Responsiveness of epigenetic aging biomarkers to longevity interventions in humans (paper 21 Aug 26)

https://www.nature.com/articles/s41591-026-04562-9

There also a linked commentary behind a paywall from Steve Horvath
https://www.nature.com/articles/s41591-026-04524-1

chatGPT(5.6paid):

Overall assessment

This is a valuable large-scale benchmarking study of how DNA-methylation biomarkers respond to putative longevity interventions. Its chief contribution is not evidence that any particular intervention slows ageing, but the creation and standardized analysis of TranslAGE: 51 longitudinal studies, 3,128 blood samples, 16 major epigenetic clocks and 94 additional methylation-derived biomarkers.

The most defensible conclusion is:

Some second-generation, technically reliable methylation biomarkers—particularly DunedinPACE and PCGrimAge—change more consistently after interventions than first-generation chronological-age clocks.

The paper does not establish that those changes represent slowed ageing, improved healthspan or extended lifespan.

1. Summary

Study objective

Epigenetic clocks are frequently proposed as short-term surrogate endpoints for longevity trials. But a biomarker must do more than predict mortality: it must also respond reproducibly when an intervention changes the underlying ageing process.

The authors therefore asked:

  • Which epigenetic clocks are most responsive to interventions?
  • Which types of intervention produce the largest methylation changes?
  • Are responses reproducible across related studies and clocks?
  • Do population characteristics affect responsiveness?
  • Can “explainable” or system-specific biomarkers indicate what biology is changing?

Dataset and analysis

The authors assembled 51 longitudinal intervention datasets containing pre- and post-intervention blood methylation measurements. These included:

  • lifestyle interventions, such as diets and exercise;
  • pharmacological interventions, including metformin, anti-TNF treatment, rapamycin, semaglutide and ketamine;
  • supplements;
  • medical procedures, including transplantation and hyperbaric oxygen.

They calculated more than 110 methylation biomarkers with a consistent computational pipeline. Sixteen broadly applicable ageing clocks were used for the principal comparisons.

Biomarker values were:

  1. residualized against chronological age;
  2. divided by an external reference standard deviation to produce standardized effect sizes;
  3. compared before and after treatment using paired tests.

Principal findings

Second-generation reliable clocks were most responsive

Mortality- and pace-of-ageing clocks generally changed more consistently than first-generation chronological-age clocks.

The strongest performers were:

  • DunedinPACE: the broadest apparent sensitivity and the largest average response;
  • PCGrimAge: the strongest overall statistical evidence;
  • GrimAgeV2, PCPhenoAge and SystemsAge: also relatively responsive.

DunedinPACE decreased significantly in 16 interventions and increased in only one. It also never significantly disagreed in direction with the other clocks, although in some studies it was the only significant clock.

By contrast, Horvath, Hannum and related first-generation clocks showed sporadic increases and decreases without a consistent overall pattern. The principal-component versions of several clocks were generally more consistent than the original versions, supporting the importance of technical reliability.

Pharmacological interventions produced the largest average changes

Pharmacological interventions produced larger standardized reductions than lifestyle, supplement or procedural categories. Lifestyle interventions also produced an average reduction, but of smaller magnitude.

Anti-TNF treatment and metformin generated some of the strongest patterns. Mediterranean-diet studies showed comparatively consistent changes across studies.

However, the authors appropriately state that this comparison concerns biomarker responsiveness, not comparative clinical efficacy.

Nineteen interventions produced aggregate decreases

Across the 16 clocks:

  • 19 interventions produced nominally significant aggregate decreases;
  • 13 remained significant after the authors’ multiple-testing correction;
  • 5 produced increases, of which 3 survived correction;
  • 26 produced no significant aggregate effect.

The paper does not provide grounds for interpreting every significant reduction as rejuvenation.

Health status strongly affected responsiveness

Clocks generally changed more in populations with diagnosed disease than in healthier populations.

PCPhenoAge, SystemsAge, GrimAgeV2 and PCGrimAge showed particularly large responses in disease cohorts. DunedinPACE was unusual in responding comparatively similarly in healthy and disease populations.

This could reflect:

  • more biological dysfunction available to reverse;
  • regression to the mean;
  • treatment of an active disease process;
  • the populations and outcomes used to train the clocks.

Different clocks responded to different intervention classes

DunedinPACE was particularly responsive to lifestyle interventions, whereas GrimAgeV2 and PCGrimAge were more responsive to pharmacological interventions.

