What a Methylation Analysis Test Measures, and What It Can Tell You

Oak
methylation epigenetics dna bioinformatics biomarkers
Studio-lit feathered dinosaur-like creature in profile against black, its feathers dotted with small beads that are either glowing pale or dark.

A methylation analysis test measures whether a cytosine followed by a guanine (a CpG site) in your DNA carries a methyl group, and at what fraction across the millions of cells in your sample. The output is a per-site value between 0 and 1 called a beta value, the fraction of DNA molecules methylated at that position. From that matrix you can estimate cell-type composition of the sample, compute epigenetic age predictors, and look at specific loci with known age or exposure associations. What it does not do is tell you how well your one-carbon metabolism is working, which is what most consumer products sold as “methylation tests” imply.

Two different products share the name

The first thing to sort out is which kind of test you are looking at, because they measure unrelated things.

A genotype-based “methylation panel” reports variants in genes such as MTHFR, MTR, MTRR, and COMT. These are single-nucleotide polymorphisms read from a genotyping array or from whole-genome sequencing. They tell you about the amino acid sequence of enzymes in folate and homocysteine metabolism. They measure zero methyl groups. Whatever you conclude from MTHFR C677T, you concluded from DNA sequence, not from DNA methylation.

A true methylation test measures the epigenetic mark itself. That requires a chemistry that distinguishes 5-methylcytosine from unmodified cytosine, because standard sequencing and genotyping read both as C. Sodium bisulfite treatment deaminates unmethylated cytosine to uracil (read as T) and leaves 5mC intact, converting an epigenetic difference into a sequence difference. Enzymatic conversion (EM-seq) achieves the same result with less DNA damage. Nanopore sequencing skips conversion entirely and calls the modification from the raw current signal.

Arrays, bisulfite sequencing, and nanopore

The human genome contains roughly 28 million CpG sites. No practical assay for an individual covers all of them at useful depth, so every method is a choice about which sites and how precisely.

Illumina methylation arrays are the workhorse. The EPIC v1 array interrogates about 866,000 CpGs and EPIC v2 about 935,000, selected for promoters, enhancers, CpG islands, and sites used by published models. Each probe reports an averaged methylation fraction over the input DNA, which makes the per-site measurement precise and the technical noise small relative to sequencing at moderate depth. Nearly every epigenetic clock in the literature was trained on 450K or EPIC data, so arrays give you direct compatibility with published coefficient sets.

Whole-genome bisulfite sequencing gives you every CpG but at a cost. You need roughly 30x post-conversion coverage for stable per-site estimates, and even then each site is a binomial draw: at 20 reads, the standard error on a beta value near 0.5 is about 0.11. Reduced-representation bisulfite sequencing (RRBS) concentrates coverage on a few million CpG-dense regions for far less money, at the price of a biased, island-heavy footprint that overlaps array content only partially.

Nanopore sequencing calls 5mC and 5hmC natively. With Dorado you basecall using a modified-base model (dorado basecaller sup,5mCG_5hmCG), then aggregate with modkit pileup into a bedMethyl file of per-strand counts. This preserves phasing, so you can read allele-specific methylation and imprinted loci directly, which no array can do. The tradeoff is per-site noise and the fact that clock coefficients trained on array beta values do not transfer cleanly to nanopore beta values without recalibration.

Our recommendation for a personal baseline is an EPIC array as the primary assay, with nanopore or WGBS added only if you have a specific question about imprinting, allele-specific methylation, or a region the array does not cover.

The pipeline from IDAT to beta values

Arrays arrive as paired IDAT files, one per color channel per sample. We would process them in R with sesame rather than the older minfi defaults, for one specific reason: sesame’s pOOBAH detection-p-value method uses out-of-band signal from Infinium I probes to build a proper background null, which masks failing probes that minfi’s detectionP lets through. openSesame(idat_dir, prep = "QCDPB") runs the full chain: probe quality masking, dye bias correction, noob background subtraction, and beta extraction.

Before trusting anything downstream, run sample-level QC. Check bisulfite conversion control probes and reject any sample below about 90 percent inferred conversion. Confirm predicted sex against reported sex from X and Y probe intensities. Use the 59 rs genotyping probes on the EPIC array as a fingerprint to verify that a follow-up sample is the same person as the baseline, which catches the single most expensive error in longitudinal work, a swapped tube.

Then deconvolve cell composition. Blood methylation is dominated by which leukocytes are in the tube, and a shift in neutrophil-to-lymphocyte ratio will move hundreds of thousands of probes. The Houseman reference-based method with the FlowSorted.Blood.EPIC reference gives you estimated fractions for CD4T, CD8T, NK, B cells, monocytes, and granulocytes. Report those fractions alongside any result, and include them as covariates in any comparison across timepoints. For site-level statistics, convert to M-values (log2(beta / (1 - beta))), which are roughly homoscedastic, and report in beta units, which are interpretable.

Which clocks to compute

Epigenetic clocks are linear models over selected CpG beta values fitted to predict chronological age or a mortality-linked phenotype. The mechanics are straightforward: the Bioconductor package methylclock takes a beta matrix and returns Horvath, Hannum, PhenoAge, skin-and-blood, and several pediatric clocks in one call, handling missing-probe imputation explicitly 1. Reviews of the field lay out the model families and the training targets clearly, which is worth reading before you interpret a number 2.

