Illumina EPIC and 450K methylation arrays behind epigenetic clocks
Prepared by: Research Desk
Prepared by: Research Desk. Sources are linked for verification.
Every epigenetic clock starts as a table of numbers: one value between 0 and 1 for each CpG site measured in a DNA sample. For most published clocks, those numbers came from an Illumina Infinium BeadChip, a microarray that reads DNA methylation at hundreds of thousands of positions at once. The Horvath multi-tissue clock, PhenoAge, GrimAge and DunedinPACE were all developed on data from these arrays [1][2][3][4].
The chip is therefore not a background detail. Which generation of array was used, how the lab processed it, and which probes it contains all feed directly into the biological age figure on a report. This explainer covers the measurement layer that sits underneath the clocks described in our guide to how epigenetic clocks work.
From 27K to EPIC v2: four generations of arrays
Illumina's methylation arrays have grown in coverage over roughly fifteen years:
HumanMethylation27 (27K): about 27,000 CpG sites, concentrated near gene promoters.
HumanMethylation450 (450K): about 485,000 sites, adding gene bodies, CpG island "shores" and other regions.
MethylationEPIC v1 (EPIC, ~850K): over 850,000 sites, retaining most 450K content and adding many enhancer regions.
MethylationEPIC v2.0: close to 930,000 probes, with new content added and a substantial set of EPIC v1 probes removed [7].
Clock development tracked these generations. Horvath's 2013 estimator was built from public datasets covering 51 tissue and cell types and about 8,000 samples, and it was restricted to CpGs shared by the 27K and 450K platforms, ending with 353 clock sites [1]. Our page on the Horvath clock covers that model in detail. Later clocks such as PhenoAge (513 CpGs) were trained mainly on 450K-era data [2], and GrimAge combined DNA methylation surrogates for plasma proteins and smoking pack-years derived from cohort data [3]; see GrimAge explained.
Step one in the lab: bisulfite conversion
A microarray cannot "see" a methyl group directly. The workaround is chemical. DNA extracted from saliva or blood is exposed to sodium bisulfite, which converts unmethylated cytosines to uracil (read as thymine after amplification), while methylated cytosines are left unchanged.
After conversion, a methylation difference becomes a sequence difference: C where the site was methylated, T where it was not. The array's probes are designed to tell those two versions apart.
Conversion is never perfectly complete, and incomplete conversion can make unmethylated sites look partly methylated. Arrays include control probes that let labs check conversion efficiency, which is one reason quality-control reporting matters when evaluating any lab.
Type I and Type II probe chemistry
Infinium arrays use two probe designs on the same chip.
Type I probes use two bead types per CpG: one matches the methylated (C) version, the other the unmethylated (T) version. Both report in the same colour channel, and the ratio of the two signals indicates methylation.
Type II probes use a single bead type per CpG. A single-base extension step adds a labelled nucleotide, and the colour of the signal (red or green) indicates whether the site was methylated. Type II needs only one bead per site, which is how later arrays fit far more CpGs onto a slide. Most 450K and EPIC content is Type II.
The two designs do not behave identically. Their value distributions differ, which is why normalisation methods that adjust for probe type are standard in array preprocessing.
Beta values: the number clocks actually use
For each CpG, the array produces a methylated signal (M) and an unmethylated signal (U). The beta value is roughly M divided by (M + U), with a small constant added to stabilise low-intensity readings. A beta of 0.10 means about 10% of DNA copies in the sample were methylated at that site; 0.90 means about 90%.
A clock is essentially a weighted sum of selected beta values passed through a formula. Small shifts in many betas, whether biological or technical, can add up to a visible change in the estimated age.
Why probes go missing between generations
When Illumina redesigns an array, some probes are dropped, often because they performed poorly or overlapped common genetic variants. That creates a practical problem for clocks: a model trained on 450K data may include CpGs absent from EPIC, or present on EPIC v1 but removed in EPIC v2.
Labs handle missing clock sites in different ways, such as imputing values from reference data or using a modified model. Each choice can shift results. Peters and colleagues characterised EPIC v2 and reported high reproducibility and strong agreement with EPIC v1 for shared probes, while documenting the probes that changed between versions and discussing implications for downstream analysis tools [7].
