Why biological clocks get our ‘true age’ wrong – and how AI could help

When I first started writing about ageing years ago, there was a buzz around something called biological clocks, also known as ageing clocks or “true age” measurements. In principle, these are quite simple: we all have a chronological age, the number of years since birth, but this doesn’t necessarily reflect how far we are down the slippery slope from birth to decrepitude.

On average, this follows a fairly predictable trajectory, with gradual declines in almost every physical and mental attribute throughout adulthood. When we judge how old somebody is, we are intuitively totting up many of these tell-tale signs we see – the wrinkles and grey hair, or changes in posture, gait, voice, mental acuity and so on.

The goal of measuring biological age is to capture this decline in a single metric, evaluated scientifically and expressed in years. The results tell us something we intuitively know: some people age better than others.

Most people are, biologically, within a few years either side of their chronological age. But the two can diverge wildly. A 56-year-old (which I am) can have the biological age of a typical 30-something (which I almost certainly don’t), while another can have the biology of a septuagenarian (ditto). Crucially, biological age can rise more slowly than chronological age and even go down.

Biological age is a useful measure. It can give individuals succinct and easy-to-understand information about their overall state of health, encourage them to make lifestyle changes and tell them whether any resulting interventions, such as diet and exercise, are working. Judging from the number of commercial companies offering biological age testing, there is substantial clamour for such information, even though it is costly.

For scientists trialling anti-ageing interventions, it is a useful tool to see what works and what doesn’t without having to wait years to observe whether their guinea pigs (human or otherwise) decline and die at different rates. And for those working on the basic biology of ageing, biological age measurements can help them understand what is going on in our bodies as we get older.

So, what’s not to like? Quite a lot, as it happens. Biological age is sound in principle, but in practice leaves a lot to be desired.

The first biological clocks were based on epigenetic markers. These are molecular tags added to or removed from nuclear DNA that influence patterns of gene expression. About a decade ago, researchers led by Steve Horvath – the father of biological clocks, based at the University of California, Los Angeles – found that while there is a lot of individual variation, epigenetic markers change predictably over an average lifespan. Measure the right ones, feed the data through a complex algorithm and, hey presto, out pops an estimate of somebody’s biological age.

But epigenetics isn’t the only way to make an estimate. In the intervening years, many other clocks have been developed based on various other biological markers, including blood proteins, the length of caps of DNA at the ends of chromosomes called telomeres, urine metabolites, facial images and chest X-rays. That wouldn’t be a problem if they all came up with roughly the same answer, but they don’t.

As just one example, we can see this in a recent analysis of a clinical trial in humans called CALERIE, which examined the impact of long-term caloric restriction – a proven anti-ageing intervention in many organisms, though whether it applies to us remains up in the air. The CALERIE trial applied five different ageing clocks to 220 adults. Two of the clocks showed a significant reduction in biological age among the calorie-restricted participants. Three of them didn’t. Which should we believe? This is a problem that bedevils anyone – whether an individual or a scientist – who uses an ageing clock.

Another problem with ageing clocks is the illusion of accuracy. Most spit out a single figure with no error bars, despite inherent levels of uncertainty in the data and the analyses. According to a recent paper in the journal npj Aging, that is just the tip of the iceberg. Overall, the researchers conclude, existing clocks don’t do what they say on the tin and they run the risk of giving people either unwarranted confidence or unnecessary anxiety about the state of their health.

Does that mean that ageing clocks are useless? Not entirely. The paper’s authors, led by Dmitrii Kriukov at the Skolkovo Institute of Science and Technology in Russia, say that “all limitations of aging clocks are hypothetically solvable”. But whether they are worth solving is another question.

In part, that is because of a new and highly promising approach coming down the pipeline. Existing ageing clocks need biological samples. The new approach doesn’t, relying instead on – you guessed it – artificial intelligence, specifically something called large health models (LHMs). These are essentially large language models – like those that power AI chatbots such as ChatGPT – trained on huge volumes of health data to predict two of the main goals of ageing clocks: an individual’s risk of dying and their risk of developing age-related diseases. A recent paper in Nature Medicine reported that this approach outperforms existing clocks.

LHMs are still in development, and while we’re still getting them up to speed, the problems with existing clocks may be solved. But the take-home message is this: if you’re tempted to have your biological age measured, think twice. Or, if you have done so already, take the results with a pinch of salt. In return, I promise that next time I’m writing a story on ageing, I’ll be much more sceptical about research that uses them. Older, wiser.