Which Epigenetics Test Is Worth Buying, and What to Do With the Output

Woolf Software
epigenetics dna-methylation biological-age bioinformatics longitudinal
A giant fallen tree's cut face glows with ringed bands studded by pale nodules, in a misty valley with distant grazing animals.

The most useful thing an epigenetics test can give you is your raw data. Choose whichever vendor will hand you the IDAT files from an Illumina EPIC array and tell you which array version they ran. Everything downstream of that is math you can reproduce yourself in about twenty minutes of R. That includes the branded biological age, the “pace of aging,” and the nine organ system scores. If a vendor will not release the IDATs, buy from one who will. The number printed on the PDF is a single point estimate carrying technical noise large enough to swamp a year of behavior change, and you cannot see that noise unless you hold the raw intensities.

What you are buying

It helps to be precise about what the assay measures and what physically arrives when the lab is finished. A methylation array reports the fraction of DNA molecules methylated at each of several hundred thousand CpG sites, the cytosine-guanine positions where methylation marks are placed. Illumina’s EPIC v1 covers about 866,000 CpGs. EPIC v2 covers roughly 935,000 and retired some legacy probes. That retirement matters, because several published clocks were trained on the older 450k platform, and a handful of their input CpGs are now missing and must be imputed from a training-set mean. Most vendor reports perform that imputation silently.

What you receive per sample is two IDAT files in Illumina’s raw intensity data format. One covers the green channel and one the red, at a few megabytes each:

206891110045_R01C01_Grn.idat
206891110045_R01C01_Red.idat

Alongside them comes a sample sheet listing the sentrix ID and position, which identify the physical slide and the well your sample occupied. That is the whole payload. The epigenetic age itself is a linear model applied to transformed beta values, and the models are published openly.

Sequencing is the alternative, and it answers a different question. Whole-genome bisulfite sequencing or enzymatic methyl-seq (EM-seq) gives you all of the roughly 28 million CpGs rather than a curated one percent, but at 10-30x coverage it costs several times an array. Every published clock was trained on array betas, so you end up extracting array-equivalent positions anyway and accepting coverage-driven noise at individual CpGs. For age clocks specifically, we would run the array. For discovery work in regions no array covers, sequence.

The clocks worth computing

Run several clocks rather than one. They disagree with each other, and that disagreement is itself informative. Each was trained on different tissues, different sample sizes, and different targets. The following are the ones we would compute, with the reason each earns a place.

  • Horvath 2013, 353 CpGs, trained across 51 tissues. Best used as a tissue-agnostic sanity check rather than a health readout.
  • Hannum 2013, 71 CpGs, whole blood only.
  • Skin and blood clock, 391 CpGs, built for fibroblasts and keratinocytes as well as buccal cells and blood. It was validated on progeria samples and ex vivo culture 1. This is the one to use if your sample is a cheek swab.
  • PhenoAge, 513 CpGs, trained not on chronological age but on a phenotypic age derived from nine clinical chemistry measures plus chronological age. It outperforms chronological-age clocks for mortality and healthspan association 2.
  • Zhang 2019 elastic net, 514 CpGs, trained on roughly 13,000 samples. It comes with a companion BLUP predictor using over 300,000 CpGs. The headline finding of that paper is precision: larger training sets produce predictors whose test-retest variation is much smaller, which is exactly the property longitudinal tracking requires 3.

If you are tracking yourself over years, give the most weight to the Zhang predictor and PhenoAge. The first-generation chronological-age clocks were optimized to predict a number you already know exactly, which limits how much they can tell you about your health.

A pipeline that runs in an afternoon

Processing your own IDATs is well within reach if you are comfortable in R, and the steps below take a single sitting. Begin with normalization and quality control using the sesame package.

library(sesame)
sesameDataCache()

# EPIC v2 IDATs in ./idats, one pair per sample
betas <- openSesame("idats", prep = "QCDPB", func = getBetas)
# Q = mask nonunique/low-quality probes, C = infer channel,
# D = dye bias correction, P = pOOBAH detection p-value masking,
# B = noob background subtraction

The pOOBAH step masks probes whose signal cannot be distinguished from out-of-band background, meaning the measurement carries no real information. On a good sample, expect a small percentage of probes to be masked. If more than a few percent fail, the DNA input or the bisulfite conversion was poor, and you should request a re-run before interpreting anything.

