What a DNA Age Test Measures, and How Much to Trust the Number

A DNA age test does not read your age out of your genome. Your germline sequence is essentially the same at 20 and at 70. What these tests measure is DNA methylation, the addition of a methyl group to cytosine in a CpG dinucleotide, at a few hundred selected sites whose methylation fraction drifts predictably with age across a population. A regression model turns those fractions into a number in units of years. A good blood-based clock predicts chronological age with a median absolute error of roughly three to five years in a held-out cohort, which means the assay is a decent population-level age estimator and a coarse individual-level one. The deviation between predicted and chronological age, the residual usually marketed as “biological age,” is the part people care about and also the part with the worst signal-to-noise.

What the assay physically measures

Almost every commercial DNA age test starts with sodium bisulfite conversion. Bisulfite deaminates unmethylated cytosine to uracil, which reads as thymine after amplification, while 5-methylcytosine is protected and still reads as cytosine. The methylation fraction at a site, the beta value, is then computed as beta = M / (M + U + 100), where M and U are methylated and unmethylated intensities or read counts and the offset of 100 stabilizes low-intensity probes. Everything downstream is a function of those betas.

The age signal is real and has been characterized independently of any commercial product. Methylation at specific loci changes systematically from birth through adulthood, with some regions gaining methylation and others losing it, and the trajectories are steep in early life and flatten later ¹. That nonlinearity is why Horvath’s pan-tissue clock transforms age before fitting: it applies a log transform below age 20 and a linear function above, then fits an elastic-net regression over 353 CpGs.

Which clocks to run, and their error

There are four generations of clock worth knowing, and they answer different questions.

The first generation predicts chronological age. Horvath’s 2013 pan-tissue clock (353 CpGs, median absolute error around 3.6 years) works across tissues including brain, buccal cells, and blood. Hannum’s clock (71 CpGs) is blood-specific and typically lands near 4 to 5 years of error. Both were trained on chronological age, so a perfect model would have zero residual and therefore zero information about health. Their usefulness comes entirely from the errors they make in a consistent direction.

The second generation trains on something other than the calendar. PhenoAge regresses methylation onto a composite of clinical chemistry values, and GrimAge regresses onto methylation-based surrogates for plasma proteins and smoking pack-years. These predict mortality and morbidity better than the first-generation clocks in cohort studies, because they were built to.

The third approach estimates rate rather than level. DunedinPACE was trained on the longitudinal slope of multiple organ-system biomarkers within a single birth cohort, and reports a pace in years of biological change per calendar year. It is the most stable of the commonly available scores on technical replicates, which matters enormously if you plan to retest.

A fourth line of work, from forensics, is worth understanding because it shows the ceiling on the pure age-prediction problem. Targeted massively parallel sequencing of a handful of loci, typically ELOVL2, FHL2, KLF14, TRIM59, C1orf132, and ASPA, reaches mean absolute errors in the three-to-four-year range on blood, using random forest or neural network regression rather than a linear model ²³. Comparisons across model families on the same sequencing data find that the choice of learner matters less than the loci and the tissue, and that error grows in older subjects ⁴. A test claiming better than about two years of error on blood is either overfitting or reporting training performance.

Arrays versus targeted sequencing

Most consumer tests run the Illumina Infinium MethylationEPIC array, either v1 (about 850,000 probes) or v2 (about 935,000). You get IDAT files, which are the raw red and green channel intensities, and that is what you should insist on receiving. Targeted bisulfite amplicon sequencing of 6 to 20 loci is cheaper and gives you per-read methylation with real depth, but it locks you into whichever clock the panel was designed for. Array data lets you run Horvath, Hannum, PhenoAge, GrimAge, and DunedinPACE on the same sample and compare. We would take the array for that reason alone.

One annoyance: the EPIC arrays dropped some of the CpGs in the original 450k-based clocks. Implementations impute the missing probes with reference cohort means, which adds error that is invisible in the output. If your report gives a single integer with no interval, that integer is hiding this.

From IDAT to a number

If you have IDATs, the pipeline in R is short:

library(sesame)
betas <- openSesame("idat_dir/")          # detection p-value masking, dyeBias, noob
library(dnaMethyAge)
methyAge(betas, clock = "HorvathS2013")
methyAge(betas, clock = "LevineM2018")    # PhenoAge
methyAge(betas, clock = "DunedinPACE")

Use sesame::openSesame rather than raw minfi::preprocessRaw. It applies detection p-value masking (default cutoff 0.05), dye-bias correction, and noob background subtraction, all of which the published clocks assume in some form. Then run cell-type deconvolution, because blood composition is the single largest confound:

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

If your neutrophil-to-lymphocyte ratio shifted between draws, your first-generation clock age shifted with it, and that has nothing to do with aging.

Where the number goes wrong

Technical reproducibility is the failure mode nobody advertises. Split one blood draw into two tubes, run them on the same chip, and first-generation clock estimates can differ by a couple of years. Principal-component versions of the clocks, which fit on PCs of the full array rather than on individual noisy probes, substantially improve replicate agreement and are what we run when tracking change over time.

