A polygenic risk score (PRS) uses genetic data to estimate an individual’s risk for a particular disease. At-risk individuals can be monitored frequently, enabling early detection of disease. This benefit is often limited, however, because many PRSs have only modest predictive power.
We hypothesize that predictive power is limited in part by natural selection. When selection acts on disease liability, it changes the frequencies of variant forms of the underlying genes. In addition, it changes the statistical association among genes at different locations within the genome. Although the latter effect is usually ignored, it is often larger than the former. This becomes important when a PRS is constructed. A PRS is built from estimates of the effect of each genetic variant on liability to the disease. The methods currently used to estimate these effects ignore all or most of the statistical associations mentioned above. Consequently, effect sizes are systematically underestimated to an extent that depends on the strength of natural selection. Thus, the PRS of a disease under strong selection should predict more poorly than that of one under weak selection.
To test this hypothesis, we will employ a new method for estimating the effect of natural selection on “quantitative traits”, that is, continuously varying phenotypes such as weight and blood pressure. Our method has greater power than alternatives, because it incorporates information about the statistical associations discussed above. We will apply this method to a series of disease-related quantitative traits from the UK Biobank in order to search for signals of selection. We will also construct a PRS for each trait. Our hypothesis predicts that the PRS of a trait under strong selection will predict more poorly than that of one under weak selection.