Biological clocks: seductive, sometimes useful, far from settled

High level summary

Biological clocks are tests that attempt to estimate how fast your body is ageing, and how much your body has aged. Most use DNA methylation, which are biochemical tags that affect how genes are read. The appeal is obvious: if you could measure biological ageing early, you might identify risk sooner, personalise prevention, and track whether treatment is genuinely improving long-term health.

The science is real, but the field is not settled. There is no single agreed measure of “biological age.” Different clocks are built to predict or estimate different things, including chronological age, current disease burden, mortality risk, or pace of ageing.

Some of the newer clocks, especially ‘mortality-trained’ models like GrimAge, are quite good at forecasting age of death. But that creates a central problem: are they measuring ageing itself, or are they serving as ‘black box’ risk calculators that repackage familiar risk into one impressive-looking number? Useful prediction is not the same thing as a pure measure of ageing.

Caution should be exercised regarding intervention claims. Some studies show clock scores improving after diet changes, calorie restriction, supplements, or lifestyle programs. But a better score does not automatically mean fewer heart attacks, less dementia, or, indeed, more years of healthy life. In medicine, moving a biomarker is interesting but real-world outcomes are what matters.

Right now, biological clocks are more useful in research, trial design, and carefully framed exploratory use than in routine care. If the result does not change the basics such as blood pressure, ApoB, blood sugar, smoking, fitness, sleep, diet quality, and body composition, it’s probably not adding much. As with many things I look into carefully, the marketing is still ahead of the medicine.

Why we want biological clocks

In geroscience, the modern definition of ageing is the progressive loss of physiological integrity. The gradual accumulation of damage, dysregulation, and reduced repair capacity across the systems that keep us alive and functioning.

Read that definition again and notice what’s missing: any reference to time.

Nothing in it mentions birthdays, or calendars, or how many years you have been alive. Ageing, nowadays, is about the state of your biology, not the distance from your birth. This raises a strange and useful question.

Can you actually get older without ageing?

Can two people be the same age and not, in any meaningful biological sense, be ageing at the same rate?

That decoupling - the gap if you will, between the years you have lived and the condition your body’s in, is exactly the gap biological clocks claim to measure.

A meaningful biological-clock test has enormous appeal. If you could measure whether someone is ageing faster than they should be (or want to be!), in theory, you could personalise their surveillance and prevention, intervene earlier and track whether a treatment is doing anything useful. That’s why this topic has been on the agenda of every longevity conference I’ve attended.

While I’m excited by the concept of biological clocks, I simultaneously find them horoscopic: they give the feeling that a hidden truth about someone’s future has been revealed. Biological clocks are more scientific than astrology, being built from real assays, cohorts and statistics. But they trigger a similar human temptation: a single number that tells us where we stand with the mess of ageing compressed into something interpretable.

I understand the seduction and, myself, bought in early. Two years ago, I used to order biological-clock testing on all of my patients. After all, I’m trying to keep patients healthier for longer, so of course I wanted a readout of both how fast they were ageing and how much they had aged.

However, many of the test results ‘felt’ completely wrong (I recognise the irony of using ‘feel’ here), or at least so difficult to reconcile against the patient’s health and behaviour, that I started to question what I was measuring and how much attention the results deserved.

Biological clocks are scientifically interesting; some seem to move with intervention and others seem able to predict (bad) outcomes. However, presently at least, they are research tools rather than clinical tools.

The problem isn’t that these clocks are measuring nothing. The problem is that we cannot say exactly what they are measuring. What was the clock trained to predict? Is it measuring ageing itself, or just repackaging risk we already know how to find? Does moving the score change anything worth doing? Does it improve hard outcomes, or only a number? These clocks were built to answer different questions from the start.

The desire for this type of tool is rational. Ageing sits underneath most of the chronic disease burden we care about, so a reliable way to identify accelerated ageing early could sit upstream of the diseases we usually wait to diagnose.

How these clocks are built and work, in concept and in practice

Most of the clocks I discuss here are built from DNA methylation patterns. Your genome is your DNA sequence; the epigenome is a layer of chemical tags that influences how it is read, and methylation is one of the main tags. Unlike the genome, the epigenome is at least partly modifiable, shifting with age and with exposures such as smoking, adiposity, diet, stress, exercise, sleep, and illness.

