Cross-population robustness of polygenic risk scores: a controlled evaluation across two biobanks and traits

Genetic variations contribute to the diversity of human traits and diseases, but their individual impact is small. polygenic risk scores (PRS) aggregate these effects across the genome to assess overall genetic risk. PRS aids in patient stratification, disease prediction, and personalized treatment, improving health outcomes. However, PRS transferability across populations is limited due to the predominantly European focus of genome-wide association studies (GWAS). Enhancing cross-population prediction accuracy is crucial for realizing PRS’s clinical potential and ensuring equitable access to effective therapies globally. Polygenic risk scores (PRS) aggregate the effects of numerous genetic variants to quantify inherited risk for complex traits and diseases. Although PRS have shown promise in risk stratification and precision medicine, their predictive performance often deteriorates when applied across populations, limiting clinical translation. While many LD-aware and multi-ancestry PRS methods have been proposed, it remains unclear how different PRS paradigms perform under realistic biobank-to-biobank transfer settings, particularly across traits with distinct genetic architectures. We utilize simulated and biobank data (Taiwan Biobank (TWB) and UK Biobank (UKB)) for height (1869 TWB individuals, 4,095,705 SNPs; 1056 UKB individuals, 339,550 SNPs) and Glaucoma (51,694 TWB individuals, 6,779,677 SNPs; 192,009 UKB individuals, 502,274 SNPs) to construct cross-population PRS models. We conducted both simulation studies and real-data analyses using the Taiwan Biobank (TWB) and the UK Biobank (UKB). Cross-population PRS models were constructed for height, a highly heritable quantitative trait, and glaucoma, a complex disease phenotype. We systematically compared population-structure-adjusted PRS (PRSadj) with LD-aware methods, including LDpred2 and Lassosum, as well as clumping and thresholding (C+T) as a baseline. Bi-directional prediction between biobanks was evaluated while adjusting for age and sex. PRS predictive performance and incremental value varied across traits, methods, and prediction directions. Under the primary C+T specification, adding PRS to the covariate model provided limited incremental predictive value for height, with Δ R 2 ranging from −0.1844 to −0.0039 across the four within- and cross-population prediction settings. For glaucoma, the incremental contribution of C+T PRS was similarly modest, with ΔAUC ranging from −0.0200 to +0.0010. PC-adjusted sensitivity analyses showed that adjustment for population structure could alter the magnitude of predictive performance estimates, particularly for height. PRS performance and cross-population portability were context-dependent across traits, transfer directions, and modeling strategies. Incremental PRS value beyond baseline covariates was often modest, and no single method showed consistent superiority across all settings. These findings support multidimensional evaluation of PRS using incremental performance, portability, and calibration rather than absolute predictive performance alone.