Epigenetic clocks

Epigenetic clocks are statistical models that estimate a person's age, or their risk of death, from the pattern of methyl groups attached to cytosines at selected positions in their genome. They are the most widely used molecular measure in aging research, cheap enough to run on stored blood or saliva and accurate enough at predicting chronological age to be startling. What they measure biologically, and whether moving a clock reading means anything, remain open.

§What a clock measures

DNA methylation at CpG dinucleotides regulates gene expression and differs systematically between tissues and between young and old individuals. A clock is built by taking methylation values at hundreds of thousands of sites, measured on a commercial array, and fitting a penalised regression that predicts a target variable from a small subset of them. The selected sites are not a mechanism. They are whichever coordinates carried the most predictive signal in the training data.

For a clock trained on chronological age, the interesting quantity is not the prediction but the residual: the difference between predicted and actual age, usually called epigenetic age acceleration. A person whose blood reads five years older than their birth certificate is, on average across large cohorts, at somewhat elevated risk of death and of several chronic diseases. The effect is real and modest.

Compared with earlier candidate measures, methylation performs well. Leukocyte telomere length, discussed under Telomeres and telomerase, correlates with age so loosely in individuals that it is close to useless as a personal readout. Methylation clocks are far tighter, which is why they displaced telomere length as the field's default molecular measure of the epigenetic hallmark within a few years.

§Generations of clocks

§First generation

Early clocks were trained directly on chronological age. Steve Horvath's 2013 multi-tissue clock used 353 CpG sites and predicted age across most human tissue types with a median error of roughly three and a half years, a result that also held for cells in culture and, notably, returned close to zero for Induced pluripotent stem cells.¹ Gregory Hannum's group published a blood-specific clock in the same year.² Because these models are optimised to reproduce the calendar, any deviation from it is treated as error by the fitting procedure, which limits how much aging-relevant signal they can retain.

§Second generation

The second wave changed the training target. DNAm PhenoAge was fitted to a composite of clinical laboratory measures and chronological age rather than to age alone.³ GrimAge went further, building methylation surrogates for plasma proteins and for smoking history and then regressing those on time to death.⁴ Both predict mortality and disease incidence considerably better than first-generation clocks, which is unsurprising given what they were trained on.

§Pace-of-aging

DunedinPACE takes a different approach again. Rather than estimating a level, it estimates a rate, trained on two decades of repeated organ-function measurements in the Dunedin birth cohort, all of whom are the same age.⁵ Its output is a speed: roughly one biological year per calendar year for an average person. Because chronological age is constant within the training set, the model cannot be fitting the calendar.

§What the epidemiology shows

In population cohorts, age acceleration on second-generation clocks associates with smoking, obesity, low socioeconomic position, chronic infection, and markers of Inflammaging. It also associates prospectively with incident cardiovascular disease, dementia and all-cause mortality, though generally with hazard ratios well below those of established clinical risk factors. GrimAge in particular carries a large share of its predictive weight in its smoking surrogate, which complicates the claim that it is measuring aging as such.

Clocks have been applied to questions about exceptional longevity, including studies of residents of the regions described under Blue Zones and of supercentenarian cohorts relevant to Maximum human lifespan. Results in such groups are limited by tiny sample sizes and by the same data-quality problems that affect the underlying age records.

§Reliability and confounding

Technical noise is a serious and underappreciated problem. Running the same DNA sample twice on the same array platform can produce clock estimates differing by more than a year, which is large relative to the effect sizes reported in intervention studies. Reconstructing clocks on principal components of the methylation data substantially improves test-retest reliability.⁶

Cell composition is the other main confounder. Blood is a mixture, and the proportions of naive T cells, memory T cells, monocytes and granulocytes shift with age and with acute illness. A clock trained on whole blood partly measures that shift. Some clocks adjust for estimated cell counts; the adjustment is imperfect.

Contested

No epigenetic clock has been shown to track a causal driver of aging rather than a downstream correlate. Methylation changes could be the mechanism, a readout of the mechanism, or an incidental consequence of cell-turnover history. The distinction is invisible to the regression.

