The Horvath clock: the 2013 multi-tissue epigenetic age estimator

Oct 23, 2026

Prepared by: Research Desk

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

Data point first: In 2013, Steve Horvath published a model that estimates age from DNA methylation at 353 CpG sites. It was built from roughly 8,000 samples drawn from 82 public array datasets covering 51 healthy tissues and cell types [1]. In test data, estimated age correlated with chronological age at about 0.96, with a median error of about 3.6 years [1]. That combination of accuracy and tissue range is why the "Horvath clock" became the reference point for almost every aging clock that followed.

This explainer covers the Horvath clock specifically: how it was built, what its output means, what the original paper reported, and where it falls short. For a broader overview of all the major clocks and how consumer tests use them, see our biological age test guide.

What the Horvath clock actually estimates

The Horvath clock outputs DNA methylation age (DNAm age). It is an estimate of chronological age based on chemical tags on DNA. Methylation at certain CpG sites (a cytosine followed by a guanine) shifts in a fairly predictable way as people get older. The clock combines readings from 353 of these sites into a single number expressed in years [1].

Of the 353 sites, 193 tend to gain methylation with age and 160 tend to lose it [1]. No single site is a reliable age marker. The weighted combination is what carries the signal.

The key design choice is the training target. The Horvath clock was trained to predict chronological age, not mortality, disease or any health outcome. That makes it a "first-generation" clock. Its main job is to answer one question: how old does this tissue look, methylation-wise? Any link to health outcomes has to be tested separately.

How it was built

The method is straightforward to describe, even if the scale was unusual for its time:

Pool public data. Horvath gathered DNA methylation array datasets from many laboratories and tissue types. The collection included blood, brain, breast, kidney, liver, lung, saliva and others, amounting to about 8,000 samples [1].
Split into training and test sets. The model was fitted on one portion of the data and evaluated on held-out samples.
Fit a penalized regression. A penalized regression (elastic net) selected a sparse set of CpGs, the 353, and assigned each a weight [1].
Evaluate across tissues. Accuracy was checked separately in many tissue types. This is where the clock stood out.

The 2013 paper positioned its multi-tissue design against earlier single-tissue predictors [1]. A model that worked across most of the body was the paper's main contribution.

What the 2013 paper reported

Beyond its headline accuracy, the original study reported several observations that shaped the field. All of them came from analyses of human (and some chimpanzee) samples and cell cultures [1]:

Stem cells read near zero. Embryonic stem cells and induced pluripotent stem cells had DNAm age estimates close to zero. Reprogramming adult cells into induced pluripotent stem cells was associated with this near-zero reading [1].
Cell culture matters. DNAm age correlated with cell passage number in culture [1].
Age acceleration is heritable. The gap between DNAm age and chronological age, called age acceleration, behaved as a highly heritable trait in the twin and family data analysed [1].
Cancer tissue looks different. Many cancer samples showed DNAm age that diverged substantially from the donor's chronological age [1].
Cross-species signal. The clock produced meaningful estimates in chimpanzee tissues [1].

None of these findings shows that changing DNAm age changes health or lifespan. They are observations about what the estimator tracks.

Age acceleration: the number most studies use

In research, the raw DNAm age is often less important than the residual. Researchers regress DNAm age on chronological age and look at whether a sample sits above or below the expected line. A positive residual is reported as accelerated epigenetic aging; a negative one as decelerated.

This metric lets researchers compare groups without the result being driven by age differences between them. Later work has examined epigenetic aging in specific disease contexts, for example a 2024 paper on epigenetic aging and rheumatoid arthritis [9]. Researchers have also studied which factors appear to influence the speed and direction of epigenetic aging, both in living organisms and in cultured cells [5].

Where the Horvath clock falls short

The clock's strengths and weaknesses come from the same design choice.

