Skip to content

What GrimAge Measures, and How to Compute It From Your Own Methylation Data

Oak
A huge shelled animal in a glowing misty forest, its back covered in spiral bands of small white nodules beside a ringed fallen trunk.

GrimAge predicts lifespan better than earlier epigenetic clocks because it was trained on a different target. First-generation clocks such as the Horvath pan-tissue clock were fit to chronological age, so their ceiling is how well chronological age itself predicts mortality. GrimAge was fit to time-to-death, using methylation-based surrogates of seven plasma proteins and of smoking pack-years as its features 1. The reported gain is meaningful: the composite was described as roughly 18% more accurate than chronological age and about 14% better than previously available methylation predictors for time-to-death 1. Accelerated GrimAge also tracked coronary heart disease, cancer incidence, comorbidity count and measures of physical functioning 1.

That is the summary. The rest of this page is about what you can and cannot do with your own data.

What the clock is built from

GrimAge is a two-stage model. In the first stage, elastic net regression predicts the plasma level of each of seven proteins from methylation beta values.

  • adrenomedullin
  • beta-2 microglobulin
  • cystatin C
  • GDF-15
  • leptin
  • PAI-1
  • TIMP-1

A separate elastic net predicts smoking pack-years from methylation. In the second stage, those eight surrogates plus chronological age and sex are combined in a Cox model of time-to-death, and the linear predictor is rescaled into units of years 1.

Two consequences follow. First, the model is a mortality risk score wearing age units, so a GrimAge of 52 for a 45-year-old means the person’s methylation profile carries the mortality hazard typical of a 52-year-old in the training population. Second, chronological age and sex are inputs. If you supply the wrong age, the output is not a corrected age, it is garbage.

The smoking component deserves attention because it carries a large share of the signal. The DNAm pack-years surrogate leans heavily on probes near AHRR and F2RL3. In a never-smoker with elevated DNAm pack-years, the component is picking up correlated inflammatory and xenobiotic-response methylation. The person has not secretly smoked.

Why the training target matters more than the algorithm

The algorithm is ordinary penalized regression. The choice of outcome is what separates clock generations. Clocks trained on chronological age converge toward an accurate age estimator, and their residuals capture whatever biology happens to deviate from the age trend. Clocks trained on mortality or on clinical biomarkers put that biology in the objective function directly.

This maps onto the broader question of what methylation drift represents. The epigenetic clock theory of ageing treats age-related methylation change as a readout of an underlying maintenance process rather than as a simple passive decay 2. Reviews of the field describe both directed, tissue-specific change and stochastic drift at CpG sites as they accumulate with age 3. Work in mammals has argued that the rate of epigenetic drift itself scales with species lifespan 4. GrimAge exploits the correlation efficiently. It does not resolve which mechanism dominates.

Generating the input: arrays versus bisulfite sequencing

For GrimAge specifically, you need an Illumina Infinium array. The practical choice is EPIC v2, which assays roughly 935,000 CpGs. EPIC v1 (about 850,000 sites) and the older 450K array also work. Ask the lab for raw IDAT files, both Grn and Red. A processed beta matrix is not enough. Without IDATs you cannot compute detection p-values or control-probe metrics, and you cannot renormalize.

Whole-genome bisulfite sequencing gives far more sites but does not let you run GrimAge as published. The clock’s probes are array probes, and array beta values and sequencing methylation ratios are not interchangeable without careful calibration. If your data is WGBS or RRBS, use a clock trained on that data type. BS-clock was developed for exactly this case, predicting age from high-resolution bisulfite sequencing rather than array beta values 5. Mitochondrial methylation is another sequencing-only domain, with reported inverse correlation between age and methylation at specific mtDNA CpG sites 6. Neither substitutes for GrimAge.

If you want one array and one sequencing run, do the array for clocks and the sequencing for discovery. Do not expect the two to agree numerically.

Running it

The awkward part is that GrimAge coefficients were never published. The CpG weights for the 2013 Horvath clock and PhenoAge sit in supplementary tables. GrimAge is available only through Horvath’s online calculator, now hosted by the Clock Foundation. You cannot reimplement it locally from the paper. Plan for that: your data leaves your machine, and you should read the terms before uploading.

What you control is the preprocessing you submit. Our approach in R:

library(minfi)
rgSet <- read.metharray.exp(base = "idats", extended = TRUE)

# QC before anything else
detP  <- detectionP(rgSet)
frac_failed <- colMeans(detP > 0.01)     # drop samples > 0.01
mSet  <- preprocessNoob(rgSet)           # noob background + dye bias
betas <- getBeta(mSet)

The calculator expects a CSV of beta values with a ProbeID column and one column per sample, plus a sample sheet carrying Age, Female (1/0) and Tissue. It applies its own BMIQ normalization against an internal gold standard, so noob is the right stopping point. Do not pre-filter probes to the clock set, and do not drop rows with NA. The calculator imputes missing clock probes from a reference mean, which is fine for a handful of sites and misleading for many. Check how many clock CpGs are absent from an EPIC v2 manifest before you trust the output.

