Which Epigenetic Clocks Actually Respond to Longevity Interventions
A massive harmonized database of 51 studies reveals which epigenetic aging clocks are most sensitive to real-world longevity interventions.
Summary
Epigenetic clocks are often used to judge whether an anti-aging intervention works, but nobody had compared them side by side. Researchers built a database of 51 human studies and ran the same 16 clocks on every one. Clocks trained to predict mortality or pace of aging, such as DunedinPACE, PCGrimAge, GrimAge V2 and SystemsAge, responded most consistently. Drug-based and lifestyle interventions moved the clocks the most. Anti-TNF drugs and Mediterranean diets gave consistent results across studies, while senolytics and one supplement blend gave mixed, conflicting signals. Responses were bigger in people with disease and depended on study duration. The work does not prove any clock is a validated stand-in for lifespan, but it gives trial designers a data-driven way to choose clocks and interventions.
Detailed Summary
Testing an aging intervention in humans could take decades, so researchers want biomarkers that change quickly and predict long-term outcomes. DNA methylation (DNAm) clocks are the leading candidates. Although they predict morbidity and mortality, little was known about whether they respond to interventions, a prerequisite for use as surrogate endpoints. Prior studies used different clocks, populations and designs, which made comparison nearly impossible.
To fix this, the authors built TranslAGE, a harmonized database of 51 longitudinal interventional studies drawn from public repositories (GEO, EMBL) and private data (TruDiagnostic). They calculated 16 prominent epigenetic clocks plus 94 other DNAm biomarkers for each study using MethylCIPHER 2.0, removed the effect of chronological age, standardized the residuals, and compared pre- and post-intervention samples with paired t-tests. Four control datasets and nine pro-aging event datasets served as sanity checks. Aging events raised epigenetic age (effect size 0.14), interventions lowered it (−0.06), both at P<0.0001.
Across interventions, 19 significantly decreased epigenetic age (13 after multiple-testing correction), 5 increased it (3 after correction), and 26 had no significant effect. Pharmacological interventions had the largest pooled effect (mean −0.093) and lifestyle interventions were also significant (mean −0.039), while supplements and medical procedures were not. Pharmacological effects were larger than lifestyle (P=0.0063) and supplements (P=0.0009).
Clock choice mattered. Reliability-optimized second-generation clocks trained on mortality or pace of aging responded best: DunedinPACE (mean −0.089, P=0.0012), PCPhenoAge (−0.089, P=0.003), GrimAge V2 (−0.081, P=0.0075), PCGrimAge (−0.078, P=0.0003) and SystemsAge (−0.059, P=0.0043). DunedinPACE fell in 16 interventions and rose in one. First-generation clocks like Horvath and Hannum changed sporadically without a clear pattern. PCGrimAge, PCHannum and PCPhenoAge agreed with other clocks 90–100% of the time. DunedinPACE never disagreed with other clocks, though it was the only significant clock in 17.6% of its hits. Clocks also differed by intervention type: DunedinPACE and DunedinPoAm38 led in lifestyle studies, while GrimAge V2 led in drug studies.
The authors proposed that a real effect should appear across multiple clocks within a study and replicate across studies. Anti-TNF therapies (in arthritis and inflammatory bowel disease) and two Mediterranean diets in healthy cohorts met both rules. Two SRW supplement studies and five senolytic studies did not, with clocks moving in opposite directions. Population health status was the strongest predictor of response: studies in people with disease showed larger effects (Z-scores 4.83 for PCPhenoAge, 4.65 for SystemsAge and 4.07 for PCGrimAge). Study duration and population characteristics also shaped responsiveness.
The findings offer a practical guide for trial design, including which clocks to prioritize to limit multiple testing and how to size studies. Caveats remain: responsiveness is not the same as proof that a clock predicts clinical benefit, effects were small, and the studies were heterogeneous.
Key Findings
- TranslAGE pooled 51 longitudinal human intervention studies, with 16 epigenetic clocks and 94 additional DNAm biomarkers calculated uniformly
- Aging events raised epigenetic age (effect size 0.14) while interventions lowered it (−0.06), both P<0.0001; 19 interventions decreased epigenetic age (13 after correction), 5 increased it, 26 had no effect
- Pharmacological interventions showed the largest pooled effect (mean −0.093) versus lifestyle (−0.039); drugs beat lifestyle (P=0.0063) and supplements (P=0.0009)
- DunedinPACE was the most responsive clock, decreasing in 16 interventions and increasing in 1 (pooled mean −0.089, P=0.0012); PCGrimAge had the strongest significance (P=0.0003)
- PCGrimAge and PCHannum agreed with other clocks 90% of the time and PCPhenoAge 100%, versus 50–66% for original Horvath and Hannum clocks
- Anti-TNF therapy and Mediterranean diets replicated across studies and clocks; five senolytic studies and two SRW supplement studies gave inconsistent directions
- Disease populations showed larger clock responses than healthy ones (Z=4.83 for PCPhenoAge, 4.65 for SystemsAge, 4.07 for PCGrimAge)
Methodology
The authors curated 51 longitudinal interventional studies with whole blood DNAm data from GEO, EMBL and private TruDiagnostic datasets, plus four control and nine pro-aging event datasets as validation. They calculated 16 clocks and 94 other DNAm biomarkers with MethylCIPHER 2.0, regressed out chronological age, and standardized residuals within each dataset. Pre- versus post-intervention changes were tested with paired t-tests, and pooled effects with one-sample t-tests, with multiple-testing correction. Linear regression related eight study-level variables, including health status, sample size, duration, age and percent female, to effect sizes, and cell-fraction adjustment was tested.
Study Limitations
Responsiveness is a prerequisite for surrogate endpoint status but does not show that clock changes predict clinical outcomes or lifespan. The studies were heterogeneous in population, intervention, duration and design, effect sizes were small, and some findings did not survive multiple-testing correction. All data came from whole blood, and potential conflicts of interest (including use of private TruDiagnostic data) were not detailed in the supplied text.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: