Polygenic risk scores perform poorly across populations

Abstract |

The vast majority of genome-wide association studies (GWAS) are performed in cohorts of European ancestry. Systematic differences in polygenic risk scores (PRS) between European and non-European ancestry populations are believed to be largely spurious. However, it is not clear whether they are completely inaccurate nor how much individual-level predictive power is lost by applying PRS based on European-ancestry GWAS to non-European ancestry populations. Finally, a quantitative understanding of the biological or statistical basis for the poor performance of PRS in non-European-ancestry populations is lacking. To test this, we explored how well PRS predict a well-studied and highly polygenic trait: height. We calculated PRS using 41 sets of independent SNPs based on distance (physical or genetic) or LD clumping and pruning methods. We first compared PRS based on effect sizes from two independent GWAS: GIANT and UK Biobank (UKB), both of which were performed in individuals of European ancestry. We replicate previous observations of population-level differences in PRS, but these results are significantly different depending on datasets and clumping strategies. Depending on clumping strategy, the average difference between 1000 Genomes European and African PRS varies from 0.57-6.75 standard deviations (SD) using GIANT and 0.48-2.16 SD using UKB summary statistics. This dependence on clumping strategy supports the idea that most of these differences are spurious. We then investigated individual-level prediction in ~7,300 African American (AA) and ~7,400 European American (EA) individuals. Using UKB effect sizes, we found that PRS explain ~1.7% of height variation in AA individuals, compared to about 5.5% of variation in EA individuals. In both AA and EA, PRS explains more of the variance in height in individuals with a higher proportion of European ancestry. Interestingly, although we find a significant positive correlation (r=0.262, p=7.3e-115) between PRS and European ancestry among AA individuals, the correlation between height and European ancestry is extremely low (r=-0.02, p=0.056) and genome-wide ancestry explains only 0.05% of the variance in height, confirming that cross-population differences in PRS do not correlate to phenotypic differences. Finally, we evaluated whether local ancestry improves prediction for non-European populations, investigated dependence on other genomic features and extended our model to other phenotypes.