The at-home biological age test has quietly become one of the wellness industry's best products. You spit into a tube, post it off, and a few weeks later a dashboard tells you that you are 38.4 while your passport insists on 44. Prices run from about $30 to well over $1,000.[5] The appeal is obvious: one number, apparently objective, apparently improvable.
A paper published in Nature Communications and picked up widely in the first days of August makes that number look considerably stranger than most buyers assume. Not wrong, exactly. But an average taken over a population that, it turns out, is not ageing together at all.
What the single-cell data showed
The work came from a group led by Hagit Masika, Tommy Kaplan and Howard Cedar at the Hebrew University of Jerusalem, with collaborators at Altos Labs and the Babraham Institute in Cambridge, ETH Zurich and the German Cancer Research Center.[1]
Rather than grinding up a tissue sample and measuring its average chemistry, they used single-cell whole-genome methylation data — readings taken one cell at a time — across a range of tissues and ages in both mice and humans. The question was simple and, in hindsight, obvious: when a tissue looks "older" chemically, is that because every cell drifted a little, or because some cells drifted a lot?
The answer was the second one. Age-related methylation did not creep up uniformly. It concentrated in a subset of cells while the rest stayed comparatively young. As Masika put it in coverage of the study, two cells sitting side by side can have completely different biological ages.[2]
The paper's own framing is blunter than most press releases: the results "challenge traditional models of homogeneous cellular ageing" and suggest ageing is "a highly individualized process at the single-cell level."
The chemistry, briefly
The specific marks involved are worth a moment, because they are the same ones your test is reading.
Polycomb CpG islands are stretches of DNA sitting near genes that a cell keeps switched off but poised — held in a ready-to-use state by a family of proteins called polycomb. Over a lifetime these regions slowly pick up methyl groups, which shuts them down more permanently. That drift is one of the most reliable molecular signatures of ageing anywhere in biology, which is exactly why it anchors so many epigenetic clocks.
What the single-cell view added is that the drift is lumpy. And the strongest predictor of which cells drifted fastest was how often they divided. Rapidly proliferating cells gained methylation quickest — a mechanistic hint that this is at least partly a copying problem, an error that compounds each time the epigenetic pattern is rewritten into a daughter cell.
When the team looked at what those fast-ageing cells were doing differently, the affected genes clustered in four areas: immune response, protein translation, tumour development and neurodegeneration. That is a striking list. It is more or less a roll call of the diseases that define later life.
Blood already told us this
Here is the part that should make the finding feel less exotic. Haematologists have been staring at mosaic ageing for over a decade without calling it that.
In 2014 two teams published back-to-back papers in the New England Journal of Medicine describing what is now called clonal haematopoiesis. Sequencing blood from 17,182 people who had no known blood disorder, Siddhartha Jaiswal's group found cancer-associated mutations expanding quietly in a fraction of blood cells.[3] The prevalence climbed steeply with age: rare below 40, then 9.5 percent of people aged 70 to 79, 11.7 percent of those aged 80 to 89, and 18.4 percent of people aged 90 to 108. A parallel Swedish analysis of 12,380 people found the same pattern — roughly 10 percent above 65 versus 1 percent below 50.[4]
These were not sick people. Their blood counts were normal. Yet carrying one of these clones was associated with an 11-fold higher risk of blood cancer, and — the finding that genuinely surprised the field — roughly double the risk of coronary heart disease and 2.6 times the risk of ischaemic stroke.
The point for our purposes is structural. A standard blood panel on those people came back clean, because a panel measures the average. The risk lived in a minority population of cells that the average could not see. The methylation study is describing the same geometry, one layer down and across every tissue rather than just blood.
That story has kept developing. A 2025 NEJM paper found these blood clones physically infiltrating solid tumours — present in 42 percent of lung cancer patients who carried them — and independently predicting worse outcomes.[6] Small rogue populations, large consequences.
What this does to your test result
Steve Horvath's original 2013 clock was built from 8,000 samples spanning 51 tissue and cell types, reading 353 CpG sites.[7] It was, and remains, a genuinely remarkable piece of work. But like every clock that followed it, it was trained on bulk tissue. The input is a blended signal from millions of cells at once.
