Epigenetic Clock

An epigenetic clock is a trained predictor, usually elastic net regression, that maps DNA methylation fractions at a fixed panel of CpG sites onto a number reported in years. The first-generation clocks (Horvath’s 353-CpG multi-tissue clock, Hannum’s 71-CpG blood clock) were trained on chronological age. Later ones (PhenoAge, GrimAge, DunedinPACE) were trained on clinical biomarkers, mortality, or rate of physiological decline, which makes them different instruments that happen to share an input matrix.

How it works

Each CpG site gets a beta value between 0 and 1: the fraction of DNA molecules at that position in that sample carrying a methyl group on cytosine. Across the genome, thousands of sites drift with age in a reproducible direction, some gaining methylation (often CpG island promoters of developmental genes), some losing it (often intergenic and repetitive regions). The drift is continuous from birth onward rather than starting at some threshold age, and it is detectable in cohorts spanning the whole lifespan ¹. A clock is just a sparse linear combination of a few hundred of these sites, with coefficients fit by penalized regression, plus a transformation so that predictions behave sensibly in children.

The accuracy question has a clean answer. Horvath’s multi-tissue clock reports a median absolute error around 3.6 years. When the training set gets large enough, the error collapses: Zhang and colleagues trained on over 13,000 samples and pushed blood prediction error to roughly two years, and they argued that this precision paradoxically limits the clock’s value as an ageing biomarker, because a model that predicts chronological age almost perfectly has little residual left to correlate with health ². The interesting quantity was never the predicted age. It is the residual, age acceleration, the difference between predicted and chronological age after regressing out chronological age. That residual associates with mortality independently of telomere length ³ and shifts with environmental exposures such as rotating night shift work ⁴.

In your own data

The input is not whole-genome sequencing. Standard short-read WGS strips methylation, because PCR does not copy the methyl group. You need one of: an Illumina methylation array (450k, EPIC v1 at ~865k probes, EPIC v2), enzymatic or bisulfite conversion sequencing, or nanopore, where the basecaller calls 5mC directly from the raw signal.

For arrays you get a pair of .idat files per sample (Grn and Red). Read them with minfi::read.metharray.exp() or SeSAMe. Normalize with preprocessNoob() or preprocessFunnorm(), then extract betas with getBeta(). Feed the matrix to the methylclock R package, which implements Horvath, Hannum, PhenoAge, and several others and will tell you how many required CpGs are missing.

For nanopore, Dorado with a 5mCG model plus modkit pileup gives you a bedMethyl file: chromosome, start, end, modified base, coverage, percent modified. Lift the clock CpG coordinates to your reference build before intersecting. Coverage matters. At 5x, a single CpG’s methylation fraction is quantized to multiples of 20 percent, and clock coefficients can be large. Nanopore also reads phase, so you can separate maternal and paternal alleles, and age-associated methylation change at imprinted loci is parent-of-origin specific in a way arrays cannot see ⁵.

Mistakes we see repeatedly. Missing probes: EPIC v2 dropped and renamed content, so some Horvath CpGs need imputation, and imputing to the training-set mean quietly biases the estimate. Cell composition: whole blood methylation is dominated by the neutrophil-to-lymphocyte ratio, so run Houseman deconvolution and include the estimated fractions as covariates, or you will measure your differential count. Tissue mismatch: a saliva sample scored with a blood-trained clock is not comparable to a blood sample. Batch: two samples run on different slides can differ by a year or more of apparent age from technical effects alone, so if you want longitudinal comparison, bank aliquots and run them together.

Limitations

The clocks are correlational. Methylation drift may be a readout of ageing processes rather than a cause of them, and the field has not resolved which ^{6 7}. Reported age acceleration is sensitive to the normalization pipeline, the reference panel, and which clock you pick, and different clocks disagree on the same sample ⁸. Deep learning clocks improve pan-tissue accuracy but do not solve the interpretation problem ⁹. Age-related methylation changes vary across tissues and between individuals, and a single timepoint tells you very little ¹⁰. Treat one number as a noisy measurement. Two or three annual timepoints on the same platform, run in the same batch, are worth more than any single score. None of this is a clinical test, and any decision about your health belongs with a physician.

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.