein assembly, reporter ion quantitation, and protein fold change calculations were done using ProteoIQ at 5% peptide and protein FDR (Premier Biosoft). Hierarchical clustering analysis was done using Cluster 3.0 [30]. Heatmaps were generated using Java Treeview [31]. Circos plots were generated as described in Krzywinski et. al. to visualize the genomic locations of the quantitated proteins [32]. For better visualization, only those regions of the genome that were quantitated in this study are shown. The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE partner repository with the dataset identifier PXD002371 [33].

Environmental interaction analysis

All analysis was performed using R scripts to parse the fold change expression data to identify proteins that show specific expression patterns in response to complex environmental stimuli. For each protein, we used linear regression to test for any association of high temperature or glycerol using a model that included main effects for glycerol and temperature and the glycerol by temperature interaction. We used the effect size estimates and ANOVA p-values (3 degrees of freedom) calculated by the lm function and adjusted the p-values for a 5% FDR using the Benjamini-Hochberg procedure for finding differentially expressed proteins [34]. We used the adjusted p-value cut-off of 0.05 to determine statistical significance. If the overall adjusted p-value was greater than 0.05, we classified the proteins as non-responders. The positive and negative signs of the effect size estimates correspond to upregulation and downregulation, respectively, showing the direction of change. The remaining proteins were further classified into environmental interaction classes based upon the effect size estimate p-values and the direction of change. If the p-value of an estimate was less than 0.05, the protein was considered differentially expressed in response to that environmental stimulus.

To test if a protein is affected by environmental epistasis, the effect size estimates for the individual high temperature (HT) and glycerol stimuli (G) were summed, the combined standard error calculated as the square root of the sum of the squared standard errors, and a two-sample t-test of independence was used to compare the summed effect size estimate to the effect size estimate for the concurrent high temperature and glycerol stimuli (HT+G). If a t-test p-value was less than 0.05, the protein was assumed to be affected by environmental epistasis.

Environmental interaction analysis of transcriptomic dataset

Normalized expression data described in Knijnenburg et. al. was downloaded [13]. The transcriptomic data were generated using haploid S. cerevisiae (CEN.PK113-7D MATa) cells grown in chemostat cultures [13]. We chose 4 culture conditions similar to our experimental design for further analysis. The culture conditions tested were: 1) with ammonium sulfate as the nitrogen source (n = 5), 2) with methionine as the nitrogen source (n = 3), 3) anaerobic conditions (n = 4), and 4) with methionine as the nitrogen source and anaerobic conditions concurrently (n = 3). Transcriptomic data from the cells grown with ammonium sulfate as the nitrogen source were used as the baseline control. The fold change was calculated by subtracting the average normalized expression data of baseline samples from the individual expression data. Finally, the genes were classified into various types of environmental interactions as described above.

Co-expression network analysis

Sparse PArtial Correlation Estimation (SPACE) was used to build protein co-expression networks and identify the hub genes [35]. To account for outliers, the data were normalized using probabilistic quotient normalization and scaled using a generalized logarithmic scaling factor [36,37]. The data were scaled and centered to have a standard deviation of 1 and mean of 0 to remove any bias in the correlation analysis [38]. We estimated the partial correlation matrix using the space.dew method implemented in the SPACE R package [35]. We selected the default value of the tuning parameter for constructing the initial network [35]. The network was visualized in Cytoscape 3.1.1 [39].

Stimuli-specific expression patterns can be used to identify proteins important for responding to the stimuli.

