Epigenetic Clocks Explained: DunedinPACE vs. Horvath vs. GrimAge

If you have ordered a biological age test in the last few years, the number you received came from an epigenetic clock. What most reports do not explain is that "epigenetic clock" is not one test. It is a family of statistical models, trained on different outcomes, that answer genuinely different questions. Two clocks run on the same saliva sample can disagree by years — and both can be correct, because they are not measuring the same thing.

This article walks through what these clocks actually measure, how the three most cited ones differ, and how to use them without over-reading a single number.

What an epigenetic clock actually measures

Your DNA sequence stays essentially fixed for life. What changes is how it is marked and read. One of the best-studied chemical marks is DNA methylation: a methyl group attached to a cytosine base, usually at a site called a CpG. These marks influence whether a gene is switched on or off in a given tissue.

Methylation patterns change with age in a way that is remarkably predictable across people. At some sites methylation rises steadily across the lifespan; at others it falls. In 2011 to 2013, researchers realized that if you measure enough of these sites and feed them into a regression model, you can estimate a person's chronological age from a blood or saliva sample with striking accuracy.

That is the foundation. An epigenetic clock is a weighted score across a set of CpG sites. What distinguishes one clock from another is which sites it uses and — more importantly — what the model was trained to predict.

First generation: Horvath and the "age" clocks

The Horvath clock, published in 2013, is the classic first-generation model. It uses several hundred CpG sites and was trained to predict chronological age across many tissue types. Its multi-tissue design was the breakthrough: the same model works on blood, saliva, brain and other tissues.

What it gives you is an estimate of how old your tissue looks methylation-wise. Subtract your chronological age and you get "epigenetic age acceleration" — a positive number means your tissue looks older than the calendar says, a negative number means it looks younger.

The strength of this generation is stability and enormous validation literature. The limitation is conceptual: the model was trained to predict calendar age, so by design its best possible performance is to reproduce a number you already know. The interesting signal is entirely in the residual — the gap between predicted and actual — and that residual is a smaller, noisier quantity than the headline number suggests.

Other first-generation clocks (Hannum, and the tissue-specific variants that followed) share this basic architecture. If you want to understand where these tests sit within a broader assessment, our overview of biological age testing covers the non-epigenetic markers that usually accompany them.

Second generation: PhenoAge and GrimAge

Researchers then asked a better question. Instead of training a model to predict calendar age, train it to predict something that actually matters clinically.

PhenoAge was trained against a composite of clinical biomarkers associated with mortality risk. GrimAge went further: it was trained to predict time to death, and it does so indirectly by first building DNA-methylation-based surrogates for plasma proteins and for smoking pack-years, then combining those surrogates.

That design choice matters. GrimAge is not really telling you how old your cells look. It is estimating the burden of accumulated damage and exposure — the methylation fingerprint of smoking, of inflammatory protein levels, of metabolic stress. In published cohort studies it is consistently among the strongest epigenetic predictors of mortality and of age-related disease incidence.

The practical consequence: GrimAge-type results tend to respond to the big, well-established exposures. Quitting smoking, sustained changes in inflammatory load, and improvements in cardiometabolic health are the kinds of shifts that show up. It is less suited to detecting whether a three-month supplement protocol did anything.

Third generation: DunedinPACE measures speed, not age

DunedinPACE is architecturally different and this is the point most often missed. It was developed from the Dunedin birth cohort, in which the same people were measured repeatedly across decades on a panel of organ-system biomarkers. The researchers first calculated each person's rate of physiological decline over time, then trained a methylation model to estimate that rate from a single blood sample.

The output is therefore not an age. It is a pace. A DunedinPACE of 1.0 means you are accumulating roughly one year of biological aging per calendar year. A value of 0.85 suggests a slower current rate; 1.2 suggests a faster one.