Clonal hematopoiesis is the subtler problem. With age, hematopoietic stem cell clones carrying somatic mutations expand and come to contribute a disproportionate share of circulating leukocytes. Detectable clonal mosaicism in blood is rare before 50 and rises steeply afterward, and it is associated with subsequent hematologic cancer ⁵. A dominant clone has its own methylation profile, so part of what a blood clock reports in an older person is clonal structure rather than diffuse cellular aging. This is a real finding about your blood, and it is also a reason not to read the residual as a whole-body measurement. If a methylation report is paired with an unexplained persistent abnormality in your complete blood count, that belongs with a hematologist, not with a self-tracking spreadsheet.

Tissue is the third issue. A blood clock tells you about leukocytes. Buccal swabs sample a different cell mixture and need a buccal-calibrated model. Cross-tissue clocks trade accuracy for generality.

What whole-genome sequencing tells you about age

Germline WGS gives you almost no age information directly. What it does carry is somatic signal: mutation burden in blood accumulates with time, and deep sequencing of aging cohorts recovers both that burden and the clonal expansions described above ⁶. Telomere length can be estimated from WGS reads with tools such as TelSeq or Telomerecat, but the estimates are sensitive to read length, coverage, and library prep, and the within-person variance is large enough that a single measurement is close to uninformative. If you want an age-related readout from sequence rather than methylation, variant allele fraction tracking of somatic mutations in blood over serial draws is the more interpretable path.

How we would use the result

Treat a single methylation age as a calibration point, not a verdict. The informative quantity is the trajectory: the same clock, the same tissue, the same lab, the same processing pipeline, sampled two to four times a year, with cell composition estimated every time and regressed out before you look at the residual. Prefer DunedinPACE or a PC-based clock for tracking, because their replicate reliability is high enough that a change of meaningful size is distinguishable from assay noise. Keep the IDATs forever, because clock models keep improving and you will want to re-score old samples with new models. No methylation age is a diagnosis, and no change in one should drive a medical decision without a clinician looking at the underlying clinical data.

Questions people also ask

Can a DNA test tell your age? It can estimate it. Methylation at a few hundred CpG sites predicts chronological age with a median absolute error of roughly three to five years in blood, whether measured on arrays or by targeted sequencing of loci such as ELOVL2 and FHL2 ²⁴.

Are DNA age tests accurate? Accurate enough for forensic bracketing and population research, less so for individual tracking. The predictive error is a few years, and the technical replicate error of first-generation clocks is a substantial fraction of that, which is why repeat testing on the same clock can move your “age” without anything biological changing.

How can I test my biological age? Order an EPIC array run on whole blood, demand the raw IDAT files, process them with sesame, and score several clocks including a second-generation one and DunedinPACE. Interpret the residual only after adjusting for estimated leukocyte composition.

Can you figure out age from DNA sequence alone? Not from the germline genome. Age-related information in sequence data comes from somatic changes, including accumulated mutation burden and clonal expansions in blood, both of which rise with age ⁶⁵.

Why does my methylation age differ between two providers? Different clocks, different normalization, different probe imputation, and different tissue. Two labs running different models on the same blood draw routinely disagree by several years, which is a property of the models rather than evidence that one is wrong.

Oak builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

1. Paraskevi Salpea, Valya R. Russanova, Tazuko H. Hirai, et al. Postnatal development- and age-related changes in DNA-methylation patterns in the human genome. Nucleic Acids Research, 2012. https://doi.org/10.1093/nar/gks312 ↩

2. Jana Naue, Huub C.J. Hoefsloot, Olaf R.F. Mook, et al. Chronological age prediction based on DNA methylation: Massive parallel sequencing and random forest regression. Forensic Science International: Genetics, 2017. https://doi.org/10.1016/j.fsigen.2017.07.015 ↩ ↩²

3. Athina Vidaki, David Ballard, Anastasia Aliferi, et al. DNA methylation-based forensic age prediction using artificial neural networks and next generation sequencing. Forensic Science International: Genetics, 2017. https://doi.org/10.1016/j.fsigen.2017.02.009 ↩

4. Anastasia Aliferi, David Ballard, Matteo D. Gallidabino, et al. DNA methylation-based age prediction using massively parallel sequencing data and multiple machine learning models. Forensic Science International: Genetics, 2018. https://doi.org/10.1016/j.fsigen.2018.09.003 ↩ ↩²

5. Giulio Genovese, Anna K. Kähler, Robert E. Handsaker, et al. Clonal Hematopoiesis and Blood-Cancer Risk Inferred from Blood DNA Sequence. New England Journal of Medicine, 2014. https://doi.org/10.1056/nejmoa1409405 ↩ ↩²

6. Galina A. Erikson, Dale L. Bodian, Manuel Rueda, et al. Whole-Genome Sequencing of a Healthy Aging Cohort. Cell, 2016. https://doi.org/10.1016/j.cell.2016.03.022 ↩ ↩²