NOVEL MULTIVARIATE BAYESIAN VARIABLE SELECTION METHODS WITH APPLICATION TO GENETIC FINE MAPPING

Abstract

Genome Wide Association Studies (GWAS) and more recently, Transcriptome Wide Association Studies (TWAS) have been extensively used to identify genomic loci of strong association with complex human traits. However, extreme correlation between genetic variants (i.e. linkage disequilibrium (LD)) and the marginal nature of such studies can make it difficult to identify causal genes or genetic variants. Furthermore, these studies are often univariate in nature, whereas phenotypes of interest such as complex human diseases are also correlated with one another. There also exists the problem of pleiotropy, where a single variant or gene may be causal for multiple phenotypes. Statistical models for such data to more accurately identify causal variants and genes require a flexible framework known as fine mapping. In this dissertation, we propose several multivariate Bayesian variable selection models to perform fine mapping. The rest of the dissertation is presented as follows: In Chapter 1, we provide an introduction and review of molecular biology, fine mapping, variable selection and Bayesian methods. In Chapter 2 we propose a multivariate Bayesian variable selection model for multi-trait fine mapping for GWAS. In Chapter 3 we extend our model in Chapter 2 to allow for higher dimensional outcomes through the use of a latent infinite factor model for phenome-wide fine mapping for TWAS. In both chapters, we apply the proposed models to multiple real-data applications (e.g. fine mapping of heritable neuroimaging features, disease conditions from electronic health record data) and evaluate the models through extensive simulations compared against existing methods. In Chapter 4, we propose a Variational Bayes approach as an alternative way to estimate the models described in Chapters 2 and 3. Finally, in Chapter 5, we discuss the work we have done and provide potential future extensions.