77. CHARACTERIZING THE UNCERTAINTY, MISCLASSIFICATION AND INCONSISTENCY OF POLYGENIC PREDICTION

Background Polygenic risk scores (PRSs) hold promise for stratifying individuals by genetic risk, yet substantial uncertainty in individual-level PRS estimates remains poorly characterized. Multiple PRSs for the same disease often yield discordant risk classifications, raising concerns about the reliability and clinical interpretability of PRS-based stratification. Here, we develop a unified inferential framework to calibrate PRS point estimates and quantify uncertainty for both quantitative traits and binary phenotypes, and to characterize how PRS accuracy, uncertainty, and pairwise correlation jointly determine misclassification and classification inconsistency. Methods Under a Bayesian polygenic prediction framework, we derive methods for slope and variance calibration of PRSs and show that prediction accuracy and uncertainty constitute complementary components of total genetic variance. We further derive analytic expressions for individual- and population-level misclassification probabilities when thresholding PRSs to identify high-risk individuals. Extending the framework to multiple PRSs, we derive multivariate models that characterize disagreement across PRSs as a function of their prediction accuracies and pairwise correlations. We additionally develop methods to quantify uncertainty in integrated PRSs constructed from weighted combinations of multiple scores and evaluate PRS integration and uncertainty-aware probabilistic thresholding for risk stratification. Results Using extensive simulations across a range of genetic architectures, GWAS sample sizes, and disease prevalences, we demonstrate that theoretical predictions closely match empirically observed misclassification and inconsistency rates. Misclassification and inconsistency were highest near decision thresholds and decreased monotonically with increasing PRS accuracy. We further show that highly accurate PRSs must necessarily be strongly correlated, imposing fundamental constraints on the degree of disagreement expected among well-performing scores. Finally, we applied the framework to five representative complex traits and diseases – body mass index, total cholesterol, coronary artery disease, type 2 diabetes, major depressive disorder – in an ancestrally diverse sample from the All of Us Research Program using multiple PRSs per phenotype. After slope and variance calibration, predicted disease probabilities closely matched empirically observed case proportions, whereas uncalibrated PRSs underestimated disease risk. Calibrated PRSs exhibited larger posterior uncertainties in East Asian and African individuals relative to European individuals, reflecting lower predictive accuracy in non-European populations. Predicted misclassification and inconsistency rates closely aligned with empirical observations across traits and thresholds. Sequential integration of complementary PRSs improved predictive performance and reduced inconsistency, while uncertainty-aware probabilistic thresholding preferentially selected individuals with higher-confidence risk estimates, thereby improving positive predictive value and reducing classification discordance. Discussion Together, these results demonstrate that instability in PRS-based classification is a predictable statistical consequence of uncertainty and establish a principled framework for incorporating uncertainty into PRS interpretation, integration, and clinical risk stratification.