First, second and third generation epigenetic clocks, explained

Oct 22, 2026

Prepared by: Research Desk

Prepared by: Research Desk. Sources are linked for verification.

Data point first: The biggest difference between epigenetic clocks is not the lab, the sample type or the number of CpG sites. It is the training target, meaning what the statistical model was taught to predict. That single design choice is why researchers group clocks into "generations." It is also why second generation epigenetic clocks such as PhenoAge and GrimAge can give a different number from a first-generation clock run on the same DNA sample.

This explainer walks through each generation, what its output means, and how the generations are used in published research. For the basics of methylation and how a consumer kit works, start with our biological age test guide.

Why clocks are sorted into generations

Every epigenetic clock is a regression model. It takes methylation levels at a set of CpG sites and combines them into one number. The model learns which sites to use, and how much weight to give each one, by fitting against a target variable in a training dataset.

Change the target and you change what the clock measures:

First generation: the target is chronological age.
Second generation: the target is a measure of health status or mortality risk.
Third generation: the target is the rate of biological change over time.

The generation labels are a convention used in reviews and commentary, not formal categories defined by any standards body. Broader reviews of biological-age measurement describe the same underlying point: what an omics-based aging measure captures depends on how it was built and validated [2].

First generation: clocks trained on chronological age

The best-known first-generation clock is Horvath's multi-tissue clock, published in 2013 [1]. Horvath combined publicly available methylation data from thousands of human samples across dozens of tissue and cell types. He then selected 353 CpG sites whose combined methylation predicted chronological age across tissues [1]. Its defining feature was that one model worked across many tissues, not just blood.

What the output means: a first-generation clock estimates age in years. If the estimate is higher than a person's calendar age, researchers call the gap "epigenetic age acceleration." It is usually calculated as the residual after regressing clock age on chronological age.

The built-in limitation: because the model was optimised to match calendar age, it was not directly trained to separate people who are aging well from people who are not. Any signal about health in the residual is a by-product, not the training goal. This is the gap that second-generation clocks were designed to address.

Second generation epigenetic clocks: trained on health and mortality

Second generation epigenetic clocks swap chronological age for a target that carries information about health and survival. The two most cited examples use different strategies.

PhenoAge: a two-step, biomarker-anchored target

Levine and colleagues built PhenoAge in two stages [3]. First, they created a "phenotypic age" score from chronological age plus a panel of routine clinical blood chemistry markers, weighted by how well they predicted mortality in a large population dataset. Second, they trained a methylation model to predict that phenotypic age score. The result was a clock based on 513 CpG sites [3]. The paper's title frames it as "an epigenetic biomarker of aging for lifespan and healthspan" [3].

GrimAge: methylation surrogates and time-to-death

Lu and colleagues took a different route for GrimAge [4]. They first built methylation-based surrogate measures for a set of plasma proteins and for smoking history (pack-years). Then they combined these surrogates with age and sex in a model trained on time to death. The authors reported that GrimAge "strongly predicts lifespan and healthspan," outperforming earlier clocks on the mortality and morbidity outcomes they tested [4].

What the output means: both clocks still report a value in years, so the number looks like a first-generation result. But the years are scaled to a health or mortality signal, not to calendar age. A GrimAge "age" and a Horvath "age" for the same person are answering different questions.

Third generation: measuring the pace of aging

DunedinPACE, published in 2022, is the clock most often described as third generation [5]. Rather than predicting a level (an age), it was trained to predict a rate. The training target was "Pace of Aging," which was derived from repeated measurements of many physiological biomarkers. These measurements were tracked in members of a single birth cohort across two decades of adulthood [5].

What the output means: DunedinPACE reports years of biological change per calendar year. A value of 1.0 corresponds to the cohort average. Values above 1.0 indicate faster-than-average change, and values below 1.0 indicate slower change. The authors also reported strong test–retest reliability. That property matters when the goal is to compare measurements taken at different times [5].

The trade-off: a pace measure describes a recent trajectory, not a cumulative state. It is not converted to an "age," which makes it harder to market as a headline number, but arguably easier to interpret correctly.

The generations side by side

	First generation	Second generation	Third generation
Example	Horvath multi-tissue [1]	PhenoAge [3], GrimAge [4]	DunedinPACE [5]
Training target	Chronological age	Phenotypic age score; time to death	Longitudinal pace of physiological change
Output unit	Years	Years (health/mortality-scaled)	Biological years per calendar year
Key strength	Works across many tissues	Stronger links to health outcomes in the original papers	Designed for change over time; reliability reported
Main caveat	Health signal is indirect	Depends on the outcomes and cohort used in training	Trained in a single cohort; no "age" output

How the generations show up in published research

Because each generation measures something different, studies often report several clocks side by side. Some examples from the research literature:

Systematic review of risk factors. Oblak and colleagues reviewed biological, social and environmental factors associated with epigenetic clock acceleration across published studies [6]. Reviews like this pool associations measured with different clocks, which is one reason findings can look inconsistent.
Population cohorts. Analyses from the U.S. Health and Retirement Study have examined how physical activity relates to epigenetic aging measures [7] and how a range of aging measures relate to cancer [8]. These are observational human studies: they report associations, not cause and effect.
Disease-specific studies. Researchers have studied epigenetic aging in relation to Parkinson's disease risk [9] and in rheumatoid arthritis [10].
Randomized trials. A randomized trial in people with HIV-associated lipohypertrophy reported, in the authors' words, that semaglutide "slows epigenetic aging" [11]. It is a single trial in a specific population. A change in a clock score in such a study is a change in a biomarker, and it does not on its own establish an effect on health outcomes.

The practical lesson for reading any such study is to check which generation of clock was used. A shift in a chronological-age clock, a mortality-trained clock and a pace measure are three different findings, even when headlines describe them all as "biological age."

Limits shared by every generation

Training cohort dependence. Each clock reflects the population, sample type and measurement platform it was built on. Performance in other groups has to be shown, not assumed [2].
Biomarker, not diagnosis. Clock outputs are research measures. Consumer epigenetic-clock tests are generally run by individual laboratories rather than sold as FDA-cleared kits; the FDA's page explains how laboratory-developed tests are overseen [12]. Our page on whether biological age tests are FDA approved covers what that means.
Association is not mechanism. Even the best-validated second- and third-generation clocks are statistical prediction models. On their own, they do not explain why methylation patterns differ between people. For a broader overview of how omics-based aging measures are built and interpreted, see the review by Rutledge and colleagues [2].

Frequently asked questions

What makes a clock second generation?

A second-generation clock is trained on a health-related target, such as a clinical-biomarker score or time to death, rather than on chronological age. PhenoAge [3] and GrimAge [4] are the standard examples. Their outputs are still expressed in years, but those years are scaled to health or mortality signals.

Is a third-generation clock more accurate than a first-generation one?

"Accurate" depends on the question being asked. Horvath's clock is built to estimate chronological age across tissues [1], while DunedinPACE is built to estimate a rate of change [5]. Neither is a better version of the other; they measure different things.

Can an epigenetic clock result go down between tests?

Clock values can differ between measurements because of real biological change, measurement noise or differences in lab processing. That is why the test–retest reliability reported for DunedinPACE [5] is relevant when comparing results over time. A single lower reading is not, on its own, evidence of slower aging.

Last updated: 19 October 2026.

Related on this site

Not medical advice. This page reports published research and regulatory records for information only.

Research-only notice: Content on this site is provided for research and educational purposes only and is not intended for human use. It is not medical advice.

Translation notice: Some content on this site may have been machine-translated using Google Translate and may contain errors.

Sources

‹ How epigenetic clocks work: DNA methylation and age estimation