The Best GrimAge Test Is the One That Gives You Your IDAT Files

If you want a GrimAge number, buy a whole-blood DNA methylation array that returns your raw IDAT files, then compute the clocks yourself. The array matters more than the vendor’s report. Any service running Illumina EPIC v2 on whole blood and handing back unprocessed IDATs will let you produce several clocks from one sample. You get GrimAge, GrimAge2, and PhenoAge, along with the principal component versions of all of them. You also get DunedinPACE. A vendor that returns only a PDF with a single “biological age” locks you out of the calculation, prevents you from re-running it when a better clock is published, and makes longitudinal comparison across years impossible.

What GrimAge measures

GrimAge is a two-stage model. The first stage trains DNA methylation predictors of seven plasma proteins and of smoking pack-years. The proteins are listed here.

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

The second stage combines those surrogates with chronological age and sex into a linear predictor of time to death, calibrated in years. GrimAge2 adds methylation surrogates for log C-reactive protein and log hemoglobin A1c.

This matters for interpretation. GrimAge is a mortality-trained composite that happens to be expressed in age units. The smoking component carries a large share of the signal in most cohorts. If you have never smoked, a substantial part of the model’s discriminatory power in the training data does not apply to you, and your score is driven by the protein surrogates instead.

Epigenetic clocks in general track mortality and morbidity better than chronological age alone, and the second-generation clocks trained on phenotype or survival outperform the first-generation ones trained on age.¹ In head-to-head comparisons against metabolomic age scores, the methylation-based markers show consistent but moderate associations with outcomes.² Treat the number as an epidemiological summary. It is not a readout of your personal state.

Which platform to run

Use the Illumina Infinium MethylationEPIC v2.0 BeadChip on whole blood drawn into EDTA tubes. It covers roughly 935,000 CpG sites and needs about 1 microgram of input DNA. The workflow uses bisulfite conversion and a two-color intensity readout per probe. The alternatives are worse for this purpose.

Whole-genome bisulfite sequencing gives broader coverage but no clock was trained on it, and cross-platform calibration of clock CpGs from sequencing to array is unreliable at the per-site level. Targeted methylation panels covering only the clock CpGs are cheap but strand you if a new clock uses sites your panel did not assay. The array is the interoperable format.

Ask the lab for four things before you pay. First, the two IDAT files per sample, Grn and Red. Second, the sample sheet with Sentrix barcode and position. Third, the array version and manifest used. Fourth, confirmation that the sample is whole blood rather than buccal or saliva, because clock coefficients and reference distributions differ by tissue.

Computing the clocks

In R, read the IDATs with sesame rather than minfi if you are on EPIC v2. The sesame pipeline handles the v2 manifest and the probe suffixes cleanly:

library(sesame)
sdfs  <- openSesame(idat_dir, prep = "QCDPB", func = NULL)
betas <- openSesame(idat_dir, prep = "QCDPB")
betas <- betasCollapseToPfx(betas)   # EPICv2 -> legacy cg IDs

The betasCollapseToPfx step is the one people miss. EPIC v2 probe identifiers carry suffixes such as cg00000029_BC21, and every published clock uses legacy identifiers. Without collapsing, your clock function silently finds none of its CpGs and imputes the lot.

For GrimAge and GrimAge2, the reference implementation is the Horvath lab’s online calculator. It takes a beta matrix plus a sample annotation file with age, sex and tissue. The coefficients for GrimAge are not distributed publicly, so local computation is not possible the way it is for Horvath 2013 or PhenoAge. Budget for that dependency. It means you cannot audit the arithmetic and you must upload a beta matrix to a third party.

Everything else you can run locally. The dnaMethyAge and methylclock R packages implement Horvath multi-tissue, Hannum, PhenoAge, Zhang and others. The principal component clocks from Higgins-Chen and colleagues ship as R objects and a calcPCClocks function. DunedinPACE is its own package and returns a rate of aging rather than an age.

Our recommendation: report PC-GrimAge alongside DunedinPACE, and treat plain GrimAge as a secondary figure. The PC versions project the beta matrix onto principal components before applying clock weights, which suppresses probe-level technical noise. Single-CpG clocks have test-retest reliability on replicate samples that is poor enough to obscure a year of change. The PC versions are substantially more reproducible. DunedinPACE is separately useful because it estimates the current rate of biological aging from a single draw rather than accumulated damage.

What moves the number, other than aging

Normalization choice moves clock output by a meaningful fraction of a year. Noob, BMIQ and funnorm will not agree. Pick one pipeline and never change it across your own timepoints, because your own longitudinal slope is the only comparison you control tightly.

