Biological Age Clocks: What a Score Can Estimate, Predict, Track, and Still Cannot Prove

Topic ·  Aging and Longevity Biology

Epigenetic clocks can be useful instruments, but only when the job of the model is clear.

A clock is a model built for a specific job

From a biological sample to a score

Epigenetic clocks begin with biology. They do not read biological age directly.

The clocks discussed in depth here use DNA methylation. Methyl groups can be added to cytosines at sites commonly called CpGs. Across a blood or tissue sample, methylation patterns carry information related to age, cell lineage, smoking and other exposures, disease, adaptation, and current physiological state. Common array assays measure hundreds of thousands of positions. A statistical model then selects or combines a much smaller feature set and assigns weights that best reproduce a chosen target.

The sequence is simple enough to state in one line:

The simplicity of the output is the point. Aging unfolds across cells, tissues, organs, immune states, metabolic systems, and functional capacities. By compressing part of that complexity into one score, a model creates comparison and reveals signal that might otherwise be difficult to see. The same compression also hides which biological dimensions carry most of that signal.

This is why an output expressed in years can be more metaphorical than it appears. The model is not counting a universal substance called biological time. It is converting a pattern into the scale that best serves its training task.

Other biological-age models use proteins, metabolites, clinical measurements, imaging, organ-specific data, or wearable signals. Their existence reinforces the same point: biological age is not one measurement category. This article focuses on DNA-methylation clocks because they have the deepest history and some of the strongest current intervention claims.

The training target sets the meaning

A clock’s training target determines what a successful result means.

A model trained to reproduce chronological age is rewarded for matching the calendar. One trained on mortality-linked features is rewarded for separating people with different future risks. A pace measure can be trained on longitudinal change across multiple physiological systems. All three outputs may be presented as forms of “biological age,” but they are not interchangeable readings of one hidden quantity.

Nor are the CpGs selected by a model automatically causes of aging. A site may improve prediction because it reflects accumulated exposure, immune-cell composition, damage, compensation, adaptation, or a downstream consequence of another process. A model can forecast an outcome without identifying the lever that should be manipulated to change it. Predictive value and causal meaning require different evidence.

The same care applies to “age acceleration.” For many clocks, acceleration is calculated as the difference between an observed score and the value expected for someone of the same chronological age, sometimes after additional adjustment. A positive residual means the model placed that sample above its reference expectation. It does not automatically mean the person is aging faster each year.

A cross-sectional residual compares one person with a reference group at one moment. A longitudinal slope asks how the same person’s score changes over time. The reference population and covariates also shape the result. Calling both quantities acceleration makes them sound more interchangeable than they are.

Three credible clocks make the distinction concrete.

Three clocks, three different jobs

Horvath, GrimAge, and DunedinPACE sit within the same broad category. Scientifically, they ask different questions.

Clock	Primary job and output	What it demonstrates	What it does not prove
Horvath pan-tissue clock	Estimates chronological age, expressed in years	The methylome contains a strong signal of elapsed time across many tissues	That the residual is a precise personal aging rate or healthspan measure
GrimAge	Predicts mortality-linked risk, with an age-like output	A methylation model can add prognostic information beyond calendar age	That the score is a literal age, an individual diagnosis, or the cause of mortality risk
DunedinPACE	Approximates a defined longitudinal pace, centered near 1.0	Blood methylation can project a multisystem rate-like construct	That 1.0 is a universal biological year or that every short-term change is interpretable

Horvath: estimating elapsed time

The 2013 Horvath pan-tissue clock was developed from roughly 8,000 samples spanning 51 tissues and cell types. Its model used 353 CpG sites. In test data, the correlation between DNA-methylation age and chronological age was about 0.96, with a median absolute error near 3.6 years, although performance varied substantially by tissue (Horvath, 2013).

This is an impressive demonstration that human methylation patterns contain a reproducible signal of elapsed time. It does not turn the gap between predicted age and calendar age into a direct measure of accumulated damage, resilience, or years of healthy life. A model can estimate its assigned target extremely well without answering a larger biological question.

GrimAge: predicting risk

GrimAge was built for a different job. It combines DNA-methylation surrogates for plasma proteins and smoking exposure in a model oriented toward mortality-related outcomes rather than chronological age alone. In its foundational work, GrimAge strongly predicted mortality, coronary heart disease, and cancer (Lu et al., 2019).

An external analysis in the Irish Longitudinal Study on Ageing makes that distinction concrete. Among 490 participants, each standard-deviation increase in GrimAge acceleration was associated with an adjusted all-cause mortality hazard ratio of 1.91, with a 95% confidence interval from 1.23 to 2.96. Horvath, Hannum, and PhenoAge measures were not significant in the same mortality model (McCrory et al., 2021).

