Biological age clocks: what the algorithm behind the number was trained on

A clock is a regression, not a reading

An epigenetic clock is a statistical model that takes chemical marks measured at selected positions in the genome and returns a number expressed in years. The model was fitted by finding a weighted combination of those marks that best predicted a known quantity in whatever training population was available. Nothing in the biology declared those positions important; they were selected because they carried predictive signal in the particular data the model saw.

The output therefore inherits every property of that training set, including its age range, its ancestry composition and the tissue that was sampled. Calling the output an age is a naming convention, and it quietly invites a reader to treat a prediction as though it were a measurement.

What the training target determines

First-generation clocks were trained to predict chronological age, so a perfect model of that kind would simply return the number on a birth certificate. Because such a model is trained to ignore deviation, the leftover residual is what gets interpreted as accelerated ageing, which is a curious way to obtain a biomarker. Later clocks were trained instead against composite measures of physiological state or against time to death within a cohort, which changes what the output means.

Those trained on outcomes track health-related variation better, and they are correspondingly less accurate at guessing chronological age from a sample. A number from one family and a number from the other are not interchangeable, even though both are reported in the same familiar units.

Why two clocks disagree about one sample

Different clocks draw on largely non-overlapping sets of genomic positions, so there is no mechanical reason for their outputs to converge on each other. They also weight those positions differently, and a mark carrying substantial weight in one model may be absent from another model entirely. When a single blood sample is run through several published clocks, the resulting estimates commonly span a range of several years.

That spread is not a malfunction; it is what happens when several models built for different targets are asked the same informal question. It does mean a single number quoted without naming the clock that produced it carries much less information than it appears to.

Tissue, timing and technical noise

Methylation patterns differ substantially between tissues, so a clock trained on blood does not straightforwardly describe liver, muscle or brain. Blood is itself a mixture, and a shift in the proportions of immune cell types can move a clock reading without any change inside the cells. Technical factors including batch, array version and sample handling introduce variation that is not biological in origin at all.

Repeat measurements taken on the same person close together therefore differ, and that test-retest variation is often comparable to the effects being claimed. Careful analyses report the size of that variation explicitly, and consumer-facing reports very rarely do.

What the number can reasonably support

At the level of populations, clock outputs correlate with health outcomes well enough to serve as useful research instruments in large cohorts. At the level of one person on one occasion, measurement noise and between-clock disagreement leave very little room for confident individual interpretation. A change observed after an intervention is especially hard to read, since it must be separated from ordinary fluctuation and from cell composition shifts.

None of this makes the field unserious, and the underlying observation that methylation changes with age in a structured way remains robust. It does mean a personal result is better treated as a curiosity than as a finding, and any health concern belongs with a clinician rather than a report.