Newer epigenetic clocks responded more consistently to longevity interventions
Researchers recalculated the same 16 clocks across 51 human intervention studies, so the results could finally be compared. Clocks built to predict mortality risk or the pace of aging responded more consistently than clocks built to estimate chronological age. Twenty-six interventions produced no clear overall effect, and a lower score still has no agreed clinical meaning.
Epigenetic age tests are one of the most widely sold things in longevity medicine. You give a blood or saliva sample, and a few weeks later you are handed a number that is meant to say how old your body is, as distinct from how old you are. Clinics throughout this directory offer them, usually with a plan for bringing the number down.
Which raises a question that nobody had been able to answer properly: when someone actually takes an intervention, does the number move? Individual studies had looked, using different clocks and different methods, and their answers could not be compared with one another. A group led from Yale School of Medicine has now gone back and recalculated the same 16 clocks across 51 human intervention studies, which is the first time these results have been put on one footing. Some clocks moved a great deal more reliably than others, and which one your clinic uses turns out to matter.
Done in people, which is the setting that counts. It still matters who was studied, how many, and for how long.
Not a new experiment. Researchers went back over studies already done and analyzed them together, which can reveal patterns no single study is big enough to show. It inherits the limits of the studies it pools.
Published in a journal after review by independent researchers in the field.
Humans. 3,128 blood samples pooled from 51 existing longitudinal intervention studies.
Secondary analysis. The authors did not run a new trial. They gathered the raw methylation data from studies already done and recalculated all 16 clocks the same way for every one.
3,128 samples across 51 interventions, with 16 clocks calculated for each.
The National Institute on Aging, part of the NIH, with fellowship support from Yale University and an Impetus Grant.
Several authors declare commercial interests in epigenetic-age testing: two are named co-inventors of a patented clock, three are employees of a testing company whose biomarker features in the analysis, and two report consulting fees from that company and others in the field. The remaining authors declare none. Some of the pooled data was private and supplied by that company.
What an epigenetic clock measures
DNA methylation is a chemical tag that sits on top of DNA and helps decide which genes are switched on. The pattern of those tags shifts through life in ways that are partly predictable, so a statistical model can be trained to read the pattern and estimate something from it. That model is the clock. It is not a measurement of your body's age. It is a prediction, and what it predicts depends entirely on what it was trained to predict.
That is the distinction the whole field turns on, and it is why 16 clocks give 16 answers. Clocks built on a different molecule can behave differently again: six protein-based clocks run over one small drug trial all moved the same way, which is the opposite of what the epigenetic clocks do here. The first clocks, such as Horvath1 and Hannum, were trained to guess chronological age from methylation. They are good at it, which is a slightly strange thing to want from a test, since you already know your age. Later clocks were trained on something more useful: PhenoAge and GrimAge were built around markers of disease risk and mortality, and DunedinPACE was built to estimate the rate at which someone is aging rather than a point on a scale. Newer ones again, such as SystemsAge and OMICmAge, try to report several biological dimensions rather than one global score.
What the researchers did
They assembled a database of 51 longitudinal intervention studies with blood methylation data taken before and after, 3,128 samples in all, and recalculated all 16 clocks from the raw data using one pipeline. That is the part that had not been done. Two studies reporting different answers previously could have been disagreeing about the intervention, or about the clock, or about the processing, and there was no way to tell which.
The interventions were a mixed group: prescription medicines including metformin, semaglutide and drugs targeting TNF in inflammatory conditions; diets and exercise programs; supplements; hyperbaric oxygen; senolytics; and a handful of surgical and transplant studies. They were also a mix of designs rather than all randomized and controlled, which is a real limit on what pooling them can show.
Which clocks responded
The clearest finding is that the clocks are not interchangeable, and the newer ones did better. Clocks trained around mortality risk or the pace of aging responded most consistently. Clocks trained mainly to predict chronological age moved least, which fits earlier work suggesting those first-generation clocks largely track random epigenetic noise rather than anything biological. Among lifestyle interventions, DunedinPACE registered significant decreases in 8 of 15; among drug interventions, GrimAgeV2 in 8 of 14.
Prescription medicines and combined lifestyle programs produced the strongest responses, anti-TNF therapies and metformin most consistently of all. Over-the-counter supplements did not produce comparable changes.
Taken across all 16 clocks together, 19 of the 51 interventions significantly lowered epigenetic age, and 13 of those held up after correcting for the number of comparisons made. Five raised it, three after the same correction. The remaining 26 showed no significant effect in that combined test. The senolytics studies illustrate why the authors set consistency rules before believing any single result: biomarkers moved in opposite directions within one study, and the same biomarker sometimes did across studies. Senolytics satisfied neither rule.
