What a younger ageing-clock reading can tell us

An ageing clock is a statistical model. Feed it measurements such as DNA methylation or blood proteins and it calculates an estimate. Some models estimate chronological age. Others were trained using mortality-related variables or changes observed over time.

The number has units that sound familiar, often years. That makes the output easy to read and easy to overinterpret. A predicted age is the model's summary of its inputs, interpreted against the data used to train it.

Horvath's 2013 methylation clock estimated age across many human tissues and cell types. Its performance showed that age-related information could be extracted from patterns of DNA methylation. The study did not establish that reducing the model's estimate through any intervention would extend life.

Age, pace and the tissue sampled

An age estimate and a pace estimate describe different quantities. A clock that predicts age can report a change in years. A pace measure attempts to describe the rate of biological change. Comparing the outputs requires understanding what each model learned.

CALERIE provides a human example of clocks responding differently. A post hoc analysis of blood samples from its randomised calorie-restriction trial detected a change in DunedinPACE, while the reported PhenoAge and GrimAge estimates did not show significant treatment effects. The authors called for follow-up to determine whether the pace result would translate into disease prevention or longer healthy life.

Reporting only the most favourable clock conceals the disagreement. Readers need the prespecified outcomes and all relevant analyses.

Blood is convenient to collect, but it is a sample of a complex body. A change in circulating cells or proteins can alter a model's input without demonstrating the same change in every organ.

The 2023 organ-age proteomics study used organ-enriched blood proteins to construct estimates associated with disease and mortality outcomes. It helped describe heterogeneity between organs. Observational association still leaves a separate question: whether an intervention-induced change in an estimate predicts the clinical effect of that intervention.

Noise and trial design

Higgins-Chen and colleagues studied technical reliability in epigenetic clocks and developed principal-component methods to reduce noise. This is relevant whenever a claimed treatment effect is close to the variation introduced by the measurement process.

Imagine taking two aliquots from one blood sample. If their clock estimates differ, the difference cannot be a biological change in the person between visits. It tells researchers about technical variation. Repeatability needs to be established before interpreting a small shift over time.

Trials also need appropriate controls. An individual whose unusually high baseline estimate decreases at a second measurement might show regression toward the average. Randomised comparison, repeat measurements and consistent sample processing help distinguish that possibility from a treatment effect.

Testing many clocks and choosing a favourable result increases the opportunity for a chance finding. Analyses should disclose which clock was chosen before the data were examined and how multiple comparisons were handled.

The remaining connection

FDA guidance distinguishes biomarkers from clinical outcomes and explains that surrogate endpoints need evidence connecting them to clinical benefit in a defined context. A marker's ability to predict risk in untreated populations is useful, but it does not by itself validate that marker for every intervention.

A clock result becomes more persuasive when an adequately controlled study shows a reproducible change alongside improvements in relevant function, and follow-up establishes persistence. The explanation should retain the model name, sample type and actual measured outcome.

“The blood-protein model estimated a lower age” is a clear finding. Whether participants aged more slowly, became healthier or will live longer requires the corresponding evidence.