Rare-variant risk scores complement common-variant polygenic scores for disease risk prediction and stratification

Abstract

Polygenic risk scores (PRSs), which aggregate genetic effects across the genome, are typically constructed from common variants and therefore do not capture a substantial component of rare genetic variation. Using whole-genome sequencing data from the UK Biobank, we develop and benchmark rare-variant PRSs (rvPRSs) across 31 complex traits and 464 disease endpoints. Although rvPRSs provide only modest average improvements in population-level prediction beyond common-variant PRSs (cvPRSs), selected phenotypes show substantial discrimination driven by large-effect genes and rare-variant association signals not tagged by common-variant GWASs. At the individual level, rvPRSs identify largely nonoverlapping sets of individuals with extreme phenotypes or elevated disease risk compared with cvPRSs. These individuals are enriched for protein-truncating, damaging missense, or regulatory variants in biologically relevant genes, including those involved in lipid metabolism, liver function, cancer susceptibility, and cardiomyopathy. Survival analyses further show that rvPRSs stratify incident disease risk beyond cvPRSs over 15 years of follow-up. Together, these findings demonstrate that rvPRSs complement cvPRSs by enhancing tail-risk stratification and improving the biological interpretability of high-risk individuals.