Split one tube of blood across a dozen labs, hand each portion to the same principal-component epigenetic clock, and the answers will land within a fraction of a year of each other. That is the reassuring picture: a test so reproducible that a dozen operators cannot make it disagree.
Now draw a second tube from the same person after breakfast. On most clocks, the number jumps—and it is that number, the before-versus-after delta, that decides whether your anti-aging drug "worked."
This week the crack has a name and a location: not accuracy, not precision, but a specific kind of reliability that most clocks fail—and that failure sits directly under the open question of whether early clock movement can serve as a trial endpoint at all.
The load-bearing distinction comes from a new bioRxiv benchmark out of Albert Higgins-Chen's lab at Yale, which ran 18 epigenetic clocks through the TranslAGE platform. It separates two things the field has quietly conflated.
Technical reliability is agreement across replicates of the same blood sample. Biological reliability is agreement across repeated samples from the same person over hours or days—the interval in which meals, stress, circadian rhythm, and minor exposures happen.
On technical reliability, nearly every clock passed. On standard EPIC and 450K replicate assays, the whole panel achieved excellent reproducibility; PC-based clocks like PCGrimAge and SystemsAge stayed robust even when slide position and DNA extraction protocol were perturbed.
Then the floor gave way. Across repeated biological samples, most clocks showed only moderate stability. PCGrimAge was the only clock to clear the good-reliability bar, ICC above 0.75.
The critical finding is that the two properties are unrelated. Technical robustness did not predict biological robustness; clocks engineered to ace the replicate test often proved biologically unstable anyway.
Technical robustness did not predict biological robustness; clocks engineered to ace the replicate test often proved biologically unstable anyway.
The meal result is the one to sit with. Samples taken before and after eating produced very different readouts in the vast majority of clocks. Only the PC version of the original GrimAge stayed in the good range; GrimAge version 2 was very poorly reliable, and PhenoAge and the original Horvath clock slid into the moderate zone.
Acute stress was worse. Under stressful conditions, not a single clock rated good. Pollution exposure degraded many into the good and moderate ranges, and even an altitude change reportedly introduced unreliability—including, in that case, in PC GrimAge.
This is not a rounding-error problem. The authors report that some clocks are so unstable their z-scores swing "across the full significance spectrum," meaning the same biological reality can be reported as significant aging, no change, or significant rejuvenation depending on when you happened to draw blood.
Older work already sketched the magnitude. In two healthy men sampled 24 times over three months with no intervention, PhenoAge reportedly ranged by at least 8.23 years, with a maximum single-day swing of 6.53 years.
Here is why this is a design flaw rather than a footnote. A biological-age readout is being proposed as a surrogate endpoint—the number that stands in for whether an intervention slowed aging. A surrogate that moves more with lunch than with the drug is not measuring the drug.
The authors state the trap plainly: "if a meal can dramatically impact a score, then an intervention study that did not control for fasting status pre- and post-intervention may detect changes that are mistakenly attributed to the intervention but actually reflect trivial changes due to meals."
They then demonstrate the damage rather than just assert it. Linking each clock's reliability to its downstream behavior, they show clocks with higher ICCs produced more stable prognostic associations and more consistent responsiveness to a real vegan-diet intervention, while unreliable clocks yielded highly variable or spurious effects.
Read that against a trial with no standardized fasting window at both timepoints. A clock could credit the diet for what a participant's breakfast did—or hide a real effect under meal noise. The effect size becomes a function of catering, not biology.
There was an obvious-seeming fix, and it backfired completely. Immune cell counts fluctuate with stress and meals, so the authors tried statistically controlling for cell composition to stabilize the readouts.
In nearly every clock and nearly every condition, the correction made reliability worse. That inversion forced a strong conclusion: "changes in immune cell composition appear to reflect meaningful biological processes that are reproducibly detected by DNA methylation aging biomarkers."
In other words, part of the "noise" was signal. You cannot subtract your way to stability, because the fluctuating component carries real biological information—which is exactly why standardized collection, not post-hoc adjustment, is the answer.
The stakes are concrete because the design is already in the field. The PROMETHEUS trial protocol, from Andrea Maier and colleagues, describes a precision-geromedicine approach that tailors exercise, nutrition, sleep, cognitive training, and nutraceuticals to each participant's biological aging profile.
Individualized dosing keyed to each person's aging profile is only as sound as the profile. If the underlying clock can move years between a fasted and a fed draw, then the "profile" driving someone's personalized regimen may partly encode when they last ate—and two people with identical aging could be sorted into different arms by breakfast timing.
None of this is a reason to abandon clocks, and the authors do not argue that. Biological instability is not unique to methylation. Inflammatory markers move with exercise, hematocrit moves with hydration, and, per lifespan.io, some proteomic and even blood dementia markers vary by time of day.
The point is that we already know how to handle assays like this: we regulate the pre-analytic conditions. Nobody reports a fasting glucose off a post-lunch draw. Clocks have simply been used as though collection context were irrelevant, when the evidence says it is decisive.
So the fix is already on the table, and it is unglamorous. Mandate standardized collection windows—fasting status and time of day fixed and matched pre- and post-intervention. Select clocks vetted for biological reliability rather than technical reliability alone, which on this benchmark reportedly narrows the field dramatically. And build an evidence framework that treats a clock endpoint like a regulated assay.
That evidence gap is the deeper problem. As precision-prevention tools outrun the frameworks meant to evaluate them, the temptation is to accept a technically clean, reproducible-looking number as validation. This benchmark shows why reproducible is not the same as trustworthy.
The honest reading of the 18-clock result is narrow and sharp. On current evidence, one clock—PCGrimAge—cleared the biological-reliability bar, and most did not. Until trials standardize collection and choose their clocks on that basis, many clock-based surrogate endpoints are measuring the morning, not the aging.
The caveats belong here too. Much of the perturbation data came from young-adult cohorts, and the study lacked separate control groups to disentangle a specific stressor from the mere passage of hours. The authors call for exactly that follow-up, plus next-generation clocks designed to be insensitive to short-term swings.
Until those clocks exist, the discipline is procedural, not technological. Fix when you draw the blood, prove the clock is stable across a real day, and only then let its movement stand in for aging. Skip those steps and the endpoint is measuring lunch.