We observed 283 proteins differentially expressed with high temperature, 379 proteins differentially expressed in response to glycerol, and 370 proteins were differentially expressed in concurrent high temperature and glycerol (Fig 1B and 1D, and S1 Table), while 41 proteins did not change in response to any of the stimuli. We selected GeneMANIA Cytoscape plugin for pathway analysis since it extends the input list of differentially expressed proteins by adding related proteins to enhance sensitivity and coverage [42,43]. It also allows using the complete proteome as the background. This helped to build a more complete picture of differentially regulated pathways. Pathway analysis of these two differentially expressed protein groups revealed the same top five pathways; none were specific to either stimulus (Fig 1C and 1E). All of the top five pathways were related to protein synthesis and translational control, suggesting that the regulation of protein synthesis is an important step in responding to environmental stimuli. Translation factors are some of the most abundant proteins in yeast and our proteomic assays are limited by the abundance of proteins in the cell. Although this could have confounded pathway analysis and led to the identification of translation associated pathways as being the most enriched, using only the differentially expressed proteins suggests that these pathways are, at the least, being differentially regulated. Furthermore, similar observations have also been made in previous studies [1,10,44]. It is noteworthy that the pathways expected to be important for responding to these stimuli, such as “protein folding” for growth at high temperature and “TCA cycle” for growth with glycerol were present farther down the list at numbers 39 and 53, respectively (S2 and S3 Tables) [45,46,47]. This mirrors a common problem in ‘omics’ studies that generate large lists of candidate genes, transcripts and proteins. The important responders are lost in a long list where a majority of differentially expressed genes or proteins is not directly responding to the stimulus. Therefore, choosing candidates for an in-depth mechanistic study becomes a challenge.

To address this problem, we devised a methodology using dominance in environmental interactions to identify proteins and pathways important for responding to a stimulus. We noticed proteomic expression patterns in which the response to one stimulus was dominant over the other. We speculated that a protein critical in responding to a stimulus will respond to that stimulus even when challenged by a competing stimulus. If this hypothesis is correct, such an environmental interaction could serve as a filter to select and identify proteins that respond specifically to the dominant environmental stimulus.

To test this hypothesis, we classified the list of 466 proteins responding to the concurrent glycerol and high temperature stimuli based upon their expression patterns. Two classes of dominant environmental interactions are possible. In one class, a stimulus reverses an expression change induced by the other stimulus (Fig 2A and 2B, top panels, rows 1 and 3). In the other class, a stimulus induces a change in expression, while the other stimulus has no effect on its own and does not change the response to the concurrent stimulus (Fig 2A and 2B top panels, rows 2 and 4). Each class is represented by two theoretical expression patterns for a total of four expression patterns for each stimulus (Fig 2A and 2B top panels).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Dominance of an environmental stimulus used to identify proteins that are important for responding to the environmental stimulus.

The color bar shows the range of fold changes. Pathway analysis was done using the GeneMANIA Cytoscape plugin [42]. Bar graphs were generated in Graphpad Prism. A) Proteins for which HT stimulus is dominant over G stimulus. The theoretical expression patterns are depicted in the top panel (Red, upregulation; green, downregulation; and black no statistically significant change in expression). The heatmap of fold changes in expression for 30 proteins for which HT stimulus is dominant over G stimulus is shown in bottom panel. B) Proteins for which G stimulus is dominant over HT stimulus. The theoretical expression patterns are depicted in the top panel. The heatmap of fold changes in expressions for 121 proteins for which G stimulus is dominant over HT stimulus is shown in bottom panel. C) Bar graph shows the–log q-value of enrichments of the top five pathways in the list of proteins for which HT stimulus is dominant over G stimulus. D) Bar graph shows the–log q-value of enrichments of the top five pathways in the list of proteins for which G stimulus is dominant over HT stimulus.

https://doi.org/10.1371/journal.pone.0134099.g002

For the environmental interactions in which the HT stimulus was dominant over the G stimulus, the p-values for all of the effect size estimates were less than 0.05. The changes for the HT and HT+G stimuli were in the same direction and differed from the G stimulus (Fig 2A, top panel, rows 1 and 3). Alternatively, the p-values for only the HT and HT+G stimuli effect size estimates were less than 0.05 and the directions of change for the HT and HT+G stimuli were the same (Fig 2A, top panel, rows 2 and 4). In all, we identified 30 proteins for which the response to the HT stimulus was dominant over the G stimulus (Fig 2A and S1 Table). We used pathway analysis to identify which protein classes were responding to the dominant stimulus. The group of proteins for which the HT stimulus was dominant included the heat shock response proteins HSP10, HSP60, SSA1, SSA2, and HSP150 (Fig 2A bottom panel, and S1 Table). Pathway analysis of these 30 proteins showed that the top five enriched pathways included protein folding, protein refolding, and unfolded protein binding (Fig 2C, S5 Table). These pathways are expected to be important for growth at higher temperatures [45,46,48,49].

