Research Brief · Measuring aging

Sex-specific aging clocks found differences—and large limits

The models exposed sex-dependent patterns across organs and health outcomes. They did not become clinical tests, individual forecasts or evidence that aging was slowed.

Lois RunePublished October 10, 2026Updated October 10, 2026

Abstract editorial graphic showing two reference paths flowing into age-gap markers, followed by a clear boundary before clinical interpretation. — **Editorial graphic.** A publication-owned visual for this article. It is not a measured result, scientific evidence, diagnosis or treatment claim.

Evidence scope

Model development plus observational and genetic association analyses

Population or model: Sex-stratified MRI, proteomic and metabolomic age models using MULTI Consortium cohorts, with downstream UK Biobank, BLSA, ADNI and A4 analyses
Design: age-prediction model development and internal holdouts, followed by prospective associations, genetic analyses and secondary reuse of trial data
Comparator: female- and male-stratified reference models, with matched sex-pooled models used for selected performance comparisons
Measured outcome: chronological-age prediction, biological age gaps, genetic and molecular associations, incident disease, mortality and cognitive trajectories
Publication status: peer-reviewed open-access Version of Record, first online September 16, 2026; Crossmark reported the document current on October 10
Funding & conflicts: NIH MULTI and NIA intramural support; reused A4 data came from a public-private-philanthropic partnership. Authors declared no competing interests.
Safety & clinical context: No intervention was tested in this clock analysis, no clinical cutoff was validated and no regulator-cleared aging test was identified.

Bottom line: Sex-stratified models revealed differences worth studying, but internal prediction and downstream associations do not make a biological-age score clinically actionable.

Not established: diagnosis, individual prognosis, a validated surrogate endpoint, treatment response, slower aging, longer healthspan or longer life.

What the researchers built

The study trained 38 sex-specific age-prediction models across 15 organ systems. Inputs included magnetic-resonance imaging, plasma proteins and metabolites. Each model learned chronological age separately in people classified as female or male in the source data.

The headline output was a “biological age gap”: model-predicted age minus chronological age. A positive gap means the model estimated an older age than the calendar says; a negative gap means younger. That subtraction is a model residual. It is not a direct reading of cellular damage, and it does not say how many years a person will live.

From reference model to association—not to diagnosis

Reference dataFemale and male samples were modeled separately.
Age modelMRI, proteins or metabolites predicted calendar age.
Age gapPredicted age minus actual age created a study variable.
AssociationResearchers tested links with later outcomes.
Clinical boundaryNo diagnosis, individual forecast or treatment effect follows automatically.

Text equivalent: the study moved from sex-stratified reference data to an age-prediction model, then calculated predicted age minus chronological age and tested that gap against health outcomes. Every step adds assumptions. The chain stops before clinical interpretation: an association does not validate a diagnostic test or show that changing the gap changes health.

Separate models changed the reference, not the meaning of sex

The rationale was straightforward: if aging-related features differ by sex, a pooled reference curve may blend distinct patterns. In matched comparisons, some sex-stratified models showed less performance loss from training to holdout data. The authors also found different genetic, protein and outcome associations across female and male analyses.

That does not establish two fixed aging programs. The paper treats sex as a binary variable and does not model hormones, gender, social exposures or changes across the life course. It also says pooled and sex-interaction models remain useful when shared biology, sample size or stability matters. The practical lesson is to report which reference population a model uses—not to assume stratification is always superior.

Prediction performance was mixed

Across within-distribution holdouts, mean absolute error was about five years. Several clocks nevertheless showed pronounced overfitting: performance worsened outside the samples used during model development. The male eye proteomic clock was a sharper warning. Its predicted ages had almost no correlation with chronological age in the holdout set (r=0.024; P=0.7), even though its average absolute error looked similar to training.

The researchers excluded clearly overfit clocks from downstream work and urged reporting more than one performance metric. But internal holdouts are not the same as external validation. A model can perform acceptably inside related source data and still fail across health systems, scanners, laboratories, ancestries or age ranges.

Later disease and mortality links were associations

In UK Biobank follow-up, some age gaps were associated with incident disease and all-cause mortality after multiple-testing correction. The report counted 12 female and 14 male clock–mortality associations that met its corrected threshold. That is useful prognostic research, not a bedside forecast. The article does not provide a validated cutoff, absolute-risk calculator or decision rule for an individual.

ADNI analyses also linked brain-age gaps with progression from mild cognitive impairment to Alzheimer’s disease. Reported hazard ratios were 1.74 in males and 2.25 in females, yet the sex difference in the Kaplan–Meier curves was not statistically significant (P=0.90). Two significant associations cannot be converted into proof that the sexes differ when the direct comparison remains uncertain.

A trial dataset did not turn the clocks into a treatment endpoint

The authors reused imaging and cognitive data from 1,055 participants in the A4 study, registered as NCT02008357. They reported differences in cognitive trajectories through week 240 when participants were grouped by brain-age gap. This was a secondary exploratory association inside an existing dataset.

A4 itself was a randomized Phase 3 test of solanezumab in cognitively unimpaired older adults with elevated brain amyloid. Its primary results found that solanezumab did not slow cognitive decline versus placebo. The 2026 clock analysis does not reverse that result, isolate a clock-mediated treatment effect or show that moving brain age would change cognition.

Genetic analyses add hypotheses, not intervention proof

The study ran bidirectional Mendelian-randomization screens between 38 age gaps and 525 disease endpoints. These analyses can probe whether inherited variants support a direction of association. They still depend on instrument strength, absence of horizontal pleiotropy and other assumptions; the authors flagged sensitivity concerns for some signals.

The genetic analyses were restricted to European-ancestry data and used genetically proxied binary sex. Few repeated measurements were available, so the work mostly compares people at one point rather than directly measuring their rate of change. Those boundaries limit transport to other ancestries and to the broader biological and social experience of sex and gender.

Funding and disclosure context

The clock study reports support from NIH MULTI award RF1AG092412 and the National Institute on Aging’s intramural program. The reused A4 study was supported through a public-private-philanthropic partnership that included NIH/NIA, Eli Lilly, the Alzheimer’s Association, Accelerating Medicines Partnership, GHR Foundation and other donors, with in-kind support from organizations including Avid and Cogstate.

The clock-paper authors declared no competing interests. That disclosure should be reported as written; it does not erase the commercial partners attached to reused trial data, and it is not a reason to discount or accept the findings on its own.

What should happen next?

Independent teams should validate the clocks in external biobanks and health systems, report calibration and subgroup performance, and test stability over repeated measurements. Any proposed clinical use would then need evidence that the score changes decisions or outcomes without causing avoidable harm.

A harder question comes after prediction: does an intervention-induced change in a clock reliably predict a meaningful change in function, disease, disability or survival? This study did not answer it. Until that evidence exists, sex-specific clocks are research tools for studying model behavior and biological hypotheses—not instructions for testing or treating an individual.

Sources and editorial method

Three research passes were completed before prose. Every substantive statement maps to S89-01 through S89-14 in the frozen evidence record. A separate evidence review followed drafting. See issue #189.

AI disclosure: AI-assisted research and drafting were used only after the evidence map was frozen. Primary identity, methods, quantitative results, limitations, funding, disclosure and publication status were checked against the evidence record. This article is educational and not medical advice.

See what aging measurements can—and cannot—tell us

Publication history

Published October 10, 2026 after three separately documented research passes, frozen atomic claim mapping and a separate evidence review. Publication status, linked-trial identity, funding, disclosures and limitations remain subject to the release-time recheck.