This site and its domain are for sale. Email the owner if you want to talk about it.

bioclocktest.com

GrimAge: The Clock Trained on Death

GrimAge was trained on death rather than birthdays, which makes it a good mortality score and an unproven target for anyone hoping to lower it.

A row of small glass blood sample vials in a metal laboratory rack on a bench, lit by soft window light.

I spent an unreasonable part of this week reading one sentence from a lab-services company’s explainer page the way a lawyer reads an indemnity clause, because it turns out to be the cleanest description of GrimAge available. If you typed grimage into a search box, here is the plain answer: GrimAge is an epigenetic clock, meaning a formula that reads chemical tags on DNA (usually from a blood sample) and outputs a number in years. It was published by Lu and colleagues in 2019, with Steve Horvath among its developers. What makes it different is what it was trained to predict. Not birthdays.

Death.

A GrimAge higher than someone’s calendar age means their blood methylation looks like the blood of people who died sooner. In a nationally representative U.S. cohort (NHANES) followed for 17.5 years, it was the strongest mortality predictor among the clocks compared. What it has not been shown to be, according to a 2026 systematic review, is a validated target that an intervention can push down with a known payoff in years of life. Those are different claims, and most of what follows is about the distance between them.

The sentence, and what a training objective is

Here is the sentence, from CD Genomics: “Second-generation clocks such as PhenoAge and GrimAge shifted the training objective from predicting chronological age to predicting clinical phenotype, disease incidence, and all-cause mortality.” It reads like a software changelog. It is actually the whole story, so let’s take it clause by clause.

Start with “training objective.” The plain version: you collect a big pile of blood samples, you attach an answer to each one, and you let a computer hunt for the combination of methylation sites that best guesses the answer. Whatever you write in the answer column, the computer learns to guess. That column is the training objective. (It is the least glamorous phrase in the sentence and the only one doing any work.)

The point is that a clock has no opinion about aging. It has an answer key. Change the key and you get a different clock out of the same DNA.

From chronological age, which was the easy part

“From predicting chronological age.” This is the first generation. The Scientist describes those clocks, Horvath’s included, as trained “by comparing DNA methylation signatures of young and old individuals to predict chronological age.” The Horvath pan-tissue clock uses 353 CpG sites and lands within about 3.6 years of a person’s real age on average, across many tissue types. As engineering, that is impressive.

As a health test, it has an obvious problem. If the answer key is your age, the perfect clock tells you your age. You already knew that. It’s a very expensive calendar.

People noticed. The same CD Genomics page says the purely chronological objective makes first-generation clocks better at capturing normal aging than the pathological kind tied to disease and death. The data agree in their own way. In the Add Health study of younger U.S. adults, sociodemographic and lifestyle factors showed up more often in clocks trained on health outcomes (PhenoAge, GrimAge, DunedinPACE) than in Horvath. And in a study of men with hypertension, heart disease, obesity and dyslipidemia, none of 19 first-generation clocks registered those conditions. GrimAge and DunedinPACE did.

What GrimAge is trained to predict, read literally

“To predicting clinical phenotype, disease incidence, and all-cause mortality.” The last item is the one that earned the name. “All-cause” means what it says: the clock does not care what killed anyone (heart attack, cancer, falling piano), only that they died. This is basically the premise of the Final Destination films, minus the set pieces and plus a great deal of regression.

So GrimAge predicts when you will die. That’s not quite right, and the precise version is stranger. It does not produce a date. It scores how much someone’s blood resembles the blood of people who died sooner, and the score comes out in years (presumably because “62” fits on a results page better than a hazard ratio does). A GrimAge “age” is a mortality score, expressed in years.

As a mortality score, it is pretty good. In the NHANES cohort, each five years of GrimAge acceleration carried a hazard ratio of 1.50 for death from any cause, against 1.15 for PhenoAge, 1.17 for Hannum and 1.12 for Horvath. GrimAge acceleration also specifically predicted cardiovascular death (hazard ratio 1.46). In people with well-controlled HIV, positive acceleration on GrimAge versions one and two was each linked to more than double the risk of serious non-AIDS events. Train a clock on death and it gets good at death. Lol, of course it does.

The smoking stand-in and GrimAge2

Two pieces of GrimAge lore come up constantly, and on both the honest answer is shorter than the lore.

