Epigenetic Clocks: How DNA Methylation Reveals Your Age

What DNA Methylation Is

DNA methylation is the addition of a methyl group to the fifth carbon of a cytosine base, usually where a cytosine sits next to a guanine. There are roughly twenty-eight million such sites in the human genome.

Methylation is a regulatory mark rather than a change to the genetic code. Heavy methylation in a gene promoter region generally associates with reduced transcription, while methylation in gene bodies has more complex effects. The pattern is what makes a liver cell a liver cell and a neuron a neuron, despite identical DNA, and it is maintained through cell division by dedicated enzymes.

The pattern is also not static. It responds to age, to environment, to nutrient availability, particularly one-carbon metabolism involving folate, B12 and methionine, to inflammatory signalling, and to exposures including tobacco smoke. That responsiveness is what makes it measurable as a biological signal and also what makes it noisy.

How a Clock Is Built

The construction is conceptually simple and worth understanding, because it explains both the power and the limits of the result.

Start with methylation measurements at hundreds of thousands of sites in thousands of individuals of known age. Apply a penalised regression, typically an elastic net, which selects a small subset of sites whose weighted combination best predicts the target while discarding the rest. What emerges is a formula: multiply the methylation fraction at each selected site by its coefficient, sum, apply a transformation, and output a value in years.

Two consequences follow directly. First, the selected sites are not necessarily causal or even biologically interesting; they are statistically useful. Second, the output inherits the properties of the training population and the target. A clock trained on chronological age in a European blood cohort is optimised for exactly that, and behaves less predictably outside it.

This is why the generation of a clock matters more than its brand name. A model trained to predict chronological age is by construction trying to discard the individual variation that carries health information, discarding it as error. A model trained to predict mortality is trying to capture it.

The Major Clocks and What They Target

Horvath multi-tissue clock (2013). The landmark first-generation clock, built on 353 sites and validated across many tissue types, which was the crucial advance: it demonstrated that the age signal is not blood-specific. Extremely accurate for chronological age. Weakest for outcome prediction, as expected from its training target.

Hannum clock (2013). A blood-specific first-generation clock developed in parallel, using a smaller site set.

PhenoAge (2018). Second generation. Trained not on chronological age but on a phenotypic age derived from nine clinical biomarkers that predict mortality. It correlates less tightly with chronological age and substantially more with outcomes.

GrimAge (2019). Second generation, and the strongest outcome predictor among the widely used clocks. Built by first creating methylation-based surrogates for plasma proteins and smoking pack-years, then combining those surrogates to predict time to death. Its association with mortality and morbidity is consistently the strongest reported, and its dependence on a smoking surrogate is worth knowing when interpreting an individual result.

DunedinPACE (2022). Third generation, and different in kind. Built from a birth cohort followed for two decades with repeated measurement of multiple organ system markers, it estimates the current rate of biological aging expressed as years of biological change per calendar year. A value of 1.0 means aging at the population-typical rate. Because it estimates rate rather than accumulation, it is the most appropriate clock for evaluating whether an intervention has changed a trajectory.

DNAmTL. A methylation-based estimator of telomere length, which predicts outcomes better than measured telomere length in several comparisons, an amusing and instructive result.

Causality-enriched clocks. A newer approach that attempts to select sites with causal rather than merely correlational relationships to aging outcomes, using genetic instruments. This is where the field is heading and is not yet mature.

Does Methylation Drift Cause Aging?

This is the central open question and it deserves an honest answer: nobody knows.

The case for causation. Epigenetic alteration is one of the recognised hallmarks of aging. Partial reprogramming with Yamanaka factors in mice resets methylation patterns and restores function in some tissues, including regeneration of optic nerve after injury, which suggests the epigenetic state is not merely a record but a determinant. Loss of epigenetic information has been proposed as a primary driver of aging, with the corollary that the information is recoverable.

The case for recording. Many clock sites sit in regions with no obvious functional relevance. Methylation drift may simply be the accumulated imprint of cell division, inflammatory exposure and metabolic history, in which case a clock is a very good odometer and changing the odometer changes nothing.

Why it matters practically. If methylation state is causal, then interventions that reset it are interventions on aging itself. If it is a record, then an intervention that lowers a clock reading without changing underlying physiology has achieved nothing except a better number. Since it is currently possible to move clock readings with short-term interventions whose long-term effects are unknown, this distinction is not academic.

Practical Limitations of Testing

Technical variability. The same sample measured twice can differ by a year or more depending on platform and processing. Principal-component-based versions of several clocks were developed specifically to reduce this, and are a meaningful improvement worth asking about.

Cell composition. Blood is a mixture, and different white cell types have different methylation profiles. A shift in the proportion of naive to memory T cells, which happens with infection or immune aging, changes the estimate independently of anything else. Good analyses adjust for cell composition; consumer reports often do not say whether they have.

Tissue specificity. A blood clock reads blood. Correlation between tissues is moderate, not high, so an unfavourable blood reading is not automatically an unfavourable brain or liver reading.

Interpretation without uncertainty. A report giving a single figure to one decimal place, with no confidence interval and no statement of which model was used, is presenting a level of precision the method does not support. This is the most common failure in the consumer market.

How to Use an Epigenetic Test Sensibly

Epigenetic testing is worth doing under conditions that make the result interpretable, and largely wasted otherwise.

• Choose the generation that matches the question. For an outcome-weighted snapshot, a second-generation clock. For evaluating an intervention, a pace-of-aging measure. A first-generation clock answers a question nobody needs answered.
• Fix the provider and the method. Cross-provider comparison is uninterpretable.
• Leave twelve to twenty-four months between tests. Shorter intervals mostly measure noise and transient state.
• Standardise the conditions. Same time of day, similar sleep and training state, not during or immediately after acute illness.
• Read direction, not decimals. A shift of half a year is inside the noise. A consistent multi-year gap across independent measures is a signal.
• Do not optimise the number. The clock is a proxy. The physiology is the target.