GrimAge explained: how the mortality-trained epigenetic clock works
Prepared by: Research Desk
Prepared by: Research Desk. Sources are linked for verification.
Most people meet GrimAge as one line on a biological-age report, often next to a Horvath age and a DunedinPACE score. The name sounds dramatic, and the design is unusual. GrimAge was not trained to guess how old someone is. It was trained to predict time-to-death [1]. That one design choice explains why GrimAge behaves differently from other clocks, why researchers rely on it, and why one GrimAge number is easy to over-read.
This explainer covers how GrimAge is built, what its output means, what changed in GrimAge2, and how research studies use it as a measurement. For a broader comparison of all the major clocks, see our biological age test guide.
Where GrimAge sits among aging clocks
Epigenetic clocks are statistical models. They read DNA methylation at selected CpG sites and combine those readings into one estimate [6][7]. What sets one clock apart from another is the training target: the outcome the model was built to predict.
First generation: chronological age. Horvath's 2013 multi-tissue clock was trained to predict calendar age across dozens of tissue and cell types, using 353 CpG sites [2].
Second generation: health-linked targets. PhenoAge (Levine et al., 2018) was trained on "phenotypic age", a composite of chronological age and routine clinical chemistry markers [3]. GrimAge (Lu et al., 2019) went a step further and trained directly on mortality [1].
Third generation: pace. DunedinPACE estimates the rate of biological change. It was trained on how fast many physiological markers changed over years of follow-up in a single birth cohort [4].
Reviews of the field describe this progression as moving from "how old is this sample?" to "how much risk does this methylation profile carry?" [6][7]. GrimAge is the clearest example of the second question.
How GrimAge is built: a two-stage design
The original GrimAge paper used methylation and follow-up data from the Framingham Heart Study offspring cohort, a long-running human cohort with recorded deaths and age-related conditions [1]. The construction has two stages.
Stage 1: methylation "surrogates" for blood proteins and smoking
The authors first built separate methylation-based estimators for a panel of plasma proteins and for smoking pack-years, a measure of lifetime cigarette exposure [1]. Each estimator is a small model that predicts a protein level (or smoking history) from methylation alone. Plasma proteins are not measured directly in a GrimAge test. Methylation stands in for them.
Seven protein surrogates made it into the final model [1]:
DNAm surrogate | What the underlying protein is broadly associated with in research |
|---|---|
Adrenomedullin (ADM) | Vascular function |
Beta-2 microglobulin (B2M) | Immune activity and kidney function |
Cystatin C | Kidney function |
GDF15 | Cellular stress signalling |
Leptin | Fat tissue signalling |
PAI-1 | Blood clotting and fibrinolysis |
TIMP-1 | Tissue remodelling |
Smoking gets its own surrogate because tobacco exposure leaves a strong, well-documented methylation signature, and it is a major mortality factor. A methylation-based pack-years estimate captures exposure even when self-reported smoking history is incomplete [1].
Stage 2: regression on time-to-death
In the second stage, the surrogates, along with chronological age and sex, were fed into a survival model trained to predict time-to-death [1]. The resulting risk score was then rescaled into years, so the output reads like an age. That rescaling is a presentation step. Underneath, GrimAge is a mortality risk score expressed in age units.
Reading a GrimAge result
Reports usually show two related numbers.
GrimAge (in years). The raw clock output. It tracks chronological age closely because age is one of the model's inputs [1].
GrimAge acceleration (AgeAccelGrim). The difference between GrimAge and what would be expected for a person's chronological age. Positive values mean the methylation profile looks higher-risk than typical for that age. Negative values mean lower-risk than typical. In the original study, this acceleration measure was associated with time-to-death, time to coronary heart disease and cancer, and other age-related conditions [1].
Three points matter when interpreting the number:
It is population-level risk, not a personal forecast. GrimAge associations come from cohort statistics. A higher score means a profile that, across a population, was linked with earlier mortality. It does not tell any individual when anything will happen.
The smoking component carries weight. Because one surrogate estimates tobacco exposure, past or current smoking can shift the score substantially. That is part of why GrimAge predicts mortality well, and also why comparing a smoker and a non-smoker on GrimAge alone says little about other processes.
Results vary by clock. Getting a "young" Horvath age and an "old" GrimAge from the same sample is not a contradiction. The clocks were trained to predict different things [6][7]. Our biological age test guide explains this mismatch in more detail.