That modifiability is the appeal. If a clock is based upon a changeable layer of biology, then in theory you might alter your pace of ageing, not just measure it. Mixed intervention results keep the idea alive, though hard-outcome evidence is still missing [9,10,12,16].

So how is a clock actually built? It starts as a statistical model trained on data. Researchers take samples from a large group of people, usually blood, extract the DNA, and measure methylation at hundreds of thousands of sites across the genome, each scored as the proportion of DNA molecules carrying a methyl mark there, from 0 to 1. Beside those profiles sits the thing they want to predict: usually each person’s chronological age, though newer clocks use who later died, or markers of disease. A machine-learning method then sifts those sites, keeps the smaller set carrying the most signal, often a few dozen to a few hundred, and gives each a weight. The finished clock is just a formula: a list of sites, each multiplied by its weight and summed.

Applying it to a new person is mechanical. You measure methylation at those same sites in their sample, feed the values into the formula, and out comes one number, their predicted age or score. Subtract their actual age and you get the headline figure, often called age acceleration: how much older or younger their biology looks than the calendar says. The output is only as good as what the clock was trained to predict, which is why two clocks run on the same sample can return different answers.

Horvath’s original multi-tissue clock was trained to estimate chronological age across tissues. It was a foundational proof of concept, with a test correlation of 0.96 and a median error of 3.6 years [1]. Hannum’s whole blood model was also foundational and showed that methylomic ageing-rate variation could be detected in blood, with men appearing to age about 4% faster than women in that dataset [2].

Later clocks looked at different things. PhenoAge was designed as a DNA methylation surrogate of ageing tied to mortality biology rather than simple calendar time [5]. GrimAge and GrimAge2 went further again, using methylation surrogates for smoking and protein pathways trained on mortality risk, which helps explain why they perform so strongly for prognosis [6,13]. DunedinPACE shifted the field again by focusing on pace of ageing rather than age level, which is why it is attractive for longitudinal and intervention research [8].

The field is plural from the start

Lots of people are talking about “your biological age” as though the field has settled on one coherent measurement, which of course it hasn’t.

A multi-tissue chronological age estimator, a whole blood model, a mortality-trained clock, and a pace-of-ageing endpoint are not all answering the same question. PhenoAge, GrimAge, GrimAge2, DunedinPACE, Horvath, Hannum, principal-component clocks, PCPhenoAge, PCGrimAge, and constructs such as Li methylation age are related, but they are not interchangeable [14,15,18]. “Your biological age” is not one measurement. It is a label sitting on top of several, and they do not always agree.

This is one reason I became more cautious clinically. A single report can look more certain than the science underneath it really is.

How real is the underlying data?

Biological clocks are not fantasy metrics (my son recently got me into fantasy baseball and it’s not working out at all). They come from real methylation training sets, real longitudinal cohorts, and real follow up data. Some are anchored to chronological age, others to physiologic or mortality related phenotypes. GrimAge and GrimAge2, as already described, derive much of their strength from those mortality-linked surrogates [6,13]. DunedinPACE was built around pace of ageing, with an intraclass correlation coefficient of 0.96 in validation work [8]. Principal-component approaches appear to improve stability, and Apsley and colleagues found that acute stress reduced probe stability [14].

So the data are real. The problem is that a real signal can still be conceptually slippery. If a clock is trained on cohorts where researchers observe who accumulates disease, who dies sooner, or which methylation signatures track smoking and inflammatory biology, then the resulting score may be informative while still meaning something different from what a patient hears when they hear the phrase “biological age”. That is where my scepticism bites.

The more powerfully a clock predicts mortality, the more we need to ask whether it has become a disguised risk-prediction tool rather than a pure measure of ageing itself.

Are some clocks closer to mortality clocks than ageing clocks?

Probably. The older clocks gave us proof that methylation patterns carry age-related information. They also showed prognostic ability, though more modestly. Marioni and colleagues found that a 5 year higher age acceleration was linked to a 21% higher mortality risk after age and sex adjustment and a 16% higher risk after broader adjustment [3]. Perna and colleagues reported that each 5 year higher Horvath delta age was associated with all-cause, cancer, and cardiovascular mortality, with hazard ratios of 1.23, 1.22, and 1.19 respectively [4].

