Polygenic risk scores (PRSs) aggregate the effects of hundreds to millions of genetic variants to predict disease risk. Their clinical uptake is limited in part because most historical genome-wide association studies (GWASs) have been performed in European cohorts, causing substantial drops in predictive accuracy when PRSs are applied to non‑European populations. Differences in genetic architecture, population structure, and allele frequencies across populations drive these accuracy gaps. Conducting new GWASs on hundreds of thousands of individuals in each underrepresented population is costly; transfer learning that augments large European GWAS with target‑population GWAS offers a practical alternative.
This study systematically evaluated how PRS performance in a non‑European target population depends on the sizes of both the target and European discovery cohorts. The authors used two large, deeply genotyped and phenotyped resources: the UK Biobank (UKB) European samples and Biobank Japan (BBJ), a cohort of nearly 200,000 Japanese individuals. They assessed eight clinically relevant traits present in both datasets: body mass index (BMI), systolic blood pressure, diastolic blood pressure, red blood cell count, white blood cell count, high‑density lipoprotein (HDL), low‑density lipoprotein (LDL) and blood glucose. SNP heritability in UKB for these traits ranged from 0.07 to 0.28.
Methods
Three approaches were compared to probe PRS transferability:
-
UKB discovery GWAS + elastic net: Trait‑associated variants were identified by running GWAS on the full UKB European sample and selecting independent variants also present in BBJ. Elastic net models were then trained on various combinations of BBJ and UKB sample sizes to produce target‑population PRSs. For each trait, 96–104 PRS models were generated across paired sample‑size combinations.
-
Meta‑analysis + elastic net: GWAS was run on the full UKB and on sample‑size‑specific BBJ subsets, the summary statistics were meta‑analyzed to select variants, and elastic net models were fit on mixed UKB/BBJ training sets.
-
PRS‑CSx: A multi‑ancestry method designed to account for cross‑population linkage disequilibrium differences. PRS‑CSx was applied to full UKB and sampled BBJ GWAS statistics, producing two scores per individual per trait; validation splits were used to find optimal linear combinations of these scores.
All PRS evaluations used the same held‑out BBJ validation set.
Main findings — more data is not always better
Model performance was quantified by Pearson correlation between predicted and observed phenotypes. European discovery data provided a useful baseline when target‑population data were scarce or absent. However, once the target population sample size reached about 15,000 individuals or more, models trained solely on target‑population data outperformed those co‑trained with external UKB data. In other words, pooling large out‑of‑population data can limit accuracy gains achievable by increasing local sample size.
This pattern appeared for all examined phenotypes, but the crossover point where local data overtakes pooled training varies by trait. The authors used cross‑population genetic correlation to quantify shared genetics. Traits with higher genetic correlation between UKB and BBJ (so‑called "conserved" traits, e.g., BMI) continue to benefit from pooling UKB data up to larger BBJ sample sizes (roughly 25–40k or more). By contrast, population‑specific traits such as HDL, LDL and blood glucose reach the crossover at much smaller BBJ sizes; UKB data are less helpful because they are farther out‑of‑distribution for these traits.
Effects of meta‑analysis and PRS‑CSx
Initial experiments limited input variants to those found in UKB, excluding variants unique to BBJ. To capture BBJ‑specific signals, the authors added two strategies: (1) cross‑population meta‑analysis of UKB and BBJ GWAS followed by elastic net, and (2) PRS‑CSx using both GWAS summary statistics.
Meta‑analysis had limited impact for conserved traits, mainly because smaller BBJ GWAS sample sizes reduced the power to discover variants. For population‑specific traits (especially HDL and LDL, and to a lesser extent blood glucose), meta‑analysis substantially outperformed single‑population discovery. The gains came largely from changing the set of variants input to the elastic net: including UKB samples during training produced modest improvement when BBJ samples numbered about 10,000 or fewer, but this advantage disappeared with larger BBJ sample sizes.
PRS‑CSx, which reweights population‑specific models, should in principle be less sensitive to trait conservation. In practice, PRS‑CSx required more data to perform well: for target sample sizes under 25k, PRS‑CSx underperformed the best elastic net model for nearly all phenotypes except BMI. As sample sizes approached ~100k, PRS‑CSx matched or exceeded the best model across almost all traits except blood glucose.
Conclusions and practical implications
A systematic assessment of cross‑population genomic prediction shows that large out‑of‑population datasets are not universally beneficial for improving PRS performance in underrepresented ancestries. European UKB data provide statistical benefit when target‑population sample sizes are small (roughly <15,000 BBJ samples), but as local sample sizes increase, inclusion of external European data can degrade accuracy. The crossover depends on trait genetics: conserved traits (e.g., BMI) retain benefits from pooled data up to larger target sizes, while population‑specific traits (lipids, blood glucose) require less external data and benefit earlier from local data.
Advanced multi‑ancestry methods like PRS‑CSx and cross‑population meta‑analysis can improve prediction for population‑specific traits, but they typically need substantial target‑population sample sizes to outperform simpler approaches. These results underscore the need both to expand local, diverse biobanks and to choose modeling strategies tailored to trait heritability and available sample sizes in order to optimize PRS performance in diverse populations.
Acknowledgments
The study authors thank Biobank Japan and collaborators at RIKEN and The Institute of Medical Science, The University of Tokyo for enabling the research, and Google collaborators Babak Behsaz, Andrew Carroll, Farhad Hormozdiari and Taedong Yun. Additional thanks to Hiroki Kayama and Joe Ledsam for institutional support, and Michael Brenner and Katherine Chou for leadership support.
