| Genetic interactions occur when the phenotypes from two-gene perturbations differ from the single-gene effects, and are the result of buffering activities of genes and pathways. High-throughput genetic interaction screens are measuring binary synthetic fitness and lethality (SFL) interactions and quantitative epistasis scores for Saccharomyces cerevisiae on a genome-scale. This dissertation focuses on developing computational methods for inferring functional modules and novel genetic interactions from the yeast genetic interaction networks alone and in conjunction with protein-protein interaction networks. These methods include probabilistic network models, Bayesian inference and graph theoretic algorithms. First, we developed Genetic Interaction Motif Finding (GIMF) to infer pathways from SFL interactions. Starting from a seed gene, GIMF iteratively builds the motif, the genetic interaction pattern shared by co-pathway genes, and pathway membership probabilities. GIMF predicted functional associations that correlated favorably with Gene Ontology annotations and modules representing known complexes and pathways. New components suggested for the dynein-dynactin pathway were supported by experimental evidence. Next, we developed graph diffusion kernels as a unified framework to infer complex/pathway membership and genetic interactions from the SFL network. The kernels significantly improved over previous best methods for both types of predictions. We achieved a 50% precision with 20-50% recall in genome-wide prediction of new genetic interactions, supported by experimental validation. We also used support vector machines to integrate kernel scores derived from the genetic and physical networks to predict SFL interactions. Third, we devised a Bayesian algorithm for inferring the probability of SFL interactions from quantitative dSLAM experimental measurements. On benchmark testsets from the BioGRID database, this method worked as well as Stouffers z-score, achieving a precision of 40% at a recall of 30%. Last, we developed a framework known as the hierarchical random graph suitable for inferring functional modules and novel interactions from heterogeneous biological graphs. We focused on inferring functional modules from normally distributed and binary edges. On representative synthetic datasets, we achieved near-perfect predictions. When applied to quantitative and binary genetic interaction data for 88 chromosome-related yeast genes, the best predictions achieved an AUC of 0.817 and an F-score of 0.660. |
