Bayesian Variable Selection Methods for Genome-Wide Association Studies with Categorical Phenotypes

Abstract

Genome-wide association studies (GWAS) attempt to find the associations between genetic markers and studied traits (phenotypes). The problem of GWAS is complex and various methods have been developed to approach it. One of such methods is Bayesian variable selection (BVS). We describe the BVS methods in detail and demonstrate the ability of BVS method Posterior Inference via Model Averaging and Subset Selection (piMASS) to improve the power of detecting phenotype-associated genetic loci, potentially leading to new discoveries from existing data without increasing the sample size. We present several ways to improve and extend the applicability of piMASS for GWAS. The first method incorporates non-genetic covariates in the BVS process for GWAS with continuous phenotype, therefore making it possible to account for population stratification. Next, we extend the method mentioned above to work with binary phenotype. Finally, we extend the piMASS method to work with ordinal phenotype. The presented methods allow the existing BVS methods reach wider applicability and higher quality of detected associations. We conduct simulation studies and compare the results to the original method piMASS to show their efficacy. We also apply two of the methods to the Alzheimer's Disease Neuroimaging Initiative 1 (ADNI1) dataset containing data on Alzheimer's patients with categorical phenotype and demonstrate the method's ease of use and applicability. Finally, we discuss the potential of the methods in GWAS and possible directions for further research.