Epigenetic clocks are the most quantitatively rigorous biomarkers of biological aging available outside specialist clinical settings. The four major clocks (Horvath, Hannum, PhenoAge, GrimAge) differ in their training data, biomarker basis, and what they actually predict. Marketing of consumer "biological age" tests collapses these distinctions. This page lays out how each clock is built, what it forecasts, what intervention reversibility data exists, and where the evidence does not yet support the strong claims being made.
What is an epigenetic clock?
DNA methylation is the addition of a methyl group to cytosine in CpG dinucleotides. The pattern across the genome is tissue-specific and changes systematically with age. Some CpGs gain methylation with age (typically in CpG islands near gene promoters), others lose methylation (typically in repetitive elements and intergenic regions). The age-dependent component is reproducible enough that a regression model trained on methylation data can predict chronological age with median error of 2 to 5 years.
A clock is a regression equation: weighted combination of methylation values at hundreds of CpG sites, summed to produce a single age estimate. The training procedure determines what the clock predicts. Train on chronological age, you get a chronological-age estimator. Train on time-to-death, you get a mortality predictor. Train on phenotypic aging biomarkers, you get a clock that captures health status better than chronological age alone.
The "epigenetic age acceleration" metric is the residual: predicted methylation age minus chronological age. Positive acceleration (DNAm age higher than chronological age) is associated with elevated mortality and disease risk in most studies. Negative acceleration (DNAm age lower than chronological) is associated with longer expected lifespan.
Horvath 2013: the multi-tissue clock
Horvath 2013 was the first published clock with broad applicability ( Horvath 2013 ). Trained on 8,000 samples across 51 tissue types, it uses 353 CpG sites and predicts chronological age with correlation 0.96 and median absolute error 3.6 years. The cross-tissue applicability was the breakthrough: the same 353-CpG model works in blood, brain, muscle, kidney, and most other tissues. The clock is sometimes called pan-tissue.
The Horvath clock was trained on chronological age, which means its primary signal is age itself. Mortality prediction is real but modest: pooled meta-analyses show hazard ratios around 1.04 per year of acceleration, weaker than the later phenotypic clocks. The clock is best understood as a chronological-age estimator that captures some but not all of the age-related biology that matters clinically.
Use cases where the Horvath clock is most informative: tissue-specific age questions (does this brain look older than expected for chronological age?), forensic age estimation, and as a baseline for studying intervention effects across multiple tissues.
Hannum 2013: the blood-specific clock
Hannum 2013 was published the same year as Horvath but trained specifically on whole blood ( Hannum et al. 2013, n=656 ). The clock uses 71 CpG sites and shows tighter age prediction in blood (correlation 0.96, median error 3.0 years) than the Horvath clock, because it does not need to handle cross-tissue variation. It is also more sensitive to short-term changes in blood biology, which can be useful for intervention studies but introduces noise from acute inflammation, infection, and recent training.
Hannum and Horvath are sometimes called "first-generation" clocks because they were trained directly on chronological age. They are good age estimators but blunt mortality predictors. Both correlate with smoking, BMI, and chronic disease status, but the correlations are weaker than for the second-generation phenotypic clocks. For modern biological-age assessment, they are typically reported alongside the newer clocks rather than as primary metrics.
PhenoAge: the phenotypic clock
Levine 2018 changed the framework by training on a phenotypic aging score rather than chronological age ( Levine et al. 2018 ). The PhenoAge phenotypic score combines 9 clinical biomarkers (albumin, creatinine, glucose, CRP, lymphocyte percent, mean cell volume, red cell distribution width, alkaline phosphatase, white blood cell count) plus chronological age into a single Mahalanobis-distance metric calibrated against time-to-death in NHANES III. The methylation clock (DNAm PhenoAge) was then trained to predict this phenotypic score.
The result is a methylation clock that captures health status more than chronological age. PhenoAge acceleration of 5 years (DNAm PhenoAge 5 years above chronological) associates with about 9 to 10% higher all-cause mortality, 19% higher coronary heart disease incidence, and meaningfully elevated cancer risk. The hazard ratios are stronger than for Horvath or Hannum, and the clock is sensitive to the intermediate biomarkers that intervention trials can move.
The biologicalx biological age tool implements the underlying PhenoAge phenotypic score (the 9-biomarker version, without methylation) because the inputs are routinely available from a comprehensive metabolic panel and CBC. The methylation version requires a DNAm sample and lab analysis (typically $200 to $400 commercially) and is more precise but less accessible.
GrimAge: the mortality clock
Lu 2019 took the methodology one step further ( Lu et al. 2019 ). GrimAge trains directly on time-to-death and incorporates DNAm surrogates for smoking pack-years and 7 plasma proteins (adrenomedullin, beta-2 microglobulin, cystatin C, GDF15, leptin, PAI-1, TIMP1). The DNAm-based "DNAmSmoking" surrogate is a particularly clever component: it captures smoking history more accurately than self-report, which removes one of the largest confounders in epidemiological studies of aging.
