03-07

Learning Predictive Models of Gene Regulation

Studying the behavior of gene regulatory networks by learning from high-throughput genomic data has become one of the central problems in computational systems biology. Most work in this area has focused on learning structure from data -- e.g. finding clusters or modules of potentially co-regulated genes, or building a graph of putative regulatory "edges" between genes -- and has been successful at generating qualitative hypotheses about regulatory networks.

Instead of adopting the structure learning viewpoint, our focus is to build predictive models of gene regulation that allow us both to make accurate quantitative predictions on new or held-out experiments (test data) and to capture mechanistic information about transcriptional regulation. Our algorithm, called MEDUSA, integrates promoter sequence, mRNA expression, and transcription factor occupancy data to learn gene regulatory programs that predict the differential expression of target genes. Instead of using clustering or correlation of expression profiles to infer regulatory relationships, the algorithm learns to predict up/down expression of target genes by identifying condition-specific regulators and discovering regulatory motifs that may mediate their regulation of targets. We use boosting, a technique from statistical learning, to help avoid overfitting as the algorithm searches through the high dimensional space of potential regulators and sequence motifs. We will report computational results on the yeast environmental stress response, where MEDUSA achieves high prediction accuracy on held-out experiments and retrieves key stress-related transcriptional regulators, signal transducers, and transcription factor binding sites. We will also describe recent results on hypoxia in yeast, where we used MEDUSA to propose the first global model of the oxygen sensing and regulatory network, including new putative context-specific regulators. Through our experimental collaborator on this project, the Zhang Lab at Columbia University, we are in the process of validating our computational predictions with wet lab experiments, with encouraging preliminary results.