Epigenetic clocks — Horvath, GrimAge, and what biological age tests actually measure

The science is real. It's also younger than most people realize, and the gap between what epigenetic clocks measure in population studies and what they can meaningfully tell any individual person is wider than the marketing usually acknowledges.

The story starts in 2013. Steve Horvath, a biostatistician and geneticist at UCLA, published a paper in Genome Biology that redefined how researchers thought about biological aging at the molecular level. Horvath's insight wasn't that DNA changes with age — that had been known — but that a specific pattern of DNA methylation, tracked across hundreds of specific sites in the genome, produces a remarkably accurate predictor of chronological age across tissue types. The correlation between his clock's predicted age and actual age in his original dataset was extraordinary, around 0.96. For a biological measurement, that was remarkable.

DNA methylation is the process by which methyl groups attach to cytosine bases at specific genomic sites, typically at cytosine-guanine dinucleotide sequences called CpG sites. Methylation patterns regulate gene expression — a heavily methylated promoter region tends to silence the gene it controls. These patterns are not random. They change in predictable ways as the body ages, some sites gaining methylation and some losing it, with a consistency across people that Horvath's analysis captured and quantified. His first clock used 353 CpG sites, identified through machine learning, to produce a predicted age. Feed it a methylation profile from any tissue — blood, saliva, brain, skin — and it would estimate how old that tissue appeared to be, in methylation terms, with impressive accuracy.

This was the founding observation of the epigenetic clock field. In the years that followed, different research groups built different clocks for different purposes, and understanding the distinctions between them matters for interpreting your results.

The Hannum clock, published the same year as Horvath's, was built specifically from blood and used a different set of 71 CpG sites. It performed similarly to Horvath on chronological age prediction in blood, though it didn't generalize across tissue types the way Horvath's multi-tissue clock did. Then came the second generation: PhenoAge, developed by Morgan Levine at Yale, published in 2018. PhenoAge was built differently. Instead of training a clock to predict chronological age, Levine trained it to predict phenotypic age — a composite measure derived from nine clinical biomarkers associated with mortality risk, including albumin, creatinine, glucose, C-reactive protein, and several others. The methylation sites PhenoAge captured were those that correlate with this mortality-predictive composite rather than simply with how many years a person has lived. The result was a clock that, in large epidemiological studies, correlates more strongly with mortality risk than Horvath's original. Someone whose PhenoAge runs significantly older than their chronological age is, in aggregate population data, at elevated risk. Someone whose PhenoAge runs younger has a more favorable risk profile.

GrimAge, developed by Ake Lu and Steve Horvath and published in 2019, pushed this logic further. Lu and Horvath trained their clock on a different target: DNAm surrogates of aging-associated plasma proteins and smoking history, calibrated to predict time-to-death. GrimAge is currently the strongest predictor of mortality and morbidity of the epigenetic clocks, showing the highest hazard ratios for all-cause mortality, cancer, coronary heart disease, and other endpoints in the large cohort studies it's been validated in. It uses 1,030 CpG sites and a more complex underlying model. When researchers compare clocks for predictive utility in longevity research, GrimAge typically performs best. That's a meaningful distinction if you're trying to understand what the number on your report implies about your actual risk trajectory.

Then there's DunedinPACE, which is different in a philosophically important way. The previous clocks are cross-sectional: they predict a biological age at a single point in time, estimated from a single methylation snapshot. DunedinPACE, developed by researchers at the University of Otago and Duke using data from the Dunedin longitudinal cohort, attempts to measure the pace of aging — a rate rather than a position. It's calibrated to data from people tracked from birth to midlife, with repeated biomarker measurements, and it estimates how fast the underlying biological machinery is aging right now. A DunedinPACE score above 1.0 means aging faster than average; below 1.0 means slower. In some ways this is the more actionable measurement: it tells you about trajectory rather than snapshot, and it may be more sensitive to lifestyle changes because it's tuned to detect rate of change rather than accumulated damage.

All of this is real science. The correlations with mortality in large prospective cohorts — UK Biobank, the Framingham Heart Study offspring cohort, the Women's Health Initiative, others — are replicated and meaningful. People with accelerated epigenetic aging carry genuinely elevated risk for a range of serious outcomes. The population-level signal is there.

