Large human-study analysis tests which epigenetic clocks respond to longevity interventions

A large analysis published August 21 in *Nature Medicine* has compared how consistently prominent epigenetic aging biomarkers change during human intervention studies. The findings could help researchers choose biomarkers for clinical trials, but they do not validate epigenetic age as a substitute for disease, disability, survival or other clinically meaningful outcomes.

The investigators assembled TranslAGE, a database of 51 public and private longitudinal intervention studies containing 3,128 pre- and post-intervention blood samples. They applied a harmonized pipeline to calculate 16 epigenetic clocks and 94 additional DNA-methylation biomarkers across the datasets.

This approach addressed a persistent comparability problem: individual studies often calculate different clocks in populations receiving different interventions for different lengths of time. A clock that appears responsive in one trial therefore may never have been tested under comparable conditions elsewhere.

What the analysis found

Clocks trained to estimate mortality risk, phenotypic aging or the pace of aging generally responded more strongly and consistently than first-generation clocks designed primarily to predict chronological age. DunedinPACE produced the largest standardized effects, while PCGrimAge produced the greatest number of statistically significant results. SystemsAge and other biomarkers with multiple interpretable components also helped indicate which physiological systems were changing.

When the studies were grouped by intervention type, pharmacological interventions produced larger average changes in DNA-methylation biomarkers than lifestyle programs, supplements or medical procedures. Lifestyle interventions also produced statistically detectable average reductions in epigenetic-age measures, although their average effect was smaller.

Those category-level findings should not be read as a ranking of treatments. The pharmacological group included very different agents and populations, including anti-TNF therapies used in people with inflammatory disease, metformin, rapamycin, semaglutide and ketamine. The researchers found that biomarkers were often more responsive in participants with disease than in healthy populations, potentially because greater baseline dysregulation left more room for measurable change.

The most reproducible patterns included changes during anti-TNF treatment across studies of arthritis and inflammatory bowel disease, and during two Mediterranean-diet studies. By contrast, studies of senolytic interventions produced inconsistent directions of change across clocks and datasets. None of these comparisons establishes that an intervention slowed organism-wide aging or should be used for longevity.

A response is not yet a surrogate

Responsiveness is a necessary property for a trial biomarker: a measure that never changes cannot capture an intervention effect. It is not sufficient to qualify the measure as a surrogate endpoint.

FDA materials define a surrogate endpoint as a substitute for a direct measure of how a patient feels, functions or survives. Establishing that role requires evidence that changes in the surrogate reliably predict changes in the clinical outcome of interest. The new analysis did not test whether short-term movement in an epigenetic clock mediated or predicted later reductions in disease, disability or death.

That distinction is especially important because newer clocks incorporate signals related to inflammation, smoking, proteins or metabolic and cardiovascular risk. A treatment could improve one of those intermediate factors and consequently move the clock without altering aging biology more broadly. A favorable clock result therefore cannot automatically be translated into years of life gained or biological-age reversal.

Important limitations

The underlying studies differed substantially in participant age, health status, sample size, duration, design and intervention. The authors characterized their pooled comparisons as descriptive rather than causal. Preprocessing and batch correction were not fully harmonized across all datasets, and some data came from private studies.

There is also no empirically established minimal clinically important difference for these clocks. Without a threshold tied to function, physiology, disease or survival, even a statistically significant change has uncertain practical meaning.

The analysis nevertheless offers a useful trial-design map. It supports prioritizing technically reliable, outcome-oriented clocks and selecting biomarkers suited to the population and intervention being studied. The next step is harder: prospective trials must show that intervention-driven clock changes track meaningful human outcomes before epigenetic aging measures can carry the evidentiary weight of surrogate endpoints.