Fig 1.
Coexpression has been calculated and is well known from intra-tissue-centered data (bottom scenarios from v to viii, generally in z-score alike procedures). Nevertheless, coexpression at their whole expression range across tissues or inter-tissue coexpression is not commonly analyzed (in Transcripts Per Millon scale, for example, top scenarios from i to iv). This estimation corresponds to a whole-organism or system-level view of coexpression, opening new opportunities for research and comprehension of gene function and regulation. Examples of discrepancies among system vs. tissue level estimations are indicated in further figures.
Fig 2.
CoGTEx data processing workflow.
GTEx samples were clustered by similitude and cleaned by removing samples showing some degree of tissue mixture, contamination, or confusion. The 42 tissue clusters were used to robustly estimate the minimum correlation from 20 random 70-sample subset selections per cluster. Pearson, Spearman, and G-statistics are calculated as coexpression measures in two modalities, at system-level gene expression and tissue z-scores. Additionally, the correlation among Sex, Age, and Ischemia co-variates are also computed. The databases are available via a web application at http://victortrevino.bioinformatics.mx:8080/cogtex. The 140 correlation matrices estimated result from 2 modalities, 20 random subsamples, 3 metrics, plus 2 modalities in 7 levels of covariates and the 3 minimums of 2 modalities. See Methods for more details.
Fig 3.
Pathway and network enrichment analysis.
Panel A shows the number of associations (hypergeometric test p < 0.05) to KEGG pathways of the top 5% of most coexpressed genes per every gene (absolute correlation) compared to the bottom 5% (which should represent zero correlation). Each boxplot is composed of the number of associations to ~200 KEGG pathways of the 33,445 genes considered. Panel B shows the AUROC per gene classifying positive and negative GIANT gene pairs from tissue-specific networks.
Fig 4.
Comparisons of coexpression estimations.
Panel A shows the distribution of the number of overlapped genes among the top 100 with COXPRESdb. Panel B shows the distribution of the median rank for the top 100 genes in COXPRESdb. Panel C shows the overlap within CoGTEx between system and tissue levels (TPM and Z-scores, respectively) of the top 100 genes. Panel D shows the similitude between overlaps in Panel C. Each dot represents a gene.
Fig 5.
Comparison of system- and tissue- level coexpression for the TP53 gene.
Left scatter plot shows the coexpression ranks of all genes highlighting specific examples shown at the right. Note the low overlap in ranks among the top 100 (within the square). TNFRSF1A gene (marked with dotted circle is shown in S1 File). At the right, the coexpression estimations for the 3 genes marked are shown. Three estimations of coexpression are shown in columns for comparisons corresponding to system- and tissue level in our results, and the equivalent to tissue-level from COXPRESdb. Point colors correspond to different tissues as indicated (see S1 File for details).
Fig 6.
Examples of non-linear or exceptional cases.
Panel A compares the G-statistic and absolute Pearson estimations for the top 1000 coexpressed genes for BRCA. Marked genes show examples of increased G-statistic generated by exceptional tissue clusters. Point colors correspond to different tissues as indicated (see S1 File for details).
Table 1.
Examples of coexpressions from literature.
Table 2.
Top 10 GO biological processes for BRCA1 coexpressed genes from an EnrichR analysis from top 100 genes coexpressed at the tissue-level or at the system-level.
Fig 7.
Comparison of gene set enrichment analyses for three selected genes and four coexpression tools.
Each heatmap shows the rank of significant gene ontology terms for biological processes resulted by a EnrichR analysis from the top 100 coexpressed genes reported by the shown coexpression tools. Colors refer to the position of the rank for the corresponding biological term. Black refers to not significant GO terms or that did not meet the criteria.
Fig 8.
Comparison of coexpression and ChIP-peaks for selected transcriptional regulators.
Each panel shows the ChIP-peaks for a regulator, in z-score at left, or TPM at right. Vertical black lines represent ChIP-peaks for the corresponding coexpressed gene. The height of each line corresponds to the cumulative number of genes showing a peak. The red line corresponds to the expected ratio, which is estimated by the total genes showing a peak divided by the 33,445 genes. Thus, for random order the height of black lines should match the red line as shown in the bottom panels for two random orders of genes (for ARID4B and FOXA1 data). Deviations from the red line should correspond to higher association between peaks and the order of co-expressed genes. To quantify the deviation, a “D” score was estimated by the sum of the absolute difference between the observed counts (cyan) and the expected counts (red) divided by the total area of the triangle. This D-score represent the fraction of the area that is deviated from the expected. The D-score of ARID4B in z-score is 0.3452 which is almost 100 times higher than random D = 0.0035. The lowest D-score correspond to 0.0259 of BRCA2 in TPM scale, which is 5.5 times higher than random (mean of 100 random order peaks in BRCA2 is D = 0.0047). Note that the three regulators used show deviations from randomness, although at different degrees.
Fig 9.
CoGTEx web database and interface.
Panel A shows the home page where a gene is first selected. A network tab can be used to extract sub-module estimations. Panel B shows an overall expression comparison among tissue clusters. Panel C shows the top/all coexpressions of the selected gene (TF, transferrin gene in the figure). A coexpression figure can be drawn in the interface and contains clickable tissue clusters to remove (right-click) or include (left-click) tissues. The web address is http://bioinformatics.mx/cogtex.