The newer clocks are far more bluntly prognostic. PhenoAge linked each additional year of phenotypic age to hazard ratios of 1.09 for all-cause mortality, 1.10 for cardiovascular mortality, and 1.20 for diabetes mortality [5]. GrimAge stood out for time to death, time to coronary heart disease, and time to cancer in its original report [6]. Over 17 years of follow up, each standard deviation higher AgeAccelPheno and AgeAccelGrim carried hazard ratios of 1.32 and 1.47 for all-cause mortality [7]. GrimAge2 then outperformed GrimAge for mortality across multiple racial and ethnic groups and also tracked coronary heart disease, lung function, and fatty liver measures [13]. In the HUNT cohorts, each 1 standard deviation higher GrimAge2 carried hazard ratios of 2.42 in HUNT2 and 2.30 in HUNT3 for all-cause mortality after adjustment for established risk factors, while DunedinPACE carried a hazard ratio of 1.99 in HUNT2 [20].

Some of the newer clocks are edging closer to mortality prediction than to simple age estimation. That is part of their appeal, but it is also where interpretation gets murky.

A mortality-trained clock can be excellent at prediction without being a pure measure of ageing. If it is compressing smoking burden, inflammation, immune-cell composition, and comorbidity into one elegant score, it may be acting less like a biological essence meter and more like a sophisticated risk index, taking risk we can already detect and repackaging it into a single number that sounds deeper, and sells better, than the ordinary tests it is built from.

Its practical value then comes mainly from forecasting outcomes better than chronological age, even as its biological interpretation becomes less clean. GrimAge and GrimAge2 sit very close to that line.

What gets lost when patients hear “countdown timer”

First, uncertainty gets lost. These are probabilistic tools, not diagnostic verdicts. They do not establish causation, destiny, or a personalised treatment algorithm.

Second, construct validity gets lost. A mortality-trained clock is not identical to a pure measure of whole-body ageing. If the score is heavily shaped by the risk factors we can already measure, then the message may be closer to “your known risk burden is accumulating” than “this is the exact pace of your biological decline”.

And proportion gets lost. The countdown-timer framing invites fatalism, drama, and score chasing. Patients can end up relating emotionally to the number rather than clinically to the work that still matters.

What the current evidence genuinely shows

Strip away the marketing and one thing holds up: as predictors, these clocks work. They forecast meaningful outcomes beyond chronological age, and in some cohorts beyond established risk factors. In one cohort, Kuo and colleagues found that combining a clock’s baseline level with how fast it was changing predicted survival about as well as a single biomarker realistically can: a C statistic of about 0.81, meaning that given two people, the model correctly identified the higher-risk one roughly four times in five. The other leading clocks landed within a hair of that [19]. No clock clearly wins.

On prognosis, the strongest performers are GrimAge and GrimAge2. DunedinPACE looks most useful as a longitudinal and intervention endpoint, and PhenoAge for population-level risk stratification. Horvath remains historically foundational, even if it is weaker for clinical use today.

The implications extend beyond mortality. For incident mild cognitive impairment and dementia, hazard ratios per 1 standard deviation of clock were 1.11 for GrimAge2, 1.07 for DunedinPACE, and 1.01 for PhenoAge, though only GrimAge2 survived correction for multiple testing [17].

So prediction is established. What it doesn’t answer is the question I actually face: what do I do differently for the patient I’m looking after?

Prediction does not equal actionability

This is where current commercial use runs ahead of the science. A score can be statistically impressive and still add little to care once I have already assessed blood pressure, ApoB or lipids, glycaemic markers, smoking status, fitness, waist size, sleep, diet, booze and body composition.

Prediction tells you a marker is associated with future outcomes. Actionability asks whether knowing the result changes management in a way that improves care. In 2026, we still do not have widely accepted clinical cutoffs, retest intervals, reimbursement norms, or treatment pathways tied to clock results.

Can you move the number, and does that matter?

It’s the question patients care about most. The answer at present is awkward. Some clocks move but it doesn’t mean the hard outcomes move with them.

