Partially observed bipartite network analysis to identify predictive connections in transcriptional regulatory networks

Here we present a Bayesian method that identifies which known transcriptional relationships in a regulatory network are consistent with a given body of static gene expression data by eliminating the non-relevant ones. The Partially Observed Bipartite Network (POBN) approach developed here is tested using E. coli expression data and a transcriptional regulatory network derived from RegulonDB. When the regulatory network for E. coli was integrated with 266 E. coli gene chip observations, POBN identified 93 out of 570 connections that were either inconsistent or not adequately supported by the expression data.POBN provides a systematic way to integrate known transcriptional networks with observed gene expression data to better identify which transcriptional pathways are likely responsible for the observed gene expression pattern.Significant effort has been invested in identifying which genes regulate the expression of which other genes in a given genome[1-3]. The bioinformatics community has collected many of these gene-gene regulatory relationships into transcriptional networks that provide a global view of how gene regulation is orchestrated. For example, TRANSFAC collects protein-DNA binding interactions to identify potential gene regulatory mechanisms[4]. Similarly, RegulonDB provides a hand annotated regulatory network for the E. coli genome[5]. As more data become available, these transcriptional regulatory networks will become increasingly complete in the sense that they will describe the set of possible mechanisms for regulating each gene.However, even with a fully complete and accurate transcriptional regulatory network, only some of the regulatory relationships will be relevant for a given cellular environment. For example, some gene regulatory mechanisms may only be used in rare cases of stress, or during a short developmental stage. In these cases, these rarely used regulatory mechanisms are correct, but largely non-predictive and as such may not be relevan