Principal-component epigenetic clocks and measurement reliability
Prepared by: Research Desk
Prepared by: Research Desk. Sources are linked for verification.
An epigenetic clock turns DNA methylation measurements into a single age estimate. If the same person tests twice and the number moves, the first question is whether that movement reflects biology or measurement. A 2022 study in Nature Aging reported that technical noise alone produced deviations of up to 9 years between replicates for six prominent epigenetic clocks [1]. The same study proposed a computational fix: principal-component (PC) clocks. This explainer covers where the noise comes from, how the PC approach works, what the reliability studies reported, and what remains unresolved.
Where the noise in CpG clocks comes from
Widely used clocks, including the Horvath multi-tissue estimator [2], PhenoAge [3], GrimAge [4] and DunedinPACE [5], are calculated from methylation values at selected CpG sites. In the original versions, each estimate is essentially a weighted combination of individual sites. Our explainer on how epigenetic clocks work walks through that construction.
The weak point is the individual site. Higgins-Chen and colleagues describe DNA methylation data as "surprisingly unreliable" when it comes to clock calculation [1]. They measured this with technical replicates: the same biological sample measured more than once. Because the biology is identical, any difference between replicate results is, by definition, measurement noise. When a clock leans on a limited set of CpGs, the error at those sites carries straight through to the final age estimate.
For background on the arrays that generate these CpG values, see Illumina EPIC and 450K methylation arrays behind epigenetic clocks.
How principal-component clocks work
The PC approach adds one step before age prediction. Instead of feeding individual CpG values into the model, researchers first run principal components analysis on CpG-level data, then use the resulting components as inputs for biological age prediction [1].
A principal component is a summary axis that captures a pattern of variation shared across many CpG sites at once. The reasoning is that random error at any single site matters less when that site is only one contributor among many to a shared signal. The authors report that the method:
entails only one additional step compared with traditional clocks [1];
requires no replicates or prior knowledge of CpG reliabilities for training [1];
can be applied to any existing or future epigenetic biomarker [1].
Using this approach, the team retrained principal-component versions of six clocks [1].
The method has also been used outside blood. Thrush and colleagues built PCBrainAge, an age predictor trained solely on cortical samples that combines principal components analysis with regularized regression. They report that this design reduces technical noise and "greatly improves test-retest reliability" [7]. The team generated DNA methylation data from multiple brain regions in a sample from the Religious Orders Study and Rush Memory and Aging Project [7].
What the reliability studies reported
Study | What was measured | Reported finding |
|---|---|---|
Higgins-Chen et al., 2022 [1] | Replicate agreement for six prominent clocks and their PC versions | Original clocks: deviations up to 9 years. PC versions: most replicates agree within 1.5 years |
Thrush et al., 2022 [7] | Brain-specific PC clock in cortical tissue | PCA plus regularized regression reduced technical noise and improved test-retest reliability |
Borrus et al., 2024 (preprint) [6] | Variability across 6 popular clocks in longitudinal datasets | Clocks trained on mortality or rate of aging captured aging events more reliably than clocks trained on chronological age |
Markov et al., 2024 [8] | Replicate agreement for stochastic epigenetic mutations | 24 to 39% of hypoSEMs and 46 to 67% of hyperSEMs not shared between replicates |
Beyond replicate agreement, Higgins-Chen and colleagues reported that the PC versions improved detection of clock associations and intervention effects, and produced reliable longitudinal trajectories in vivo and in vitro [1].
The Borrus study, posted as a bioRxiv preprint, addresses a related problem. With dozens of clocks available, different clocks may not agree on the effect of an intervention, and some apparent changes may simply be false positives caused by technical noise [6]. The authors measured variability between 6 popular clocks across longitudinal datasets containing either an aging intervention or an age-accelerating event, and compared them with the same clocks retrained for high test-retest reliability [6].
