April 2008
To appear in: Bayesian Modelling in Bioinformatics, (D. Dey, S. Ghosh and B. Mallick, eds.), Taylor-Francis, 2008
Much recent and current expression genomics research has been devoted to the discovery of lists of genes showing differential expression across some known phenotypes, or that demonstrate an ability to stratify patients into different risk groups. While studies of this type have been shown to be useful in many applications, they have typically relied on the use of clustering and other simple methods to make the transition between in vitro and in vivo studies. The overall framework of Bayesian factor regression modelling utilising sparsity priors, as implemented in the BFRM software and exemplified in a range of recent studies, provides a formal, encompassing framework for such trans-study analysis. We describe and exemplify the strategy of in vivo profiling of in vitro defined gene expression signatures of controlled biological perturbations or environmental changes. A main focus is on cancer genomics and the connections between gene expression summaries of the outcomes of controlled experiments in cultured cells with physiological and clinical studies on human tumour data. Bayesian factor regression modelling and its uses in profiling an in vitro signature gene set across multiple human cancer data sets is a powerful and statistically sound approach that has been used to effect in studies in cancer genomics, and other areas of both basic and human disease related biology. This paper describe application to profile the E2F oncogene across across three distinct cancer types and sample data sets. The analysis illustrates the ability of factor profiling to (a) define factors that relate to underlying biological pathway interconnections and suggest novel directions for functional evaluation, (b) identify biologically interesting contrasts across cancer types, and (c) identify clinically useful predictive factors as potential biomarkers of cancer subtypes, survival, and other outcomes. The methodological framework of Bayesian regression factor modelling under sparsity priors for both designed experiments and observational samples enables and drives the overall strategy.