Leveraging pleiotropy for joint analysis of genome-wide association studies with per trait interpretations

We introduce pleiotropic association test (PAT) for joint analysis of multiple traits using genome-wide association study (GWAS) summary statistics. The method utilizes the decomposition of phenotypic covariation into genetic and environmental components to create a likelihood ratio test statistic for each genetic variant. Though PAT does not directly interpret which trait(s) drive the association, a per trait interpretation of the omnibus p-value is provided through an extension to the meta-analysis framework, m-values. In simulations, we show PAT controls the false positive rate, increases statistical power, and is robust to model misspecifications of genetic effect.Additionally, simulations comparing PAT to three multi-trait methods, HIPO, MTAG, and ASSET, show PAT identified 15.3% more omnibus associations over the next best method. When these associations were interpreted on a per trait level using m-values, PAT had 37.5% more true per trait interpretations with a 0.92% false positive assignment rate. When analyzing four traits from the UK Biobank, PAT discovered 22,095 novel variants. Through the m-values interpretation framework, the number of per trait associations for two traits were almost tripled and were nearly doubled for another trait relative to the original single trait GWAS.Author summary: Genome-wide association studies have identified tens of thousands of genetic variants associated with complex traits. An ever increasing number of associated variants are shown to affect multiple traits, a phenomenon known as pleiotropy. We propose a method that leverages this genetic architecture and uses summary statistics to perform an omnibus association test between one genetic variant and a set of traits. Simulations show that the method properly controls for type-I errors and increases statistical power. In addition to a powerful omnibus test, we also enable a per trait interpretation of the associations by extending the m-value framework to account for the correlation structure between traits. This framework enables a significant increase in the identification of per trait effects.