This argues against treating all clocks as interchangeable measures of a single latent “biological age”.

Explainable biomarkers offered apparent biological specificity

SystemsAge, OMICmAge and their component measures indicated system-specific responses. Examples included:

  • smoking cessation → strongest apparent lung response;
  • metformin → inflammatory, metabolic and brain components;
  • gastric bypass → metabolic component;
  • vegan diet → inflammatory and musculoskeletal components.

Across seven dietary interventions, musculoskeletal SystemsAge declined in all seven, while inflammation and metabolic components responded in six. Methylation proxies for triglycerides were responsive in six diets.

Structured versus stochastic methylation change

First-generation Horvath age changes correlated with their estimated stochastic component, whereas PhenoAge changes showed little relationship to its stochastic component.

The authors interpret this as evidence that second-generation biomarkers preferentially capture structured, biologically directed changes rather than random methylation drift.

That is an interesting observation, but “nonstochastic” does not necessarily mean “causal ageing”.

2. What is genuinely novel?

The main novelty: standardized cross-study recalculation

Previous intervention papers generally selected their own preferred clocks. This creates a fragmented and publication-biased literature in which one cannot tell whether discordant results arise from the intervention, the population or selective biomarker reporting.

The authors instead recalculate the same large biomarker panel across all accessible datasets. This permits something close to a head-to-head comparison of clocks.

That is the paper’s strongest and most useful innovation.

TranslAGE as an infrastructure contribution

The harmonized TranslAGE database is itself an important research product. If maintained and expanded, it could support:

  • prospective biomarker selection;
  • power calculations;
  • replication analysis;
  • comparison of newly developed clocks;
  • eventual validation against clinical outcomes.

Joint analysis of responsiveness, reliability and concordance

The paper distinguishes three properties that are often conflated:

  • whether a biomarker changes;
  • whether it changes consistently with other biomarkers;
  • whether technically more reliable versions behave better.

The finding that principal-component and updated clocks generally outperform their originals provides useful empirical guidance for trial design.

Intervention-dependent biomarker selection

The demonstration that lifestyle and pharmacological interventions are optimally detected by somewhat different clocks is novel and practically important. It challenges the idea that one universal clock should serve every intervention.

Incorporation of explainable component scores

The systematic comparison of whole-body clocks with organ-, pathway-, protein- and metabolite-related methylation scores is more informative than reporting a single number called biological age.

It suggests a possible future trial architecture:

  • one robust global ageing measure;
  • several intervention-relevant system scores;
  • conventional biochemical and functional outcomes.

Population health as an explicit effect modifier

The systematic demonstration that diseased cohorts yield larger clock responses is important. It implies that a clock’s apparent responsiveness cannot be separated from the population in which it is tested.

3. Critique

A. Responsiveness is not surrogate validity

This is the central conceptual limitation.

A useful surrogate must satisfy at least three distinct conditions:

  1. it predicts a clinically meaningful outcome;
  2. an intervention changes it;
  3. the intervention-induced change predicts the intervention’s effect on that clinical outcome.

This paper primarily tests condition 2. It does not demonstrate condition 3.

A clock might fall after treatment because it detects:

  • reduced inflammation;
  • altered blood-cell composition;
  • improved glucose or lipid metabolism;
  • reversal of a disease state;
  • a direct pharmacological effect on methylation;
  • a technical or sampling effect.

None of these necessarily means that the rate of organismal ageing has slowed.

The correct description is therefore responsive biomarker, not validated surrogate endpoint.

B. The study largely treats pre–post change as the intervention effect

The analysis uses paired pre- versus post-intervention comparisons, even though some source studies apparently contained control groups.

A pre–post difference can be caused by:

  • secular change;
  • seasonal variation;
  • acute illness or recovery;
  • altered medication;
  • regression to the mean;
  • repeated sampling or batch effects;
  • natural ageing during follow-up.

For randomized studies, the appropriate causal estimand is normally the difference in change between intervention and control groups, preferably with baseline adjustment—not whether the intervention arm alone changed significantly.

This is probably the paper’s most important statistical weakness. A reanalysis restricted to randomized controlled trials, using treatment-by-time effects, would be much more persuasive.

C. Correlated clocks are repeatedly treated as observations

Several analyses test whether the collection of 16 clock effect sizes for one intervention differs from zero. But those clocks are strongly correlated, share CpGs, share training outcomes and sometimes derive from one another.

Consequently, the effective sample size is much smaller than 16. Treating clock values as if they were independent can generate overly small P values.