For prognosis, GrimAge can therefore be more useful than a clock trained on chronological age. That does not make it the body’s true age. It means the model captures a combination of signals that helps predict risk in studied populations. Prediction is valuable, but it is not direct measurement of one causal aging process, nor is it, by itself, an individual diagnosis.

DunedinPACE: approximating a defined pace

DunedinPACE starts with a longitudinal target. The Dunedin birth cohort included 1,037 participants. Researchers built a physiological pace measure from change in 19 indicators assessed at ages 26, 32, 38, and 45, then trained a 173-CpG blood methylation model to reproduce that construct. The pace target was centered near 1.0 in the development cohort (Belsky et al., 2022).

A DunedinPACE value of 1.10 is therefore best read as an estimated pace about 10% above that cohort-derived reference. It is not a universal finding that a person accumulated exactly 1.10 biological years during one calendar year. The measure has been associated with morbidity, disability, and mortality in external datasets, which makes it a credible operational pace construct. Its unit still inherits the population, indicators, and time course used to create it.

None of these clocks fails simply because it cannot do the others’ jobs. The mistake is asking all three to behave like equivalent stopwatches. Once the target is clear, disagreement becomes less mysterious. It may reflect sensitivity to different dimensions of aging biology and different time scales.

The score belongs to a tissue, a population, and a moment

No score floats free of the biology that produced it.

Blood is biologically rich. It carries age-related, immune, exposure, inflammatory, and risk-linked signals. It is also a shifting mixture of neutrophils, lymphocytes, and other cell populations. When that mixture changes, the methylation pattern changes with it. Some of this variation can complicate longitudinal comparison. Some may be part of the biology the clock was designed to capture.

That dual role matters. Cell composition is not simply laboratory noise, but neither is it a complete report on the brain, muscle, vasculature, liver, or every other tissue. Horvath’s cross-tissue results revealed both a shared age signal and substantial tissue-specific differences. The continued development of tissue-, cell-, and organ-specific clocks gives another reason not to treat one blood value as the measured age of the entire body (Horvath, 2013; Wyss-Coray & Topol, 2026).

Calibration also belongs to a population. Age range, ancestry, geography, cohort history, health status, and social exposure can shape methylation distributions and model performance. Recent work found reduced age-prediction accuracy for several clocks in specific genetically admixed cohorts, with poorer performance especially among participants with substantial African ancestry (Cruz-Gonzalez et al., 2026).

That finding should not be universalized to every population or reduced to genetic determinism. Training-set composition, environmental exposure, social conditions, cohort effects, and the structure of the validation sample can all contribute.

These influences do not empty a score of meaning. They define the conditions under which its meaning is strongest. A result is most defensible when the named clock, tissue, population, and collection conditions resemble the setting in which its performance was established.

And then there is the moment of collection. A blood draw captures a biological state at a particular time. Meals, acute stress, infection, sleep disruption, activity, medication timing, altitude, pollution, and short-term cell shifts can affect repeated measurements. Before a change becomes evidence of a new aging trajectory, the comparison must be stable enough to carry that conclusion.

A stable assay is not necessarily a stable person

Reliability sounds reassuringly singular. It is not.

Those experiments can produce very different answers.

In a 2022 analysis of technical replicates, some historical clocks produced differences approaching nine years when essentially the same sample was measured again. Principal-component versions sharply reduced that error and achieved very high same-sample reliability (Higgins-Chen et al., 2022). This mattered because it showed that some apparent biological movement in older implementations could come from the measurement architecture itself.

But reducing assay noise does not freeze the biology of the person who supplied the sample.

A 2026 comparison of 18 epigenetic aging biomarkers examined technical replicates alongside repeated biological samples from small studies designed around acute exposures such as meals, stress, altitude, and pollution. Most pooled technical intraclass correlations were above 0.9. Pooled repeated-sample correlations were commonly around 0.4 to 0.7, and no clock achieved excellent short-interval biological stability across the pooled analysis. Technical and biological reliability were almost unrelated, with a correlation of 0.0168 (Sehgal et al., 2026a).

In plain language, a laboratory can reproduce the measurement very well while the value still changes across short-term conditions. Some movement may be a real transient biological response rather than assay error. It still should not automatically be read as a persistent aging trajectory. Decimal precision on a report is not proof that the instrument can resolve biological change to one tenth of a year.

This is where two technical-sounding thresholds become practical.

