A systematic comparison of methods for decoding the genetic architecture of phenotypic traits

A central pursuit of genetics is to infer the genetic basis of phenotypic variation. Fundamental to this challenge is the elucidation of genetic architecture, comprised of the number, location and relative effect of genes on a phenotypic trait. Many methods have been developed to resolve this covariation in genes and traits, and in this thesis we will systematically compare three: the standard variant-by-variant Genome-Wide Association Study (GWAS), a well-known variable selection approach (Least Absolute Shrinkage and Selection Operator, or LASSO), and a novel method we call the E-value. To achieve this, we propose a new framework for isolating a discrete set of factors influencing the performance of such methods so that their individual and synergistic impacts can be quantified. In this framework, we control population structure, population size, the number of variants, and the number and heritability of trait-associated variants. Then for each configuration of these parameters, over many replicates, we generate genotype and phenotype and quantify how well their simulated relationship is resolved by each method. Although GWAS remains a common choice for initial analyses, our results indicate that LASSO is far better suited to the challenge of resolving genetic architectures. In contrast, our new method — the E-value — performed unexpectedly poorly. The framework we contribute, and the inference it empowers, enhances our understanding of the methodologies used to resolve genotype-phenotype relationships. We hope this translates to improvements in future studies seeking to understand or leverage the genetic basis of important traits such as human disease.