Also estimate blood cell composition and keep it alongside the result:

library(FlowSorted.Blood.EPIC)
cc <- estimateCellCounts2(rgSet, referencePlatform = "IlluminaHumanMethylationEPIC")

Several GrimAge components correlate with neutrophil and lymphocyte fractions. A recent infection or an unusual differential will move your number. Draw blood in a consistent state and log the complete blood count from the same draw.

Reading your number

GrimAge acceleration is the residual from regressing GrimAge on chronological age within a cohort. For one person there is no cohort, so the residual depends entirely on the reference population the provider uses. An acceleration of +3 years is a statement about a reference distribution. It is not a physical quantity.

Technical noise is the second limit. Array clocks have non-trivial test-retest variability, and single measurements separated by weeks can differ by a year or more without any biological change. If you care about trajectory, plan at least three timepoints six to twelve months apart and fit a slope. Treat one measurement as a weak prior. Principal-component versions of these clocks were developed specifically to cut that measurement error and are preferable when a provider offers them.

The third limit is interpretive. GrimAge was validated as a population-level predictor of time-to-death and disease incidence 1. It is not a diagnostic test, and an elevated value does not identify a condition or indicate what to do about it. If your result concerns you, bring it and your standard blood panel to a physician. Nothing in a methylation array substitutes for clinical evaluation.

Questions people also ask

Does DNA methylation cause aging?

Not established. Methylation change is tightly correlated with age and predicts mortality, and the epigenetic clock theory of ageing proposes that clock sites report on an underlying maintenance process rather than merely accompanying it 2. Comparative work suggesting that drift rates scale with species lifespan is consistent with a causal role 4. The evidence is not sufficient to call methylation a cause of aging in humans.

Can DNA methylation clocks predict chronological age?

Yes, and quite precisely. Chronological-age clocks reach median absolute errors of a few years in blood, and methods built for high-resolution bisulfite sequencing report comparable accuracy on sequencing data 5. GrimAge is not one of these, since it targets mortality and takes age as an input.

Can DNA tell how old a person is?

Methylation can, within a few years. The underlying DNA sequence cannot, because it does not change appreciably with age. Forensic age estimation uses methylation assays rather than genotyping for that reason.

What diseases are linked to DNA methylation?

Accelerated GrimAge was associated with coronary heart disease, cancer incidence, higher comorbidity counts and reduced physical functioning 1. Separately, DNA methyltransferase activity and methylation state have been implicated in age-associated cognitive function 7. These are associations across populations. They are not individual diagnoses.

Footnotes

  1. Ake T. Lu, Austin Quach, James G. Wilson, et al. DNA methylation GrimAge strongly predicts lifespan and healthspan. Aging, 2019. https://doi.org/10.18632/aging.101684 ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  2. Steve Horvath, Kenneth Raj. DNA methylation-based biomarkers and the epigenetic clock theory of ageing. Nature Reviews Genetics, 2018. https://doi.org/10.1038/s41576-018-0004-3 ↩ ↩2

  3. Shuang Jiang, Yuchen Guo. Epigenetic Clock: DNA Methylation in Aging. Stem Cells International, 2020. https://doi.org/10.1155/2020/1047896 ↩

  4. Andrew R. Mendelsohn, James W. Larrick. Epigenetic Drift Is a Determinant of Mammalian Lifespan. Rejuvenation Research, 2017. https://doi.org/10.1089/rej.2017.2024 ↩ ↩2

  5. Congcong Hu, Yunxiao Li, Longhui Li, et al. BS-clock, advancing epigenetic age prediction with high-resolution DNA methylation bisulfite sequencing data. Bioinformatics, 2024. https://doi.org/10.1093/bioinformatics/btae656 ↩ ↩2

  6. Shakhawan K. Mawlood, Lynn Dennany, Nigel Watson, et al. Quantification of global mitochondrial DNA methylation levels and inverse correlation with age at two CpG sites. Aging, 2016. https://doi.org/10.18632/aging.100892 ↩

  7. Di Cui, Xiangru Xu. DNA Methyltransferases, DNA Methylation, and Age-Associated Cognitive Function. International Journal of Molecular Sciences, 2018. https://doi.org/10.3390/ijms19051315 ↩

Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.