Adaptive boosting for pre-trained polygenic risk scores

A peer-reviewed study in Nature Communications, published August 29, 2026 by Jie Hu, Raelynn Chen and colleagues at the University of Pennsylvania, Cedars-Sinai and partners, addresses a weakness in existing polygenic-risk models: they can miss some predictive signals.

The team's contribution is a fine-tuning framework using adaptive variable selection and boosting, validated on data from the All of Us, eMERGE and Penn Medicine Biobank programmes.

The practical application is that pre-trained clinical risk models can be improved without rebuilding them from zero.

The limitations matter: gains vary by setting, and performance and fairness must be tested again for each population and each clinical use before deployment.