How many interventions moved nothing at all
That figure of 26 is worth slowing down on, because it is easy to read as something it is not. The authors tested each intervention by taking its effect on all 16 clocks together and asking whether the group of them differed from zero. Twenty-six interventions did not. That is a statement about the 16 clocks as a set. It is not a statement about any one of them, and an intervention can fail that combined test while still moving a clock or two on its own.
The paper does not report how often that happened, but the data to work it out is published. Supplementary Table 4 lists every one of the 51 interventions against every one of the 16 clocks, giving an effect size where that clock responded significantly and marking the rest. Reading it against Supplementary Table 5, which holds the combined test:
- Of the 26 interventions with no significant combined effect, 16 moved at least one clock on its own. One of them, a senolytics combination, moved nine.
- Ten of the 51 interventions moved no clock at all. Those include exercise, semaglutide, omega-3 fatty acids, folate, and vitamin B12 with folate.
Those two counts are ours and not the paper's. We produced them by cross-referencing the two supplementary tables the authors published, and the method is simple enough to repeat: take the 26 interventions marked not significant in Table 5, find each of them in Table 4, and count how many have at least one clock carrying an effect size rather than the not-significant marker. The paper is open access and both tables download from it, so this is checkable rather than something to take on trust. The paper's own figure for that paragraph is 26, and it answers the other question.
Fifty-one squares in three groups: 25, then 16, then 10.
We are reporting this because an earlier version of this article got it wrong in the direction that mattered. It said no clock detected a change in 26 of the 51 studies, which treated the combined test as though it were a per-clock result. The number for that sentence is 10.
Responding and being useful are different questions
A clock that reliably moves is a better research instrument than one that does not. It is not the same as a clock that tells you something about your life.
Three questions are worth holding apart. Did the measurement change: for some clocks and some interventions, clearly yes. Did the person become healthier: not established here, because this analysis looked at biomarker readings and not at health outcomes. Does a lower reading mean less disease, more healthy years or a longer life: unknown. The authors are explicit that there is no agreed threshold for what counts as a clinically important change, so nobody can tell you what a given shift is worth, and it is not known whether short-term movement tracks anything long term.
Healthy people and patients did not behave the same way
This is the finding most likely to matter to someone reading this, and it is easy to miss in the paper. Of eight study-level characteristics the authors tested, the health of the population was the one most strongly associated with whether biomarkers responded at all.
In groups of people with a disease, several clocks responded significantly and survived correction for multiple comparisons. In healthy groups, only a few showed any response, and none of those survived the same correction. Most of the people buying these tests from a clinic are in the second group. The interventions that moved the clocks most were also drugs given to patients being treated for something, rather than optimization programs sold to well people.
If you are considering one of these tests
This analysis is genuinely useful for that decision, because it puts every clock on the same footing and says which ones responded. That turns a vague question into three sharp ones. Which clock does this clinic use, and was it one of the ones that responded consistently. Was the evidence for it generated in people like me, or in patients being treated for a disease. And what would the clinic actually do differently depending on the result, because a number that changes nothing is a number you are paying to look at.
A clock reading is a research measurement still being validated. It is most useful as one line in a longer trend, read alongside established health measures, and it is not a substitute for age-appropriate screening. Our explainer on biological age and organ age covers how these scores are built and why two of them can disagree about the same person.
Sources
- Nature Medicine: Responsiveness of epigenetic aging biomarkers to longevity interventions in humans. Sehgal R, Borrus D, Armstrong JF, et al., published 21 August 2026, open access. Aggregate intervention results in Supplementary Table 5; the intervention-by-biomarker matrix in Supplementary Table 4; healthy against disease populations in Supplementary Tables 13 to 15.
- Yale School of Medicine news release, 21 August 2026. Used for framing, not as independent validation.
- bioRxiv: DNAm aging biomarkers are responsive: Insights from 51 longevity interventional studies in humans. Posted October 2024, the earlier preprint of the peer-reviewed paper rather than a separate result.
Corrections
- September 13, 2026
An earlier version was headlined "In 26 of 51 studies, not one epigenetic clock detected a clear change", and said the same in the body. That overstated the paper by more than a factor of two. Twenty-six interventions produced no significant effect when all 16 clocks were tested together, which is a different question from whether any single clock responded. Cross-referencing the study's two relevant supplementary tables, 16 of those 26 moved at least one clock on its own, and 10 of the 51 interventions moved no clock at all. The article now reports the paper's figure of 26 for the combined test, and gives our count of 10 as ours, with the working published.
Prepared by the LongevityClinicsHub editorial team. Evidence last checked September 2026. How we research and source these guides.
Enter your ZIP code and we'll point you at the nearest city we've researched.
This report is independent editorial content, written to orient, not to diagnose. It is not medical advice, and a research result is not a treatment you can act on. Where we describe what a company, a clinic or a researcher claims, the claim belongs to them.
Educational reading, not medical advice. Talk to a clinician about your own health.