S4-Multi: enhancing polygenic score prediction in ancestrally diverse populations

While polygenic scores (PGSs) have shown promise in advancing precision medicine by capturing the additive effects of common germline variants on inherited disease risk, they are presently limited by reduced performance outside of European-origin populations. We extend our previously developed Bayesian polygenic model (PGM) method, select and shrink with summary statistics (S4), to improve prediction accuracy in ancestrally diverse populations. We benchmark this multi-ancestry extension (S4-Multi) against alternative methods on both simulated and biobank data predicting type 2 diabetes, breast cancer, colorectal cancer, asthma, and stroke. In simulation tests, we find that S4-Multi achieves 169% improvement on average over its single ancestry S4 counterpart at prediction in non-European target populations. S4-Multi matches or exceeds top performing methods across the ancestry continuum. In biobank tests, we find that the top-performing PGM method varies considerably by target ancestry and phenotype, with S4-Multi achieving comparable performance to top multi-ancestry methods overall. However, S4-Multi does so while including between 9% and 77% fewer genetic variants relative to competing models, suggesting potential for robust performance in clinical settings with limited available genomic data.