Using Local Genetic Correlation Improves Polygenic Score Prediction Across Traits

Abstract

IntroductionThe predictive utility of polygenic scores (PGS) is steadily increasing as genome-wide association studies (GWAS) increase in sample size and diversity, and as PGS methodology is further developed. Multivariate PGS approaches incorporate GWAS results for secondary phenotypes which are genetically correlated with the target phenotype. These improve prediction over using PGS for only the target phenotype. However, previous methods have only considered the genome-wide estimates of SNP-based heritability (h2SNP) and genetic correlation (rg) between target and secondary phenotypes. In this study, we assess the impact of local h2SNP and rg within specific loci on cross-trait prediction. MethodsWe evaluate PGS using three target phenotypes (depression, intelligence, BMI) in the UK Biobank, with GWAS summary statistics matching the target phenotypes and 14 genetically correlated secondary phenotypes. PGS SNP-weights were derived using MegaPRS. Local h2SNP and rg were estimated using LAVA. We then evaluated PGS after reweighting SNP-weights according to local h2SNP and rg estimates between the target and secondary phenotypes. Elastic net models containing PGS for multiple phenotypes were evaluated using nested 10-fold cross validation. ResultsModelling target and secondary PGS significantly improved target phenotype prediction over the target PGS alone, with relative improvements ranging from 0.8-12.2%. Furthermore, we show reweighting PGS by local h2SNP and rg estimates can enhance the predictive utility of PGS across phenotypes, with additional relative improvements of 0.2%-2.8%. Reweighting PGS by local h2SNP and rg improved target phenotype prediction most when there was a mixture of positive and negative local rg estimates between target and secondary phenotypes. ConclusionModelling PGS for secondary phenotypes consistently improves prediction of target phenotypes, and this approach can be further enhanced by incorporating local h2SNP and rg estimates to highlight relevant genetic effects across phenotypes.