Which means a bulk clock cannot distinguish between two very different bodies. In one, every cell has drifted modestly. In the other, ninety percent of cells are pristine and ten percent are far down an accelerated path toward the immune, tumour and neurodegeneration signatures the new paper flagged. Both can return the same epigenetic age. Only one of them is the situation you would want to know about.
This sits on top of problems the field already acknowledges. Idan Shalev at Penn State and Abner Apsley at the University of Illinois wrote earlier this year that dozens of clocks now exist, that they frequently disagree on the same person, that saliva and blood samples can produce substantially different results, and that no gold-standard method exists across laboratories.[5] Their verdict was that these tests are useful for researchers comparing groups and not yet useful for individuals comparing themselves. The mosaic finding adds a reason that is not about measurement noise at all — it is about what the average is capable of representing in principle.
The honest limits
Some restraint is warranted, and the paper invites it.
This is a computational reanalysis of existing single-cell methylation datasets, not a fresh clinical cohort. Single-cell methylation sequencing is still sparse and technically noisy — you recover a fraction of the genome from each cell — and separating true biological variation from that noise is the central methodological difficulty. Much of the underlying data is mouse. And the human illustration that made the coverage, a comparison of black and white hairs from one 53-year-old man, is a vivid demonstration rather than evidence of anything on its own.
Most importantly: nobody has shown that measuring the fast-ageing fraction predicts disease better than the bulk number does. That study has not been run. The claim on the table is that the current metric is structurally blind to something real, not that a better metric already exists.
It is also worth saying that this does not obviously hand you a new thing to buy. If you were hoping the answer is a supplement that targets the fast-ageing minority, there is no such product and no trial supporting one. The most closely related idea — selectively clearing a small population of damaged cells — is what senolytics are attempting, and that field is still short of convincing human outcome data.
So what should you actually take from it
Two things, we think.
The first is a recalibration of confidence. If you have paid for a biological age result and been quietly pleased or quietly alarmed, the appropriate response to a three-year gap in either direction is mild interest, not a change of plan. The number is an average of a mosaic, measured with a method that has no cross-lab standard. Treat it as a soft signal. Our own review of what biological age reversal claims can support reaches a similar place from a different direction.
The second is more interesting. If ageing really is driven by a minority of cells that diverge from the rest, the long-term therapeutic target changes shape. It becomes less about slowing a uniform process everywhere and more about finding and dealing with outliers — which is, notably, how oncology already thinks. That is also the frame in which partial cellular reprogramming starts to look like a plausible mechanism rather than a headline, and it fits what we know about how the immune system's own decline concentrates in particular cell populations.
None of that changes what you do on a Tuesday. The unglamorous inputs — sleep, resistance training, aerobic base, not smoking — still have vastly better evidence behind them than anything in this article. What has changed is our picture of the thing they are acting on. It was never one clock ticking. It was several billion, running at different speeds.
Common questions
Do all cells in the body age at the same rate?
No. The 2026 Nature Communications analysis found that most cells stay relatively young while a smaller subset accumulates age-related methylation much faster, with the fastest-dividing cells leading. A tissue is better pictured as a mosaic of biologically younger and older cells than as a population ageing in step.
Does this mean biological age tests are useless?
Not useless, but easy to over-read. Epigenetic clocks were built on bulk tissue, so the number is an average over millions of cells and cannot reveal whether a fast-ageing subset is present. Add that dozens of clocks exist and often disagree on the same person, that saliva and blood can give different answers, and that no cross-laboratory standard exists, and the result is a soft signal rather than a diagnosis.
What is polycomb CpG island methylation?
Polycomb CpG islands are DNA regions near genes that a cell keeps switched off but ready to use. With age they gradually gain methyl groups, locking them shut. This drift is one of the most consistent molecular signatures of ageing and underpins several epigenetic clocks — and the single-cell data show it concentrates in a subset of cells rather than accumulating evenly.