Figure 3: (CAN MATT VERIFY THAT I USED THE FINAL NETWORK?) Summarising the network by GO term.  (A) In order to look for enrichments for pairs of GO terms that are associated with one another, we counted the frequency of a certain GO term being upstream of each other GO term.  To determine how likely we were to observe this by chance, we shuffled the GO terms, but kept the network structure, and calculated a null distribution from which to calculate a p-value.  (B) A heatmap of the resulting p-values from the procedure in (A), clustered using the default parameters of hclust.  We repeated this twice, once with a network that contained all the genes and once with a network that contained genes encoding DNA binding proteins only.  This is showing the table for the latter. It is possible to show this as a network by only including edges with a p-value below a threshold.  We show this for (C) the network with all genes and (D) the network only containing DNA binding proteins.

Analysis of network edges that are consistent with TF binding data

Next, we found a subset of regulatory associations that is supported both by our inferred network and by DNA binding data (via DAP-seq from (O’Malley et al. 2016)).  We find that the resulting network forms eight disconnected sub-networks, as shown in Figure 4A (in addition to a set of disconnected edges, see Table S5– not done ).
These include networks involved in temperature response, either via HSFs (blue) or C-REPEAT/DRE BINDING FACTOR 2 (CBF2) and SHORT VEGETATIVE PHASE (SVP) (pink).  Interestingly, both these subnetworks are up-regulated in elevated temperatures, although the HSF-driven network has raised expression in phyAphyBcry1cry2 , while the CBF2/SVP subnetwork has increased expression in prr5prr7prr9 .  Previously, when we analysed the entire set of DNA binding genes, we had found that there is an early burst of gene expression that is elevated at 27oC, which was upregulated in phyAphyBcry1cry2to a greater extent than prr5prr7prr9 .  Since this cluster of genes included HSF1A, it may be that this early burst of gene expression drives the expression of the subnetwork containing HSFB2B, HSP90, and HSP60.  Indeed, we see that this subnetwork experiences reduced expression in hsf1abcd .
Additionally, we find a subnetwork of genes (orange) that include a number of genes that can respond to ABA signalling within an hour of signalling (HB6, HB7, CHX17, HAI1, and ATAF1).  To put this result in context, out of the approximately 30,000 total genes in Arabidopsis , there are less than 300 genes with significantly increased expression (FDR<0.001) after 1 hour exposure to ABA (Song et al. 2016).
Another subnetwork is centred on MYBS2 (brown), which is also known to increase the sensitivity to ABA (Chen et al. 2016).  The HB6/HB7 centred subnetwork has a number of genes with increased expression inphyAphyBcry1cry2 , while the MYBS2-centred subnetwork is up-regulated in prr5prr7prr9 .   Both of these networks are up-regulated in elevated temperature.
phyAphyBcry1cry2 and prr5prr7prr9 also have opposite effects in the network that includes GBF6, BBX31, ERF39, MYB13 and HAT2, which includes genes involved in biosynthesis and metabolism, as well as cold tolerance.  These genes are up-regulated in prr5prr7prr9, but downregulated in phyAphyBcry1cry2.