Polygenic scores in disease prediction: evaluation using the relevant performance metrics

We examine the performance of polygenic scores in screening and disease prediction using metrics which are firmly established for non-genetic tests but surprisingly, rarely used in evaluations of polygenic scores. Using reported metrics from the Polygenic Score Catalog (odds and hazard ratios, area under the receiver operating characteristic curve or C-index), we calculate the sensitivity or detection rate (DR5) for polygenic score cut-offs that define a 5% false positive rate (FPR). To examine clinical and public health relevance, we use information on disease incidence to calculate the odds of being affected given a positive result (OAPR; the ratio of true to false positives) both for an individual and a population. Polygenic scores typically detect 7.6-16.4% (and therefore miss 83.6-92.4%) of affected individuals with a 5% FPR. For a polygenic score for coronary artery disease (CAD) with a DR5 of 12% (88% of cases missed), the OAPR is 1:3.75 if used in a population with an average 10-year CAD risk of 10%, and 1:41 in a population with an average 10-year CAD risk of 1%. For a polygenic score for breast cancer also with a DR5 of 12% and a population average 10-year risk of 1 in 65 (odds =1:64), odds are reduced to 1:91 for a woman with a polygenic score at the 25th centile and increased to 1:53 for a woman with a result at the 75th centile. Analysis using the relevant metrics reveals the weak predictive performance of polygenic risk scores, which limits their effectiveness in screening and disease prevention.