Epigenetic Clocks Explained: How DNA Methylation Estimates Age

This article explains the biology and statistics behind epigenetic clocks; it does not interpret any individual test result.

The short answer

An epigenetic clock is a statistical model that estimates a person's age from chemical marks on their DNA, specifically a process called DNA methylation. Researchers measure methylation at hundreds of selected sites across the genome, then combine those measurements with weights learned from large datasets to produce an age estimate in years. The first clocks were trained to predict chronological age and did so to within a few years [Horvath 2013, Hannum 2013]. Later "second-generation" clocks were trained instead on health and mortality outcomes, so they aim to capture biological condition rather than simply guess the date on a birth certificate [Levine 2018, Lu 2019]. A further development, the pace-of-ageing measure, estimates how fast a person is ageing rather than how old they are [Belsky 2022].

These tools are genuinely useful in research, where they predict differences in disease risk and mortality across large groups [Marioni 2015]. They are also limited: technical noise can shift a reading by several years, the clocks measure correlation rather than mechanism, and whether any intervention can reliably "turn back" a clock in a way that translates to better health remains an open question [Bell 2019, Higgins-Chen 2022]. This article is the technical companion to our consumer-facing explainer of what biological age tests measure and our practical guide to how to test your biological age.

What DNA methylation actually is

DNA methylation is one of the better-characterised forms of epigenetic regulation, sitting within the broader category of epigenetic alterations described in the hallmarks of ageing [López-Otín 2023]. In mammals it usually involves the addition of a methyl group to a cytosine base where that cytosine is followed by a guanine, a pairing written as a CpG site. The human genome contains tens of millions of these sites.

Methylation does not change the DNA sequence itself. Instead it acts as a layer of annotation that influences whether nearby genes are switched on or off, and it behaves differently depending on where it sits [Jones 2012]:

• CpG islands, which are dense clusters of CpG sites often found near gene start sites, are typically unmethylated when the associated gene is active.
• Gene bodies and regulatory regions away from start sites show more variable methylation, and the relationship to gene activity is less straightforward.
• Repeat sequences across the genome are usually heavily methylated, which is thought to help keep them quiet.

The important point for clocks is that methylation patterns are not static. They shift in characteristic, partly predictable ways as we age. Some sites gain methylation over the years; others lose it. These shifts are measurable on commercial microarrays that read methylation at hundreds of thousands of CpG sites at once, which is what made clock-building possible.

How a clock is built

An epigenetic clock is, at heart, an exercise in supervised machine learning rather than a biological theory. The general recipe is consistent across the major clocks:

1. Collect a training dataset of DNA methylation profiles from many people whose chronological age, and sometimes health outcomes, are known.
2. Select the informative sites. Of the hundreds of thousands of CpG sites measured, only a subset carries a useful age signal. A penalised regression method (commonly elastic net) keeps the most predictive sites and discards the rest.
3. Assign weights. Each retained CpG site gets a coefficient. The weighted sum of methylation values, plus a calibration step, produces the age estimate.
4. Test on independent data the model has not seen, to check it generalises.

What a clock predicts depends entirely on what it was trained against. Train it on chronological age and it estimates chronological age. Train it on a composite of clinical markers, or on time to death, and it estimates something closer to biological condition. This training target is the single most useful thing to understand about any clock, and it is the basis for the "generations" described below. The same logic applies to reading any predictive model in health research, which is why it helps to understand how to read a clinical trial and how to evaluate a meta-analysis before drawing conclusions from clock studies.

The generations of clocks

It is helpful to group the main clocks by what they were trained to predict, because that determines what their output can reasonably be taken to mean.

First generation: predicting chronological age

The Horvath and Hannum clocks, both published in 2013, were the breakthrough demonstrations. Horvath's multi-tissue predictor was built from roughly 8,000 samples spanning 51 tissues and cell types and estimated chronological age with a median error of about 3.6 years, while remaining applicable across most tissues rather than blood alone [Horvath 2013]. Hannum's clock was trained on whole blood from several hundred adults and achieved comparable accuracy in that single tissue [Hannum 2013].

A subtle but important point: because these clocks were optimised to predict chronological age, the better they get at that task, the less room remains for them to reflect anything else. Their value for biology comes not from the estimate itself but from the residual, the gap between predicted and actual age, discussed below.

Second generation: predicting health and mortality

The second-generation clocks changed the training target. PhenoAge was trained against a "phenotypic age" derived from clinical chemistry and blood-count markers, then translated into methylation space; it predicted all-cause mortality, several cancers and physical functioning more strongly than the first-generation clocks [Levine 2018]. GrimAge went further, building methylation-based surrogates for several plasma proteins and for smoking pack-years, then combining them against time to death. In its development cohorts GrimAge predicted lifespan and time to coronary heart disease more accurately than chronological age or earlier clocks [Lu 2019].

