A family-aware All of Us benchmark for polygenic prediction and genomic machine learning

Polygenic risk scores (PRSs) have improved substantially in predictive performance, yet results remain difficult to compare across studies that differ in cohorts, phenotype definitions, and evaluation procedures. Comparison to machine-learning methods is harder still, since their performance depends strongly on how data are partitioned and whether related individuals leak across splits. Here, we introduce a family-aware benchmark for polygenic prediction built on version 9 of the All of Us Research Program (AoU). The benchmark assigns entire family components to fixed discovery, calibration, and test partitions, reserves first-degree relatives for within-family evaluation, and separates models trained only on AoU discovery data from those using external GWAS summary statistics. This enables direct comparison between conventional PRS pipelines and machine-learning models trained under identical data budgets. We report results for 14 binary disease traits. Our strongest scores explain 5.7-24.9% of liability-scale variance, with particularly strong prediction for Alzheimer's disease, type 2 diabetes, prostate cancer, glaucoma, and inflammatory bowel disease, and match or exceed published commercial and academic scores for most traits. Across 12 traits with interpretable within-family estimates of within-sibship-to-population effect ratio, 11 confidence intervals included a ratio of 1. Type 2 diabetes showed an attenuation ratio of 0.86 [0.76, 0.97].