Validation of fracture-derived polygenic scores with FRAX for fracture risk prediction in postmenopausal women

We evaluated whether adding a genetic risk score derived from forearm fracture improves fracture prediction beyond standard clinical tools. Incorporation of GPS resulted in a small increase in time-dependent AUC, with improvement in NRI. Decision curve analysis suggested potential clinical benefit across certain threshold ranges. BACKGROUND: Previous studies have incorporated estimated bone mineral density (eBMD)-derived polygenic scores into FRAX to improve fracture risk prediction, indirectly capturing genetic susceptibility through bone mineral density. However, the genetic architectures of fracture and BMD only partially overlap, suggesting that fracture-specific genetic risk may provide complementary and more direct biological information. With the availability of recently released forearm fracture GWAS summary statistics, we developed Bayesian genome-wide polygenic scores (GPS) derived from fracture risk and evaluated whether integrating fracture-derived Bayesian GPS into FRAX could enhance clinical fracture risk prediction. METHODS: We constructed two Bayesian GPS using recently released UK Biobank forearm fracture GWAS summary statistics (GWAS Catalog study ID: GCST90281273) with two Bayesian frameworks: polygenic risk score-continuous shrinkage (PRS-CS) and summary-data-based Bayesian regression with continuous shrinkage and functional annotation integration (SBayesRC). The fracture-derived Bayesian GPS were integrated into FRAX to derive fracture-specific GPS-FRAX models. Model performance was evaluated in 10,135 postmenopausal women from the Women's Health Initiative (WHI) using time-dependent area under the receiver operating characteristic curve (AUC), Brier score, net reclassification improvement (NRI), calibration analysis, and decision curve analysis (DCA). RESULTS: Compared with the FRAX-CRF model (time-dependent AUC = 0.683), fracture-derived Bayesian GPS-FRAX models demonstrated modest improvements in discrimination, with time-dependent AUCs of 0.693 for PRS-CS and 0.690 for SBayesRC. Overall reclassification proportions were low (1.58% for SBayesRC and 1.82% for PRS-CS), but NRI was significantly improved, with overall NRI estimates of 2.20% (95% CI, 0.83 to 3.58%) for SBayesRC and 2.72% (95% CI, 1.34 to 4.19%) for PRS-CS. These findings indicate incremental predictive gain beyond clinical FRAX factors despite modest increases in discrimination. CONCLUSIONS: Incorporation of fracture-derived Bayesian GPS into FRAX resulted in a modest but statistically significant improvement in model discrimination, as reflected by a small increase in AUC. In addition, improvements in NRI and higher estimated net benefit in decision curve analysis suggest incremental clinical utility. However, the overall magnitude of improvement remained limited, indicating that the added predictive value beyond established clinical risk factors is modest. Further evaluation in more diverse populations is warranted.