Characterizing Cellular Phenotypes via Bayesian Regression in the Gene Ontology

• Activity Number: 72
• Type: Contributed
• Date/Time: Sunday, August 9, 2015 : 4:00 PM to 5:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Paula J. Griffin* and Tatsunori Hashimoto and Edo Airoldi
• Companies: Boston University School of Public Health and MIT and Harvard University
• Keywords: Biological networks ; Bayesian hierarchical models ; gene ontology ; gene expression

Abstract

A common task in genome- and proteome-wide studies is to characterize the driving forces behind cellular phenotypes. Current approaches based on clustering and enrichment often fail to provide interpretable gene groups, and do not consider the impact of uncertainty in gene set assignment. To address these shortcomings, we propose a biological-function oriented regression based on the gene ontology (GO) to characterize the importance of gene groups to cellular phenotypes. We model the expression of sets of genes involved in biological functions based on the the experimental setting (such as a gene knockouts or limited nutrient access) and phenotypes within a Bayesian hierarchical formulation. Linear responses are combined according to membership to GO terms, whose priors depend upon their relationship in the ontology. In addition to providing estimates of GO term involvement in phenotype, our model can predict responses for missing genes, and its goodness-of-fit aligns well with the accuracy of its predictions. We demonstrate the efficacy of functional regression in data from an experiment that limits instantaneous cellular growth rate by nutrient restriction.