Polygenic risk scores (PRSs) predict disease risk from genetic variants, but their clinical use is limited because historical genome-wide association studies (GWASs) have focused on European cohorts, causing accuracy drops in non-European populations. Differences in genetic architectures and allele frequencies contribute to this issue, and conducting new GWASs is costly. Transfer learning from existing European-centric GWAS, augmented with target-population data, offers a solution.

This study evaluated PRS performance in a target non-European population, varying the sizes of both target and European populations across eight clinical traits. Using the UK Biobank (UKB) and Biobank Japan (BBJ), which includes nearly 200 thousand Japanese individuals, the researchers conducted dataset ablation experiments. Traits included body mass index (BMI), blood pressure, cholesterol levels, and blood cell counts.

Three methods were tested: direct variant transferability using UKB discovery GWAS with elastic net, meta-analysis combining Japanese data during variant discovery, and PRS-CSx for handling linkage disequilibrium diversity. Models were evaluated on held-out BBJ samples. Results showed that more data is not always better for cross-population PRS prediction, providing empirical guidelines for optimizing predictive performance in target populations.