Home > Biological age tests: how epigenetic clocks work and what your number means
Biological age tests: how epigenetic clocks work and what your number means
Prepared by: Research Desk. Sources are linked for verification.
You mail off a saliva sample or a finger-prick of blood, and a few weeks later an app tells you you’re “biologically 38” when your passport says 45. Is that number science? Partly. The clocks behind these tests are some of the most exciting tools in aging research. They’re also widely misunderstood. This guide explains how they’re built, what each one actually predicts, and how much you can read into your own result.
Contents
The core idea: methylation as a biological record
Your DNA sequence stays essentially the same throughout life. But chemical tags sit on top of it and help control which genes are switched on. The most studied of these tags is DNA methylation: a methyl group attached to cytosine, usually at sites called CpGs, where a cytosine is followed by a guanine.
Methylation at some CpG sites changes predictably with age. Measure enough of these sites, use statistics to find the best combination, and you have an “epigenetic clock”.
How a clock is built (in four steps)
Collect samples. Thousands of DNA samples from people of known age, and often with known health outcomes.
Measure methylation. Usually with arrays that read hundreds of thousands of CpG sites at once.
Choose a training target. This is the crucial design decision. Is the clock trained to predict chronological age, or a measure of physiological decline, or time to death?
Fit a model. Usually penalised regression, which picks out the most predictive CpGs and assigns each a weight.
The training target in step 3 defines what the clock measures. That’s why different clocks disagree.
Generation 1: predicting chronological age
Horvath’s multi-tissue clock (2013)
Training target: chronological age.
Data: Developed from about 8,000 samples across 82 datasets, covering 51 healthy tissues and cell types. The final clock uses 353 CpG sites [1].
Key findings [1]:
DNA methylation age was close to zero for embryonic and induced pluripotent stem cells.
It correlated with how many times cells had divided in culture.
“Age acceleration” (methylation age minus actual age) was highly heritable.
Across 20 cancer types, tumour tissue showed significant age acceleration, averaging 36 years.
Horvath proposed that DNA methylation age reflects “the cumulative effect of an epigenetic maintenance system” [1].
What it tells you: how old your tissue looks in methylation terms. It’s an impressive age estimator. But because it was trained to match calendar age, it’s less sensitive to differences in health between people of the same age.
Generation 2: predicting health and death
PhenoAge (2018)
Training target: a “phenotypic age” built from clinical blood markers and chronological age. That phenotypic age was itself designed to capture differences in mortality risk [2].
Key finding: PhenoAge strongly outperformed earlier measures in predicting all-cause mortality, cancers, healthspan, physical functioning and Alzheimer’s disease [2]. Higher PhenoAge relative to actual age was associated with more activity in inflammatory and interferon pathways [2].
GrimAge (2019)
Training target: time to death. GrimAge was built in two stages. First, researchers developed DNA-methylation “surrogates” for seven plasma proteins (including PAI-1 and GDF-15) and for smoking pack-years. Then they combined those surrogates into a lifespan predictor [3].
Key findings [3]:
In large validation data, GrimAge stood out among the clocks then available for predicting time to death, time to coronary heart disease and time to cancer.
Age-adjusted GrimAge (“AgeAccelGrim”) was strongly associated with the number of age-related conditions a person had.
It was related to CT measures of fatty liver and visceral fat, and to age at menopause.
Disclosure: The University of California holds patent applications on PhenoAge and GrimAge, with the developers named as inventors [2][3]. This is common in the field.
Generation 3: measuring the pace of aging
DunedinPACE (2022)
Training target: the rate of decline. Researchers followed members of the Dunedin Study birth cohort (born 1972–73) and tracked 19 indicators of organ-system integrity at four time points over two decades. They then distilled that two-decade “Pace of Aging” into a single-time-point blood test [4].
Design choice that matters: CpG probes with low test-retest reliability were excluded during development [4].
Key findings [4]:
High test-retest reliability.
Associated with morbidity, disability and mortality.
