Integrating large-scale functional genomic data to dissect the complexity of yeast regulatory networks

Abstract

A key goal of biology is to construct networks that predict complex system behavior. We combine multiple types of molecular data, including genotypic, expression, transcription factor binding site (TFBS), and protein-protein interaction (PPI) data previously generated from a number of yeast experiments, in order to reconstruct causal gene networks. Networks based on different types of data are compared using metrics devised to assess the predictive power of a network. We show that a network reconstructed by integrating genotypic, TFBS and PPI data is the most predictive. This network is used to predict causal regulators responsible for hot spots of gene expression activity in a segregating yeast population. We also show that the network can elucidate the mechanisms by which causal regulators give rise to larger-scale changes in gene expression activity. We then prospectively validate predictions, providing direct experimental evidence that predictive networks can be constructed by integrating multiple, appropriate data types.

Figure 1

A generic approach to identifying clique-communities in the PPI network. A) Hierarchical clustering over the clique-clique similarity matrix heatmap derived from a network of the 492 k-cliques with k ≥ 5. Cliques in the rows and columns are sorted by an agglomerative hierarchical clustering algorithm. The clique network clearly displays strong modularity under hierarchical organization. Each of the colored bars along the top horizontal and left vertical axes represents a network module. B) The clique community network. Each node represents a clique and each link indicates that the two connected cliques have a similarity greater than 0.5 based on the Dice similarity measure. Of the clique communities represented in this plot, 74% overlap the set of stable protein complexes.

Figure 2

eQTL hot spot 4 subnetworks. A) Subnetworks in BN_full enriched for expression traits linked to eQTL hot spot 4. Of the 3,662 genes comprising BN_full, 203 link to eQTL hot spot 4. There are 309 genes comprising the three subnetworks shown here, and 170 of these genes link to eQTL hot spot 4 (red nodes), a nearly 10-fold enrichment over what was expected by chance (empirical p < 10⁻⁸). LEU2 and ILV6 were identified as the primary causal regulators for the large subnetwork, as described in the text. ILV6 is supported as causal for GCN4 in the large subnetwork. B) The ILV6 knockout signature is enriched in the large eQTL hot spot 4 subnetwork. Of the 635 genes in the ILV6 knockout signature, 432 were represented in BN_full, and 129 of these overlapped the large eQTL hot spot 4 subnetwork (colored nodes), a nearly 4-fold enrichment over what was expected by chance (empirical p < 10⁻⁸). The red and green nodes represent genes that are up- and down-regulated, respectively, in the ILV6 knockout signature (note GCN4 is up-regulated).