Epigenetic Clocks: How DNA Methylation Measures the Rate of Aging

What are epigenetic clocks?

DNA methylation — the addition of a methyl group to cytosine at specific positions in the genome — changes in a predictable way with age. Steve Horvath showed in 2013 that methylation patterns at a few hundred CpG sites can predict a person's chronological age to within a few years. These became the first "epigenetic clocks" — essentially regression models trained on blood and tissue samples.

First-generation clocks (Horvath 2013, Hannum 2013) were optimized to predict chronological age. Their "acceleration" — the gap between the model's biological age and calendar age — is associated with various diseases, but predictive power remains modest. Second-generation clocks are trained directly on mortality and disease data: GrimAge incorporates plasma proteins linked to mortality; PhenoAge uses clinical biomarkers. The third generation, DunedinPACE, measures not accumulated age but the current rate of aging.

How well does GrimAge predict mortality?

Key study: the NHANES cohort (1999–2002), n=2,105, median follow-up 17.5 years, 998 deaths. For each additional 5-year GrimAge acceleration (i.e., the model's biological age is 5 years ahead of calendar age):

• All-cause mortality: HR 1.50 (95% CI 1.32–1.71; p<0.0001)
• Cardiovascular mortality: HR 1.55 (95% CI 1.29–1.86; p<0.0001)
• Cancer mortality: HR 1.37 (95% CI 1.00–1.87; p=0.049)

Of all clocks compared (Hannum, Horvath, PhenoAge, Vidal-Bralo), only GrimAge significantly predicted cardiovascular mortality. GrimAge2, the updated version, in a retrospective cohort (n=1,942, median follow-up 208 months) showed HR 1.07 per unit of acceleration for all-cause, cardiovascular and cancer mortality — a linear relationship with no apparent threshold effect.

DunedinPACE: what does it mean to measure pace, not age?

GrimAge shows accumulated damage — how many biological years have been "lived." DunedinPACE addresses a different question: how fast is a person aging right now? Developed at the University of Dunedin (New Zealand) from a 45-year longitudinal cohort, it reflects changes in 19 biomarkers — lung, kidney, liver, immune, and metabolic function — between measurement time points.

A value of 1.0 means average pace of aging for one's age; 0.8 means aging 20% slower, 1.2 means 20% faster. In the Framingham Heart Study Offspring Cohort (n=2,296, ages 25–101) faster DunedinPACE was associated with worse cognitive function at baseline and steeper cognitive decline over time (B = −0.37 to −0.45; p<0.001). The association held after adjusting for education.

An important property of DunedinPACE is its sensitivity to current living conditions. This makes it potentially useful for evaluating intervention effects, although most data remain observational or come from small samples.

What slows epigenetic aging?

Among non-pharmacological factors, physical activity is the best studied. A cross-sectional analysis of the Health and Retirement Study (HRS, n=3,873) compared active and inactive participants on three epigenetic measures. Physically active participants showed DunedinPACE 0.05 years/year lower (95% CI: −0.06; −0.04) adjusted for age, sex, race, education, health status and smoking. Analogous associations were found for GrimAge and PhenoAge.

Smoking, obesity and chronic stress are consistently associated with epigenetic clock acceleration across many studies. Diet and sleep are less studied; data on individual nutrients are contradictory.

Key caveat: most data come from cross-sectional studies. Lower DunedinPACE in physically active people may reflect a healthier baseline phenotype rather than an effect of exercise. Randomized trials with epigenetic clocks as the primary endpoint remain scarce.

Limitations: clocks measure but do not explain

Some researchers criticize clocks for "circularity": if a model is trained on disease and mortality data, its acceleration predicts those same outcomes not because it measures something fundamentally new, but because those outcomes are already encoded in the model's weights. A systematic comparison of 14 epigenetic clocks against 174 diseases (Nature Communications, 2025) showed that second- and third-generation clocks predict disease significantly better than first-generation ones, but their advantage over each other varies by specific outcome.

DNA methylation measured in blood reflects the state of blood-forming cells, not the whole organism. Clocks calibrated on blood may diverge from the state of brain, muscle or vascular tissue. This is a fundamental limitation for interpreting individual test results.