GrimAge is currently the strongest mortality predictor among the four clocks. Per-year of GrimAge acceleration, hazard ratios for all-cause mortality run around 1.10 in pooled cohorts. Per 5-year acceleration, the hazard ratio is roughly 1.5. GrimAge also predicts lifespan, healthspan (defined as time to first major chronic disease), and time-to-cancer better than the first-generation clocks.
The trade-off: GrimAge is the most "downstream" of the clocks. It captures mortality risk well precisely because it is trained on it, but this also means a GrimAge reading reflects the integrated load of smoking, inflammation, vascular dysfunction, and metabolic stress more than a fundamental aging-rate signal. For mortality prediction, this is a feature; for understanding which biological process to intervene on, it is a limitation.
Can you reverse your epigenetic age?
The intuitive question is whether interventions can move a person's clock back. The evidence is preliminary and mixed:
- Caloric restriction. Small trials in humans (CALERIE-2 substudy, 12% calorie restriction over 2 years) showed roughly 0.5 year deceleration in PhenoAge over the intervention period. The effect was small and the trial was not designed primarily for clock readouts.
- Exercise. Cross-sectional data shows fitter adults have lower epigenetic age acceleration. Longitudinal intervention trials are sparse; effects on clocks are modest (0.5 to 1 year deceleration over 6 to 12 months in small trials).
- Comprehensive lifestyle intervention. A 2021 small RCT (n=43, 8 weeks of diet+exercise+sleep+stress intervention) reported about 3 year Horvath clock deceleration in the intervention group. The trial was small, single-blinded, and the effect size is larger than other reports, raising replication concerns.
- Pharmacology. Mannick 2018 demonstrated rapamycin improved immune function in elderly adults via TORC1 inhibition ( Mannick et al. 2018, n=264 ); methylation clock readouts were not the primary endpoint. Whether mTOR inhibitors decelerate epigenetic clocks is being studied but no large trial has reported.
The deeper question is whether clock deceleration translates to mortality reduction. Mendelian randomization studies suggest the methylation patterns that drive clock acceleration are partially causally upstream of mortality, but the causal weight remains uncertain. An intervention that decelerates a clock by 2 years on paper is not yet proven to extend life by 2 years.
How do you interpret a biological age test result?
Most consumer DNAm-age tests report Horvath, Hannum, PhenoAge, and GrimAge in some combination. Reasonable interpretation:
- Look at GrimAge for mortality prediction. It is the most validated for that purpose. Per-year acceleration matters more than the absolute number.
- Use PhenoAge as the actionable clock. Its inputs map to clinical biomarkers (CRP, glucose, lymphocytes) that intervention trials can move. Tracking PhenoAge over time gives a more responsive readout than GrimAge.
- Treat single-timepoint readings cautiously. Test-retest variability for most clocks is 1 to 3 years on identical samples. Year-over-year tracking with the same lab and protocol is more informative than absolute values.
- Avoid overweighting Horvath/Hannum alone. They are good age estimators but weaker mortality predictors. Marketing that emphasizes a single Horvath number is underdosing the more informative clocks.
- Recognize the methylation-vs-phenotypic distinction. A non-methylation PhenoAge calculator (using just the 9 blood markers) captures most of what the methylation version captures, at much lower cost.
What the clocks cannot tell you
Three claims around epigenetic clocks deserve scrutiny:
- "My biological age is X, so I will live to Y." Population-level hazard ratios do not translate cleanly to individual prognosis. A 5-year clock acceleration shifts the population mortality curve modestly; the individual variance is enormous.
- "This intervention reversed my biological age by N years." Test-retest noise plus regression to the mean make single-trial individual deltas unreliable. Aggregate trial data on intervention effects is what matters.
- "Methylation clocks measure aging directly." They measure correlates of aging. Whether the methylation patterns are upstream causes or downstream consequences of aging biology is partially settled (some are causal, some are not) and varies by clock.
Practical synthesis
For an adult interested in tracking biological aging:
- Start with the cheap proxy. The non-methylation PhenoAge score (9 standard blood markers + age) captures most of what methylation PhenoAge captures. Run a comprehensive metabolic panel, CBC, and CRP. Use the biological age calculator.
- Add a methylation clock once or twice. A baseline DNAm reading establishes the GrimAge and PhenoAge methylation values. Repeat in 1 to 2 years to track delta.
- Track CRP, fasting glucose, lipid panel, and lymphocyte percent more frequently. These move on intervention timescales (weeks to months) and feed directly into PhenoAge.
- Do not over-react to single readings. Variance is real, and the actionable interventions (exercise, sleep, weight management, cardiovascular risk reduction) are the same regardless of clock readout.
