Sparse graphical models for cancer signalling

Protein signalling networks play a key role in cellular function, and their dysregulation is central to many diseases, including cancer. Recent advances in biochemical technology have begun to allow high-throughput, data-driven studies of signalling. In this thesis, we investigate multivariate statistical methods, rooted in sparse graphical models, aimed at probing questions in cancer signalling. First, we propose a Bayesian variable selection method for identifying subsets of proteins that jointly in uence an output of interest, such as drug response. Ancillary biological information is incorporated into inference using informative prior distributions. Prior information is selected and weighted in an automated manner using an empirical Bayes formulation. We present examples of informative pathwayand network-based priors, and illustrate the proposed method on both synthetic and drug response data. Second, we use dynamic Bayesian networks to perform structure learning of context-specific signalling network topology from proteomic time-course data. We exploit a connection between variable selection and network structure learning to efficiently carry out exact inference. Existing biology is incorporated using informative network priors, weighted automatically by an empirical Bayes approach. The overall approach is computationally efficient and essentially free of user-set parameters. We show results from an empirical investigation, comparing the approach to several existing methods, and from an application to breast cancer cell line data. Hypotheses are generated regarding novel signalling links, some of which are validated by independent experiments. Third, we describe a network-based clustering approach for the discovery of cancer subtypes that differ in terms of subtype-specific signalling network structure. Model-based clustering is combined with penalised likelihood estimation of undirected graphical models to allow simultaneous learning of cluster assignments and cluster-specific network structure. Results are shown from an empirical investigation comparing several penalisation regimes, and an application to breast cancer proteomic data.EThOS - Electronic Theses Online ServiceEngineering and Physical Sciences Research Council (EPSRC)GBUnited Kingdo