The next step is estimating cell composition, which you need in hand before computing any clock, for reasons the following section explains.

library(FlowSorted.Blood.EPIC)
counts <- estimateCellCounts2(rgSet, referencePlatform = "IlluminaHumanMethylationEPIC",
                              cellTypes = c("CD8T","CD4T","NK","Bcell","Mono","Neu"))

With clean betas and cell estimates available, the clocks themselves are a few lines.

library(dnaMethyAge)
methyAge(betas, clock = "HorvathS2013")
methyAge(betas, clock = "LevineM2018")   # PhenoAge
methyAge(betas, clock = "ZhangQ2019")

Finally, compare the vendor’s reported age against the one you recomputed. If the two differ by more than a few tenths of a year on the same clock, the vendor is applying an undisclosed calibration or normalization step, and you should know that before you begin trending the number over time.

Where the number goes wrong

Several distinct sources of error can move an epigenetic age estimate without any durable change in your biology, and knowing them is what separates a usable measurement from a decorative one.

The first is cell composition. A blood methylation profile is a weighted average over leukocyte subsets whose methylomes differ considerably. A shift in the neutrophil-to-lymphocyte ratio caused by an infection two days before the draw will move your estimated age even though nothing about your underlying biology has changed durably. Formal assessments of clocks as aging biomarkers treat cell-type adjustment as a baseline requirement rather than an option 4. Buccal swabs present the same problem in a different form, since a swab collects a mixture of epithelial cells and infiltrating leukocytes, and adding cell-type-specific signatures to the model measurably improves buccal age prediction 5.

Tissue is the second source. Prediction error differs by tissue, and a clock trained on blood and applied to saliva is a different measurement rather than a slightly noisier version of the same one. Use the skin and blood clock for buccal and swab samples 1.

Technical noise is the third, and in consumer testing it is the dominant failure mode. Two aliquots of the same blood draw will return different ages when run on different chips. The reported year-over-year change is often smaller than that replicate spread. Larger training sets reduce the problem 3, and principal-component reformulations of the classic clocks reduce it substantially as well. If you can afford it, buy two samples per timepoint and average them. At minimum, ask the vendor to place your longitudinal samples on the same chip, since slide and position are real batch effects.

Population and ancestry form a fourth consideration. Age-associated CpGs and the magnitude of age acceleration are not identical across ancestry groups, and studies in middle-aged African American and white participants have found both novel age-associated sites and group differences in acceleration 6. Most clocks were trained on predominantly European-ancestry cohorts, so interpret your own result with that in mind.

Last comes the interpretation of the residual itself. Epigenetic age acceleration is a residual from a regression, and its mechanistic meaning remains unsettled. The field has not resolved whether methylation drift causes aging, passively reads out cell division and immune composition, or reflects some mixture of both 7. Reviews of the prediction methods are direct on this point, describing the clocks as well-validated predictors first and mechanistic claims second 8. Counterintuitive results are common as a result. Elite athletes, for instance, showed modified epigenetic aging patterns that did not reduce to a simple “younger” readout across all markers 9.

What a result can and cannot tell you

Having covered the failure modes, it is worth stating plainly what remains once they are accounted for. A result can tell you where you sit in a distribution, and with repeated sampling and cell-type adjustment it can tell you whether your residual is trending in one direction. That is a real measurement, with real associations to mortality and healthspan in large cohorts 2.

What it cannot tell you is why. No epigenetic age result identifies a condition, and none of it should be read as a reason to start or stop anything medical. If a result prompts a concern, that conversation belongs with a physician who can order actual diagnostics. The same caution applies to organ-specific age panels, where the organ labels derive from CpGs correlated with organ-related traits in cohort data and do not constitute an assessment of any organ in you.

