An unbiased comparison of 14 epigenetic clocks in relation to 174 incident disease outcomes

Why it matters

Dozens of epigenetic clocks now exist, and companies sell tests built on them, but most have been checked against only a few diseases in small groups of people. If you want to use DNA methylation to judge whether someone is aging faster, or whether a treatment works, you need to know which clocks actually predict illness. This study is a large, even-handed comparison that helps sort the useful clocks from the less useful ones. Because clocks are now used as outcomes in supplement and drug trials, including the multivitamin trial in this list, knowing which clocks track real disease risk tells us which trial results deserve attention.

What they did

The researchers measured 14 widely used clocks in 18,859 people and followed them for 10 years. They tested each clock as a predictor of 174 newly diagnosed diseases and of death from any cause. Their statistical models adjusted for lifestyle and socioeconomic factors. They also asked whether adding a clock to traditional risk factors made disease prediction more accurate, and whether the results differed by sex or smoking.

What they found

Second- and third-generation clocks, which are trained on health outcomes or pace of aging, clearly outperformed first-generation clocks, which were trained only to predict age; the authors say the older clocks have limited use for disease. There were 176 associations that passed a strict significance threshold. For 27 diseases, including primary lung cancer and diabetes, a clock's link to the disease was stronger than its link to death. In 32 of the 176 findings, adding the clock to a model with traditional risk factors raised classification accuracy by more than 1%. There was little evidence that the clocks worked differently by sex or by smoking status. The authors see the most promise for respiratory and liver conditions.

Limits

Accuracy gains were modest, often around 1%, so the clocks add to rather than replace standard risk factors. The findings are associations within one population and follow-up period, and the abstract does not show that changing a clock changes disease risk. The authors list further limitations in the full paper.

Key terms

First-generation clock

A clock trained to predict calendar age, such as Horvath or Hannum.

Second/third-generation clock

Clocks trained on health, mortality or pace of aging, such as GrimAge, PhenoAge or DunedinPACE.

Hazard ratio

How much more often an event happens in one group than another over time.

Bonferroni correction

A stricter significance threshold used when many tests are run at once.

Incident disease

A disease diagnosed for the first time during follow-up.

Classification accuracy

How often a model correctly sorts people into those who will and will not get a disease.