Agreement across platforms also varies probe by probe. In matched cartilage samples run on both 450K and EPIC, Cheung and colleagues reported strong overall correlation between arrays, but many individual probes correlated weakly across platforms [6]. Kling and colleagues, validating EPIC on fresh-frozen and formalin-fixed paraffin-embedded tumour samples, reported that EPIC data were highly consistent with 450K data overall [5]. High sample-level correlation and probe-level precision are different questions, and clocks depend on specific probes.
Technical noise and clock precision
Running the same DNA sample twice on an array does not give identical betas. Individual probes, especially those with intermediate methylation, can wobble between technical replicates. Because clocks combine hundreds of probes, that noise propagates into the age estimate.
Clock developers have responded in different ways. The DunedinPACE team restricted candidate CpGs to probes with good test-retest reliability before model building, aiming for a measure that performs consistently across 450K and EPIC data [4]. Other research groups have proposed modelling approaches designed to dampen probe-level noise. Our page on biological age test accuracy discusses what replicate variability means for reading two results side by side.
What this means when evaluating a test provider
Platform details are rarely on the front of a consumer kit, but they are reasonable questions for any lab:
Which array and version? EPIC v1 and v2 differ in content, and a clock may be computed slightly differently on each.
How are missing clock CpGs handled? Imputation, modified models and published versions can diverge.
What preprocessing and normalisation is applied? Probe-type adjustment and quality filtering affect betas.
Is technical replicate data published? Repeatability figures matter more than headline correlations with age.
What is the lab's regulatory status? Most consumer epigenetic tests operate as laboratory-developed tests. FDA's 2024 rule on LDTs was vacated, and FDA formally implemented that vacatur in September 2025 [8][9]. CLIA certification, administered by CMS, addresses laboratory quality standards rather than the scientific validity of an aging claim [10]. Our page on whether biological age tests are FDA approved covers this further, and the biological age test guide covers interpreting the clocks themselves.
Further reading
Reading a methods section or a lab white paper is much easier with a grounding in microarray design, normalisation and statistical modelling. Useful textbooks tend to fall into three groups: introductory epigenetics, applied bioinformatics for array and sequencing data, and statistics for high-dimensional data.
Neutral criteria for choosing one: publication date (anything predating EPIC will not cover current arrays), whether worked code examples are included, coverage of normalisation and batch effects, and whether the author explains bisulfite chemistry rather than assuming it.
A newer edition is generally more useful than an older classic for array-specific content, because probe content and preprocessing tools have changed with each generation.
How listings and commissions work on this site is described on the about page.
Frequently asked questions
What is the Illumina EPIC methylation array?
It is a BeadChip microarray that measures DNA methylation at over 850,000 CpG sites (EPIC v1) or roughly 930,000 probes (EPIC v2) after bisulfite conversion [7]. It succeeded the 450K array and is widely used in epigenetic clock research.
Does it matter which array a lab uses?
It can. Probe content differs between 450K, EPIC v1 and EPIC v2, so some clock CpGs may be missing and need handling by imputation or a modified model [6][7]. These choices can shift an estimated age slightly.
What is a beta value?
A beta value is the estimated fraction of DNA copies methylated at a CpG site, calculated from the methylated and unmethylated signals. It runs from 0 to 1, and clocks combine weighted beta values to produce an age estimate.
Why do two runs of the same sample differ?
Probe signals carry technical noise from conversion, hybridisation and scanning, so betas vary slightly between replicates. Clocks built on reliability-filtered probes, such as DunedinPACE, were designed partly with this issue in mind [4].
Related on this site
How epigenetic clocks work: DNA methylation and age estimation
Biological age test accuracy: reliability, noise and repeat results
First, second and third generation epigenetic clocks, explained
Biological age tests: how epigenetic clocks work and what they mean
Are biological age tests FDA approved? How aging tests are regulated
Last updated: 23 October 2026.
Not medical advice. This page reports published research and regulatory records for information only.
Research-only notice: Content on this site is provided for research and educational purposes only and is not intended for human use. It is not medical advice.
Translation notice: Some content on this site may have been machine-translated using Google Translate and may contain errors.