Showed faster aging in young adults who had experienced childhood adversity.
Effect sizes were similar to GrimAge, and DunedinPACE added predictive information beyond GrimAge for some outcomes.
What it tells you: a speedometer rather than an odometer. A DunedinPACE value of 1.0 corresponds to one year of biological aging per calendar year. Values above 1 suggest faster aging and values below 1 slower.
Side-by-side
Clock
Year
Trained to predict
Output
Best for
Horvath
2013
Chronological age
Years
Research across tissues [1]
PhenoAge
2018
Clinical phenotypic age
Years
Morbidity and mortality risk [2]
GrimAge
2019
Time to death
Years
Mortality and disease prediction [3]
DunedinPACE
2022
Rate of physiological decline
Rate (years per year)
Tracking change over time [4]
Why your results can differ between tests
If you take two biological-age tests and get two different numbers, that’s expected. The main reasons:
Different clocks. They’re trained on different targets, so they measure different things.
Different tissues. Many clocks were developed on blood. Saliva contains a different mix of cells, which affects the readings.
Different lab platforms. Array versions and processing pipelines vary.
Technical noise. Individual CpG measurements have error. Clocks built without filtering out unreliable probes can shift noticeably between repeat runs of the same sample. That’s one reason DunedinPACE’s developers excluded low-reliability probes [4].
Proprietary tweaks. Some companies use modified or unpublished algorithms.
What a single result can and can’t tell you
It can:
give you a rough, research-grade signal about how your methylation profile compares with population patterns;
provide a baseline if you plan to repeat the same test the same way.
It can’t:
diagnose a disease;
give a precise personal lifespan forecast. The predictions are statistical associations across groups [2][3][4];
prove that any intervention changed biological age. Short-term changes may reflect noise, shifts in blood cell mix, or real change, and a single test can’t tell them apart.
Using clocks to judge interventions: proceed carefully
Clocks are increasingly used as outcomes in trials. They’re useful as early signals. But no clock has yet been shown to be a validated surrogate endpoint, meaning one where improving the score reliably improves real outcomes such as disease or survival. Until then, a better clock score is a hypothesis, not a result.
Practical checklist before testing
Pick a test that names its clock or clocks and cites the published papers.
Prefer clocks validated on the same sample type the test uses.
If you’ll track change, use the same company, clock and sample type, ideally at the same time of day.
Read the privacy policy carefully.
The regulatory side, including why most of these tests aren’t FDA-reviewed, is covered in our regulation explainer.
Myths we hear most often
“A younger biological age means I’ll live longer.” Not necessarily. Clocks like GrimAge and PhenoAge predict outcomes across groups of people [2][3]. Your personal future depends on many things no clock captures, including accidents, genetics outside the measured sites, and access to care.
“All clocks measure the same thing.” They don’t. Horvath’s clock was trained on calendar age [1], PhenoAge on a clinical measure of physiological age [2], GrimAge on time to death [3], and DunedinPACE on the rate of decline across organ systems [4]. Asking which one is “right” is like asking whether a thermometer or a barometer is the correct weather instrument.
“If my score improved after three months, the intervention worked.” Maybe, but a single before-and-after pair can’t separate real change from measurement noise, seasonal shifts or changes in the mix of cells in your sample. Researchers use control groups for exactly this reason.
“Saliva and blood tests are interchangeable.” Saliva contains a different mix of cells from blood. A clock developed on blood may behave differently on saliva.
What researchers still argue about
Mechanism. Does methylation drift cause aging, or is it a record of it? Horvath proposed it reflects an “epigenetic maintenance system” [1], but the causal question is open.
Surrogate status. Can a clock change stand in for real outcomes in trials? Not yet established.
Diversity of training data. Clocks trained mostly on some populations may be less accurate in others.
Which clock for which job. Odometer-type clocks and speedometer-type clocks answer different questions [4].
Last updated: 18 October 2026.
Not medical advice. This page reports published research and regulatory records for information only.