A minimal detectable change is the difference needed to exceed expected technical error and within-person variation under a specified protocol and confidence level. A minimal clinically important difference goes further. It identifies a magnitude linked to a meaningful difference in health, function, disease, or survival. Current epigenetic clocks do not have generally validated, clock-specific thresholds that allow a short personal retest to be translated confidently into either kind of change.

Without those thresholds, several possibilities remain entangled. A lower second value might reflect persistent biological change. It might reflect technical variation, a real but transient response to sampling conditions or immune state, or regression toward the average after an unusually high first result. Standardizing the clock, assay pipeline, collection conditions, and interval improves the comparison. It cannot supply a clinical meaning that has not been validated.

This does not make longitudinal tracking impossible. In the InCHIANTI cohort, 699 adults had repeated DNA-methylation measurements and up to 24 years of follow-up. When baseline level and longitudinal slope were modeled together, changes in most of the assessed clocks were associated with mortality independently of chronological age and sex (Kuo et al., 2026). This observational result suggests that multi-year trajectories can contain information that a single baseline score misses.

Time scale changes the scientific question. A standardized trajectory over years can be informative even when a three-month consumer retest remains difficult to interpret. Reliability is not a permanent label attached to a clock. It belongs to a particular comparison, under particular conditions, over a particular interval.

When interventions move clocks, disagreement can be the result

If uncontrolled personal retests are hard to read, randomized trials offer a cleaner test of responsiveness. They can balance baseline differences, standardize follow-up, and compare a planned intervention with a control. The evidence now shows that selected clocks do respond. It also shows why each response must be named precisely.

CALERIE and COSMOS do not move the same clocks

In a post hoc DNA-methylation analysis of CALERIE, a randomized trial, 220 healthy adults without obesity were assigned to caloric restriction or an unrestricted control diet, and 197 contributed usable longitudinal methylation data. Participants in the intervention group achieved average caloric restriction of about 11.9% over two years. Relative to control, caloric restriction reduced DunedinPACE, with standardized between-group effects of -0.29 at 12 months and -0.25 at 24 months. Investigators interpreted the magnitude as roughly a 2% to 3% slowing of that named pace measure. PCPhenoAge and PCGrimAge did not change significantly (Waziry et al., 2023).

COSMOS produced a different pattern. In a prespecified analysis of 958 older adults followed for two years, daily multivitamin-mineral use was associated with a between-group difference in yearly change of -0.113 years for PCGrimAge, with a 95% confidence interval from -0.205 to -0.020, and -0.214 years for PCPhenoAge, with a 95% confidence interval from -0.410 to -0.019. PCHannum, PCHorvath, and DunedinPACE were null. Cocoa extract was null across all five measures (Li et al., 2026).

If every clock measured the same universal quantity, this mismatch would look like failure. Once the clocks are understood as instruments with different targets, features, time scales, and sensitivities, the mismatch becomes informative. Caloric restriction may have affected signals emphasized by the DunedinPACE construct, while the multivitamin intervention produced small changes in features captured by the two principal-component age measures. The trials cannot identify one universally correct clock because that is not a scientifically coherent contest.

The claim each trial supports is narrower and more useful: under the tested conditions, the intervention changed the outputs of particular named clocks. CALERIE did not prove a 2% to 3% gain in healthspan. COSMOS did not show that participants gained a known number of healthy months. Both trials provide evidence of biomarker responsiveness, not a direct conversion into clinical benefit.

TranslAGE: responsiveness is real and heterogeneous

The 2026 TranslAGE analysis widened the lens. Researchers harmonized 51 longitudinal intervention studies, recalculated 16 prominent clocks, and assessed 94 additional DNA-methylation biomarkers. On the authors’ standardized effect-size scale, longevity-oriented interventions produced a pooled effect of about -0.06, while pro-aging events produced a pooled effect of about +0.14. Responses varied across interventions, populations, and biomarker families. DunedinPACE was among the most frequently responsive clocks (Sehgal et al., 2026b).

That is meaningful progress. A field once dominated by cross-sectional age correlations and prognosis now has evidence that selected biomarkers move in expected directions after human interventions and adverse events. Responsiveness can help identify promising instruments for future trials.

It is not the end of validation. The studies were heterogeneous, average effects were modest, and the analysis did not establish that short-term changes predict later changes in disease, function, healthspan, or lifespan. Nor did it establish a clinically important difference that converts the size of a clock effect into a known health benefit.

Statistical significance is not a personal or clinical threshold

A small effect can be statistically credible. That does not make it trivial. It also does not make its practical meaning known.

Statistical significance asks whether an observed average difference is difficult to explain by chance under a model. It does not ask whether the effect exceeds expected variation in one person’s retest. It does not reveal how much function improved, how many cases of disease were prevented, or how many healthy years were gained.