In CALERIE, long term caloric restriction reduced DunedinPACE by about 2 to 3%, but there was no significant change in PhenoAge or GrimAge and no proof that changing the clock reduced clinical events [12]. In a multidomain lifestyle intervention in frail older adults, DNAm PhenoAge improved and frailty and function improved, but there were no hard outcome data showing that the clock change translated into fewer deaths, cardiovascular events, or dementia cases [16].

There are smaller encouraging results. A vitamin D3 randomised trial linked 4,000 IU per day to a 1.85 year decrease in Horvath age and 2,000 IU per day to a 1.90 year decrease in Hannum age [9]. A diet and lifestyle pilot randomised trial reported that the treatment group finished 3.23 years younger than controls on Horvath DNAmAge [10]. Those findings are interesting. They don’t prove that the intervention slowed clinically meaningful ageing in a way that reduced hard outcomes.

Other studies are much less flattering. A randomised metformin and weight-loss trial in breast cancer survivors found no clear signal by intervention arm across GrimAge, PhenoAge, and related measures [11]. DIRECT PLUS showed mixed results, with no global between-arm clock effect, although greater Green Mediterranean diet adherence was linked to lower relative age change and about 8.9 months of favourable Li methylation age difference [15]. In the MACRO 12 month weight-loss diet trial, baseline associations were strong, but after intervention most associations attenuated and there was no evidence that clocks mediated intervention effects [18].

So the most defensible sentence is this: clock movement has not been proven to mean fewer deaths, fewer heart attacks, or fewer dementia cases. If your biological age goes down, you may have improved a biomarker. Unfortunately, you haven’t proven that you bought extra years of healthy life.

How could clock testing change medicine if the field matures?

I don’t want to undersell the upside. If the field matures, clocks could become useful in preventative medicine. A robust, reproducible test that adds meaningful risk information beyond standard assessment could identify people whose biology is diverging before conventional disease labels appear. That could sharpen earlier risk stratification.

A reliable longitudinal clock could also help monitor intervention response without forcing us to wait years for hard events. DunedinPACE is particularly interesting here because it was built around pace of ageing and has strong technical reliability, which is why it keeps emerging as the most plausible longitudinal endpoint in the brief [8,14].

The trial-design implications follow. If ageing interventions are going to be tested seriously, researchers need endpoints that move on a practical timeframe. A mature clock could help decide whether an intervention deserves a large outcome trial or should be abandoned early.

The danger is premature over-medicalisation. If we treat immature technology like established truth, we turn healthy but worried people into permanent patients. We create more testing, more retesting, more supplements, more off-label drugs, more fear, and more medical theatre without proof of benefit. Biological clocks could genuinely change medicine.

False reassurance, psychological impact, and why pre-test counselling should come first

The most under-discussed harm here is false reassurance. A young biological-clock score can make a patient downplay high ApoB, hypertension, smoking, poor fitness, dysglycaemia, poor sleep, or excess central adiposity. A flattering result can become an excuse to ignore proven risk.

An older score creates the opposite problem. It can trigger anxiety, shame, and fatalism without offering a validated treatment pathway. The commercial setting adds another risk because it makes upselling easier. Once a patient has been told they are biologically older than expected, they are vulnerable to the next offer, whether that is supplements, hormones, off-label drugs, or serial retesting.

There is also a psychological distortion specific to these tests. They invite people to optimise the score rather than optimise the underlying risk. That is not a subtle difference. It can pull behaviour away from the fundamentals that matter.

The evidence here is thin. There is little direct research on the psychological impact of clock testing itself, so this argument runs on principle and analogy, not data. But the risk is obvious enough that patients should have pre-test counselling first, even if they are only testing out of curiosity.

That counselling does not need to be elaborate. It does need to cover what the test can and cannot tell you, that results are probabilistic rather than diagnostic, that a good score does not override conventional risk factors, that a bad score does not by itself justify aggressive treatment, that repeatability and platform variation matter, and that interpretation should sit beside a standard clinical assessment rather than replacing it.

