Blood Sample Trajectories Years Before Diagnosis Predict Osteoporosis Risk

Osteoporosis is a disease of slow erosion. Bone mineral density declines gradually over decades, often silently, until a fracture announces what screening scans might have caught earlier. Current clinical tools, including fracture risk assessment scores and dual-energy X-ray absorptiometry, require clinical visits and measure the outcome of bone loss rather than its progression. Now, a team of Danish researchers has shown that the key to earlier warning may already be sitting in freezer archives: repeated blood donations collected years before anyone knew a donor would develop the disease. By analyzing longitudinal, pre-diagnostic plasma samples from the Danish Blood Donor Study, the investigators demonstrated that tracking how metabolites change over time within a single person predicts osteoporosis far better than comparing single snapshots between people.

The study, published in iScience, exploited a uniquely powerful resource. The Danish Blood Donor Study encompasses more than 2.7 million plasma samples from over 165,000 healthy donors enrolled since 2010. By linking this biobank with national health registries, the researchers identified donors who were later diagnosed with osteoporosis after their first donation. For each case, they retrieved up to six pre-diagnostic samples spaced eight to fourteen months apart, alongside multiple samples from matched controls who never received an osteoporosis diagnosis. The result was a genuinely longitudinal dataset: 78 participants contributing 418 samples in the longitudinal arm, and a separate cross-sectional comparison set of 120 participants with single time points.

Crucially, the cohort captured early-stage disease. Ninety-six percent of the cases were diagnosed without fracture events, and eighty percent continued donating blood after their diagnosis, indicating they remained healthy enough for routine donation. This matters because post-diagnostic samples are notoriously confounded: fractures themselves, treatments such as bisphosphonates and denosumab, and lifestyle changes after diagnosis all distort the molecular picture. By working strictly with samples drawn before diagnosis and before treatment, the team sidestepped much of this contamination and asked a cleaner question: does the blood chemistry of future osteoporosis patients already drift differently years in advance?

The analytical platform was untargeted metabolomics using ultra-performance liquid chromatography coupled to high-resolution quadrupole time-of-flight mass spectrometry, run in both positive and negative electrospray ionization modes. After rigorous cleaning, normalization, and filtering, the datasets retained 3,221 features in positive mode and 1,429 in negative mode. Principal component analysis revealed a sobering reality of single-timepoint studies: cases and controls overlapped almost completely across the first six principal components, and samples from one individual were as dispersed as samples across the entire dataset. In other words, within-person variability was comparable to between-person variability, which is precisely why cross-sectional biomarker hunts so often stall.

To turn that liability into an asset, the researchers computed the slope of each metabolite’s trajectory across each participant’s five or six donations and fed those slopes into LASSO logistic regression, a regularized machine learning approach that shrinks uninformative features to zero. The contrast with the cross-sectional design was striking. Cross-sectional models, using only the latest sample before diagnosis, performed barely better than chance, with area under the receiver operating characteristic curve (AUC) values of 0.63 for negative ionization data and 0.57 for positive data. The longitudinal slope-based models reached an AUC of 0.76 for negative ionization data and 0.67 for positive data, despite using fewer independent participants. Random forest models fared poorly across all settings, suggesting the disease-related metabolic changes are predominantly linear and additive rather than complex nonlinear patterns.

The team then probed how far back in time the signal extended. By sequentially removing the most recent pre-diagnostic samples and re-evaluating the model, they showed detectable metabolic progression one to two years before diagnosis, with AUC around 0.80 when the most recent 200 days of data were removed and 0.65 at roughly 600 days. This degradation pattern establishes a potential screening window for population-level monitoring. Functional data analysis, a more flexible smoothing approach, slightly underperformed the simple slopes, likely because five or six time points provide too little temporal resolution for functional smoothing to shine. The parsimonious slope representation proved less prone to overfitting while still capturing the rate and direction of metabolic change.

