Most Epigenetic Clocks Are Technically Reliable But Biologically Unstable

Epigenetic clocks — biological age estimates derived from patterns of DNA methylation in blood — have surged in popularity as tools for tracking aging, predicting disease risk, and measuring the impact of lifestyle interventions. Their clinical promise rests on a fundamental assumption: that a measurement today will closely reflect a measurement taken under similar conditions tomorrow. This study from Yale University's Higgins-Chen lab systematically tests that assumption for the first time across 18 different clocks, covering five generations of clock design, using the TranslAGE platform previously validated across 51 intervention studies.

The researchers distinguished two separate forms of reliability. Technical reliability captures whether the same blood sample gives the same clock reading when processed twice in the laboratory — varying array type, slide position, or DNA extraction protocol. Biological reliability captures whether two blood draws from the same person, taken hours or days apart under conditions of ordinary daily life (meals, mild stress, environmental exposures), yield consistent readings. Four technical datasets (including ADNI with 196 participants and GSE174422 with 128 participants) and four biological datasets (including a stress-test cohort of 34 young adults sampled four times over five hours, a high-altitude exposure cohort at 5,260 m, and a diesel-exhaust exposure cohort) were analyzed using intraclass correlation coefficients (ICC) as the primary reliability metric.

On technical reliability, the majority of clocks performed well, with most exceeding ICC > 0.9 — categorized as excellent. PCGrimAge, SystemsAge, and GrimAge were consistently among the most stable. Earlier-generation clocks such as Hannum and PhenoAge showed lower but still acceptable ICCs in the 0.7–0.8 range. However, some clocks showed meaningful sensitivity to slide position and DNA extraction protocols, a finding with direct practical implications for multi-site studies. Biological reliability told a very different story: nearly all clocks showed only low to moderate stability across short-interval repeated measures, with ICCs often dropping substantially when cell-type composition was statistically adjusted. Critically, a clock's technical ICC did not predict its biological ICC — the two dimensions were essentially independent.

The downstream consequences of this instability were directly quantified. Using ADNI longitudinal data, the team showed that clocks with higher biological reliability produced more consistent and statistically stable associations with future cognitive decline, while lower-reliability clocks yielded variable and sometimes contradictory associations across subsamples. In intervention analyses — including a vegan diet study — clocks with poor biological reliability produced effect estimates that fluctuated substantially across analytical conditions, risking both false positives and false negatives. In contrast, PCGrimAge and related PC-based clocks yielded more reproducible intervention effect estimates.

The practical implications are significant. Any clinical trial using epigenetic clocks as endpoints must account for biological noise introduced by uncontrolled daily factors. An uncontrolled pre-meal versus post-meal blood draw, or differential stress levels at baseline versus follow-up, could generate apparent clock changes that are entirely artifactual. The authors argue that improving biological reliability — not just technical reproducibility — must become a design priority for next-generation aging biomarkers. PC-based clocks represent the current best option, but even these showed imperfect biological stability, pointing to the need for fundamentally new approaches.