For the environmental interaction in which the G stimulus is dominant, we saw a similar set of patterns as described above except the G stimulus dominates the HT stimulus (Fig 2B, top panel). There are 121 proteins for which the response to the G stimulus was dominant over the HT stimulus (Fig 2B, bottom panel and S1 Table). The group of proteins for which the G stimulus was dominant includes metabolic enzymes such as CDC19, ACO1, and LSC1. (Fig 2B, bottom panel, and S1 Table). Pathway analysis of these 121 proteins showed that the top five pathways included the oxidation-reduction process, the generation of precursor metabolites and energy, and the tricarboxylic acid cycle (Fig 2D, and S6 Table). All of these three pathways are expected to be important for respiratory growth [47,50,51]. Consistent with our hypothesis, pathway analysis of proteins that respond to a dominant environmental stimulus reveals a functional relationship to the response to the stimulus. High temperature has a dominant effect on proteins involved in protein folding, while glycerol has a dominant effect on proteins involved in respiratory metabolism. These results show the practical applications of using dominant environmental interactions to identify proteins that respond to specific stimuli and that are directly involved in the cell’s response to that stimulus.

Analysis of expression patterns reveals that environmental interactions mirror gene interactions.

In addition to the dominant interactions of concurrent environmental stimuli, we observed other classes of environmental interactions that mirror gene interactions. First, we observed a class of proteins whose abundance either increased or decreased in response to both the individual stimuli as well as the concurrent stimuli (Fig 3A). This is similar to gene pairs in which both the individual mutants as well as the double mutant have the same phenotype. We classified these proteins as non-specific environmental responders. This class is represented by two theoretical expression patterns: activated or repressed (Fig 3A, top panel and S1 Table). For these non-specific environmental response modules, the p-values for all the effect size estimates were less than 0.05 and the directions of change were the same (Fig 3A, top panel). We identified 175 proteins that correspond to these patterns, and pathway analysis revealed that they are largely involved in protein synthesis and translational control (Fig 3A, bottom panel and 3D, and S7 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Proteins in different environmental interaction classes and the corresponding enriched pathways after concurrent G and HT stimuli.

The color bar shows the range of fold changes. Pathway analysis was done using GeneMANIA Cytoscape plugin[42]. Bar graphs were generated in Graphpad Prism. A) Non-specific environmental response (NER) proteins to individual and concurrent HT and G environmental stimuli. The theoretical expression patterns are shown in the top panel. The fold changes of 175 NER proteins are shown as a heatmap. B) The theoretical expression patterns for discordant environmental interaction are shown in the top panel. The fold changes of 41 proteins are shown as a heatmap. C) The theoretical expression patterns for suppression environmental interaction are shown in the top panel. The fold changes of the 58 proteins affected by suppression are shown as a heatmap. D) Bar graph shows the–log q-value of enrichments for the top 5 pathways for the non-specific environmental response proteins. E) Bar graph shows the–log q-value of enrichments for the top 5 pathways in the list of proteins affected by discordant environmental interaction. F) Bar graph shows the–log q-value of enrichments for the top 5 pathways in the list of proteins affected by suppression environmental interaction.

https://doi.org/10.1371/journal.pone.0134099.g003

We also observed proteomic responses to concurrent environmental stimuli similar to gene interactions in which the two single mutants are wild-type or have one phenotype, while the double mutant has a different phenotype (Fig 3B). This class includes proteins whose expression was either decreased or unchanged after a single stimulus but was increased if both stimuli were applied concurrently. The class also includes proteins whose expression was either increased or unchanged after a single stimulus but was decreased by the concurrent stimuli. We classified this environmental interaction group as a discordant class. There are eight theoretical expression profiles in the discordant environmental interaction class (Fig 3B, top panel). For the discordant environmental interaction, the p-value for the HT+G concurrent stimuli effect size estimate was less than 0.05 and the directions of change for either the HT or G stimuli were not the same as HT+G. We identified 41 proteins that show discordance (Fig 3B, bottom panel and S1 Table). They are mainly involved in protein synthesis and metabolic pathways (Fig 3E, and S8 Table).