Generalizability received unusually careful treatment. In a hold-one-batch-out validation, in which entire analytical batches containing matched case-control pairs were excluded from training and used only for testing, the negative ionization model still achieved an AUC of 0.70, statistically significant against chance. This design ensures complete independence between training and validation data and directly addresses a well-known reproducibility crisis in clinical metabolomics, where promising biomarkers frequently fail external validation. The authors note that batch-wise validation is rare in the metabolomics literature, and its explicit demonstration here strengthens confidence that the model captures biology rather than technical artifacts. Longitudinal designs also carry an inherent robustness: features fluctuating due to technical confounders show inconsistent slopes within individuals and are naturally down-weighted by the model.

Among the features the model selected, one molecule stood out. Hippuric acid, a co-metabolite produced jointly by host metabolism and gut microbes, showed high coefficients in both ionization modes with consistent regulation and effect size. Healthy controls showed a time-dependent increase in hippuric acid, while future osteoporosis cases showed a decrease. This aligns with prior literature indicating that hippuric acid rises with age but remains low in people with degenerative diseases or frailty. Mechanistic work by other groups suggests hippuric acid inhibits osteoclast differentiation through the GPR109A receptor and suppresses inflammation-driven bone resorption, processes particularly relevant after menopause when declining estrogen shifts bone marrow toward fat production. The researchers caution, however, that hippuric acid also tracks visceral fat and cardiovascular risk, so lifestyle confounding through diet and microbiome composition cannot be excluded, and no mechanistic experiments were performed in this study to separate association from causation.

Several known osteoporosis-related molecules were conspicuously absent from the feature set, including 25-hydroxyvitamin D, estrogen metabolites such as 17β-estradiol, and bone resorption markers like deoxypyridinoline and pyridinoline. The authors attribute these absences to analytical selection bias and annotation limitations typical of untargeted metabolomics, not to their absence from the underlying biology. Other limitations deserve emphasis: diagnoses relied on registry codes rather than direct bone density measurements, the predominantly female and Northern European donor population limits generalizability, and controls were matched on health status rather than including disease controls, which could produce false positives in broader populations. The cost and complexity of repeated sampling also constrain real-world deployment, and the authors note that longitudinal sampling is most feasible in patient-scarce diseases where recruitment is the bottleneck.

Even with these caveats, the conceptual contribution is substantial. The authors propose shifting interpretation from binary classification toward acceleration scoring, analogous to how cholesterol and blood pressure guide cardiovascular prevention. A metabolic acceleration score derived from routine blood samples already collected for other purposes could flag individuals on steep bone-loss trajectories while they remain asymptomatic, enabling early lifestyle interventions to preserve bone mineral density. With osteoporosis affecting more than ten percent of the Danish population and populations worldwide aging rapidly, the Danish Blood Donor Study, with thousands of participants carrying over a decade of donation history, offers a template for extending this framework to other degenerative diseases. The message resonates beyond bone: when it comes to reading the body’s chemistry, watching how things change may matter more than measuring where they stand.

Subject of Research: Longitudinal pre-diagnostic blood metabolomics for predicting osteoporosis risk

Article Title: Longitudinal pre-diagnostic blood samples predict osteoporosis risk

Article References: Lassen, J. K., Nielsen, K. L., Hindhede, L., Mikkelsen, S., Kjerulff, B., Ostrowski, S. R., Sørensen, E., Mikkelsen, C., Pedersen, O. B. V., Bruun, M. T., Aagaard, B., Johannsen, M., Erikstrup, C., & Villesen, P. (2026). Longitudinal pre-diagnostic blood samples predict osteoporosis risk. iScience, 29(11), Article 117737. https://doi.org/10.1016/j.isci.2026.117737

Image Credits: AI Generated

DOI: 10.1016/j.isci.2026.117737

Keywords: osteoporosis, metabolomics, biomarkers, longitudinal study, Danish Blood Donor Study, hippuric acid, bone mineral density, machine learning, LASSO regression, plasma, early diagnosis, gut microbiome