Skip to content

How accurate is DNA age testing?

Oak
Two identical feathered reptile specimens on black plinths under studio light, each with a row of pale beads along the spine, spaced differently.

A good DNA methylation clock predicts chronological age from blood to within roughly 2 to 4 years median absolute error, and the best small-panel models reach about 2.1 years. That number is the accuracy of a regression against your birth date. It says almost nothing about the accuracy of the “biological age” figure a consumer test reports, because that figure is the residual left over after the prediction, and residuals are where all the measurement noise lives. If you take the same tube of blood and split it into two aliquots, some commercial clocks will return ages several years apart. So the short answer is that the underlying measurement is quite good and the product built on top of it is often not.

What the test measures

Your DNA sequence barely changes over a lifetime. What changes is methylation: a methyl group attached to the cytosine in a CpG dinucleotide, which shifts the local regulatory state without altering the base. Hundreds of CpG sites drift in a direction consistent enough with age that you can fit a regression to them. Forensic work established this early, showing that a handful of loci track age closely enough to estimate a donor’s age from a blood stain 1.

Most commercial tests read these sites on an Illumina array, either the EPIC (about 850,000 CpGs) or its v2 successor (about 935,000). Each probe returns a beta value between 0 and 1: the fraction of molecules methylated at that site, computed as M/(M+U+100). A clock is then a weighted sum over a subset of those betas, usually fit by elastic net.

The panel size varies enormously. Horvath’s multi-tissue clock uses 353 CpGs. Hannum’s blood clock uses 71. A cohort-based approach using just 30 sites reached a median prediction accuracy of 2.1 years on held-out blood samples, outperforming larger models 2. More sites does not mean more accuracy.

Accuracy against chronological age

This is the number that is well characterized, because the ground truth is unambiguous. Published blood clocks land in the 2 to 4 year range for median absolute error in adults. Horvath’s original multi-tissue model reported around 3.6 years across tissue types, and blood-specific models do better because they avoid the tissue-composition problem.

Accuracy is not uniform across the lifespan. Methylation changes fast in childhood and then slows, so clocks are tighter in children and loosen after 60. Population differences also matter. A study in an Arab cohort found strong age associations at known CpGs but also independent associations with sex and smoking, which means a model fit in one population carries that population’s covariate structure with it 3. If you are not from the training cohort, expect the error bars to widen.

Why “biological age” is a different question

Here is the tension that consumer marketing skips. A clock is trained to predict chronological age. The biological age claim comes from the residual: you are 50, the model says 45, so you are “aging slowly.” But if the model were perfect the residual would be zero and there would be nothing to sell. Every bit of biological signal in that residual coexists with every bit of technical noise.

Nothing in that residual has a birth certificate to validate against. Its meaning has to come from outcome associations in cohort studies, and those associations are population-level effects with wide individual variance. Second-generation clocks like PhenoAge and GrimAge were trained on clinical biomarkers and mortality rather than on age alone, which makes their residuals more interpretable. They are still not diagnostic. If a result concerns you, that is a conversation for a clinician. A dashboard is not the place for it.

The reliability problem, which matters more than accuracy

Test-retest reliability is measured by the intraclass correlation coefficient (ICC) across technical replicates: the same DNA, run twice. For several first-generation clocks the ICC sits well below 0.9, which translates to swings of a few years between identical samples. Given that the effect sizes people care about are also a few years, this is disqualifying for tracking yourself over months.

Two fixes help. Principal component clocks, which project onto PCs of the methylome before regressing, push ICC above 0.9 by averaging away probe-level noise. Running technical replicates and taking the mean also works, and it costs one extra array. We would do both.

Everything else that moves your number

Cell composition is the largest non-aging driver in blood. Granulocytes, CD4 T cells and B cells have different methylomes, so a shift in your differential count shifts your apparent age. A recent infection can do this. Any serious analysis deconvolves cell proportions and reports them alongside the age estimate.

Batch effects come next. Chip position, scan date and DNA input quantity all leave fingerprints. If your two timepoints were run on different chips six months apart, part of any difference you see is the chip. Sample handling matters too: DNA from a cheek swab is mostly buccal epithelium and needs a tissue-matched clock. A blood clock will not do.