Finally, we observed suppression, in which a protein’s abundance changed in response to a single stimulus, yet the change was suppressed by the simultaneous application of the second stimulus (Fig 3C). This class is similar to gene interactions in which double mutants show the wild-type phenotype [52,53]. The suppression class is represented by eight theoretical expression patterns (Fig 3C, top panel). For suppression environmental interactions, the p-value for the HT+G effect size estimate was more than 0.05, and the p-value for at least one of HT and G stimuli effect size estimates was less than 0.05. We identified 58 proteins that are affected by suppression (Fig 3C, bottom panel and S1 Table). Pathway analysis revealed that metabolic pathways are most affected by suppression (Fig 3F and S9 Table).

A large fraction of the proteome is affected by environmental epistasis

An important feature of genetic epistasis is that the modulating effects of multiple genes are not always independent of each other [20,21,22,54,55]. In many cases, non-independence is diagnostic of a functional relationship between genes [20,22,54]. Genetic epistasis is used to test if the effects of genetic elements are independent. Genetic epistasis occurs when the effects are not independent. We tested if the effects of these two individual environmental stimuli were independent of each other for individual proteins in the proteome. Similar to the mathematical approach to genetic epistasis, we measured the response of each protein and classified a response as influenced by environmental epistasis if the sum of the effects of the individual stimuli for a protein was not equal to the response to the concurrent stimuli (t-test, p-value ≤0.05) [20,21,22]. We used log₂ fold change as the measure of the effect of a stimulus. From our list of 466 quantitated proteins, 240 proteins were affected by environmental epistasis (S1 Table). Pathway analysis of these proteins revealed that a majority of the enriched pathways are involved in protein synthesis and translational control (Fig 4A and S10 Table). The topmost enriched pathways included cytoplasmic translation, cytosolic ribosome, and structural constituent of ribosome (Fig 4A and S11 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Environmental epistasis in the proteomic response to concurrent stimuli.

Pathway analysis was done using the GeneMANIA Cytoscape plugin[42]. Bar graphs were generated in Graphpad Prism. A) Bar graph shows the–log q-value of enrichments of the top 10 pathways in the list of proteins affected by epistasis (purple) and their–log q-value in the list of proteins not affected by epistasis (orange). B) Bar graph shows the–log q-value of enrichments of the top 10 pathways in the list of proteins not affected by epistasis (orange) and their–log q-value in the list of proteins affected by epistasis (purple).

https://doi.org/10.1371/journal.pone.0134099.g004

Pathway analysis of the 226 proteins not affected by environmental epistasis revealed a large number of metabolic pathways (Fig 4B and S1 and S10 Tables). It is interesting to note that the distribution of pathways affected by environmental epistasis is different from those that are unaffected. Protein synthesis and translational control seem to be disproportionately affected by environmental epistasis compared to other pathways. These pathways have previously been found to change in response to the changes in the growth rate [56,57]. If the effects of the two stimuli on the growth rate are not independent, it could explain the observed environmental epistasis. To test the independence in the effects of the two stimuli on the growth rate, we determined the doubling times under the same conditions. The change in the doubling times was used to measure the effect of a stimulus. Our data shows that the effects of high temperature and glycerol on the growth rate are additive and, therefore, independent of each other (S4 Fig). Further studies are required to elucidate the functional significance of the environmental epistasis.

A number of genetic epistasis subtypes have been defined based upon the mathematical models used to measure the expectation of a phenotype in double mutants [54,55,58,59,60]. Four most commonly used definitions are (1) additive, (2) multiplicative, (3) minimum, and (4) log [55,58]. Although we used only the additive definition for developing the concept of the environmental epistasis in this study, future studies can be performed to compare the results obtained using different definitions.

Environmental interactions and epistasis regulate mRNA levels.