Where I land now

I am more sceptical than I used to be, but not dismissive. The contrarian point is not that biological clocks are nonsense. It is that the strongest current clocks may be most useful precisely where the marketing becomes least romantic. They may be closer to compressed risk-prediction tools than mystical ageing meters. That still leaves room for genuine clinical value in the future. It simply means we should stop pretending the future has already arrived.

For routine patient care, these tests are not essential. If a highly engaged patient wants one as a curiosity or research-style metric, that is not ridiculous. I would frame it carefully. I would not let a single score drive supplements, hormones, off-label drugs, or aggressive treatment. I would be cautious about short-interval serial retesting. I would interpret the result alongside standard clinical assessment, never instead of it.

The practical question remains very plain. Does this test change anything worth doing? For most people today, the answer is still no. The priorities remain stubbornly conventional: control blood pressure, address ApoB and lipid burden, improve glycaemic health, stop smoking, reduce drug and alcohol consumption, build fitness, improve diet quality, protect sleep, and sort out body composition. Those steps are less glamorous than a methylation score. They are still where the real clinical leverage sits.

References

1. Horvath. 2013. Foundational multi tissue DNA methylation clock with test set correlation 0.96 and median error 3.6 years. PMID: 24138928.

2. Hannum et al. 2013. Foundational whole blood age modelling paper; methylomic ageing-rate variation, with men appearing to age about 4% faster than women. PMID: 23177740.

3. Marioni et al. 2015. A 5 year higher delta age linked to higher all-cause mortality risk after age and sex adjustment and broader adjustment. PMID: 25633388.

4. Perna et al. 2016. Horvath and Hannum age acceleration associated with all-cause, cancer, and cardiovascular mortality. PMID: 27274774.

5. Levine et al. 2018. PhenoAge associated with all-cause and cause-specific mortality. PMID: 29676998.

6. Lu et al. 2019. GrimAge stood out for time to death, coronary heart disease, and cancer. PMID: 30669119.

7. Li et al. 2021. AgeAccelPheno and AgeAccelGrim associated with all-cause mortality over 17 years. PMID: 34808433.

8. Belsky et al. 2022. DunedinPACE associated with mortality and incident chronic-disease morbidity; ICC 0.96. PMID: 35029144.

9. Vitamin D3 supplementation randomised trial. 2019. Small decreases in Horvath and Hannum age with supplementation. PMID: 30256915.

10. Diet and lifestyle pilot randomised trial. 2021. Treatment group finished 3.23 years younger than controls on Horvath DNAmAge. PMID: 33844651.

11. Randomised metformin and weight-loss trial in breast cancer survivors. 2021. No clear signal by intervention arm across GrimAge, PhenoAge, and related measures. PMID: 34920739.

12. CALERIE trial. 2023. Long-term caloric restriction reduced DunedinPACE by about 2 to 3% without hard outcome proof. PMID: 37118425.

13. Lu et al. 2022. GrimAge2 validation for mortality and morbidity across multiple racial and ethnic groups. PMID: 36516495.

14. Apsley et al. 2023. DunedinPACE and principal-component clocks showed greater probe stability; acute stress reduced stability. PMID: 37393564.

15. DIRECT PLUS diet trial. 2023. Mixed methylation-clock findings; greater Green Mediterranean diet adherence linked to lower relative age change and about 8.9 months of favourable Li methylation-age difference. PMID: 37743489.

16. Multidomain lifestyle intervention in frail older adults. 2026. DNAm PhenoAge improved, with frailty and function improvement but no hard outcome data. PMID: 41677077.

17. Nguyen et al. 2026. GrimAge2, DunedinPACE, and PhenoAge assessed for incident mild cognitive impairment and probable dementia; only GrimAge2 survived Bonferroni correction. PMID: 41721741.

18. MACRO 12 month weight-loss diet trial. 2025. Baseline associations strong, but clocks did not mediate intervention effects. PMID: 40922554.

19. Kuo et al. 2026. Baseline plus longitudinal change produced the highest survival C statistics for GrimAge2, GrimAge, PhenoAge, and DunedinPACE. PMID: 41826710.

20. Sun et al. 2026. GrimAge2 and DunedinPACE strongly predicted all-cause mortality in HUNT cohorts after adjustment for established risk factors. PMID: 42163406.