Almost Free Enhancement of Multi-Population PRS: From Data-Fission to Pseudo-GWAS Subsampling

Abstract

Many multi-population polygenic risk score (PRS) methods have been proposed to improve prediction accuracy in underrepresented populations; however, no single method outperforms other methods across all data scenarios. Although integrating PRS results across multiple methods and populations may lead to more accurate predictions, this approach may be limited by the availability of individual-level tuning data to calculate combination weights. In this manuscript, we introduce MIXPRS, a robust PRS integration framework based on data fission principles, to effectively combine multiple multi-population PRS methods using only genome-wide association study (GWAS) summary statistics from multiple populations. Specifically, MIXPRS employs SNP pruning to mitigate linkage disequilibrium (LD) mismatch between the training GWAS summary statistics and LD reference panels, and utilizes non-negative least squares regression to robustly estimate PRS combination weights. Extensive simulations and real-data analyses involving 22 continuous traits and four binary traits across five populations from the UK Biobank and All of Us datasets demonstrate that MIXPRS consistently outperforms the existing methods in prediction accuracy. Because MIXPRS relies solely on GWAS summary statistics, it enjoys broad accessibility, robustness, and generalizability for underrepresented populations.