Leveraging pleiotropy to improve genetic risk prediction across diseases

BackgroundPolygenic scores (PGSs) have shown promise in predicting disease risk, but their predictive accuracy remains limited for many complex diseases. Leveraging the shared genetic architecture among correlated traits may improve prediction performance. MethodsWe developed a flexible framework for constructing multi-trait PGSs by integrating candidate PGSs (N=2,651) derived from publicly available GWAS summary statistics (N=51)--using single-trait, MTAG-all, and MTAG-pairwise approaches. Multi-trait PGS models were trained using Elastic Net regression in the UK Biobank (N = 307,230 individuals) and validated in both an internal set of UKB individuals (N = 39,122) and an external, All of Us (N = 116,394), cohort. We further evaluated the utility of multi-trait PGSs in risk prediction with non-genetic factors, interactions, and genetic subgroup identification. ResultsMulti-trait PGSs significantly improved risk prediction for eight diseases, with AUC gains ranging from 1.56% to 5.45% compared to optimal single-GWAS PGSs. Selected scores mainly consisted of genetically correlated phenotypes. Multi-trait PGSs further enhanced predictive performance and stratification when integrated with non-genetic factors. Significant interactions were identified between multi-trait PGS for peripheral artery disease (PAD) and modifiable risk factors such as smoking and waist-to-hip ratio (WHR). A clustering analysis uncovered genetically distinct subgroups with meaningful phenotypic variation, including a chronic kidney disease (CKD) subgroup enriched for diabetes- and obesity-related traits. ConclusionOur multi-trait PGS framework improves disease prediction by capturing cross-trait genetic effects and enables personalized risk assessment through integration with non-genetic exposures, interactions, and subgroup identification. This approach offers a scalable and generalizable tool for advancing precision medicine.