Low transferability of height polygenic risk scores in admixed ancestry populations

Abstract

Polygenic risk scores (PRS) summarize the results of GWAS into a single number that can predict quantitative phenotype or disease risk. One barrier to the use of PRS in clinical practice is that the majority of GWAS come from cohorts of European ancestry, and predictive power is lower in non-European ancestry cohorts. There are many possible reasons for this decrease; here we show that differences in allele frequencies, LD patterns, and phenotypic variance across ancestries are unlikely to be driving this pattern. We focus on PRS for height in cohorts with admixed African and European ancestry, which allows us to test for ancestry-related differences in PRS prediction while controlling for environment. We first show that that the predictive power of height PRS increases linearly with European ancestry (partial R2 ranges from 0.02-0.12 for 0-100% European ancestry). We replicate this pattern with effect sizes re-estimated within sibling pairs, ruling out residual population structure. This pattern persists when PRS is computed using subsets of SNPs in regions of both high and low LD and ancestry-related differences in effect size are not correlated with local recombination rate. This suggests that differences in LD are not a major driver of low transferability. Next, we show that frequency differences of associated variants between African and European ancestry backgrounds explain only up to 11% of the observed reduction in predictive power and that there is no association between ancestry and phenotypic variance, indicating that the reduction in PRS predictive power cannot be explained by causal variants that are specific to the African ancestry background. Finally, we see a modest improvement in prediction when using a multi-PRS approach that includes ancestry-specific effect sizes in the PRS. We conclude that the reduced predictive power in non-European ancestry populations is largely explained by differences in causal effect sizes across these ancestries.