It was trained on chronological age. A model that predicts calendar age very well has, almost by definition, limited room left to capture differences in health. Later clocks were built specifically to address this:

PhenoAge was trained on a composite of clinical blood biomarkers. It was reported to outperform earlier measures in predicting outcomes including all-cause mortality and healthspan [2].
GrimAge combined DNA methylation surrogates of plasma proteins and smoking history. It was reported as a strong predictor of lifespan and healthspan [3].
DunedinPACE was designed to estimate the pace of aging from a single blood sample, rather than an accumulated age [4].

Measurement context. Methylation array results can depend on platform and data processing. An estimate is only as comparable as the methods behind it, which is worth keeping in mind when lining up numbers from different sources.

Tissue context still matters. "Multi-tissue" means the clock works reasonably across tissues, not that every tissue gives an identical reading for the same person.

A number is not a diagnosis. DNAm age is a research estimate. On its own, it does not identify a disease or predict a specific person's future.

Beyond humans

The multi-tissue approach turned out to be portable. Horvath and colleagues later applied DNA methylation aging methods to other species, including a combined methylation and transcriptomic study in horses [6]. Cross-species work like this helps researchers test whether methylation aging reflects conserved biology rather than a quirk of human data.

The Horvath clock in consumer testing

Some consumer biological-age services report a Horvath-type estimate alongside newer clocks. Tests designed, manufactured and used within a single laboratory are described by FDA as laboratory-developed tests [7]; some consumer epigenetic-age services describe their assays this way. The regulatory framework for these tests shifted recently: in September 2025, FDA published a final rule implementing a court's vacatur of its 2024 rule on laboratory-developed tests [10]. Lab quality standards in the US fall under the Clinical Laboratory Improvement Amendments, administered by CMS [8]. CLIA certification addresses how a lab operates; it does not validate what an aging clock means. Our page on whether biological age tests are FDA approved covers the regulatory picture in detail.

For anyone reading a report, three practical checks apply:

Which clock produced the number? A Horvath estimate and a DunedinPACE score answer different questions.
Which tissue was sampled? Saliva and blood readings are not directly interchangeable.
Is age acceleration reported, and how? A raw DNAm age without context says little.

This site focuses on how aging is measured rather than on any intervention. Our approach is explained on the about page.

Choosing a book on epigenetic clocks

The Horvath clock is a good test of whether a book on aging science is worth the money. A strong book explains clearly what this 2013 model does and does not show. A weak one turns it into a promise. Useful comparison criteria:

Publication date. Books written before 2022 cannot cover all three of PhenoAge (2018), GrimAge (2019) and DunedinPACE (2022). Check the edition date against the clocks you want to understand.
Primary citations. The best titles cite the original papers, such as the 2013 Genome Biology study, rather than press coverage.
Clear distinction between clock generations. Look for an explicit explanation that first-generation clocks predict calendar age, while later ones were trained on health-related targets.
Statistical honesty. Good books explain error margins, the meaning of age acceleration, and why results can differ between clocks and tests.
Restraint about interventions. Be wary of titles that present a single small study as proof that biological age can be wound back.
Format and depth. Textbooks suit readers who want methods detail. Popular-science titles suit readers who want the narrative and context first.

Reading the original 2013 paper alongside any book remains the most reliable way to check how faithfully the book describes the clock [1].

Frequently asked questions

How many CpG sites does the Horvath clock use?

The Horvath clock uses 353 CpG sites. Methylation at 193 of them tends to rise with age, and at 160 it tends to fall [1]. The sites were selected by penalized regression from far larger array data.

Is the Horvath clock the same as a consumer biological age test?

No. The Horvath clock is a published statistical model; a consumer test is a lab service that may run this model, newer clocks, or proprietary ones. FDA describes tests designed, manufactured and used within a single laboratory as laboratory-developed tests [7], and some consumer epigenetic-age services describe their assays this way.

Why might the Horvath clock and GrimAge give different numbers?

They were trained on different targets. The Horvath clock was built to predict chronological age [1], while GrimAge was built around methylation surrogates linked to lifespan and healthspan [3]. Different targets can produce different estimates from the same sample, and neither number is "wrong".

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

‹ First, second and third generation epigenetic clocks, explained