How we would run it on our own data

Get the raw IDAT files. If a vendor will not give them to you, you cannot audit anything they tell you. From there, in R:

library(sesame)
betas <- openSesame("idats/", prep = "QCDPB", func = getBetas)

The QCDPB preparation chain does channel inference, dye bias correction, detection p-value masking (pOOBAH) and background subtraction (noob). Drop probes with detection p above 0.05 in more than 5% of samples, and drop the known cross-reactive and SNP-overlapping probes.

Then estimate cell fractions with EpiDISH before computing any clock:

library(EpiDISH)
cf <- epidish(betas, ref.m = centDHSbloodDMC.m, method = "RPC")$estF

For the clocks themselves, biolearn in Python and dnaMethyAge in R both implement the published coefficient sets, so you can compute Horvath, Hannum, PhenoAge, GrimAge and DunedinPACE from one beta matrix and compare them. They will disagree. That disagreement is the most informative output you will get, because it tells you how much of your number is model choice.

Two practical rules. Never compare an age from one lab to an age from another. And treat the first timepoint as a baseline with no interpretation attached, because a single draw has no slope.

Does whole-genome sequencing tell you your age?

Not directly, and this is a common confusion. Your germline sequence is fixed. What sequencing can measure is mutation accumulation: de novo mutations in offspring rise with parental age, and large-scale whole-genome studies have quantified that relationship 4. That reports on your parents at conception. It does not report on you now. Somatic mutation burden in blood does climb with age, but calling it reliably takes deep or single-cell sequencing that no consumer product offers.

Sequencing earns its place for other reasons. It reads coding variants at base resolution, which is why population studies use it to find disorders that targeted panels miss 5. Pair it with methylation and you get two independent layers: what you were born with and what has drifted since.

Questions people also ask

How accurate is DNA age testing compared to my real age? Blood methylation clocks typically predict chronological age to within 2 to 4 years median absolute error, and a compact 30-site model reached about 2.1 years on held-out samples 2. Accuracy is better in children and degrades in the very old.

Why did two tests give me different biological ages? Different clocks use different CpG panels and different training cohorts, so they measure related but distinct things. Add array noise and cell composition differences and a gap of several years between two vendors is unremarkable.

Can I retest to see if something I changed is working? Only with care. Use the same lab and the same clock, run technical replicates, and treat differences smaller than the assay’s reliability as noise. Three or four timepoints over a year tell you far more than two.

Is a cheek swab as good as blood? It is a different tissue with a different methylome. A clock trained on blood applied to buccal cells will be biased, so the model has to be tissue-matched.

Should I act on a high biological age result? Treat it as one noisy measurement. It is not a finding. If it prompts real concern, bring it and your standard blood work to a physician.

Footnotes

  1. Shao Hua Yi, Yun Shu Jia, Kun Mei, et al. Age-related DNA methylation changes for forensic age-prediction. International Journal of Legal Medicine, 2014. https://doi.org/10.1007/s00414-014-1100-3 ↩

  2. Miri Varshavsky, Gil Harari, Benjamin Glaser, et al. Accurate age prediction from blood using a small set of DNA methylation sites and a cohort-based machine learning algorithm. Cell Reports Methods, 2023. https://doi.org/10.1016/j.crmeth.2023.100567 ↩ ↩2

  3. Shaza B Zaghlool, Mashael Al-Shafai, Wadha A Al Muftah, et al. Association of DNA methylation with age, gender, and smoking in an Arab population. Clinical Epigenetics, 2015. https://doi.org/10.1186/s13148-014-0040-6 ↩

  4. Na Qin, Juncheng Dai, Yue Jiang, et al. Large-scale whole-genome sequencing reveals the landscape and health implications of de novo mutations. Nature Medicine, 2026. https://doi.org/10.1038/s41591-026-04585-2 ↩

  5. Dale L. Bodian, Elisabeth Klein, Ramaswamy K. Iyer, et al. Utility of whole-genome sequencing for detection of newborn screening disorders in a population cohort of 1,696 neonates. Genetics in Medicine, 2016. https://doi.org/10.1038/gim.2015.111 ↩

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.