Genotype-Phenotype Similarity Testing and Methods for Integrating Multiple Data Sources in Genetic Association

Genome-wide association studies (GWAS) have been successful in identifying SNP associations for many complex traits. However, for any single trait, published and validated SNPs typically explain only a small part of the genetic heritability estimated to be explained by genetic variation. New methods are needed that can detect small effect sizes and overcome challenges presented by true biological architecture, including causal variants that are not directly genotyped, and allelic heterogeneity. Here we propose a likelihood ratio test (LRT) approach in two forms to perform SNP-set analysis in GWAS. The first form compares the similarity of traits between pairs of unrelated individuals to that of genotypes. The second form jointly tests trait-genotype similarity and regional maximal trait-SNP association. In simulation studies, these methods show favorable power compared to popular alternative approaches. The basic idea of trait-similarity approaches is that, for genomic regions harboring loci that affect a trait, individuals with similar genotypes/haplotypes should have more similar traits. Therefore, as a nonparametric alternative to likelihood-based testing, we consider a statistic that directly correlates genetic and trait similarity. Due to the comparison of dependent pairs of individuals, standard distributional approximations for correlation coefficient testing do not apply. Monte Carlo approaches to evaluating significance can be applied, but can be prohibitive for the extreme testing thresholds required in the genomic setting. We describe an approach using the first four exact permutation moments of the permutation distribution of our statistic, and propose a moment matching statistic for SNP-set analysis. Finally, we discuss challenges in performing GWAS meta-analyses for differing diseases, and for joint testing of association with multiple traits (pleiotropy). A key issue is that the alternative hypothesis should be restricted to joint effects, and standard testing procedures are not adequate. We propose practical schemes to (i) create statistics that are sensitive only to the intended alternative, and (ii) cyclic shift genome-wide testing, to avoid overly conservative family-wise error control.Doctor of Public Healt