Can Human Genetic Evidence Predict Drug Success? New Analysis Offers a More Sober Answer

Drug development has increasingly relied on human genetics in recent years: if a gene has a clear association with disease risk, then in theory, designing a drug to target it should have a better chance of hitting the cause of disease. This intuition has supported many investment and R&D strategies, but a new analysis points out that although genetic evidence is associated with drug approval, it may not by itself be a reliable predictor of success.

The preprint, posted on arXiv on June 12 and not yet peer reviewed, analyzed 26,278 “drug target-disease” pairs in the Open Targets and ChEMBL databases. The study showed that pairs with human genetic associations had a higher drug approval rate than those without such evidence; in the pair-level analysis, the odds ratio was 3.25, indicating a clear association between the two.

But after the authors further adjusted the unit of analysis, the picture became more nuanced. When the analysis was shifted to the drug-target level, avoiding amplification of the signal from the same gene repeatedly appearing across multiple disease pairs, the overall odds ratio fell to 2.79. The effect was especially evident in oncology: a figure that had appeared to reach 6.72 at the pair level fell to 2.71 after adjustment. This suggests that some seemingly strong differences across therapeutic areas may partly come from the structure of the data itself, rather than entirely from biological effects.

Another key issue is data leakage. After separately testing six categories of evidence, the study found that literature mining contributed almost most of the classification model’s performance. However, approved drugs and popular targets often generate more subsequent papers, and these publication records may mix “knowledge that appeared only after success” into the prediction model. After excluding literature mining, other evidence still performed above the benchmark, but the signal shrank, indicating that genetic data is neither entirely useless nor the crystal ball some may imagine.

The study also examined drugs approved after 2015 and found that the association between genetic evidence and success could still be reproduced. However, the authors also noted that genetic evidence alone increased the model’s absolute AUPRC by only about 1.0 percentage point, and the best model’s calibration performance was also unsatisfactory. In other words, it may improve the average success rate of a research portfolio, but it is unlikely to tell an R&D team whether a specific candidate target will truly pass clinical and regulatory tests.

The value of this analysis may lie not in overturning the role of genetics in drug development, but in drawing its boundaries. The authors identified 1,433 genetically supported target-disease pairs that are still in Phase 1 or Phase 2 clinical stages, which could serve as a resource for subsequent hypothesis generation. But all results are observational analyses and cannot prove that genetic evidence itself causes drug success. For the biotechnology industry, this is a pragmatic reminder: human genetics can help narrow the search space, but it cannot replace disease biology, clinical design, and safety judgment.