How to Read a DNA Methylation Report
You need the raw data and not the PDF. That means either a pair of IDAT files per sample from an Illumina array (*_Grn.idat and *_Red.idat), a Bismark coverage file from bisulfite sequencing, or a bedMethyl file from nanopore or PacBio reads. You also need R 4.3 or newer with Bioconductor and roughly 16 GB of RAM for array work. If you are handling sequencing-scale data, use a machine with disk space measured in tens of gigabytes. Everything below is measurement and interpretation. Nothing here is medical advice. Any decision about vitamins, drugs, or clinical follow-up belongs with a physician.
1. Work out what your report measured
The word “methylation” labels two unrelated products. Most consumer reports that discuss MTHFR C677T, MTR A2756G, MTRR, CBS, and COMT V158M are genotype reports. They read DNA sequence variants in genes whose protein products participate in one-carbon metabolism. They do not measure a single methyl group anywhere in your body.
The other kind of report measures DNA methylation directly: the fraction of cells in a sample carrying a methyl group on the cytosine of a specific CpG dinucleotide. This is an epigenetic measurement, it is tissue-specific, and it changes over your lifetime. Reports of this kind talk about beta values and CpG probe IDs like cg16867657. They also report epigenetic age and pace of aging.
Check the methods section for the platform name. Illumina Infinium MethylationEPIC v1 (about 866,000 probes), EPIC v2 (about 935,000), or the older HumanMethylation450 array mean measured methylation. A Taqman or microarray genotyping chip, or a saliva SNP panel, means genotype. If your report names only genes and risk alleles and never names a CpG, you are holding a genotype report and the rest of this guide applies only in part.
2. Get the raw files and confirm the manifest
Ask the lab for IDATs. A vendor that will not release them is a vendor whose numbers you cannot check. For sequencing-based assays, ask for the aligned BAM with MM and ML tags (modified-base calls) or the per-CpG coverage table.
For arrays, confirm which manifest applies. EPIC v2 renamed and retired a large number of probes, so a v1 analysis pipeline pointed at v2 IDATs will silently drop or mismatch probes. In R:
library(sesame)
sesameDataCache()
sdfs <- openSesame("idat_dir/", func = NULL) # list of SigDF objects
sdfPlatform(sdfs[[1]]) # "EPICv2", "EPIC", or "HM450"
Confirm the probe count matches the platform you expect before anything else. A truncated IDAT pair, which happens when a transfer is interrupted, shows up here as a short array rather than as an error later.
3. Run QC before you read a single beta value
Three failures account for most bad methylation data, and all three are visible in minutes. The first is incomplete bisulfite conversion, where unmethylated cytosines fail to convert to uracil and therefore read as methylated. Arrays carry dedicated conversion control probes. Sequencing runs should include a spiked-in unmethylated lambda phage genome, and the conversion rate should exceed 99%.
The second is failed probes, where signal is indistinguishable from background. The third is sample mix-up, which the array’s sex-chromosome signal and its 59 SNP probes will catch.
library(sesame)
qc <- lapply(searchIDATprefixes("idat_dir/"), function(pfx)
sesameQC_calcStats(readIDATpair(pfx)))
do.call(rbind, lapply(qc, as.data.frame))[, c("frac_dt", "mean_intensity",
"InfI_switch_R2G", "RGdistort")]
Look at frac_dt, the fraction of probes failing detection by pOOBAH (p-value out-of-band array hybridization). Anything above roughly 10% is a sample I would not interpret. Then check predicted sex and the SNP fingerprint against your other data:
inferSex(sdfs[[1]])
getBetas(sdfs[[1]])[grep("^rs", names(getBetas(sdfs[[1]])))] # 59 SNP probes
The SNP probes give a genotype fingerprint you can compare across timepoints. If two samples from the same person disagree on these, the samples are not both yours.
4. Normalize, then compute beta and M values
A beta value is the proportion of methylated signal at one CpG: beta = M / (M + U + 100), where M and U are the methylated and unmethylated intensities and 100 is a stabilizing offset. Betas run from 0 to 1 and are what reports display. They are heteroscedastic, meaning variance depends on the mean, so statistical testing should use M values instead: M = log2((M + 1) / (U + 1)).
I would use sesame for array preprocessing rather than minfi, because its masking and detection calls handle EPIC v2 and low-input samples better. One call does the whole pipeline:
betas <- openSesame("idat_dir/", prep = "QCDPB", func = getBetas)
# Q: mask known problem probes C: infer Infinium-I channel
# D: nonlinear dye-bias correction P: pOOBAH detection masking
# B: noob background subtraction
dim(betas) # probes x samples, NA where masked
Masked probes come back as NA. Do not impute them with zero. For clock computation, impute with a reference cohort mean for that probe, and record how many probes you imputed.
