Figure 1.
Overview of the workflow and organization of the manuscript.
Figure 2.
Experimental factors affecting differentially-expressed genes (A), occurrence of different kinds of differentially-expressed links (B), and examples of types of links (C–E).
A. Fraction of variance [92] of transcriptional change for the whole set of analyzed transcripts (blue) and for cyp1a (red), associated with different experimental factors. While developmental changes dominated, dioxin modulated as many as 2009 genes at one or more time points in the experiment (detected by 2-way ANOVA on factors “DAY” and “TREATMENT”, applied to individual genes). 95% confidence intervals are too small to be seen in most cases. cyp1a regulation was largely treatment-dependent (the horizontal legends are identical to the whole set chart). B. Distribution of specific functional links inferred in the FunCoup network by analysis of microarray data. We discovered dioxin-enabled, dioxin-sensitive, and dioxin-resistant links, along with developmental stage-specific links. The majority of FunCoup links did not overlap with any of these categories, either because a gene was absent from the microarray chip, or because its transcription was not perturbed in the experimental conditions. The number of dioxin-sensitive links is smaller than the number of dioxin-enabled links, and both are much smaller than the number of dioxin-resistant links. Differentially expressed links were detected here via ANOVA (combined with Pearson correlation metric) by considering a third factor “GENE” and its interaction with “DAY” and “TREATMENT”. Note the logarithmic scale of the vertical axis. C, D, E. Visualization of specific link types. All 3 gene pairs were significantly co-expressed (PLC equaled 0.876, 0.946, and 0.754, respectively). Plot axes indicate relative transcript abundance of the genes (log10 ratio of the experimental value to the transcript abundance in the reference pool). C. For this correlation, the condition- and day-specific points were mixed randomly. The reason for the high co-expression of the two genes (both of which code for ribosomal proteins) was probably not associated with the parameters examined in our experiment, and this link would be characterized as “Ambiguous” rather than “Developmental” or dioxin-altered. D. The transcript levels for these two genes (coding for proteins involved in chromosome maintenance and protein folding) were strongly synchronized according to day, so the link was classified as “Developmental.” E. Co-expression was observed under the dioxin treatment, but absent in the control (PLC 0.9032 and 0.1286, respectively), so the link was classified as “dioxin-enabled”. These genes are uncharacterized—but now known to lose functional coupling after dioxin exposure. The differentiation between C, D, and E would not be possible without deeper variance decomposition in ANOVA. See Methods S1 for details.
Figure 3.
FunCoup sub-networks with genes and links affected by dioxin.
A. Network region around the originally (on day 1) altered CYP1A includes links to gene groups altered later (as discovered with jActiveModules and temporal GO relation analyses): calmodulins (triangles), members of heme biosynthesis (crosses), and neuronal development (stars) pathways. B. Network region with three adjoining clusters enriched in differentially expressed links. cnn2 (calponin 2) is co-expressed with many surrounding genes only after dioxin exposure (cluster P-916), while ung (uracil DNA glycosylase; cluster N-609), on the contrary, is normally co-expressed with its network neighbors but not after dioxin exposure. Meanwhile, several “border” genes lost functional coupling to genes from clusters N-609 and N-63, while gaining coupling to cluster P-916 genes, after dioxin exposure. The clusters are named P-* when containing positive (dioxin-enabled) links, N-* for negative (dioxin-sensitive), or PN-* for both positive and negative links. C. Members of cluster N-632 include DNA/RNA processing, cell division, and embryonic forebrain patterning genes. Interactions: Grey, FunCoup links at confidence FBS>9 (Fig. B) and FBS>3 (elsewhere) (every blue and brown link must overlap with a FunCoup link at FBS>3). For every sub-network, only a sub-set of links was retrieved, filtering for confidence and relevance to query genes with algorithm (1) (see description at http://funcoup.sbc.su.se/algo.html#noislets). Brown, clustered (co-expression + FunCoup) links caused by dioxin treatment; Blue, clustered (co-expression + FunCoup) links disabled by dioxin treatment; Green, developmental-stage specific co-expression. Color lines at A indicate data types evident of functional coupling and are explained in right panel of the screenshot. Nodes: Diamonds: members of respective clusters (i.e., genes participating in a DEL). Squares: other genes with measured expression in the course of the experiment; Circles: other genes having evidence from orthologs in FunCoup but for which expression data was lacking in our experiment. Shades of red and green, up- and down-regulation (main factor “DIOXIN TREATMENT”) with dioxin, respectively; genes missing microarray data colored grey. All modules can be found and manipulated at http://funcoup.sbc.su.se/zfish_supplementary.html, and GO enrichment analysis of modules significantly enriched in at least one GO biological process are presented in Data File S5. Both clusters and individual genes can be accessed at http://funcoup.sbc.su.se/zfish.html. See Methods S1 for details of link inference and other analysis.
Figure 4.
Network of biological processes based on genes dioxin-altered on sequential days.
Nodes of the network are defined as GO “biological process” categories that include one or more differentially expressed genes in the course of the experiment. Thus if a GO category is enriched on day 1 in differentially expressed genes, and linked by an unexpectedly high number of FunCoup interactions to a second GO category that on day 2 is enriched in differentially expressed genes, we infer that this temporal relationship is indicative of causality. Network interactions (edges) represent the sum of interactions between the gene members of the two connected GO categories. Node color represents the fraction of the genes in that node that are regulated by dioxin on any day (green is low, red is high). Edge thickness and opacity represent the number of gene-gene links between two categories and χ2 enrichment score for the likelihood that this pair of categories is enriched in links, respectively. Edge color and arrows show timing of differential expression between gene-gene pairs in respective GO categories.
Figure 5.
Many genes in an altered (jActiveModules) neighborhood identified on Day 2 are involved in glycolysis.
DNA repair and maintenance as well as muscle-related cytoskeletal proteins are also part of this module. Red node color indicates upregulation and green indicates downregulation, with the faintest coloration indicating a 1.3-fold change and darker coloration indicating greater change. Node border color indicates significance; the faintest blue indicates pα = 0.05, with darker blue indicating greater statistical significance (by Rosetta Resolver).