Small, durable biomarker changes could eventually prove useful, especially in trials where clinical aging outcomes take years to emerge. The honest interpretation is not that small effects have no value. It is that their biological and clinical scale has not yet been calibrated. Fractions of an epigenetic year should not become months of life gained simply because the output uses a familiar unit.

The intervention evidence supports a carefully positive conclusion: selected epigenetic aging biomarkers are responsive. It does not yet support the larger claim that lowering one demonstrates that aging itself was reversed.

A responsive biomarker is not yet a healthspan surrogate

This is where the most ambitious claims collapse three separate questions into one.

Prognosis asks: Do people with different scores have different future risks?

Response asks: Does an intervention change the score?

Surrogacy asks: When an intervention changes the score, does that change reliably predict how much the intervention changes an outcome that matters?

GrimAge can be prognostic without being a surrogate. CALERIE and COSMOS can demonstrate response without proving benefit. To use a clock in place of disease, function, or survival in a trial, researchers would need evidence that treatment effects on the clock reliably track treatment effects on those clinical outcomes within a defined context of use.

The distinction is plain but demanding. “People with better scores tend to live longer” is not the same claim as “treatments that improve the score cause people to live longer.” The first describes an association among individuals. The second requires concordance between treatment effects across multiple independent trials, ideally involving more than one intervention mechanism.

Other pieces must also fit: analytical performance across laboratories, standardized biological sampling, clock-specific minimal detectable change, an outcome-linked clinically important difference, replicated and persistent response, portability across relevant tissues and populations, and evidence that knowing the result improves decisions. Current biomarker-validation frameworks treat these as distinct jobs, not as one leap from correlation to clinical use (Moqri et al., 2023, 2024).

In a review of public FDA biomarker-qualification and EMA qualification-opinion sources through August 31, 2026, no epigenetic aging clock was identified as established or qualified to replace morbidity, function, healthspan, or survival outcomes. Clocks can still serve as exploratory trial endpoints and candidate response biomarkers. Surrogate status is not a verdict on whether a biomarker is interesting. It is a demanding claim about what that biomarker can safely stand in for.

Causal interpretation requires one more separation. A CpG can improve prediction because it marks exposure, damage, adaptation, cell composition, or a downstream response. Causal-inference work has reported that established clocks were not enriched for CpGs identified as potentially causal for aging-related traits, although that analysis depends on the assumptions of Mendelian randomization (Ying et al., 2024). Lowering a score therefore does not reveal which mechanism changed, nor does it prove that editing the model’s features would reproduce the health association.

For the same reason, a biological-age result should not be used alone to diagnose accelerated aging, choose a treatment, add a supplement, or decide that a personal protocol worked. Conventional clinical measures, functional outcomes, and established risk factors remain more directly actionable. A clock may add research information, but the decision utility of acting on an isolated score has not been established.

The field does not need to be dismissed to maintain this boundary. Responsiveness is a necessary step toward a useful intervention biomarker. It is simply not the same step as validated benefit.

Five questions to ask before believing a biological-age claim

A precise score can be useful when the claim around it grows no larger than the evidence underneath it. Before interpreting a headline, trial, or personal report, ask five questions.

Claims of a “true biological age,” organ-by-organ ages from one blood sample, or routine optimization through frequent retesting skip several of these questions. They turn a model output into a personal verdict before reliability, calibration, clinical meaning, and decision utility have been established.

A lower score can be real, scientifically interesting, and useful for research. Multi-year trajectories may prove informative. Responsive biomarkers may eventually help make longevity trials faster and better. None of that requires pretending the score is a stopwatch.

Taking biological-age clocks seriously means asking exactly what each one can do. Useful does not mean literal. A model can estimate, predict, track, or respond without proving that the whole person became younger or that healthspan increased. The science is not weakened when the claim fits the clock’s validated job. That is when the score becomes most useful.

Frequently asked questions

What is an epigenetic clock?

An epigenetic clock is a statistical model that uses DNA methylation patterns from a defined biological sample to estimate a target such as chronological age, mortality-linked risk, or a pace-of-aging construct. Its output reflects the clock’s training target, sample, reference population, and assay.

Are biological age clocks accurate?

Accuracy depends on the job. Some clocks estimate chronological age closely, while others are more useful for predicting risk or approximating a defined pace. Good performance in one role does not establish reliability for short-term personal tracking or validity as a healthspan endpoint.

Why do two biological age clocks give different results?

They may use different CpG sites, weights, tissues, reference populations, and training targets. Disagreement can therefore reflect different biological sensitivities and time scales, not simply that one clock is right and the other is wrong.