The paper acknowledges correlation when discussing multiple testing, but that does not fully solve the deeper problem of pseudoreplication. A hierarchical model should explicitly represent:

  • participants nested within studies;
  • clocks nested within biomarker families;
  • interventions nested within categories;
  • covariance among clocks.

D. Pooling selected “longevity” interventions creates directional circularity

The database includes interventions because they have been hypothesized to improve longevity or healthspan. The authors then ask whether, pooled together, their clocks tend to decrease.

That selection makes an average downward direction less surprising. It does not independently validate either the interventions or the clocks.

A stronger design would include:

  • treatments known to improve health but not ageing;
  • biologically neutral interventions;
  • harmful exposures;
  • treatments producing acute inflammation;
  • interventions with known long-term mortality effects in opposite directions.

The paper includes some controls and “pro-ageing events”, which is helpful, but the overall intervention set remains selected around a presumed anti-ageing direction.

E. Intervention-category comparisons are heavily confounded

The claim that pharmacological interventions produce larger responses does not establish that drugs are intrinsically more geroprotective.

The pharmacological group disproportionately includes:

  • patients with active disease;
  • interventions treating inflammation or metabolic dysregulation;
  • populations with high baseline abnormality;
  • potentially different study durations and sample-processing practices.

Because diseased populations themselves show larger clock responses, the pharmacological advantage may partly reflect population composition rather than intervention class.

The authors use univariate study-level regressions, but these cannot adequately disentangle correlated study characteristics. With only 51 heterogeneous studies—and only 10 disease studies—multivariable conclusions remain fragile.

F. Disease improvement may be mistaken for ageing reversal

Anti-TNF treatment provides the clearest example. Reducing systemic inflammation in inflammatory bowel disease or arthritis will predictably alter:

  • leukocyte populations;
  • CRP-associated biology;
  • methylation proxies trained on inflammatory and mortality-related variables.

A fall in GrimAgeV2 or SystemsAge could therefore mean “less active inflammatory disease”, which is clinically beneficial, without demonstrating a change in the fundamental rate of ageing.

Indeed, the strong responsiveness of generation-2 clocks may arise precisely because they contain proxies for modifiable risk factors. That makes them sensitive clinical-state markers, but not necessarily deeper measures of ageing.

G. Blood-cell composition remains a difficult confounder

The authors report that responsiveness largely persisted after regression of methylation-inferred cell fractions. This is reassuring but not decisive.

Inferred fractions:

  • are themselves calculated from methylation;
  • incompletely represent immune-cell subtypes and activation states;
  • may remove genuine biology or fail to remove compositional confounding;
  • cannot substitute for measured flow-cytometric counts.

This is particularly important for anti-inflammatory drugs, transplantation, smoking cessation and disease recovery.

H. “Explainable” does not necessarily mean mechanistic

A methylation score trained to predict a protein, metabolite or organ-related phenotype is not a direct measurement of that molecule or organ.

For example, a methylation-derived “brain” or “kidney” score measured in blood could reflect systemic correlates of those systems rather than biological change inside the brain or kidney.

Thus the organ-system results are best described as hypothesis-generating signatures, not evidence that metformin rejuvenated the brain or hyperbaric oxygen rejuvenated the lung.

They require validation against:

  • measured metabolites and proteins;
  • organ function;
  • imaging;
  • tissue-specific molecular assays;
  • clinical outcomes.

I. Significance and concordance are partly circular

The authors regard agreement among clocks as evidence that a finding is bona fide. But related clocks may agree because they share:

  • CpGs;
  • training datasets;
  • mortality-related inputs;
  • immune and inflammatory signals.

Agreement among correlated instruments is not equivalent to independent replication.

Conversely, disagreement need not mean that one result is false: two clocks may genuinely measure different biological constructs. The paper partly recognizes this, particularly in the intervention-category analysis, but its “rules” for bona fide effects sometimes overstate what concordance can establish.

J. The stochastic/causal terminology is stronger than the evidence

A reproducible, structured methylation change is not automatically causal in ageing. It may be:

  • downstream of disease;
  • compensatory;
  • environmentally induced;
  • a stable correlate of another process.

Similarly, DamAge and AdaptAge are model-derived constructs, not experimental demonstrations that particular methylation sites cause damage or adaptation. The paper’s longitudinal results support differential responsiveness of these components, but not their causal interpretation.

K. Multiple testing remains debatable

The authors estimate an effective number of independent tests using PCA and argue that conventional correction would be excessively conservative.

