Genetic prediction using polygenic risk scores often works unevenly across human populations. Studies led by Alicia R. Martin at the Broad Institute show that scores built from large European-ancestry genome-wide association studies lose predictive accuracy when applied to non-European groups, with reductions in performance that can be roughly two- to five-fold for African ancestry cohorts compared with Europeans. The NHGRI-EBI GWAS Catalog team at EMBL-EBI documents that the discovery datasets underlying many scores remain heavily skewed toward European participants, a central reason for this gap.
Why performance varies
Several measurable causes explain the drop in accuracy. Differences in allele frequency and linkage disequilibrium structure across populations change how well variants discovered in one group tag causal variation in another. Effect size estimates from a European-dominated discovery sample can be biased by population-specific environmental exposures and genetic background, so the same variants explain less trait variance elsewhere. Admixed populations create additional complexity, because local ancestry can interact with both allele frequency and environmental context, further degrading portability.
Consequences and remedies
The practical consequence is a risk of exacerbating health inequities: clinical tools built on biased scores may underperform or misclassify risk for people from African, Indigenous, or many Latin American ancestries, undermining equitable access to genomic medicine. Addressing this requires expanding diverse recruitment, investing in ancestry-specific and multi-ancestry GWAS, and improving methods that account for local ancestry and heterogeneous effect sizes. Community engagement and capacity building in underrepresented regions matter as much as statistical fixes; researchers and institutions must work with local investigators and participants to ensure culturally appropriate study design and benefit sharing.
Methodological improvements such as transfer learning, ancestry-aware weighting, and meta-analysis across diverse cohorts show promise, but their success depends on richer, well-phenotyped datasets from multiple world regions. Without that investment, polygenic tools will remain most accurate for populations already overrepresented in genetic research, limiting their global utility and risking further territorial and social inequities in genomic medicine.