This distinction has real consequences for how you use it:

1. Horvath-type clocks report an accumulated state — decades of history baked into one number. That history cannot be undone, which limits how responsive the number is to anything you do now.
2. DunedinPACE reports current velocity. In principle, a change in how you live should be reflected sooner, because the measure is about the present trajectory rather than the accumulated past.
3. GrimAge reports predicted risk, weighted heavily toward exposures with strong mortality signal.

Several randomized and observational studies have used DunedinPACE specifically because a pace measure is more plausible as a short-to-medium-term outcome than an accumulated-age measure. The available evidence suggests it is more sensitive to intervention than first-generation clocks, though "more sensitive" is not the same as "sensitive enough to guide individual decisions."

The limitations nobody puts on the report

Be clear-eyed about what these tests are not.

1. They are not diagnostic. No epigenetic clock diagnoses a disease, and none is a substitute for standard clinical workup. A result cannot tell you that you have, or will get, any specific condition.
2. Technical variability is real. Re-running the same sample, or running two samples taken days apart, can shift results. Some clocks are more reproducible than others, and reported test-retest variation is large enough that small differences between two timepoints should not be interpreted as change.
3. There is no industry standard. Different providers use different arrays, different normalization pipelines and sometimes proprietary modifications of published clocks. Results from two companies are not directly comparable, and neither are results from the same company after a methodology update.
4. Tissue matters. A saliva sample and a blood sample contain different cell populations, and cell composition itself influences methylation readouts. Some models correct for this; not all do.
5. Population representation is uneven. Many training cohorts skew toward particular ancestries and age ranges, and performance outside those groups is less well characterized.

How to use them sensibly in a longevity protocol

The single most useful shift is to stop treating an epigenetic clock as a verdict and start treating it as a trend line.

1. Measure repeatedly, not once. A single value tells you almost nothing actionable. Three or four measurements over two to three years, from the same provider using the same methodology, tell you something about direction.
2. Standardize the conditions. Same laboratory, same sample type, similar time of day, and ideally not during acute illness, immediately after an unusually hard training block, or in the middle of significant sleep deprivation.
3. Match the clock to the question. If you want a risk-weighted picture, a GrimAge-type output is more informative. If you want to track whether current habits are moving the needle, a pace measure is the more logical choice.
4. Interpret alongside conventional markers. Fasting insulin, ApoB, hsCRP, VO2 max, grip strength and sleep quality are cheaper, better validated, and more directly actionable. The epigenetic result is context, not a replacement. This is why we treat it as one input among many in our approach to longevity assessment.

What has been associated with better results

The interventions with the most consistent associations in the published literature are unglamorous and overlap almost entirely with general health advice.

Smoking cessation has the clearest relationship, particularly with GrimAge-type clocks, which is expected given that smoking exposure is explicitly encoded in the model. Regular physical activity, and cardiorespiratory fitness specifically, are associated with slower measured pace of aging across multiple cohorts. Sleep quality and duration show consistent associations, which our piece on sleep and longevity explores in more depth. Dietary patterns higher in whole plant foods and lower in ultra-processed intake are associated with more favorable readings, as is maintaining metabolic health markers within healthy ranges.

Caloric restriction has been studied in at least one randomized human trial using a pace measure, with a modest but statistically detectable slowing reported. That result is frequently over-quoted; the effect size was small, and it has not been established that the change translates into longer life.

Note the wording throughout: associated with. These are observational and, in a few cases, randomized associations at the group level. None establishes that moving your epigenetic number causes you to live longer. That causal link is the central open question in the field, and it is not yet answered.

The bottom line

Epigenetic clocks are a genuinely interesting research tool that arrived in the consumer market before the interpretive framework did. Used as a repeated measure, under standardized conditions, alongside conventional biomarkers, and with the specific clock matched to the specific question, they can add useful signal to a long-term health strategy. Used as a one-off number to be celebrated or panicked about, they add noise.

If you are considering testing, the most valuable conversation is not "which company has the best marketing" but "what decision would I make differently depending on the result" — and that conversation belongs with a clinician who can interpret it alongside the rest of your picture.

This article is educational and does not constitute medical advice. Testing and interpretation should be considered together with qualified medical guidance.