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

Polygenic risk scores (PRS) hold promise for stratifying inherited disease risk, yet their clinical adoption remains constrained by poor transferability across populations and cohorts. To bridge this gap, we propose a pragmatic framework that recalibrates externally derived PRSs to local datasets through machine learning-based optimization. This post-PRS calibration approach integrates multiple published PRSs for a given disease and reweighs them based on predictive performance in the target cohort, generating composite locally optimized PRS that is fine-tuned to local disease risk. We test local optimization on three diseases currently under active clinical trials for PRS utility (coronary artery disease [CAD], breast cancer [BC], and prostate cancer [PC]). Local optimization achieves substantial C-statistic improvements across diseases in MGB Biobank, with C-statistics improving by an average of 19.8% for CAD, 9.4% for PC, and 8.3% for BC. The locally optimized CAD PRS identifies 27.7% of CAD cases compared with 17.8% by the top-performing published CAD PRS, and improvements in case enrichment are more pronounced for breast cancer (20.8% vs. 7.1%) and prostate cancer (16.4% vs. 8.8%). This low-complexity post hoc calibration approach enables scalable implementation of PRS across diverse health systems without new GWAS discovery or complex upstream algorithmic innovations. By treating PRS as adaptable rather than static, the locally optimized PRS framework provides a practical blueprint for clinical adoption of PRS in diverse local settings and helps mitigate disparities stemming from one-size-fits-all applications of externally derived PRS. A locally optimized framework that integrates multiple published polygenic risk scores substantially improves disease risk prediction over individual scores, providing a practical strategy for enhancing genetic risk assessment across diverse diseases.