For nanopore data, the equivalent step is a pileup. modkit turns per-read modification calls into per-site fractions:
modkit pileup sample.bam sample.bedmethyl \
--ref GRCh38.fa --cpg --combine-strands \
--filter-threshold 0.9 --threads 16
The output is bedMethyl. The columns you care about are the modification code (m for 5mC), valid coverage, and percent modified. You also want the counts of modified, canonical, and failed calls. Treat any site with valid coverage below about 10 reads as unusable, because the binomial noise on a fraction estimated from five reads swamps any effect you would want to see.
5. Estimate cell composition before interpreting anything from blood
This is the step consumer reports skip, and it is the step that decides whether a result means anything. Whole blood is a mixture of cell types with sharply different methylomes. A shift in neutrophil-to-lymphocyte ratio moves thousands of CpGs and will move most clock outputs. Age-related methylation change is itself partly tissue-specific and partly shared across tissues, which is why the mixture matters so much 1.
Deconvolution recovers the proportions from the methylation data itself:
library(FlowSorted.Blood.EPIC)
library(minfi)
rgSet <- read.metharray.exp("idat_dir/")
counts <- estimateCellCounts2(rgSet, compositeCellType = "Blood",
processMethod = "preprocessNoob",
referencePlatform = "IlluminaHumanMethylationEPIC",
cellTypes = c("CD8T","CD4T","NK","Bcell","Mono","Neu"))
round(counts$counts, 3)
Compare these fractions against a same-day complete blood count if you have one. They should agree within a few percentage points. When you track a CpG or a clock across timepoints, regress out the cell fractions or at least plot them alongside, so you can see whether a jump in epigenetic age coincides with a jump in neutrophils.
6. Read the epigenetic clock section
Epigenetic clocks are regression models trained to predict chronological age or a mortality-linked phenotype from a few hundred CpG betas. Horvath’s pan-tissue clock uses 353 CpGs and achieves a median absolute error near 3.6 years across many tissues. The broader framework behind these models, including why they behave as biomarkers of biological rather than calendar time, is laid out by Horvath and Raj 2.
Compute clocks yourself rather than trusting a vendor’s implementation, because coefficient sets and normalization choices differ:
library(dnaMethyAge)
methyAge(betas, clock = "HorvathS2013")
methyAge(betas, clock = "LevineS2018") # PhenoAge, 513 CpGs
methyAge(betas, clock = "DunedinPACE") # rate, not an age
Three points govern interpretation. First, age acceleration is a residual. It is your predicted age minus your chronological age, adjusted against a reference population. Without that reference distribution the number has no scale, so “+4.2 years” means little unless you know the cohort standard deviation.
Second, clocks are tissue-dependent. A clock trained on blood will not read a buccal or tissue sample correctly, and tissue-specific clocks outperform general ones in their target tissue 3. Clocks also behave differently at the extremes of the lifespan, and performance in infancy and childhood is not the same as in adults 4.
Third, most clocks are opaque by construction. Newer work maps clock CpGs onto aging pathways to make the output interpretable rather than a bare number 5. Until your report does that, treat the age value as a summary statistic whose components you cannot inspect.
Absolute values also depend on the population used for training. Models built and validated on Japanese cohorts, for instance, were developed specifically because reference distributions differ across populations and array types 6. If your ancestry is far from the training cohort, expect a systematic offset rather than a clean read.
7. Check the specific loci a report makes claims about
When a report highlights individual CpGs, go look at them. Pull the probe by ID, check its detection p-value, and check whether nearby probes agree.
cpg <- "cg16867657" # ELOVL2 promoter, a well-known age-associated site
betas[cpg, ]
Two categories of locus-level claim are worth your attention. Imprinted loci, where methylation depends on which parent contributed the allele, change with age in ways that only phased long-read data can resolve properly 7. Global hypomethylation of repetitive elements such as LINE-1 and Alu is a real age-associated phenomenon, and it is measured by assays quite different from a standard array 8.
Exposure claims deserve more skepticism. Epigenome-wide association studies of environmental exposures do find differentially methylated sites, but the per-CpG effect sizes are small and were detected in cohorts of thousands 9. A single individual’s beta value at one such CpG carries almost no information about that individual’s exposure.
