All of Us diversity and scale yield context-dependent improvements in polygenic prediction

Polygenic risk scores (PRSs) trained on multiancestry data can improve prediction in under-represented groups, but large linked genetic and health datasets capturing broad human diversity remain limited. Using 245,388 whole-genome sequences from the All of Us research program (AoU) together with UK Biobank data, we developed multiancestry PRSs for 32 traits and diseases. We evaluated how ancestry, methodology and genetic architecture influenced PRS performance across ancestrally diverse AoU participants. Increased diversity in the AoU improved PRS accuracy for several traits, especially in under-represented populations. However, maximizing sample size by meta-analyzing AoU and UK Biobank was not universally optimal: for less polygenic traits, AoU-only training performed best in African ancestry participants, consistent with ancestry-enriched effects. Individual PRS accuracy declined linearly with increasing ancestry divergence from the discovery GWAS, but this decay was attenuated using multiancestry training data. These findings underscore the value of more representative biobanks for equitable PRS performance.