How epigenetic clocks work: DNA methylation and age estimation

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

How do epigenetic clocks work? In short, an epigenetic clock is a statistical model. It reads chemical marks at a few hundred selected spots in the genome, multiplies each reading by a weight, adds the results, and outputs a number in years or in "years per year." Everything interesting about a clock comes from three choices made when it is built: which samples it learns from, which sites it is allowed to use, and what outcome it is trained to predict. This explainer covers that machinery step by step. For a broader overview of the main clocks and how to read a consumer result, see our biological age test guide.

Step 1: The raw signal is DNA methylation at CpG sites

DNA methylation is the addition of a methyl group to a cytosine base. In human DNA this happens mostly where a cytosine is followed by a guanine, a so-called CpG site. Reviews of the field describe methylation at many CpG sites as shifting in a consistent direction with age. At some sites it tends to rise and at others it tends to fall [2][3].

At any single CpG site, a sample contains millions of cells. Each copy of that site is either methylated or not. Laboratory arrays therefore report a fraction for each site. This is often called a beta value, running from 0 (no copies methylated) to 1 (all copies methylated) [8][9]. A typical array reads hundreds of thousands of CpG sites at once. That gives a clock-builder a very wide table: a small number of samples (rows) and a huge number of candidate predictors (columns).

Step 2: Penalized regression picks a small set of sites

With far more CpG sites than samples, ordinary regression would simply memorize the training data. Clock developers use penalized regression instead. The most common choice is elastic net, which adds a penalty that pushes most site weights to exactly zero. The result is a compact formula: a short list of CpG sites, each with a positive or negative weight, plus an intercept.

The best-known example is Horvath's 2013 multi-tissue clock. It was trained on about 8,000 samples from 82 Illumina methylation array datasets covering 51 healthy tissues and cell types. The final model used 353 CpG sites [1]. Elastic net selected those 353 from a much larger pool. The other sites were not "unimportant." They were simply not needed once the selected ones were in the model.

Two practical consequences follow from this design:

• Sites are interchangeable to a degree. Many CpG sites carry overlapping age information. A different training set could select a different list with similar accuracy. So the specific sites in a clock are not a map of "the aging genes" [2][7].

• The formula is only valid on comparable data. A clock is tuned to a particular array platform, data processing method and sample type. Changing any of these can shift the output, a challenge highlighted in methodological reviews [7].

Step 3: The training target defines what the clock means

This is the design decision that most readers miss. A clock predicts whatever it was trained to predict. Reviews group clocks partly by this choice [3][8][9].

A clock trained on chronological age, like the 2013 multi-tissue model, is in effect an age estimator. Its "error" (the gap between predicted and actual age) is what people call age acceleration. Clocks trained on health-linked targets were designed so that the gap carries more information about physiology and outcomes. The GrimAge authors reported that their measure strongly predicted lifespan and healthspan measures in the cohorts they studied [5]. The PhenoAge paper reported associations with mortality and age-related conditions [4]. DunedinPACE goes a step further. It does not output an age at all. It estimates a rate, trained on how fast a panel of biomarkers changed within individuals in the Dunedin Study [6].

This is why a single sample can produce several numbers that seem to disagree. Each clock answers a different question.

Step 4: Tissue, population and sample type shape the estimate

A clock is applied to DNA from a specific tissue. Blood, saliva and cheek swabs contain different mixes of cell types, and methylation differs by cell type. Methodological reviews flag cell composition as a major source of variation in clock outputs [7][3].

Some clocks are built specifically for one sample type. CheekAge, for example, was trained on buccal (cheek) samples. Its developers reported that it was also predictive of mortality when applied to human blood data [11]. The source population matters too. A 2024 study developed methylation clocks using Chinese cohorts. This reflects a wider concern that models trained mostly on one population may not transfer cleanly to others [10].

Age estimation from methylation is not limited to aging research. Forensic scientists use methylation-based age prediction to estimate the age of an unknown person from a biological trace. That field favours small, targeted marker sets and different laboratory methods [12].

Why the clocks tick: what is and isn't known

The statistical method explains how a clock produces a number. It does not explain why methylation tracks age. Horvath and Raj proposed an "epigenetic clock theory of ageing." It links clock behaviour to processes such as development and the maintenance of cellular systems over time [2]. Field and colleagues reviewed proposed causes and consequences of methylation change with age, and the evidence linking clock readings to biology [3].

One cell-level observation from the original multi-tissue paper often comes up in discussion. Embryonic stem cells and induced pluripotent stem cells received estimated ages near zero [1]. That is a finding in cultured cells about how the model scores those cell types. It is not evidence about any intervention in people.

Known limitations of the method

Reviews of the field list several recurring issues [7][8][9]:

• Technical noise. Small measurement differences at individual CpG sites can add up across hundreds of weighted sites. That can move an estimate between runs on the same sample.

• Platform dependence. Formulas built on one array version may need adjustment on another.

• Cell-type mixture. Shifts in the proportions of cell types in a sample can change a clock reading without any change inside individual cells.

• Population and age range. Accuracy can drop outside the age range and populations represented in training.

• Interpretation. A change in a clock reading is a change in a statistical estimate. What it implies for any individual's future is an open research question.

These limits matter for anyone comparing a commercial result against the published version of a clock. Lab methods and regulatory oversight for these tests are covered in are biological age tests FDA approved?.

Further reading

Epigenetic clock science moves quickly. DunedinPACE dates from 2022, and population-specific and cheek-swab clocks appeared in 2024 [6][10][11]. Readers who want depth beyond a single explainer often turn to books on aging biology and epigenetics. A few neutral criteria are useful when comparing them: the publication date, whether the author cites primary studies, how clearly the book separates measurement from speculation, and whether it explains training targets rather than presenting all "biological age" numbers as the same thing.

A good companion title explains DNA methylation from first principles and handles statistical models with appropriate caution. It should also acknowledge open questions about mechanism. Format matters too. Textbooks give methods detail, while general-audience books trade detail for narrative. How we handle commercial links is described on our about page.

Frequently asked questions

How do epigenetic clocks work in one sentence?

They measure the methylation fraction at a selected set of CpG sites, multiply each by a weight learned through penalized regression, and sum the results into an age or pace estimate [1][8]. The weights come from training data in which the target, such as chronological age, mortality or a biomarker-based pace, was already known [4][5][6].

Why do different clocks give different results on the same sample?

Each clock was trained on a different target, uses a different set of CpG sites, and was built on different reference samples [3][7]. A chronological-age clock and a mortality-trained clock are answering different questions, so disagreement is expected rather than a sign that one is broken.

Can a clock be used on saliva or cheek samples instead of blood?

Some clocks were trained on multiple tissues, and some were trained specifically on cheek swabs [1][11]. Because cell composition differs between sample types, a clock can produce shifted estimates when applied to a tissue unlike its training data [7].

Last updated: 19 October 2026.

Related on this site

• BioAge Peptide home

• Are biological age tests FDA approved? How aging tests are regulated

• Biological age tests: how epigenetic clocks work and what they mean

• GrimAge explained: how the mortality-trained epigenetic clock works

• Biological age test accuracy: reliability, noise and repeat results

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.