Fracture-linked genetic scores combined with FRAX improve fracture prediction in postmenopausal women

For millions of postmenopausal women, the decision to begin bone-protecting medication hinges on a single number: a 10-year fracture probability generated by FRAX, the clinical risk calculator that has anchored osteoporosis guidelines for nearly two decades. But FRAX has always had a blind spot. It weighs age, body mass index, smoking, prior fractures and other lifestyle factors, yet it knows nothing about a woman’s DNA—even though heredity accounts for an estimated 50 to 80 percent of the variation in bone strength between individuals. Now, a new study published in Archives of Osteoporosis suggests that closing that gap, even partially, is possible.

Researchers led by Anqi Liu, Jianing Liu and Qing Wu report that adding a genome-wide polygenic score—derived not from bone density measurements, but directly from actual fracture outcomes—produced a modest yet statistically significant improvement in FRAX’s ability to identify which women would suffer a major fracture within a decade. The findings, drawn from more than 10,000 participants in the Women’s Health Initiative, offer both a proof of concept for genetically enhanced fracture prediction and a sobering reminder of how difficult it is to move the needle beyond well-established clinical risk factors.

The central innovation of the study lies in where the genetic signal comes from. Previous attempts to graft genetics onto FRAX have relied on polygenic scores built from genome-wide association studies of bone mineral density, or its heel-based surrogate, estimated BMD. That approach captures genetic susceptibility only indirectly, through a trait that is correlated with fracture but not identical to it. The genetic architectures of bone density and fracture overlap, but only partially—a distinction that matters, because a woman can fracture a wrist without ever having crossed the diagnostic threshold for low bone density.

To capture the fracture-specific component of genetic risk, the team turned to freshly released summary statistics from a UK Biobank genome-wide association study of forearm fractures, comprising more than one million genetic variants. Forearm fracture was a deliberate choice: it is one of the four clinical events that define a major osteoporotic fracture, and a prior wrist fracture is among the strongest predictors of future osteoporotic breaks. From this discovery dataset, the researchers constructed two genome-wide polygenic scores using Bayesian shrinkage methods—polygenic risk score continuous shrinkage, known as PRS-CS, and its more elaborate cousin, SBayesRC.

The statistical machinery behind these scores is worth unpacking. Traditional polygenic score construction, often called clumping and thresholding, selects a sparse subset of variants by statistical significance and can mishandle linkage disequilibrium—the non-random co-inheritance of neighboring variants across the genome. Bayesian shrinkage methods take a different route. PRS-CS places a continuous shrinkage prior on every SNP’s effect size, combining a global shrinkage parameter with locus-specific parameters, and infers posterior effect estimates through Gibbs sampling while borrowing a European-ancestry linkage disequilibrium reference panel from the UK Biobank. SBayesRC extends this framework with an annotation-modulated mixture prior, allowing the expected magnitude of each variant’s effect to depend on functional genomic annotations—essentially letting the model down-weight variants in genomic regions unlikely to matter and up-weight those in functionally significant territory. Both methods are fully Bayesian and tuning-free, requiring no manual selection of hyperparameters, which the authors note improves reproducibility and reduces analyst degrees of freedom.

With the genetic scores in hand, the researchers built what they call GPS-FRAX models. The baseline comparator was FRAX computed from clinical risk factors alone—age, body mass index, prior fracture, parental history of hip fracture, smoking, alcohol use, glucocorticoid exposure, rheumatoid arthritis and secondary causes of osteoporosis—without any bone density input. Because FRAX outputs a bounded probability between zero and one, the team logit-transformed it before inserting it into a Fine-Gray subdistribution hazard model, a survival model specifically designed to handle competing risks. This mattered here: 218 women in the cohort died before sustaining any fracture, and ignoring those deaths would have inflated apparent fracture risk. The models also retained age as an independent covariate and, crucially, included a genetic-score-by-age interaction term, allowing the weight of inherited risk to shift across the lifespan—an acknowledgment that genetic susceptibility to bone loss may express itself differently in a 55-year-old than in a 75-year-old. The first ten principal components of genetic ancestry were included as covariates to guard against confounding by population stratification.