The interpretation is where people go wrong. First-generation clocks such as Horvath were trained to predict chronological age across many tissues, so a well-calibrated result means the model works, not that you are healthy. Second-generation clocks trained against mortality and clinical chemistry carry more signal about outcomes. More recent work decomposes the blood methylome into system-specific scores, estimating separate aging rates for eleven physiological systems from a single test, which is more informative than one scalar 3. Clock acceleration is genuinely responsive to biological state: HIV-1 infection measurably accelerates the Horvath estimate in brain and blood tissue 4. Mechanistically, the field still debates whether clock CpGs are causal drivers of aging or downstream readouts of stochastic drift and cell-composition change, and that ambiguity constrains what any single measurement means 5 6.

Practically: treat one clock reading as a noisy measurement with a standard error of a couple of years, not a verdict. Serial measurements on the same array version, same lab, same tissue, and same season are the only ones worth comparing, and a small self-experiment reporting clock change over months should be read with the same skepticism you would apply to any uncontrolled n-of-1 trial 7. If you want a cheap, high-signal single locus, ELOVL2 promoter methylation tracks chronological age tightly enough to be used in forensic age estimation from a targeted pyrosequencing or amplicon assay 8.

Failure modes worth knowing

Tissue matters more than almost anything else. Blood, buccal, and saliva methylomes differ substantially, and a clock trained on blood will give a biased number on a buccal swab unless it was recalibrated for that tissue 9. Saliva is a variable mix of leukocytes and epithelial cells, and that mixing ratio changes with collection technique, so it is the worst choice for longitudinal comparison.

EPIC v2 dropped and renamed a fraction of v1 probes and introduced duplicate probes for some sites, so clock models referencing v1 IDs need explicit remapping, and imputing missing clock CpGs from a reference mean quietly shrinks your result toward the population average. Batch effects from array position and scan date are large, and a longitudinal design where baseline and follow-up sit on different chips confounds change with batch.

What this cannot tell you

A methylation profile is a measurement, not a diagnosis. Tumor-specific methylation markers, imprinting disorders, and clinically actionable epigenetic findings require validated diagnostic assays run in an accredited laboratory and interpretation by a physician or a clinical geneticist. If a research-grade result suggests something medically significant, the correct next step is a clinician and a confirmatory clinical test, not a change to anything you take or do.

Questions people also ask

What does methylation testing tell you? It tells you the fraction of DNA molecules methylated at each measured CpG in a specific tissue at a specific moment. From that you can estimate the cell-type makeup of the sample, compute epigenetic age and pace-of-aging predictors, and inspect individual loci with established age or exposure associations.

Is a MTHFR test a methylation test? No. It genotypes a variant in a folate-pathway enzyme. It reads DNA sequence and reports nothing about methyl marks on your genome.

How accurate is epigenetic age? Multi-tissue clocks predict chronological age in blood with a median absolute error of a few years, and the deviation from chronological age is the quantity people care about, so the test-retest reliability of the assay matters as much as the training accuracy 2.

Can I compute clocks from nanopore or WGBS data instead of an array? Technically yes, by extracting the clock CpGs from a bedMethyl or .cov file, but the coefficients were fit on array beta values with their own probe-specific biases, and sequencing beta values at typical coverage are noisier, so the result is not directly comparable to published distributions.

Does a methylation clock work in tissues other than blood? Tissue-specific clocks exist and perform well within the tissue they were trained on, including vascular tissue in animal models 10. Cross-tissue application of a blood-trained model produces systematic offsets.

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. Dolors Pelegí-Sisó, Paula de Prado, Justiina Ronkainen, et al. methylclock : a Bioconductor package to estimate DNA methylation age. Bioinformatics, 2020. https://doi.org/10.1093/bioinformatics/btaa825 ↩

  2. Daniel J. Simpson, Tamir Chandra. Epigenetic age prediction. Aging Cell, 2021. https://doi.org/10.1111/acel.13452 ↩ ↩2

  3. Raghav Sehgal, Yaroslav Markov, Chenxi Qin, et al. Systems Age: a single blood methylation test to quantify aging heterogeneity across 11 physiological systems. Nature Aging, 2025. https://doi.org/10.1038/s43587-025-00958-3 ↩

  4. Steve Horvath, Andrew J. Levine. HIV-1 Infection Accelerates Age According to the Epigenetic Clock. Journal of Infectious Diseases, 2015. https://doi.org/10.1093/infdis/jiv277 ↩

  5. Adam E. Field, Neil A. Robertson, Tina Wang, et al. DNA Methylation Clocks in Aging: Categories, Causes, and Consequences. Molecular Cell, 2018. https://doi.org/10.1016/j.molcel.2018.08.008 ↩

  6. 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 ↩

  7. Josh Mitteldorf. A Clinical Trial Using Methylation Age to Evaluate Current Antiaging Practices. Rejuvenation Research, 2019. https://doi.org/10.1089/rej.2018.2083 ↩

  8. Renata Zbieć-Piekarska, Magdalena Spólnicka, Tomasz Kupiec, et al. Examination of DNA methylation status of the ELOVL2 marker may be useful for human age prediction in forensic science. Forensic Science International: Genetics, 2015. https://doi.org/10.1016/j.fsigen.2014.10.002 ↩

  9. Zimeng Guan, Jiaqi Wang, Zidong Liu, et al. Epigenetic Age Estimation by Detecting DNA Methylation Status in Buccal Swabs. ELECTROPHORESIS, 2024. https://doi.org/10.1002/elps.202400075 ↩

  10. Ewelina Pośpiech, Anna Bar, Aleksandra Pisarek-Pacek, et al. Epigenetic clock in the aorta and age-related endothelial dysfunction in mice. GeroScience, 2024. https://doi.org/10.1007/s11357-024-01086-3 ↩