§Response to interventions

The strongest use case for clocks would be as a short-cut endpoint, letting a trial read out in a year rather than in decades. See Aging biomarkers for the general problem and the regulatory bar. The evidence that clocks respond to interventions in humans is thin.

The most-cited human result is the TRIIM study, in which nine men received growth hormone with DHEA and Metformin and the TAME trial and showed a mean reduction of a few years in several clock readings alongside thymic changes.⁷ It had no control group and nine participants. In the CALERIE randomised trial of Caloric restriction, stored samples showed a small slowing of DunedinPACE in the restricted arm but no significant effect on PhenoAge or GrimAge.⁸ That pattern, one clock moving and others not, recurs across studies and is not well explained.

Epigenetic reprogramming moves clock readings dramatically in cells and in mice, which is the strongest demonstration that the readings are manipulable. It is also the clearest case where a clock delta and a functional benefit could come apart, since the same treatment that resets methylation can erode cell identity.

§Beyond humans

The Mammalian Methylation Consortium constructed clocks that work across dozens of mammalian species using conserved CpG sites, allowing the same measure to be applied to a mouse, a bat and a bowhead whale.⁹ These cross-species clocks are useful for comparative work on Negligible senescence and for testing whether interventions such as Rapamycin or dietary restriction slow methylation change in animals whose lifespans can actually be measured. Mouse clocks have become a routine secondary endpoint in geroprotector studies, including work on Senolytics and heterochronic blood exchange, with the same interpretive caveat as in humans.

§Consumer testing

Several companies sell direct-to-consumer methylation age tests, through the same mail-in channels as the direct-to-consumer laboratory panels. The assays are usually real, but the reported precision often exceeds what the underlying reliability supports, the clocks used are frequently proprietary and unpublished, and a result expressed as a single number invites over-interpretation of what is a noisy estimate with wide individual uncertainty. Repeat testing of the same person over short intervals can produce swings that reflect assay variation rather than any change in Biological age. Guidance from professional bodies has been consistently cautious about clinical use.

§Outlook

Clocks are firmly established as research instruments and firmly unestablished as endpoints. The work that would change that is unglamorous: standardised assays, published reliability statistics, replication of intervention effects across independent cohorts, and demonstrations that a clock change predicts a later functional change in the same person. Competitions such as XPRIZE Healthspan have deliberately chosen functional endpoints instead, on the view that measured muscle, cognitive and immune restoration is what the geroscience argument ultimately has to deliver.

§See also

• Biological age
• Aging biomarkers
• Epigenetic reprogramming
• Hallmarks of aging
• Geroscience hypothesis
• Caloric restriction
• Healthspan

§References

Footnotes

1. paperHorvath, S. "DNA methylation age of human tissues and cell types." Genome Biology, 2013.↩Fitted to publicly available methylation datasets; the accuracy reported is against chronological age, the variable the model was trained on.

2. paperHannum, G. et al. "Genome-wide methylation profiles reveal quantitative views of human aging rates." Molecular Cell, 2013. ↩

3. paperLevine, M.E. et al. "An epigenetic biomarker of aging for lifespan and healthspan." Aging, 2018. ↩

4. paperLu, A.T. et al. "DNA methylation GrimAge strongly predicts lifespan and healthspan." Aging, 2019. ↩

5. paperBelsky, D.W. et al. "DunedinPACE, a DNA methylation biomarker of the pace of aging." eLife, 2022. ↩

6. paperHiggins-Chen, A.T. et al. "A computational solution for bolstering reliability of epigenetic clocks." Nature Aging, 2022. ↩

7. paperFahy, G.M. et al. "Reversal of epigenetic aging and immunosenescent trends in humans." Aging Cell, 2019.↩An uncontrolled study of nine men, run to test thymus regeneration rather than to test an intervention against a clock.

8. paperWaziry, R. et al. "Effect of long-term caloric restriction on DNA methylation measures of biological aging in healthy adults from the CALERIE trial." Nature Aging, 2023.↩An analysis of stored samples from a trial designed to test caloric restriction, not to validate the clocks.

9. paperLu, A.T. et al. "Universal DNA methylation age across mammalian tissues." Nature Aging, 2023. ↩