Markov and colleagues show that the reliability problem is not limited to clocks. Stochastic epigenetic mutations (SEMs) are outlier methylation patterns proposed as aging biomarkers. Across 3 datasets, a large share of them did not replicate [8]. The factors linked to unreliable SEMs included blood cell type composition, probe beta-value statistics, genomic location and the presence of SNPs. A machine-learning filter built on these factors enhanced reliability in two independent validation datasets [8]. The cell-composition point connects to our page on epigenetic clocks and blood cell composition.
What PC clocks do not settle
Reliability is not the same as meaning. A PC clock gives a more repeatable number, but what that number tracks still depends on the training target. Horvath's clock was built to estimate age across tissues [2]. PhenoAge [3] and GrimAge [4] were developed against lifespan and healthspan outcomes. DunedinPACE estimates the pace of aging [5]. Our page on first, second and third generation epigenetic clocks covers these differences in detail.
Clocks still disagree. The Borrus preprint notes that different clocks may not agree on the effect of an intervention [6]. Lower noise in any one clock does not remove that problem.
Availability is uneven. Published PC versions exist for particular clocks [1]. A results report may not state whether a lab used an original CpG-level clock or a PC version.
Reading a consumer result with reliability in mind
Consumer epigenetic age tests may be offered as laboratory-developed tests [9]. In September 2025, a Federal Register notice set out the implementation of the vacatur of FDA's rule on laboratory-developed tests [10]. Laboratories that test human specimens operate under CLIA [11], which concerns laboratory quality rather than whether a clock measures aging. Our page Are biological age tests FDA approved? explains this distinction.
For a broader walkthrough, see our biological age test guide and biological age test accuracy, which covers repeat results more generally. To learn how sources are chosen and cited here, see about this site.
Tools and further reading
The replicate data above translate into practical criteria for choosing a biological age test or biomarker panel. A well-documented test makes these points easy to check before ordering:
Technical-replicate figure. Whether the lab publishes how far two runs of the same sample differ.
Clock versions reported. Original CpG-level clocks, PC-based versions, or both, stated by name.
Error range on change. Whether a change between two tests is shown with an error range; a shift smaller than a clock's replicate spread is hard to interpret.
Cell-composition handling. Whether blood cell composition is estimated or adjusted for, given that it was among the factors linked to unreliable SEM calls [8].
Sample type. Saliva or blood, and whether the reported clocks were developed on that sample type.
Price transparency. Whether the full cost, including any retest or subscription, is shown up front.
Data privacy. Whether methylation data are stored, shared with third parties or deleted on request, and where that policy is published.
Turnaround. The stated time from sample receipt to report.
A test that answers most of these questions in its public documentation gives a clearer basis for reading any single score.
Frequently asked questions
What are PC clocks?
PC clocks are versions of epigenetic clocks that first compress CpG-level methylation data into principal components and then predict age from those components [1]. The extra step is designed to reduce the influence of technical noise at individual sites.
How large is technical noise in standard epigenetic clocks?
In replicate testing of six prominent clocks, technical noise produced deviations of up to 9 years [1]. The retrained PC versions brought most replicates within 1.5 years of each other [1].
Do PC clocks change what a clock predicts?
Each PC version is retrained to approximate the original clock's target [1]. The reported gain is in repeatability; whether every PC version tracks its original exactly is a separate question that a single abstract does not settle.
Does the reliability issue affect other methylation biomarkers?
Yes. In a study of stochastic epigenetic mutations, many outlier calls were not shared between technical replicates. A filter based on cell composition, probe statistics, genomic location and SNPs enhanced reliability in validation data [8].
Related on this site
Biological age tests: how epigenetic clocks work and what they mean
Are biological age tests FDA approved? How aging tests are regulated
Biological age test accuracy: reliability, noise and repeat results
First, second and third generation epigenetic clocks, explained
GrimAge explained: how the mortality-trained epigenetic clock works
Last updated: 26 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.