Fine-scale population structure confounds genetic risk scores in the ascertainment population

Genetic risk scores (GRS), also known as polygenic risk scores, are a tool to estimate individuals liabilities to a disease or trait measurement based solely on genetic information. They have value in clinical applications [1] as well as for assessing relationships between traits and discovering causal determinants of complex disease [2, 3]. However, it has been shown that these scores are not robust to differences across continental populations [4, 5] and may not be portable within them either [6]. Even within a single population, they may have variable predictive ability across sexes and socioeconomic strata [7], raising questions about their potential biases. In this paper, we investigated the accuracy of two different GRS across population strata of the UK Biobank [8], separated along principal component (PC) axes, considering different approaches to account for social and environmental confounders. We found that these scores did not predict the real differences in phenotypes observed along the first principal component, with evidence of discrepancies on axes as high as PC45. These results demonstrate that the measures currently taken for correcting for population structure are not sufficient, and the need for social and environmental confounders to be factored into the creation of GRS.