Epigenetic clocks estimate a person’s biological age from chemical tags on DNA, chiefly methyl groups added at particular positions. The first clocks were trained to guess calendar age. Later ones were trained on health and mortality or on how fast a person is aging, and they behave differently. A 2023 randomized trial found that two years of calorie restriction slowed one pace-of-aging clock slightly, and left two other clocks unchanged. That mixed result sums up the field: promising measuring sticks, not yet validated stand-ins for how long or how well someone will live.

This explainer covers what a clock measures, how the best-known ones differ, what recent trials show, and what has to happen before a clock can serve as a trial endpoint. It is general science information, not medical advice.

What is actually being measured

DNA methylation is a reversible chemical mark. In the human genome it occurs mainly at sites where a cytosine base is followed by a guanine, called CpG sites. The pattern of marks shifts with age and responds to cell type, smoking, inflammation and other influences. A clock is a statistical model: it takes methylation levels at a chosen set of CpG sites, weights them, and outputs a number.

The key phrase is “trained to predict.” A model is only as meaningful as the target it was trained on. Train it on birth certificates and it tracks calendar time. Train it on mortality risk and it tracks something closer to health. Train it on change across years in one person’s physiology and it tracks speed.

The main clocks compared

Clock Trained to predict Notable facts from the source papers
Horvath clock (2013) Chronological age 353 CpG sites; built across 51 tissues and cell types; correlation about 0.96 with age in test data, median error about 3.6 years
DNAm PhenoAge (2018) A composite clinical “phenotypic age” 513 CpG sites; each extra year of PhenoAge was tied to about 4.5% higher all-cause mortality risk after adjusting for chronological age
DunedinPACE Pace of aging over time Designed to read speed rather than a single age; used in the CALERIE trial analysis

The Horvath clock, described in Genome Biology in 2013, was a landmark because one model worked across so many tissues. Its limitation follows from its goal. Because it is tuned to calendar age, it is excellent at telling a 20-year-old from a 70-year-old and less informative about who among 70-year-olds is healthiest.

Morgan Levine, Steve Horvath, Luigi Ferrucci and colleagues addressed that in their 2018 paper in Aging. They first built a “phenotypic age” from clinical measures in NHANES III, a U.S. survey of 9,926 adults with mortality follow-up, then trained a methylation predictor of it using blood samples from 456 people. In five independent cohorts, the resulting clock predicted death better than earlier clocks. People in the fastest-aging 5% had roughly 1.62 times the mortality hazard of the average person. The authors also noted that the underlying clinical measure predicted outcomes even better than the methylation version, which is a useful reminder that a clock approximates something we can already measure more directly.

What the CALERIE trial showed

The Comprehensive Assessment of Long-term Effects of Reducing Intake of Energy trial randomized 220 healthy adults without obesity to 25 percent calorie restriction or normal eating for two years. A post hoc analysis led by Rubina Waziry and Daniel Belsky, published in Nature Aging in 2023, looked at blood methylation clocks.

The result was narrow. DunedinPACE showed slower aging in the restriction group, with standardized effects of -0.29 at 12 months and -0.25 at 24 months, which the paper’s Results section translates to a roughly 2 to 3 percent reduction in pace. Changes in the PhenoAge and GrimAge clocks did not differ significantly between the groups. The authors themselves describe the effect sizes as small, and the analysis was post hoc, meaning it was not the trial’s original plan.

What does this tell us? A diet can shift a methylation measure in a randomized setting, which is notable. It does not show that anyone will live longer or avoid disease because of that shift. Whether a 2 to 3 percent change in a pace score means anything for outcomes is a different question that this study was not built to answer.

Why clocks are not yet accepted endpoints

In 2024, members of the Biomarkers of Aging Consortium published a review titled Validation of biomarkers of aging in Nature Medicine. Their central point: biomarkers could serve as surrogate endpoints in trials of interventions meant to promote healthy aging, but there is no consensus on how they should be validated before clinical use. They highlight predictive validity for aging-related outcomes in population studies, along with comparability and generalizability across studies and groups.

Three practical problems follow.

  • Surrogate status. A surrogate must be shown to track the outcome that matters, and an intervention must move the surrogate and the outcome together. For clocks, few such links have been demonstrated.
  • Disagreement between clocks. CALERIE found one clock responded while two others did not. Different training targets yield different answers about the same person.
  • Noise and context. Methylation varies with blood cell composition, sample handling and platform. A change for one person between two tests may or may not exceed that noise.

The state of longevity trials

Few trials of anti-aging interventions run long enough to count deaths and disease. That is the reason researchers want a validated surrogate. In the absence of one, the large prizes and programs have turned to functional measures. The XPRIZE Healthspan competition, a $101 million, seven-year program launched in 2023, judges therapies by whether they restore muscle, cognitive and immune function in people aged 50 to 80, with a goal of at least a 10-year functional benefit and a treatment that takes a year or less to deliver.

The page lists 20 finalist teams and ten Milestone 2 awardees who each receive $1 million, with announcements dated August 11, 2026. That is a competition structure, not a set of published efficacy results, and none of the interventions have been shown to extend human lifespan on the strength of that listing alone.

A note on who was studied

The CALERIE participants were healthy adults without obesity, tested over two years. PhenoAge was built from U.S. survey data and a few hundred methylation samples. Neither tells us how a clock behaves in people with chronic illness, in different ancestries, or over decades. Each of these gaps is one reason the consortium review stresses generalizability, and each is a reasonable thing to ask about before trusting a headline built on a clock.

How to read a consumer biological-age number

Commercial tests report a “biological age” derived from a clock. Treat it as an interesting research-grade readout rather than a diagnosis. If two companies disagree, that is expected, since they may use different clocks. If a number changes after a lifestyle shift, it may reflect real biology or measurement noise. Nothing in the studies above supports making a medical decision from a single reading, and a clinician can interpret blood work that actually has established clinical meaning, such as blood pressure, lipids and glucose.

Questions people ask

Is biological age the same as real age?

No. Calendar age counts years since birth. Biological age is a model-based estimate of how worn or healthy the body’s systems look, and the answer depends on which model you use. Two clocks can give the same person two different numbers without either being wrong, because they were built to answer different questions.

Can I lower my score, and does that mean I am healthier?

A score can move. CALERIE shows that a sustained diet change moved one clock slightly. What nobody has shown is that moving a clock by itself reduces disease. The score may be a symptom of the underlying biology, a marker that rides along with it, or partly an artifact of how blood cells shift under a given treatment.

Why do the clocks disagree?

Because they were trained on different targets, in different people, with different CpG sites. Horvath’s model learned calendar age, PhenoAge learned a mortality-linked health score, and DunedinPACE aims at speed. Agreement is not guaranteed even in the same blood sample.

What would make clocks more useful

  1. Trials with a clock measured at the start and the end, plus real clinical outcomes tracked over many years.
  2. Standardized methods across labs, so a result in one study can be compared to another.
  3. Validation across ancestries and ages, not only in the groups already best studied.
  4. Evidence that lowering a clock score through an intervention actually lowers disease or mortality.

Until that evidence exists, epigenetic clocks are best understood as sharp research instruments. They tell scientists something real about biology and help them design better experiments. They do not yet tell any individual how many years they have left.

Sources