Pace of ageing: measuring rate rather than level

DunedinPACE takes a different angle again. Rather than estimating how old someone is, it estimates how fast they are ageing. It was developed in the Dunedin birth cohort by tracking within-person change across 19 indicators of organ-system integrity over roughly two decades, then finding a methylation signature of that rate of decline [Belsky 2022]. A reading of 1.0 corresponds to one year of biological ageing per calendar year; higher values indicate faster ageing. This makes it conceptually distinct from age-estimating clocks and potentially more sensitive to recent or ongoing change.

What "epigenetic age acceleration" means

The clinically interesting quantity is rarely the raw age estimate. It is the difference between a clock's prediction and the person's actual chronological age, usually called epigenetic age acceleration. A positive value means the clock reads older than the calendar; a negative value means it reads younger.

This residual is where the association with health outcomes lives. In a meta-analysis of four longitudinal cohorts of older adults, a five-year higher methylation age (relative to chronological age) was associated with roughly a 21 percent higher risk of death from any cause, and the association persisted after adjusting for factors such as smoking, education, blood pressure, diabetes and known genetic risk [Marioni 2015]. Several points follow:

• Age acceleration is a population-level association, not a personal prognosis. It describes average risk differences across many people, not what will happen to any single individual.
• Different clocks capture partly different signals. Two clocks can disagree about whether the same person is "older" or "younger" than their years, because they were trained on different targets [Bell 2019].
• Acceleration is a correlate, not a cause. The clocks were built by selecting whatever CpG sites happened to predict age or outcomes; they were not built to identify the biological machinery driving ageing, and the mechanisms behind most clock sites remain poorly understood [Bell 2019].

This distinction between prediction and mechanism is central, and easy to lose sight of when a single number is presented as "your biological age."

Reliability and reproducibility

A measurement is only as useful as it is repeatable, and here the clocks have a documented weakness. When the same blood sample is processed twice, the methylation readings differ slightly because of technical variation in the laboratory assay. For several prominent clocks, that technical noise alone has been shown to produce differences of up to several years, in some cases up to around nine years, between technical replicates [Higgins-Chen 2022]. A swing of that size can easily exceed the change an intervention might plausibly produce, which matters a great deal when clocks are used to track individuals over time or to read out the effect of a treatment.

Researchers have proposed fixes. One influential approach recalculates the clocks from principal components of the methylation data rather than from individual CpG sites; the principal-component versions brought most replicate pairs into agreement within about 1.5 years and improved the detection of genuine longitudinal change [Higgins-Chen 2022]. Other recommendations from the field include standardising laboratory procedures, reporting which clock and which version was used, and being cautious about comparing readings across platforms [Bell 2019].

Practical implications for anyone reading a single consumer test result:

• A one-off reading carries meaningful measurement uncertainty; a difference of a few years from your chronological age may reflect noise rather than biology.
• Comparing results from different providers, platforms or clock versions is often not meaningful.
• Tracking change over time is only informative if the same assay and clock are used and the change exceeds the known noise floor.

Can interventions change a clock reading?

This is the question that drives much of the public interest, and it is the one where the evidence is thinnest. It is important to separate two distinct claims: that an intervention can move a clock reading, and that moving the reading reflects a real improvement in health or ageing. The second does not automatically follow from the first.

Small early studies have reported reductions in epigenetic age. The TRIIM trial, which combined recombinant growth hormone with two other drugs in nine men over a year, reported a mean epigenetic age roughly 1.5 years lower than baseline alongside immune changes [Fahy 2019]. But the study was small, uncontrolled and not designed to establish that the change in clock reading corresponded to a durable health benefit; its authors framed it as preliminary. Given the measurement-noise issues described above, results of this size warrant particular caution [Higgins-Chen 2022].

The honest summary is that whether clock readings are reliably modifiable, and whether modifying them improves health outcomes, remains unresolved. The field's own consensus review stresses that clocks are not yet validated surrogate endpoints for ageing interventions and should not be treated as such [Bell 2019]. For now the clocks are best understood as research instruments and population-level risk markers, not as personal dashboards that prove a regimen is working. The broader trade-offs between living longer and living well are covered in our piece on healthspan versus lifespan, and the wider longevity research hub puts these tools in context.

What to take away

Epigenetic clocks are a real scientific achievement. They show that ageing leaves a legible chemical signature on the genome, and that the gap between predicted and actual age tracks differences in disease and mortality risk across populations [Horvath 2013, Marioni 2015]. The second-generation and pace-of-ageing measures were a thoughtful step forward, trading raw accuracy at predicting birthdays for stronger links to health [Levine 2018, Lu 2019, Belsky 2022].

At the same time, the clocks are statistical predictors rather than mechanistic readouts; they carry measurement noise that can rival the signal of interest; they disagree with each other; and the claim that they can be reliably and meaningfully reversed is not yet established [Bell 2019, Higgins-Chen 2022]. Read in that light, an epigenetic clock is a useful research tool and an interesting personal data point, not a verdict on your health or a measure of how well any particular intervention is working.

Related Proco pages

• How to test your biological age
• The hallmarks of ageing
• Healthspan vs lifespan
• How to read a clinical trial

Sources

If you are considering a biological age test or acting on one, discuss the result with a qualified healthcare professional who can place it in the context of your full health picture.