Epigenetic clocks are weather, not climate

Three of my own papers contributed to the literature on epigenetic age estimation. I was an early enthusiast. I am, today, more cautious — not because the clocks are wrong, but because they are routinely asked to do something they were never designed to do.

What the clocks measure

An epigenetic clock is a regression model. It takes the methylation level at a few hundred CpG sites, weights them, and produces a single number that correlates strongly with chronological age across a heterogeneous training population. The first-generation clocks (Horvath 2013, Hannum 2013) were trained to predict chronological age. The second-generation clocks (PhenoAge, GrimAge) were trained to predict mortality and morbidity, and they outperform the first generation on those endpoints. The third-generation principal-component-based clocks are technically more reliable but conceptually similar.

All of these are correlational instruments. None of them is a measurement of aging in the sense that, say, a thermometer is a measurement of temperature. The output of a clock is the answer to the question "how does the methylation pattern of this sample compare to the methylation patterns of samples in the training set, weighted by their phenotypic age?" That is not the same question as "how biologically old is this person?"

The weather analogy

Climate is the long-run statistical structure of the atmosphere. Weather is what is happening today. The two are related — climate is, in part, the integral of weather — but they are not the same thing, and the methods we use to measure them are different. A thermometer reading on a Tuesday in March is weather. A 30-year temperature trend is climate.

Epigenetic clocks measure weather. They tell you about the current state of a particular tissue's methylation landscape on the day the sample was taken. That landscape is sensitive to short-term inputs: an acute infection, a recent vaccination, a poor week of sleep, a recent fast, a recent course of glucocorticoids. The clocks can move by 0.5–2 "years" within a fortnight in response to these inputs. Aging itself — the climate — is a much slower trajectory underneath that weather.

Why this matters for self-quantifiers

If you take a single epigenetic-age measurement, draw a line through it and a measurement six months later, and infer that an intervention you started three months ago is responsible for the difference, you are doing the equivalent of comparing the temperature on two arbitrary days and inferring climate change. The signal you want is real, but the noise on the measurement is at least as large as the signal you are trying to detect over that timescale.

The minimum sensible measurement protocol, in my view, is: three baseline measurements over six months, then the intervention, then three follow-up measurements at quarterly intervals over a year. That is six samples, roughly $2,400 at current commercial prices, and only minimally informative. A more useful protocol nests the methylation readout inside a panel of cheaper, faster biomarkers (hs-CRP, HbA1c, ALT, IGF-1, lymphocyte count) that move on shorter timescales and respond to the same upstream biology. The methylation result then provides one slow line in a chart of faster lines.

What I look at instead

• Resting heart rate variability (HRV), measured nightly. Cheap, sensitive, integrates autonomic and inflammatory state.
• Hs-CRP every 3 months. Cheap, sensitive, the clearest single index of systemic inflammation.
• HbA1c every 6 months. Cheap, sensitive, integrates glucose excursions over ~3 months.
• Grip strength every 6 months. Cheap, sensitive, integrates neuromuscular function.
• DXA every 12 months. Moderately expensive, sensitive, integrates lean mass and visceral adiposity.
• Methylation clock once a year, at most.

I do not own stock in any methylation testing company; I do co-own one of the patents underlying GrimAge. None of which alters my view that an annual clock measurement nested in a richer biomarker panel is the only sensible use of the technology in 2026.

What the clocks are good for

They are extraordinarily useful in epidemiology. The PREDIMED replication of GrimAge as a mortality predictor in 7,000 participants is the kind of question the clocks were designed to answer. They are useful in clinical trials with large sample sizes, where measurement noise averages out across hundreds of participants. They are useful as a sanity check on a long-running individual protocol, taken once a year.

They are not a dashboard. They are not a feedback signal for short-term intervention. They are, today, somewhere between a useful research tool and a marketing instrument. The line between those two uses is exactly the line between annual measurement and quarterly measurement, and exactly the line between treating the result as a single data point inside a wider panel and treating it as the panel itself.