Does a lower epigenetic age mean aging was reversed?

No. A lower output may be a real biomarker change, but it does not show by itself that the whole body became younger, that a causal aging mechanism was reversed, or that healthspan increased.

Can a biological age test show whether a supplement, diet, or exercise plan worked?

A controlled study may show that an intervention changed a named clock. For an individual, interpretation also requires reliable sampling, a clock-specific threshold, persistence, replication, and a link to a meaningful outcome. Those conditions are not generally established for short personal retests.

How often should biological age be retested?

There is no generally validated retesting interval for using epigenetic clocks to guide personal decisions. Short-term changes can reflect technical error, transient physiology, or persistent biology, and current clocks lack broadly validated minimal detectable and clinically important change thresholds.

References

1. Horvath, S. (2013). DNA methylation age of human tissues and cell types. Genome Biology, 14, R115. https://doi.org/10.1186/gb-2013-14-10-r115
2. Lu, A. T., Quach, A., Wilson, J. G., et al. (2019). DNA methylation GrimAge strongly predicts lifespan and healthspan. Aging, 11, 303-327. https://doi.org/10.18632/aging.101684
3. McCrory, C., Fiorito, G., Hernandez, B., et al. (2021). GrimAge outperforms other epigenetic clocks in the prediction of age-related clinical phenotypes and all-cause mortality. The Journals of Gerontology: Series A, 76, 741-749. https://doi.org/10.1093/gerona/glaa286
4. Belsky, D. W., Caspi, A., Corcoran, D. L., et al. (2022). DunedinPACE, a DNA methylation biomarker of the pace of aging. eLife, 11, e73420. https://doi.org/10.7554/eLife.73420
5. Wyss-Coray, T., & Topol, E. J. (2026). Biological aging clocks in health and disease. Nature Medicine, 32, 2383-2394. https://doi.org/10.1038/s41591-026-04495-3
6. Cruz-Gonzalez, S., et al. (2026). Methylation clocks fail to generalize across genetically admixed individuals. eLife, 14, RP105343. https://doi.org/10.7554/eLife.105343.3
7. Higgins-Chen, A. T., Thrush, K. L., Wang, Y., et al. (2022). A computational solution for bolstering reliability of epigenetic clocks. Nature Aging, 2, 644-661. https://doi.org/10.1038/s43587-022-00248-2
8. Sehgal, R., Borrus, D. S., Gonzalez, J., et al. (2026a). Biological versus technical reliability of epigenetic clocks and implications for disease prognosis and intervention response. Aging Cell, 25, e70635. https://doi.org/10.1111/acel.70635
9. Kuo, P. L., Moore, A. Z., Tanaka, T., et al. (2026). Longitudinal changes in epigenetic clocks predict survival in the InCHIANTI cohort. Nature Aging, 6, 534-540. https://doi.org/10.1038/s43587-026-01066-6
10. Waziry, R., Ryan, C. P., Corcoran, D. L., et al. (2023). Effect of long-term caloric restriction on DNA methylation measures of biological aging in healthy adults from the CALERIE trial. Nature Aging, 3, 248-257. https://doi.org/10.1038/s43587-022-00357-y
11. Li, S., Hamaya, R., Zhu, H., et al. (2026). Effects of daily multivitamin-multimineral and cocoa extract supplementation on epigenetic aging clocks in the COSMOS randomized clinical trial. Nature Medicine, 32, 1012-1022. https://doi.org/10.1038/s41591-026-04239-3
12. Sehgal, R., Borrus, D., Armstrong, J. F., et al. (2026b). Responsiveness of epigenetic aging biomarkers to longevity interventions in humans. Nature Medicine. https://doi.org/10.1038/s41591-026-04562-9
13. Moqri, M., Herzog, C., Poganik, J. R., et al. (2023). Biomarkers of aging for the identification and evaluation of longevity interventions. Cell, 186, 3758-3775. https://doi.org/10.1016/j.cell.2023.08.003
14. Moqri, M., Herzog, C., Poganik, J. R., et al. (2024). Validation of biomarkers of aging. Nature Medicine, 30, 360-372. https://doi.org/10.1038/s41591-023-02784-9
15. U.S. Food and Drug Administration. Biomarker Qualification Program. FDA Biomarker Qualification Program
16. European Medicines Agency. Qualification of novel methodologies for medicine development. EMA qualification guidance
17. Ying, K., Liu, H., Tarkhov, A. E., et al. (2024). Causality-enriched epigenetic age uncouples damage and adaptation. Nature Aging, 4, 231-246. https://doi.org/10.1038/s43587-023-00557-0