Cell composition is the largest biological confounder in blood. Naive CD8+ T cell fraction declines with age and drives a large part of the raw methylation signal. Estimate proportions with the Houseman deconvolution against a blood reference, and report them next to the clock output:

library(FlowSorted.Blood.EPIC)
counts <- projectCellTypeWB(...)  # Gran, Mono, Bcell, CD4T, CD8T, NK

Clonal hematopoiesis is a specific and underappreciated confounder. Expanded mutant blood cell clones alter the methylation profile of the sample. In elderly Danish twins, carriers of clonal hematopoiesis showed measurable differences in epigenetic age acceleration relative to non-carriers.³ After roughly age 60 a rising GrimAge in blood may reflect clonal dynamics in the hematopoietic compartment rather than systemic aging. That is a question for a hematologist. Your analysis script cannot answer it.

Acute illness, recent infection and inflammation shift the CRP surrogate in GrimAge2. Draw in a stable state and fasted, at the same time of day. Ideally use the same season each time.

Reading the result

Age acceleration is the residual from regressing predicted age on chronological age within a cohort. For a single person there is no cohort, so the lab substitutes a reference distribution. Ask which one. A residual of plus three years against a Framingham-derived reference and against a healthy volunteer cohort are different statements.

Effect sizes in the literature are real but modest at the individual level. In 45,091 US military veterans, PTSD was associated with advanced epigenetic age across several clocks.⁴ Congenital heart disease in young adults shows measurable acceleration relative to controls.⁵ Body roundness index associates with epigenetic aging in the Health and Retirement Study.⁶ Associations with day-to-day cognitive performance are weaker and less consistent.⁷ None of these support reading a one-year change in your own score as a change in your health.

Use GrimAge as one longitudinal series among several. Two or three timepoints a year apart from the same lab and the same pipeline tell you more than one number does. Plot them against your proteomics and blood chemistry. Anything that looks clinically concerning belongs in front of a physician who can order confirmatory testing.

Questions people also ask

What does a GrimAge test cost?

Consumer methylation panels that include GrimAge typically run a few hundred dollars per draw, and the Horvath lab calculator itself is free if you already hold a beta matrix. Prices change often, so compare on whether IDATs are returned rather than on the headline figure.

Is GrimAge better than the Horvath clock?

For predicting mortality and disease, yes. Horvath 2013 was trained to predict chronological age across tissues, so it is a good age estimator and a weaker outcome predictor. Second-generation clocks trained on survival or phenotype consistently outperform it on health outcomes.¹

Can I compute GrimAge from my 23andMe or whole-genome data?

No. Genotyping and sequencing read DNA sequence. Methylation clocks read chemical modification of cytosines, which requires bisulfite conversion and a separate assay.

How reproducible is a GrimAge score?

Replicate samples from the same blood draw do not return identical values, and single-CpG clocks vary more than the principal component versions. Run the PC clocks if you plan to track change over time.

Footnotes

1. A. Baki Yildirim, Duygu T. Yildirim, Hilal Akalin, et al. Epigenetic Clocks: Biological Aging, Disease, and Clinical Applications. The EuroBiotech Journal, 2026. https://doi.org/10.2478/ebtj-2026-0014 ↩ ↩²

2. Lieke M. Kuiper, Harmke A. Polinder-Bos, Daniele Bizzarri, et al. Evaluation of epigenetic and metabolomic biomarkers indicating biological age. 2022. https://doi.org/10.1101/2022.12.05.22282968 ↩

3. Mette Soerensen, Morten Tulstrup, Jakob Werner Hansen, et al. Clonal Hematopoiesis and Epigenetic Age Acceleration in Elderly Danish Twins. HemaSphere, 2022. https://doi.org/10.1097/hs9.0000000000000768 ↩

4. Erika J. Wolf, Rui Zhang, Xiang Zhao, et al. PTSD is Associated with Advanced Epigenetic Age in 45,091 US Military Veterans. 2026. https://doi.org/10.64898/2026.08.07.26359954 ↩

5. Parag N. Jain, Beryl C. Zhuang, Joanne Whitehead, et al. Epigenetic age acceleration in young adults with congenital heart disease. Clinical Epigenetics, 2026. https://doi.org/10.1186/s13148-026-02049-5 ↩

6. Sihan Chen, Bo Chen, Biao Ran, et al. The association between Body Roundness Index and epigenetic aging: insights from the Health and Retirement Study. 2026. https://doi.org/10.21203/rs.3.rs-9587275/v1 ↩

7. Daisy V Zavala, Natali Dzikowski, Shyamalika Gopalan, et al. Epigenetic and chronological age: Associations with cognitive performance in daily life. 2023. https://doi.org/10.31234/osf.io/sre8p ↩