GrimAge2: what changed in 2022
In 2022, the same research group published GrimAge version 2 [5]. The structure stayed the same, with methylation surrogates combined into a mortality model. Two new protein surrogates were added:
logCRP: a methylation estimate of (log-transformed) high-sensitivity C-reactive protein, an inflammation marker.
logA1C: a methylation estimate of (log-transformed) hemoglobin A1C, a marker of longer-term blood glucose.
The authors evaluated GrimAge2 across multiple cohorts and ethnic groups. They reported that it predicted mortality more accurately than the original version and remained associated with a range of age-related conditions [5]. When a report or paper says "GrimAge", it is worth checking which version was used. Scores from the two versions are not interchangeable, and comparing a GrimAge1 result to a GrimAge2 result is not a like-for-like change.
Known limitations
Methodological reviews point to several issues that apply to GrimAge and to clocks in general:
Technical noise. Methylation arrays have measurement error. Small differences between two samples can come from processing, not biology [8].
Cell-type composition. Blood is a mixture of cell types, and shifts in that mixture change methylation readings. Separating composition effects from age-related signal is a known statistical challenge [8].
Correlation, not mechanism. Clocks are selected for predictive accuracy. The CpG sites they use are not necessarily causal drivers of aging, so a change in a clock score does not by itself show a change in an underlying biological process [7][8].
Training population. A model trained in one cohort may perform differently in groups that differ by ancestry, tissue type or age range. That is part of why GrimAge2 was tested across more diverse cohorts [5][7].
Tissue. GrimAge was developed on blood methylation data [1]. Results from saliva or other sample types may not be directly comparable.
How research studies use GrimAge as an endpoint
Because GrimAge tracks mortality risk, trials increasingly report it as a secondary outcome. One example is the DO-HEALTH trial, a randomized human trial in older adults aged 70 and over. It tested vitamin D, omega-3 and a home exercise program in a factorial design over three years [9]. The investigators measured several clocks, including PhenoAge, GrimAge, GrimAge2 and DunedinPACE. They reported that the omega-3 group showed modestly smaller increases on several of these clocks, on the order of months over the three-year period [9].
That study also shows how to read clock endpoints carefully. The reported differences were small. Each clock responded differently. A change in a methylation score is a change in a measurement, not a demonstrated change in lifespan. The same caution applies to any intervention study that reports GrimAge, including the peptide and metabolic-drug studies covered elsewhere on this site. Ask which version was used, the sample size, the follow-up length, and whether the result held across more than one clock.
Can consumers get a GrimAge score?
Some commercial epigenetic tests include GrimAge or GrimAge2 in their reports. These are generally offered as laboratory-developed tests (LDTs), not as FDA-cleared devices [10]. Our page on whether biological age tests are FDA approved explains what that status means and what lab certification does and does not cover.
Frequently asked questions
Is GrimAge the same as biological age?
Not exactly. GrimAge is a mortality-risk score rescaled into years, while other clocks define "biological age" through different training targets [1][6]. Each clock captures a different slice of what researchers mean by biological age.
What is the difference between GrimAge and GrimAge2?
GrimAge2 keeps the original design but adds methylation surrogates for C-reactive protein and hemoglobin A1C [5]. The authors reported stronger mortality prediction across several cohorts, and the two versions produce scores that are not directly comparable.
Why can GrimAge and Horvath age disagree on the same sample?
Horvath's clock was trained to predict chronological age [2], while GrimAge was trained on time-to-death [1]. Different training targets select different CpG sites, so the two models can legitimately give different answers from identical data.
Further reading
Epigenetic clocks are a fast-moving research area, and printed explanations date quickly. When choosing a book on aging biology or epigenetics, a few neutral criteria make comparison easier:
Publication date. GrimAge appeared in 2019, and GrimAge2 and DunedinPACE in 2022. Older titles may predate the newer clocks entirely.
Primary citations. Look for endnotes that point to peer-reviewed papers, so specific claims can be checked.
Training-target clarity. Strong explanations distinguish clocks trained on calendar age, on clinical markers, on mortality and on pace.
Correlation versus mechanism. Books that explain that clock scores are predictive measurements, not proof of biological change, tend to age better.
Format and level. Textbooks suit readers who want statistical detail. Popular-science titles suit a first overview. Check whether an ebook or audio edition includes the figures.
How this site sources pages and earns revenue is described on our about page.
Last updated: 19 October 2026.
Related on this site
Are biological age tests FDA approved? How aging tests are regulated
Biological age tests: how epigenetic clocks work and what they mean
Not medical advice. This page reports published research and regulatory records for information only.
Research-only notice: Content on this site is provided for research and educational purposes only and is not intended for human use. It is not medical advice.
Translation notice: Some content on this site may have been machine-translated using Google Translate and may contain errors.