JointPRS
Summary¶
JointPRS is a data-adaptive hierarchical Bayesian framework designed for multi-population polygenic risk score prediction using summary statistics. It leverages genetic correlation structures across diverse ancestries to improve prediction accuracy, particularly for underrepresented populations that have limited sample sizes. The framework is highly flexible, supporting automated estimation (JointPRS-auto), summary statistic meta-analysis (JointPRS-meta), and validation-based tuning (JointPRS-tune).
Methodology¶
JointPRS models the joint distribution of effect sizes across populations, incorporating linkage disequilibrium (LD) matrices and population-specific genetic correlations. The framework addresses a key limitation of prior methods — such as PRS-CSx and LDpred2-multi — by using a data-adaptive shrinkage scheme that adapts to the sample sizes and genetic architectures of the discovery cohorts.
The framework supports three operational modes depending on target cohort data availability: - JointPRS-auto: Runs fully automatically without requiring individual-level validation or tuning data, estimating shrinkage hyperparameters directly from GWAS summary statistics. - JointPRS-meta: Integrates multi-population summary statistics with meta-analytic approaches before applying shrinkage. - JointPRS-tune: Leverages a small individual-level tuning/validation set to optimize hyperparameters, achieving maximum prediction accuracy.
Performance¶
In large-scale benchmarking on 22 quantitative and 4 binary traits across five continental populations (utilizing UK Biobank and All of Us data), JointPRS outperformed several state-of-the-art PRS methods: - JointPRS-auto achieves an average increase in predictive (R^2) of 17% to 27% over PRS-CSx-auto in diverse ancestral cohorts. - The framework demonstrates robust performance even when cross-population genetic correlation is low or sample sizes in non-European cohorts are extremely small.