Bayesian Iterative Conditional Stochastic Search (BICOSS) for GWAS

• Activity Number: 107 - SPEED: Statistical Methods, Computing, and Applications Part 1
• Type: Contributed
• Date/Time: Monday, August 8, 2022 : 8:30 AM to 10:20 PM
• Sponsor: Section on Statistics in Genomics and Genetics
• Author(s): Jacob Williams* and Marco Ferreira
• Companies: Virginia Polytechnic Institute and State University and Virginia Tech
• Keywords: GWAS

Abstract

We propose a Bayesian analysis for genome-wide association studies (GWAS) that selects significant single nucleotide polymorphisms (SNP) using an iterative procedure. Initially, BICOSS finds a set of significant SNPs by fitting as many linear mixed models as the number of SNPs, with each model containing one SNP and random effects to account for kinship correlation. The second stage performs a stochastic search through model space with a genetic algorithm, where each model is a linear mixed model with kinship random effects, and may include multiple SNPs from the candidate set obtained from the first stage. Third, conditional models are computed in the same framework as the first stage however now conditioned on the SNPs from the model with the highest approximate posterior probability. The second and third steps are iterated until no significant SNPs are discovered from the third stage. A simulation study shows that BICOSS dramatically reduces false discovery rate while increasing statistical power. We illustrate the application of BICOSS with the analyses of two real datasets.