Although, we identified the environmental interactions using quantitative proteomic data, we speculated that this framework would be applicable to any quantifiable readout including transcriptomic and phenotypic traits. In pioneering studies using chemostat cultures of S. cerevisiae, Knijnenburg et al. measured the transcriptional response of yeast to multiple, concurrent environmental stimuli [13]. They found linear regression models of expression for the vast majority of genes required a combinatorial interaction term[13]. This suggests the change in transcription of most genes cannot be explained by simply adding the effects of the individual stimuli. Based on our proteomic results, we hypothesized that environmental epistasis plays a role in determining the state of the transcriptome as well.

To test if our environmental interaction and epistasis models are observed in the transcriptomic responses to concurrent stimuli, we analyzed Knijnenburg dataset which measured the transcriptomic responses of yeast cells growing in carbon limited chemostat cultures [13]. In the experiment, two concurrent stimuli were applied: (1) a change in nitrogen source from ammonium sulfate to methionine and (2) a change from aerobic to anaerobic growth (Fig 5A and S12 Table) [13]. The data showed 564 transcripts were affected by environmental epistasis, while 5987 transcripts were not affected (p-value ≤0.05) (S12 Table). In contrast to our proteomic analysis, pathway analysis of the transcripts affected by environmental epistasis revealed enrichment for pathways including microbody, peroxisome, and phytosteroid metabolic process (S13 Table). This could be because of the differences between the strains, stimuli, and culture conditions used in the transcriptomic and our proteomic studies. Similar to our proteomic analysis, we observed dominant environmental interactions in the expression of the transcripts (Fig 5B and 5C and S12 Table). Nitrogen source was dominant for 281 transcripts (Fig 5B and S12 Table). Pathway analysis of these transcripts identified pathways involved in methionine metabolism such as sulfur amino acid metabolic process, sulfur compound metabolic process and methionine metabolic process (Fig 5D and S14 Table). Similarly, anaerobic growth was dominant for 938 transcripts (Fig 5C and S12 Table). Pathway analysis of these differentially expressed transcripts showed enrichment of pathways involved in energy production such as cellular respiration, mitochondrial membrane and respiratory chain (Fig 5E and S15 Table). We also observed the same environmental interaction classes in their transcriptomic data as in our proteomic data, including non-specific environmental response, discordance, and suppression (S12 Table). These results strongly suggest that environmental interactions play a significant role in regulating the biochemical state of cells.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Environmental interactions affect transcriptomic profiles as well.

Normalized expression data from Knijnenburg et. al. 2009 was used for the analyses. The transcriptomic data used in the study used haploid S. cerevisiae cells (CEN.PK113-7D MATa) grown in carbon limited chemostat cultures under 4 conditions – 1) ammonium sulfate as nitrogen source (n = 5), 2) methionine as nitrogen source, NS stimulus (n = 3), 3) Anaerobic condition, AN stimulus (n = 4), and 4) methionine as nitrogen source under anaerobic conditions NS+AN stimulus (n = 3) [13]. Fold changes were calculated from normalized expression data using average normalized expression data from the five replicates of growth with ammonium sulfate as the baseline. The color bar shows the range of fold changes. Pathway analysis was done using GeneMANIA Cytoscape plugin[42]. Bar graphs were generated in Graphpad Prism. A) A heatmap of fold changes of the complete transcriptomics dataset consisting of 6551 transcripts. B) A heatmap showing the fold changes for 281 transcripts for which NS stimulus is dominant. C) The–log q-value of enrichment for the top 5 pathways enriched in the list of transcripts for which NS stimulus is dominant. As anticipated, pathways expected to be involved in metabolization of methionine are enriched. D) A heatmap showing the fold changes for 938 transcripts for which AN stimulus is dominant. E) The–log q-value of enrichment for the top 5 pathways enriched in the list of transcripts for which AN stimulus is dominant. As anticipated, pathways expected to be involved in energy production are enriched.

https://doi.org/10.1371/journal.pone.0134099.g005

Coexpression network analysis shows community structures are guided by environmental interaction and epistasis.