That can be reasonable, but concerns remain:

  • numerous related analyses and subgroup comparisons are performed;
  • the effective-test framework does not eliminate researcher degrees of freedom;
  • nominal results are discussed even when they fail correction;
  • normality is assumed for one-sample tests over sometimes small and heterogeneous sets of study effects.

A permutation framework or hierarchical false-discovery model would be more convincing.

L. Reproducibility is incomplete

The open MethylCIPHER code and TranslAGE resource are strengths. However:

  • some private TruDiagnostic datasets require controlled access;
  • OMICmAge and DNAmEMRAge calculation involves restricted/request-based components;
  • several investigators have financial or intellectual-property relationships involving TruDiagnostic or SystemsAge.

These interests do not invalidate the findings, but independent replication using fully public datasets and independently implemented clocks is especially important.

4. What the paper establishes—and what it does not

Reasonably supported Not established
Reliable second-generation clocks are more responsive than first-generation clocks That lowering a clock slows ageing
DunedinPACE and PCGrimAge are promising trial biomarkers That either is a validated surrogate endpoint
Different interventions affect different methylation constructs That one intervention is superior for longevity
Disease status strongly modifies measured responsiveness That large responses in disease represent rejuvenation
Component scores can generate mechanistic hypotheses That blood-derived organ scores measure rejuvenation within those organs
Anti-TNF and Mediterranean-diet effects show useful reproducibility That these methylation changes mediate healthspan or lifespan benefits

Bottom line

This is an important biomarker-comparison and research-infrastructure paper, rather than a validation of epigenetic clocks as measures of rejuvenation.

Its most useful practical recommendation is to avoid old chronological-age clocks as primary endpoints and instead use a prespecified portfolio containing:

  • DunedinPACE or PCGrimAge;
  • an intervention-relevant explainable/system score;
  • measured physiological and functional outcomes;
  • a randomized control group.

The decisive next study would test whether randomized treatment-induced clock changes predict subsequent changes in multimorbidity, frailty, function or mortality—and whether they statistically mediate the treatment’s clinical benefit. Until that is shown, these clocks should be treated as responsive molecular indicators, not as proof that biological ageing has been reversed.

what does it say about rapamycin

The paper includes rapamycin as one of the pharmacological interventions, but it does not present rapamycin as one of its strongest or most convincing findings.

What it reports is essentially that:

  • Rapamycin was included among the 51 longitudinal interventions analysed.
  • Its effects were assessed across the same panel of 16 principal epigenetic clocks.
  • It produced changes in some methylation biomarkers, but the response was not as broad or consistent as that seen with anti-TNF therapy or metformin.
  • The authors therefore do not single out rapamycin as an intervention with a clearly reproducible epigenetic “rejuvenation” effect.
  • Rapamycin contributes to the general finding that pharmacological interventions, taken as a category, produced larger methylation changes than lifestyle, supplement or procedural interventions.

Importantly, the paper is not testing whether rapamycin improves healthspan clinically. It is testing whether methylation clocks change after intervention. Therefore, even if individual rapamycin-associated clocks decreased, this would show only biomarker responsiveness—not that rapamycin reversed ageing.

The paper’s broader message is especially relevant to rapamycin trials: the result depends heavily on which clock is selected. A rapamycin study reporting one favourable clock while other reliable second-generation clocks remain unchanged would be weak evidence and could represent a chance or clock-specific effect. A stronger future rapamycin trial should prespecify DunedinPACE, PCGrimAge and relevant system-specific biomarkers, while also measuring function, inflammation, metabolism and adverse effects against a randomized control group.

I have not checked this response as I don’t myself worry too much about epigenetic clock results, but I uploaded the supplementary tables and asked:

given the supplementary tables what is the detailed result on rapamycin by dose

The supplementary tables show an unfavourable dose-related pattern: “rapamycin dose 1” produced no overall change, whereas “rapamycin dose 3” significantly increased—rather than decreased—the aggregate epigenetic-age signal.

Results by dose

Positive effect sizes mean movement towards greater epigenetic age; negative values mean apparent age reduction.