Our recommendation follows from all of the above. Buy an EPIC-based test that releases IDATs and run sesame and dnaMethyAge yourself. Adjust for cell counts, sample twice per timepoint, and treat any single number as provisional until you have three or four timepoints. Forensic and research work has converged on the same lesson, that carefully selected marker sets with explicit error characterization beat a black-box score 10.

Questions people also ask

Are epigenetic tests worth it? They are worth it if you will process the raw data and sample repeatedly. A one-off consumer report gives you a number with an uncharacterized error bar. Its most common use is comparing your “biological age” to your chronological age a single time, which is the least informative thing you can do with it.

Is epigenetic testing covered by insurance? Generally no. Methylation age tests are laboratory-developed tests sold direct to consumers rather than FDA-cleared diagnostics, and there is no established billing pathway for them. Clinically ordered methylation testing for specific imprinting or tumor-classification questions falls into a separate category and can be covered.

How expensive is epigenetic testing? Consumer array tests run roughly $200 to $500. Running an EPIC array through a research core, if you can send your own DNA, is typically in the low hundreds per sample. EM-seq or WGBS at usable coverage costs several times that.

Why is epigenetics controversial? There are two reasons. The measurement is noisy and heavily influenced by cell composition and batch effects, so reported changes are often technical in origin 4. And the causal story is unsettled, which means marketing language about reversing your biological age runs far ahead of what the models support 7.

Woolf Software 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. Steve Horvath, Junko Oshima, George M. Martin, et al. Epigenetic clock for skin and blood cells applied to Hutchinson Gilford Progeria Syndrome and ex vivo studies. Aging, 2018. https://doi.org/10.18632/aging.101508 ↩ ↩2

  2. Morgan E. Levine, Ake T. Lu, Austin Quach, et al. An epigenetic biomarker of aging for lifespan and healthspan. Aging, 2018. https://doi.org/10.18632/aging.101414 ↩ ↩2

  3. Qian Zhang, Costanza L. Vallerga, Rosie M. Walker, et al. Improved precision of epigenetic clock estimates across tissues and its implication for biological ageing. Genome Medicine, 2019. https://doi.org/10.1186/s13073-019-0667-1 ↩ ↩2

  4. Morgan E Levine. Assessment of Epigenetic Clocks as Biomarkers of Aging in Basic and Population Research. The Journals of Gerontology: Series A, 2020. https://doi.org/10.1093/gerona/glaa021 ↩ ↩2

  5. Monika Eipel, Felix Mayer, Tanja Arent, et al. Epigenetic age predictions based on buccal swabs are more precise in combination with cell type-specific DNA methylation signatures. Aging, 2016. https://doi.org/10.18632/aging.100972 ↩

  6. Salman M. Tajuddin, Dena G. Hernandez, Brian H. Chen, et al. Novel age-associated DNA methylation changes and epigenetic age acceleration in middle-aged African Americans and whites. Clinical Epigenetics, 2019. https://doi.org/10.1186/s13148-019-0722-1 ↩

  7. Adam Li, Zane Koch, Trey Ideker. Epigenetic aging: Biological age prediction and informing a mechanistic theory of aging. Journal of Internal Medicine, 2022. https://doi.org/10.1111/joim.13533 ↩ ↩2

  8. Daniel J. Simpson, Tamir Chandra. Epigenetic age prediction. Aging Cell, 2021. https://doi.org/10.1111/acel.13452 ↩

  9. Magdalena Spólnicka, Ewelina Pośpiech, Jakub Grzegorz Adamczyk, et al. Modified aging of elite athletes revealed by analysis of epigenetic age markers. Aging, 2018. https://doi.org/10.18632/aging.101385 ↩

  10. Alberto Montesanto, Patrizia D’Aquila, Vincenzo Lagani, et al. A New Robust Epigenetic Model for Forensic Age Prediction. Journal of Forensic Sciences, 2020. https://doi.org/10.1111/1556-4029.14460 ↩