Summary
A medRxiv preprint introduces GenSep, a framework that estimates how distinct disease subtypes could be using genetic information and compares that benchmark with polygenic risk scores. Across 18 subtype pairs, the authors report that current scores captured a small share of the estimated genetic separation.
Researchers have introduced GenSep, a statistical framework for estimating how well genetic information could distinguish between clinically defined disease subtypes. In a medRxiv preprint, they applied it to 18 subtype pairs in UK Biobank and compared its theoretical benchmark with the performance of polygenic risk scores (PRSs).
The distinction matters because a weak score can have two explanations: the subtypes may share much of their genetic architecture, or the available score may not capture the genetic differences accurately. GenSep is designed to estimate how much separation is theoretically available from additive genetic effects, then assess how much subtype-specific PRSs recover.
A theoretical benchmark for genetic distinction
A PRS combines information about many genetic variants into a single score associated with a trait or disease. A case–case area under the curve (AUC) measures how well a score distinguishes people with one subtype from people with another. An AUC of 0.5 corresponds to chance-level ranking, while 1.0 represents perfect separation.
GenSep estimates an “oracle” case–case AUC: the separation that would be attainable if the true additive genetic values for both subtypes were known. This is a benchmark, not a score available for routine testing. Comparing it with the AUC from current PRSs gives the researchers a way to distinguish limited underlying genetic separation from limited prediction accuracy.
For the 18 subtype pairs, estimated oracle AUCs ranged from 0.605 to 1.000. The observed AUC for current PRSs had a median of 0.530, with a range of 0.510 to 0.754. The authors estimate that these scores recovered a median 4.1% of the genetic separation variance available to a perfect predictor; the range was 0.3% to 15.5%.
Larger studies and broader comparisons
The authors report that oracle AUC estimates were generally consistent across UK Biobank, All of Us, FinnGen and the Million Veteran Program, and across European, African and Admixed American ancestry groups.
More discovery data did not automatically produce stronger subtype discrimination in the comparisons reported. PRSs trained using FinnGen data, despite discovery samples 2.6 to 3.9 times larger, performed no better than scores trained internally. By contrast, using substantially larger consortium genome-wide association studies raised case–case AUC by an average of 0.06, with increases ranging from 0.003 to 0.089.
The findings suggest that the route to better genetic distinction may differ by subtype pair: some may be limited by genetic overlap, while others may offer more room for improved prediction. GenSep is a research framework for assessing that distinction, rather than evidence that genetic scores can yet guide clinical decisions. The report is a medRxiv preprint, posted on September 23, 2026.