Here is where honesty requires slowing down.

Population-level correlation and individual-level prediction are different things. The hazard ratios associated with epigenetic age acceleration in large studies are meaningful but modest at the individual level. In a 50,000-person study, finding that people with accelerated GrimAge have 1.5 times the mortality risk over fifteen years is a real finding. It means the clock is measuring something biologically relevant. It does not mean that your individual score of 47.6 instead of 44 tells you anything with precision about your personal risk trajectory. The effect sizes, once you move from population statistics to the level of one person, become much less actionable. Biological noise is substantial. Methylation measurements themselves have technical variability across laboratories and platforms. Two tests from the same person at the same time can yield slightly different results.

The intervention question is even more complicated. The consumer experience has moved rapidly from "this is a research tool" to "reduce your biological age by 5 years." That framing is, to be direct, more aspirational than rigorous. The intervention studies showing meaningful epigenetic clock reduction are limited and methodologically variable. Most are small. Most are unblinded. Most involve comprehensive lifestyle interventions — diet, exercise, sleep, stress reduction, sometimes hormonal or supplemental protocols — making it impossible to attribute any directional change to a specific variable. The TRIIM trial, conducted by Greg Fahy and published in Aging Cell in 2019, is one of the more widely cited. Nine men over twelve months received a protocol including recombinant human growth hormone, DHEA, metformin, vitamin D, and zinc. On Horvath's multi-tissue clock, average biological age decreased by about 2.5 years relative to the expected increase during the treatment period. It generated legitimate scientific interest. It also enrolled nine participants, had no placebo control, and could not isolate which elements of the protocol drove the change, or whether the changes were durable. Larger, more rigorous trials are underway. The baseline evidence for specific interventions reliably and durably reducing epigenetic age in humans is, at this writing, limited.

This is not a reason to dismiss the technology. It's a reason to hold it correctly.

What epigenetic clocks are genuinely useful for, even now, is longitudinal tracking of directional trends in yourself. A single test produces a number of uncertain individual precision. A test every six to twelve months produces a trajectory. If your GrimAge or PhenoAge is moving consistently in one direction relative to your chronological age — accelerating or decelerating — that directional signal is more meaningful than any single snapshot. It's the same logic as watching a series of fasting glucose tests over three years rather than interpreting a single result. The trend tells you something the cross-section cannot.

They're also useful for identifying large discordances that warrant attention. A person whose biological age runs five or more years ahead of chronological age in a well-validated clock, especially GrimAge, has a meaningful signal worth investigating with their prescribing provider — not because the clock is a diagnosis, but because it's consistent with population data showing elevated risk, and because the variables that drive epigenetic age acceleration are often addressable: chronic inflammation, metabolic dysfunction, sleep disruption, sustained psychological stress, smoking, obesity. These are not mysteries. The clock is, in effect, an aggregated molecular readout of how those factors have been expressing in your tissue over time.

What they cannot do is precisely determine your biological age in an individually actionable sense, predict with accuracy whether you'll develop a specific disease, or confirm that a particular intervention you've taken has worked. The "I took this supplement for three months and my biological age dropped by two years" narrative circulates widely. The study design needed to confirm that narrative — properly controlled, adequately powered, with the right clock for the question being asked — almost never exists for the specific intervention being claimed.

There's a more fundamental epistemological point here. Biological age is not a thing that exists the way chronological age exists. Chronological age is a count. Biological age is a model — a set of weights applied to methylation patterns, trained on population data to predict outcomes. Different clocks produce different numbers for the same person. The same person tested twice in the same week may get slightly different results. Averaging multiple clocks probably gives a more stable estimate than any single one. And the number, whatever it is, describes a probabilistic risk landscape, not a biological fact.

The people who will get the most from this technology are those who treat it as one layer of a broader picture — longitudinal biomarker tracking, alongside lipid panels, inflammatory markers, glucose and insulin dynamics, hormonal profiles — rather than a singular verdict. A biological age number that sits alongside a comprehensive metabolic panel, a DEXA scan, VO2 max testing, and detailed sleep data means something different than a biological age number delivered in isolation. Context is the measurement.

The clock is pointing at something real. The error is in mistaking a pointer for a destination.