Benchmarking the Accuracy of Polygenic Risk Scores and their Generative Methods

Risk prediction models provide empirical recommendations that ultimately aim to deliver optimal patient outcomes. Genetic information, in the form of a polygenic risk score (PRS), may be included in these models to significantly increase their accuracy. Several analyses of PRS accuracy have been completed, nearly all focus on only a few diseases and report limited statistics. This narrow approach has limited our ability to assess as a whole whether PRSs can provide actionable disease predictions. This investigation aims to address this uncertainty by comprehensively analyzing 23 diseases within the UK Biobank. Our results show that including the PRS to a base model containing age, sex and the top ten genetic principal components significantly improves prediction accuracy, as measured by ROC curves, in a majority 21 of 23 diseases and reclassifies on average 68% of the individuals in the top 5% risk group. For heart failure, breast cancer, prostate cancer and gout, decision curve analyses using the 5% risk threshold determined that including the PRS in the base model would correctly identity at least 60 more individuals who develop the disease for every 1000 individuals screened, without making any incorrect predictions. Analyses that included disease-specific risk factors, such as Body-Mass Index, and consider time of disease onset found similar PRS benefits. The improved prediction accuracy was translated to 10 instances in which medications/supplements and 94 instances in which lifestyle modifications lead to significantly greater reduction in disease risk for individuals in the top PRS quintile compared to the bottom PRS quintile. Finally we provide guidance for tailored, future PRS generation by comprehensively ranking methods that generate PRS weights and identifying genome wide association study characteristics that influence PRS predictions. The unification of significantly enhanced disease predictions, novel risk mitigation opportunities and improved methodological clarity indicate that PRSs carry far greater clinical impact than previously known.