Introducing Elements of High Performance Computing and Two-Way Epistasis into Genome-Wide Association Studies with Stepwise Model Selection

Although the genome-wide association study (GWAS) has emerged as a predominant approach for quantifying associations between genome-wide marker sets and many important traits, the statistical models that are typically used suffer from two major drawbacks. Specifically, these models test only one marker at a time and they only consider additive effects. As such, the typical GWAS model is not capable of quantifying the simultaneous contributions of multiple interacting genomic loci to phenotypic variation. To address these two drawbacks, we are developing a java-based program that augments a publicly available stepwise model selection module by also considering two-way epistatic interactions between marker pairs for inclusion into the final statistical model. With such a modification, it will become possible for researchers to use a GWAS model that can encapsulate many relationships between the genome and phenotype that are possible with various genetic architectures. Because the incorporation of epistatic effects introduces a substantial computational burden, we are also introducing high-performance computing approaches to this stepwise model selection program. In this talk, we use a subset of the US maize nested association mapping (NAM) population to demonstrate both the biological insight that is possible by incorporating two-way epistatic effects and the reduction of computational burden that can be realized through multithreading. We also discuss the direction we would like to take this stepwise model selection program so that it can accommodate the large-scale genomic data that are becoming increasingly available.