Quantifying the genetic separability of disease subtypes

Genetic prediction of clinically defined disease subtypes could advance precision medicine, but polygenic risk scores (PRSs) often show limited ability to distinguish subtypes, either because the subtypes share genetic architecture or because available PRSs are inaccurate. We introduce GenSep, a statistical framework that estimates the oracle case-case AUC attainable if the true additive genetic values of both subtypes were known, and the proportion of this genetic separation recovered by subtype-specific PRSs. Across 18 subtype pairs in UK Biobank, oracle AUCs ranged from 0.605 to 1.000, whereas PRSs trained within UK Biobank achieved a median observed case-case AUC of only 0.543 (range, 0.521-0.752), reaching a median of only 26.8% (range, 9.7-50.4%) of the above-chance discrimination attainable at the oracle AUC. Oracle AUCs were generally consistent among UK Biobank, All of Us and FinnGen, and among European, African and Admixed American ancestry groups within the Million Veteran Program. PRSs trained on FinnGen GWAS, despite 2.5- to 3.8-fold larger discovery samples, showed no significant improvement over internally trained PRSs, whereas consortium-scale GWAS improved case-case AUC in three of four overlap-free comparisons but remained far below the oracle AUC. GenSep distinguishes subtype pairs whose discrimination is limited by intrinsic genetic overlap from those whose prediction is limited by current PRS accuracy, identifying where larger studies and better predictors could improve genetic discrimination.