Are biological age tests accurate? The number depends on the clock

Biological age tests do the easy job well and the hard job badly. Working out how old you already are is the easy job: models built from a handful of CpG sites, spots in your DNA that pick up a chemical tag as you age, estimate a donor's chronological age accurately [1]. The second number is what people pay for.

That second number is the gap between the age your DNA pattern implies and the age on your passport. Whether it holds still between two tests is the part the marketing skips.

Often it does not. Six widely used clocks were run across the same repeat samples, covering both interventions meant to slow ageing and events known to age people. Clocks trained to predict chronological age threw up changes that disappeared when the same data was rerun through steadier versions of those identical clocks [6]. The authors read those as false alarms, and note that some had already been published. Their analysis was posted as a preprint in 2024 and had not been through peer review.

Clocks built on death rates rather than birthdays came through the same tests consistently [6]. DunedinPACE, one of them, was built by first throwing out the DNA spots that gave inconsistent readings between repeat samples, and it repeats well as a result [7].

Those clocks also predict death across large groups. Among 1,771 adults in a German population study, 458 of whom died over 17 years [8]:

• a step up the mortality-risk score came with about 73% higher risk of dying
• a step up GrimAge came with about 47%
• a step up PhenoAge came with about 32%
• a step up an ordinary frailty checklist came with about 31%, nearly matching two of the clocks without reading any DNA

A step here is roughly how far apart two typical people in that study sat.

The practical size is modest. Take two of those adults at random, one of whom died first: age, sex and blood-cell mix alone put them in the right order 71 times out of 100, adding the mortality score and the frailty checklist reached 76, and adding both clocks on top reached 77 [8].

Clocks also disagree with each other inside one person. Among 216 adults with long-suppressed HIV followed for 10 years, GrimAge went with roughly four times the risk of dying, other clocks in the same blood flagged different outcomes, and none predicted heart or circulatory events [9].

Tissue changes the answer too. A heart-specific clock tested in 94 patients having a valve replaced and 289 having bypass surgery did not agree with a blood clock in the same people on the same day [2].

In laboratory-grown human cells, epigenetic ageing ran separately from cell senescence, telomere shortening and DNA damage, the processes it is usually assumed to capture, and tracked nutrient sensing and mitochondrial activity instead [4]. So the clocks are reading something, and what that something is remains partly open.

DNA methylation is a chemical tag attached to the DNA at CpG sites, often near the start of a gene, where the amount of tagging relates to how active that gene is [1]. The pattern moves over a lifetime, and people's patterns grow more different from one another as they age [1].

A commercial biological age test is a formula built by finding the age-related pattern in other people's samples, then applying it to yours. Both weaknesses follow from that.

A clock knows only the tissue it was trained on, which is why a heart-derived clock and a blood-derived one disagree in the same patient [2]. And a model fitted to predict age is not automatically a model that measures damage: in human cells, epigenetic ageing was not the same process as senescence or telomere shortening [4].

Reading thousands of DNA spots also carries measurement noise, and where that noise lands changes the answer. Rebuilding the same clocks to give steadier repeat readings was enough to make several published findings disappear [6].

No paper here tested a consumer kit or a clinic-branded biological age report. These are research clocks run in laboratories, so how any particular product on the market performs is not something this evidence settles.

The nearest thing is CheekAge, a cheek-swab clock built from 8,045 adults by the company behind it and checked against its own and public datasets [10]. It repeats well and moves with body weight, smoking and alcohol intake. Nobody has tested it as a purchased product.

The false-alarm analysis is a 2024 preprint, meaning it had not been peer reviewed when posted [6]. It argues that several published clock results should be re-examined, which is not the same as those results having been overturned.

The populations are narrow, and mostly ill:

• the stroke findings come from ischaemic stroke patients rather than healthy adults [3]
• the heart clock was built in people already booked for cardiac surgery [2]
• the depression comparison used blood from people diagnosed with major depressive disorder [5]
• the 10-year mortality figures come from 216 people living with HIV on long-term treatment [9]
• the work on what clocks actually track was done in human cells grown in a laboratory [4]

The German mortality figures describe how a group sorted over 17 years, not what one person's reading means for that person [8]. The review covering methylation and healthy human ageing was published in 2015 [1], before several of the clocks now sold commercially were built.

No study here shows that a biological age result changes what a clinician would do, or that moving the number lengthens anyone's life.

Biological age testing is one of the services listed in the Longevinsights clinics directory, offered by longevity clinics in the United States, the United Kingdom and continental Europe, usually bundled into a wider biomarker panel. A first appointment means giving a blood or saliva sample and filling in a health questionnaire. The sample goes to a laboratory that measures tagging at thousands of CpG sites, and a report comes back with a biological age in years and its gap against your real age.

Read against the papers above, the age estimate rests on well-replicated science [1], and the gap is the part that moves with the clock, the tissue and the noise in the measurement [2][5][6].

So it is worth asking which clock the laboratory runs before paying. A report from a clock trained on death rates has evidence behind it that a report from a clock trained to guess birthdays does not [6][8]. And a second test a year later is only worth comparing if it comes from the same clock in the same laboratory.