The validation population consisted of 10,135 postmenopausal women drawn from several Women’s Health Initiative genomic sub-studies, followed for an average of 16.4 years and analyzed over a 10-year horizon aligned with FRAX’s framework. During that window, 765 women—7.55 percent of the cohort—suffered a major osteoporotic fracture of the hip, spine, wrist or proximal humerus. The cohort’s mean age was 64.3 years, and women who fractured were, unsurprisingly, older on average than those who did not.

So how much did genetics add? Measured by the workhorse metric of predictive discrimination—the time-dependent area under the receiver operating characteristic curve at ten years—the gains were small but consistent. The clinical-risk-factor-only FRAX model achieved an AUC of 0.683. Adding the PRS-CS-derived genetic score pushed that to 0.693; the SBayesRC version reached 0.690. In reclassification analysis using the fixed 20 percent treatment threshold endorsed by the U.S. National Osteoporosis Foundation, only 1.58 to 1.82 percent of women changed risk categories when genetics were added. Yet the net reclassification improvement—a metric that weighs correctly moved individuals against incorrectly moved ones—was significantly positive: 2.20 percent for SBayesRC and 2.72 percent for PRS-CS, with confidence intervals comfortably excluding zero. Decision curve analysis, which quantifies net clinical benefit across a range of risk thresholds from zero to 25 percent, likewise suggested the genetic models delivered incremental value in clinically relevant threshold ranges.

The authors are careful about what these numbers do and do not mean. An AUC shift of one percentage point will not rewrite screening guidelines overnight, and the study’s own conclusions emphasize that the added predictive value beyond established clinical risk factors remains modest. The women in the cohort were overwhelmingly of European ancestry, and polygenic scores are notorious for poor portability across ancestries—a limitation the researchers flag directly, calling for evaluation in more diverse populations. There is also a subtle statistical wrinkle: because the coefficient of the FRAX probability was re-estimated within the Women’s Health Initiative cohort rather than fixed at its conventional value, the GPS-FRAX models are best understood as cohort-recalibrated FRAX models augmented with genetics, not as a drop-in replacement for the clinical calculator itself. A sensitivity analysis using offset-based models, with the FRAX coefficient fixed at one, was conducted to probe this distinction.

Even so, the study’s framing represents a meaningful conceptual shift in the field. By building polygenic scores from fracture itself rather than from bone density proxies, the researchers tested a more direct biological hypothesis: that the genetics of actually breaking a bone includes information—about bone quality, geometry, fall mechanics, and perhaps traits not yet measured—that density-based scores leave on the table. The forearm fracture GWAS summary statistics made that test feasible at scale for the first time, and the answer, at least in this population, is that the direct signal adds something real, if small, on top of the clinical model. A subgroup analysis among 689 women with available DXA bone density measurements further compared clinical-FRAX, FRAX-with-BMD, FRAX-with-BMD-plus-genetics, and genetics-plus-BMD-plus-age models, probing whether the genetic contribution survives once imaging data enter the picture.

The research, conducted under Institutional Review Board approval at Ohio State University using data accessed through the Database of Genotypes and Phenotypes, lands at a moment of genuine ferment in osteoporosis prediction. Related work published this year has shown that Bayesian bone-density-derived polygenic scores can also enhance FRAX, and earlier studies have demonstrated incremental gains from genome-wide scores built on associated SNPs. The accumulating evidence sketches a likely trajectory: genetic risk scoring will not replace clinical calculators, but it may progressively slot into them, particularly for the large group of postmenopausal women sitting in the ambiguous “gray zone” of intermediate risk, where the decision to treat or wait remains genuinely uncertain. For those women, even a small, statistically robust improvement in the precision of a 10-year forecast could change what gets prescribed—and, potentially, which fractures never happen. The next test, the authors make clear, is whether the same forensic use of fracture genetics holds up in populations whose genomes the discovery data never saw.