Recalibrating Polygenic Risk Scores to Local Data Sharpens Disease Risk Prediction

Polygenic risk scores promise to flag people at elevated inherited risk of common diseases, but scores built in one population often lose accuracy when applied in another. A new study in Nature Communications proposes a practical fix: instead of generating new scores from scratch, clinicians can recalibrate existing published scores to their own local patient data before using them.

The framework, developed by researchers including Min Seo Kim and Akl C. Fahed, integrates multiple published polygenic risk scores for a given disease and reweighs them according to how well each performs in the target cohort. Using machine learning-based optimization, the method produces a composite, locally optimized score that is fine-tuned to the disease risk patterns of the local population.

The team tested this post-PRS calibration approach on three diseases where clinical trials of polygenic score utility are currently underway: coronary artery disease, breast cancer, and prostate cancer. These conditions were chosen because they are the focus of active efforts to bring polygenic scoring into routine clinical care.

Working within the Mass General Brigham Biobank, the researchers found substantial improvements in predictive performance. The C-statistic, a standard measure of risk prediction accuracy, improved by an average of 19.8 percent for coronary artery disease, 9.4 percent for prostate cancer, and 8.3 percent for breast cancer compared with the best available published scores.

Case identification improved as well. The locally optimized coronary artery disease score flagged 27.7 percent of cases in the cohort, compared with 17.8 percent for the top-performing published score. The gains were even more striking for the cancers: the optimized breast cancer score identified 20.8 percent of cases versus 7.1 percent for the published score, and the optimized prostate cancer score identified 16.4 percent versus 8.8 percent.

These results are measured in a single health system’s biobank, and the reported improvements reflect predictive performance within that setting rather than demonstrated clinical outcomes. The authors describe the approach as a low-complexity, post hoc calibration step that requires no new genome-wide association studies and no complex upstream algorithmic innovations, making it feasible for diverse health systems to adopt.

The central idea is to treat polygenic risk scores as adaptable rather than static. Recalibration to local data offers a way to mitigate the disparities that arise from one-size-fits-all applications of externally derived scores.

The framework arrives at a moment when polygenic risk scores are being evaluated in clinical trials for coronary artery disease, breast cancer, and prostate cancer. By showing that a simple reweighting of existing scores can meaningfully improve case identification in a local cohort, the study provides what the authors call a practical blueprint for clinical adoption of polygenic scoring in diverse local settings.

Limitations remain. The reported C-statistic gains and case enrichment figures come from one biobank, and the study is an early version shared ahead of final editorial processing. Whether the same calibration gains translate to other health systems, other ancestries, and other diseases will require further evaluation. Still, the approach offers a pragmatic path for institutions that want to use published polygenic scores without commissioning new discovery research.

For now, the message is straightforward: published polygenic risk scores need not be accepted as fixed products. Reweighting them against local data can sharpen their ability to identify individuals at high risk, potentially easing one of the main barriers that has kept these scores out of routine clinical practice.

Subject of Research: Machine learning-based recalibration of polygenic risk scores to local biobank data improves disease risk prediction.

Article Title: Fine-tuning of polygenic score to local data improves risk prediction and clinical implementation

Article References: Kim, M. S., Sui, Y., Patel, A. P., Cho, S. M. J., Koyama, S., Paruchuri, K., Truong, B., Shim, I., Wang, M., Bitar, F. F., Chen, Q., Ellinor, P. T., Natarajan, P., & Fahed, A. C. (2026). Fine-tuning of polygenic score to local data improves risk prediction and clinical implementation. Nature Communications. https://doi.org/10.1038/s41467-026-77770-z

Image Credits: AI Generated

DOI: 10.1038/s41467-026-77770-z

Keywords: polygenic risk scores, risk prediction, coronary artery disease, breast cancer, prostate cancer, precision medicine, machine learning, biobank, genetic risk, clinical implementation, Fine-tuning, polygenic