8. Set up the longitudinal comparison
One methylation profile is a coordinate. Two or more make a trajectory, and the trajectory is where the value sits. Store betas as a matrix with probes as rows and timepoints as columns, keep the QC statistics and cell fractions alongside, and never mix preprocessing versions inside one comparison.
stopifnot(identical(rownames(t1), rownames(t2)))
delta <- t2[, 1] - t1[, 1]
hist(delta, breaks = 200) # centered near zero if processing matched
sd(delta) # your technical noise floor
That standard deviation is the number to beat. Any single-CpG change smaller than it is noise. Individual probes have poor test-retest reliability, while multi-CpG composites like clocks are far more stable, which is another reason to read the composite and not the component. Run replicate samples from the same blood draw at least once so the noise floor is measured rather than assumed.
Common problems
Beta values clustering at 0.5 across many probes usually indicates a cross-reactive or SNP-affected probe rather than genuine hemimethylation. Mask the published problem-probe lists, which sesame’s Q prep code applies by default.
An epigenetic age that jumps several years between consecutive draws is nearly always technical. Check in order: array batch, cell composition, detection failure rate, and probe imputation count. Real biological change on that scale over months is not what the assays show.
Missing clock CpGs are common on EPIC relative to the 450k arrays the older clocks were trained on. Every implementation imputes them, and the imputation method changes the answer by a fraction of a year. Report the count of imputed probes next to the age estimate.
Saliva and buccal samples contain a variable mix of epithelial cells and leukocytes. Applying a blood clock to them produces numbers that look plausible and are not. Use a tissue-appropriate model or do not report an age at all.
Finally, a genotype-based “methylation panel” cannot tell you whether your methylation is working. MTHFR C677T homozygosity is common and its downstream effects vary widely between people. Questions about folate status, homocysteine, or B-vitamin intake are clinical questions. They need blood biomarkers and a physician. A gene report cannot answer them.
Footnotes
-
Kenneth Day, Lindsay L Waite, Anna Thalacker-Mercer, et al. Differential DNA methylation with age displays both common and dynamic features across human tissues that are influenced by CpG landscape. Genome Biology, 2013. https://doi.org/10.1186/gb-2013-14-9-r102 ↩
-
Steve Horvath, Kenneth Raj. DNA methylation-based biomarkers and the epigenetic clock theory of ageing. Nature Reviews Genetics, 2018. https://doi.org/10.1038/s41576-018-0004-3 ↩
-
Leonardo D. Garma, Sonia Pernas, Bartomeu Fullana, et al. A breast tissue-specific epigenetic clock provides accurate chronological age predictions and reveals de-correlation of age and DNA methylation in tumor-adjacent and tumor samples. Epigenetics, 2026. https://doi.org/10.1080/15592294.2026.2714582 ↩
-
Anne K. Bozack, Sheryl L. Rifas-Shiman, Diane R. Gold, et al. DNA methylation age at birth and childhood: performance of epigenetic clocks and characteristics associated with epigenetic age acceleration in the Project Viva cohort. Clinical Epigenetics, 2023. https://doi.org/10.1186/s13148-023-01480-2 ↩
-
Rajarshi Mandal, Ning Xie, Gil Alterovitz. Interpretable epigenetic clock links aging pathways and disease-specific methylation profiles. npj Aging, 2026. https://doi.org/10.1038/s41514-026-00476-5 ↩
-
Tatsuma Shoji, Yui Tomo, Ryo Nakaki. Prediction of biological age and blood biomarkers from DNA methylation profiles measured by the methylation screening array: development and validation of models on Japanese data. Epigenetics Reports, 2026. https://doi.org/10.1080/28361512.2026.2715974 ↩
-
Brynja Sigurpalsdottir, Guillaume Holley, Sverrir Þ. Sverrisson, et al. Nanopore sequencing identifies parent-of-origin specific age-associated methylation changes at imprinted loci in the human genome. Nature Communications, 2026. https://doi.org/10.1038/s41467-026-76240-w ↩
-
Pornrutsami Jintaridth, Apiwat Mutirangura. Distinctive patterns of age-dependent hypomethylation in interspersed repetitive sequences. Physiological Genomics, 2010. https://doi.org/10.1152/physiolgenomics.00146.2009 ↩
-
Thanh T. Hoang, Cancan Qi, Kimberly C. Paul, et al. Epigenome-Wide DNA Methylation and Pesticide Use in the Agricultural Lung Health Study. Environmental Health Perspectives, 2021. https://doi.org/10.1289/ehp8928 ↩
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.