The first is smoking. GrimAge has a reputation for containing a methylation-based estimate of smoking pack-years: a guess at how much someone has smoked, made from their DNA chemistry and used instead of asking them. The logic is sensible on its face (smoking kills people, and people misremember their own habits, or whatever). But the material behind this piece does not describe that component, its weight in the formula, or how closely it tracks real smoking histories. I can explain why a death-trained clock would want a smoking stand-in. I can’t tell you how this one is built.

That isn’t settled here.

The second is GrimAge2. It exists, in the sense that studies report it: the HIV cohort above used both versions, and a study of organophosphate pesticide metabolites ran GrimAge alongside GrimAge2. In that pesticide study, the estimated mixture effect rose at higher exposures for Horvath, Hannum and PhenoAge but not for either GrimAge, and the effect sizes for the GrimAge clocks were modest. What the second version changed is not in my sources, so I won’t guess. (Sequels, in my experience, tend to follow originals that did well.)

What moves GrimAge, and why trials keep asking

Does it move? Sometimes. In adults with HIV, starting antiretroviral therapy partly reversed acceleration on Horvath, Hannum, GrimAge and PhenoAge, but GrimAge and PhenoAge were still significantly elevated compared with HIV-negative controls after two years. In the CALERIE trial, 25% calorie restriction over two years slowed DunedinPACE by about 3% but produced no statistically significant change in GrimAge. The small TRIIM study, on the other hand, reported a two-year decrease in GrimAge relative to chronological age. (GrimAge responded to one study’s attentions and not the other’s, a pattern we would describe less politely in a person.)

Each party here is acting sensibly. Researchers want a measure linked to outcomes people actually care about, and GrimAge is one. Test sellers want a number customers understand, and an age is that number. Trial designers want an endpoint that shows up in two years rather than thirty, and a blood draw does. The planning meeting more or less writes itself: “Hey, so the mortality endpoint needs a few decades and thousands of people. What if we just ran the clock?” (That quote is invented, though not much.)

The 2026 systematic review is where that meeting meets reality. It found no trial-level validation of any epigenetic clock as a surrogate endpoint, rated the certainty of evidence very low across the board, and concluded the clocks are best treated as exploratory or adjunctive markers of response. I mean, it’s fine. That is what early biomarkers are. It just means a falling GrimAge in a trial is a falling GrimAge, and nobody has yet shown it is a falling death rate.

There is also the matter of what gets measured. Almost all clocks in commercial tests are trained on blood, and the same article in The Scientist notes that GrimAge’s links to disease and death come from blood studies, adding that “this linkage has not been established for saliva-based clocks.” Anyway, the death part of the clock was learned from blood.

And the only way to check whether lowering GrimAge actually helps is to go back to the answer key and count deaths, which is the one column every trial participant signed up hoping to stay out of.

Sources

  1. The Sociodemographic and Lifestyle Correlates of Epigenetic Aging in a Nationally Representative U.S. Study of Younger Adults
  2. Epigenetic Clocks Power Biological Age Data, But Why Do Results Differ Across Tests? | The Scientist
  3. Choosing Epigenetic Clock Models: Horvath, Hannum, PhenoAge, GrimAge and Beyond - CD Genomics
  4. Epigenetic age acceleration changes 2 years after antiretroviral therapy initiation in adults with HIV: a substudy of the NEAT001/ANRS143 randomised trial.
  5. Associations of Organophosphate Pesticide Metabolite Mixtures with Epigenetic Age Acceleration
  6. Epigenetic age acceleration and mortality risk prediction in U.S. adults
  7. Epigenetic signatures of cardiometabolic risk in men: accelerated aging and differential methylation replicated across cohorts
  8. Prognostic Value of Blood Epigenetic Biomarkers of Aging in Persons With Well-Controlled HIV-1 Infection
  9. Four Aspects to Interpret Epigenetic Clock Results - CD Genomics
  10. EFFECT OF LONG-TERM CALORIC RESTRICTION ON THE PACE OF BIOLOGICAL AGING IN HEALTHY ADULTS FROM THE CALERIE TRIAL
  11. Concordance between changes in epigenetic clocks and changes in clinical outcomes in geroscience intervention trials: A systematic review
  12. Reversal of epigenetic aging and immunosenescent trends in humans

Bio Clock Test explains published research. It is not medical advice and it does not diagnose anything. We never recommend a test, a supplement or a treatment. Talk to a clinician about your health.

Read next