Coexpression networks link together proteins whose expression levels are regulated in the same way [61,62]. As a consequence, coexpression network analysis can be used to determine if the abundances of proteins affected by environmental epistasis are regulated differently than the proteins that are not affected by environmental epistasis. To explore the protein modules whose expression changes are correlated with each other, we built a coexpression network using the merged proteomic responses from both individual and concurrent stimuli using the Sparse PArtial Correlation Estimation approach (SPACE) (Fig 1A) [35]. An edge, representing coexpression, was introduced between two proteins if the correlation between them was above the average of the correlation matrix. To validate the network, we first tested the power law structure of the reconstructed network [35,63]. The reconstructed network followed the power law distribution. The power law parameter α was approximately 4, which is close to the empirically observed value of 3.45 [63]. Next, we repeatedly reconstructed the network by varying the tuning parameter around the default value and fitting the network to the power law distribution. We found that the reconstructed network follows the power law distribution and that the power law parameter was in the range of 3.75. We identified the sub-graph spanned by the top 1% of highly connected nodes. We found that the Jaccard similarity score of these highly connected nodes was 0.83 on the scale of 0 to 1. Therefore, these nodes were classified as hub nodes, which is one of the characteristic features of power law networks. There were 7 hub nodes based upon the above criterion. Next, we checked the significance of the identified hubs using the Wilcox Rank sum test and found that the hub community is statistically significant (p-value = 0.04) [64]. Finally, we compared the reconstructed network with BioGrid protein interaction data and found that approximately 30% of the edges are previously known interactions and that these interactions were found in every reconstructed network when we varied the tuning parameter to estimate the partial correlation matrix [65]. The final coexpression network consisted of 329 nodes with at least one neighbor and a total of 359 edges (Fig 6A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Coexpression network based on all the quantified proteins and all conditions (G stimulus, HT stimulus, and HT+G stimulus).

Proteins are depicted as nodes. Nodes that are coexpressed are connected with an edge. The coexpression network was generated with SPACE algorithm using fold change [35]. Network visualization and analysis was done in Cytoscape 3.1.1 [39]. Bar graphs were generated in Graphpad prism. A) All nodes that have at least one edge. Nodes affected by environmental epistasis are highlighted in purple. The circular layout was used to generate the initial network graphics in Cytoscape 3.1.1. Far-flung communities of inter-connected nodes were manually brought together, while preserving the inner community structure, for better visualization. B) The largest community in the coexpression network. Most of the proteins affected by environmental epistasis are members of a subgraph (top circle) that predominantly contains other proteins that are also affected by environmental epistasis. A similar trend is observed with the proteins not affected by environmental epistasis (bottom circle). C) The numbers of three types edges: 1) both nodes are affected by environmental epistasis (199 edges), 2) neither of the nodes are affected by environmental epistasis (85 edges), and 3) only one of the nodes is affected by environmental epistasis (75 edges). Proteins affected by epistasis are predominantly connected to proteins that are also affected by epistasis. D) Number of edges that connect nodes to other nodes within the same environmental interaction classes (299) or between the classes (60). Co-regulatory connections between proteins are predominantly between those of the same class.

https://doi.org/10.1371/journal.pone.0134099.g006

The largest community within this network includes 205 nodes and 249 edges, with two clearly separate sub-graphs connected by a single node (Fig 6B). Interestingly, one sub-graph consists predominantly of proteins affected by environmental epistasis while the second sub-graph consists of proteins not affected by environmental epistasis. Within the global coexpression network, we observed that proteins affected by epistasis were more likely to be linked with each other than with proteins that are not affected by epistasis and vice versa (Fig 6A). There are 199 edges between two proteins affected by epistasis and 85 edges between two proteins not affected by epistasis. However, only 75 edges involved proteins of both types (Fig 6C). This structural organization of the coexpression network suggests that the responses of proteins affected by environmental epistasis are controlled by a different mechanism than the responses of those not affected by environmental epistasis.

Previous studies indicate that proteins linked in a coexpression network are likely to function in the same pathway [61]. We hypothesized that the grouping of proteins upon classification into environmental interaction classes might be driven by their functional associations. If true, we would expect to find more edges in the coexpression network between proteins within the environmental classes. Indeed, we found this result in this network. Our data show that 299 of the edges (83%) are between proteins in the same environmental interaction class, while only 60 are between proteins in different classes (Fig 6D).