Clock Dose 1 Dose 3
Horvath1 +0.064 −0.090
Horvath2 +0.210 +0.178
Hannum +0.005 +0.251
PC-Horvath1 −0.009 +0.045
PC-Horvath2 −0.081 −0.002
PC-Hannum −0.004 +0.125
DNAmEMRAge +0.240* +0.006
OMICmAge +0.080 +0.018
PhenoAge +0.040 −0.009
GrimAgeV1 −0.046 +0.229*
GrimAgeV2 −0.040 +0.218*
PC-PhenoAge −0.009 +0.115
PC-GrimAge −0.098 +0.044
SystemsAge −0.016 +0.037
DunedinPoAm38 −0.006 +0.137
DunedinPACE +0.059 +0.109
Mean across 16 clocks +0.0243 +0.0883
One-sample test P = 0.3088 P = 0.0026
95% CI −0.0249 to +0.0736 +0.0362 to +0.1404

*Nominally significant individual pre–post result according to Supplementary Table 4; the supplied table does not give its exact individual P value.

Dose 1

The aggregate result was:

  • mean standardized change: +0.0243;
  • 95% CI: −0.0249 to +0.0736;
  • (t(15)=1.054);
  • P = 0.3088.

Therefore, dose 1 produced no statistically detectable overall change across the 16 clocks.

There was one nominally significant individual result:

  • DNAmEMRAge increased by +0.240.

All other clocks were nonsignificant. Several reliable second-generation clocks moved slightly favourably—PCGrimAge −0.098, PC-Horvath2 −0.081 and GrimAgeV2 −0.040—but none was individually significant.

This is therefore best interpreted as a null result with one isolated adverse signal, not evidence of epigenetic rejuvenation.

Dose 3

The aggregate result was:

  • mean standardized change: +0.0883;
  • 95% CI: +0.0362 to +0.1404;
  • (t(15)=3.610);
  • P = 0.0026;
  • partial (\eta^2=0.465).

Thus, the set of clocks shifted significantly in the direction labelled by the paper as epigenetic-age acceleration.

Two individual clocks were nominally significant:

  • GrimAgeV1: +0.229
  • GrimAgeV2: +0.218

This is more concerning than an isolated first-generation-clock result because GrimAgeV2 is one of the reliable second-generation mortality-risk clocks that the paper otherwise recommends. In addition:

  • DunedinPACE increased by +0.109;
  • DunedinPoAm38 increased by +0.137;
  • PCPhenoAge increased by +0.115;
  • Hannum increased by +0.251;
  • Horvath2 increased by +0.178.

These were not individually significant, but their broadly positive direction explains the significant aggregate result. Only Horvath1, PC-Horvath2 and PhenoAge moved slightly downwards.

What are “dose 1” and “dose 3” in milligrams?

A major reporting problem is that neither the supplied supplementary tables nor the main paper translates these coded labels into milligrams. They should not automatically be equated with the PEARL trial’s 5 mg and 10 mg arms.

The TranslAGE metadata identify them as small private retrospective datasets:

  • dose 1: TruD29, nine individuals;
  • dose 3: TruD31, seven individuals;
  • weekly compounded rapamycin over approximately one year.

The current TranslAGE database also lists a “dose 2” dataset, TruD30 with eight individuals, but it is absent from the paper’s supplied effect-size tables. TranslAGE dataset explorer

That means the apparent dose response rests on extremely small samples, and the exact dose coding needs confirmation from the authors or underlying TruDiagnostic metadata.

Important statistical qualification

The aggregate P value is obtained by treating the 16 correlated clocks as the 16 observations in a one-sample t-test. It is not a participant-level dose-response test.

This is problematic because the clocks:

  • are calculated from the same seven or nine people;
  • share methylation sites and biological inputs;
  • are strongly correlated;
  • therefore do not constitute 16 independent replications.

Nor does the analysis appear to compare each dose with a contemporaneous placebo group. These are pre–post retrospective cohorts. Consequently, it cannot establish that rapamycin caused the increase, or that dose 3 was significantly worse than dose 1. A formal dose-by-time interaction was not reported.

Bottom line

Within these supplementary data:

  • Dose 1: essentially neutral; no evidence of epigenetic age reduction.
  • Dose 3: a statistically significant aggregate shift in the adverse direction, with significant increases in GrimAgeV1 and GrimAgeV2.
  • Dose-response conclusion: suggestive of a higher-dose adverse signal, but not proven because the cohorts are tiny, non-randomized, apparently lack placebo comparisons, use undisclosed dose coding and were not directly compared statistically.

So the detailed result is not that rapamycin improved epigenetic ageing. If anything, the higher coded dose generated one of the paper’s adverse biomarker signals—but the analysis is too weak to infer that higher-dose rapamycin biologically accelerates ageing.