Cyrus Reed: Iris, genuine question before we start — if someone handed you a number that said your body was biologically three years younger than last year, would you trust it?
Iris Holm: Depends entirely on what the number measured.
Cyrus Reed: Which is — okay, that's the whole thing. Because people are getting these results right now. Commercial tests, someone does caloric restriction for a year, score goes from 50 to 47, they read it as: my body got younger. And the test is measuring something real — DNA methylation, chemical flags on your genome at specific sites that shift with age — so it's not nothing. But 'my body got younger' and 'a pattern of chemical flags changed' are not obviously the same claim.
Iris Holm: What's the flag analogy actually buying us here?
Cyrus Reed: So — your DNA stays fixed your whole life, right, the sequence doesn't change. But sitting on top of it are these methyl groups, little chemical tags at cytosine-guanine spots, CpG sites. They influence how genes get read without rewriting the code. And as you age, that pattern drifts in a way that's predictable enough that Steve Horvath, in 2013, trained a model on 353 of those sites and could guess your chronological age within about 3.6 years across multiple tissue types.
Iris Holm: Chronological age. Not health. Not mortality. Your birthday.
Cyrus Reed: Exactly — and that's the pivot I want to sit in for a second, because it's doing a lot of work. He built a clock that tells you how old you are. The clocks that came after it — GrimAge, DunedinPACE — are trained on completely different targets. Mortality risk. Pace of physiological decline. And we're using all of them under the same umbrella word: biological age. But they're not measuring the same thing.
Iris Holm: Measuring age versus predicting death. That's a category jump, not a refinement.
Cyrus Reed: And nobody tells you that when you open the app and see your score dropped three years.
Iris Holm: So the category jump is invisible because the raw signal is identical. Same CpG sites, same methylation read, different question being asked of the data.
Cyrus Reed: That's exactly the trap — okay, so Horvath's 2013 clock, first generation, it's asking: given this methylation pattern across 353 sites, what year was this person born? That's it. It's almost forensic. Then GrimAge comes along, second generation, and it's trained on mortality outcomes — who died, when — and PhenoAge is trained on aging phenotypes, disease risk. And then DunedinPACE, which is — wait, this one's different in kind, not just degree — it's not giving you a static biological age number at all. It's measuring how fast you are aging right now, like a speedometer versus an odometer.
Iris Holm: Speedometer versus odometer. That's actually clarifying.
Cyrus Reed: And if they're all called epigenetic clocks — I mean, why are they? Is it just branding? Because a researcher who sees their GrimAge score drop one year is looking at roughly a 10% reduction in associated mortality risk in some cohorts. That's a health intervention framing. Someone whose Horvath number drops is just... closer to their birthday again, technically. Those are not the same celebration.
Iris Holm: Okay — why the shared name, then?
Cyrus Reed: Because the input is genuinely the same. DNA methylation at CpG sites — that's the raw material for all of them. Horvath, Hannum, GrimAge, DunedinPACE — all reading the same layer of the genome. The divergence is in what you trained the model to predict. And that distinction just... doesn't survive the journey from the paper to the press release.
Iris Holm: Picture someone doing a caloric restriction protocol — say they're six months in, they get a new blood draw, they see their epigenetic age number drop. Which clock are they even looking at? Does it matter?
Cyrus Reed: It matters completely — because if it's first-generation, the drop is saying almost nothing about mortality. If it's GrimAge, the drop is tied to actual survival data. If it's DunedinPACE, the number isn't even a year count, it's a pace per year of biological change. Those require totally different interpretations of the exact same felt experience of 'my score went down.'
Iris Holm: And epigenetic age acceleration — the gap between what the clock reads and your actual age — that's the metric carrying the clinical weight. Positive acceleration linked to frailty, disease risk, all-cause mortality. But only if the clock you're using was trained on those outcomes.
Cyrus Reed: But here's the part that breaks the clean version of that — because it's not just caloric restriction doing this. There's a Frontiers in Genetics study that basically compiled everything shown to move next-generation clock readings in humans, and the list is... wait, okay: exercise, caloric restriction, plant-rich diet, omega-3s, semaglutide, pitavastatin, multivitamin-multimineral supplements, umbilical cord plasma, ketamine. Ketamine and a multivitamin are on the same list.
Iris Holm: Those have nothing in common mechanistically.
Cyrus Reed: Nothing! Rapamycin hits mTOR signaling, ketamine is working on — actually, I don't even know how to place ketamine in an aging framework, that's a dissociative anesthetic operating on completely different neural pathways. And yet they all lower the same score. Which means either there's one unified aging signal underneath all of it, or — and this is the one that unsettles me — the clock is measuring something generic enough that almost anything stressful or restorative can shift it.
Iris Holm: Generic enough that a biomarker of everything is effectively a biomarker of nothing specific.
Cyrus Reed: And here's the concrete version of why that matters — picture a 52-year-old runner in Oslo. Six months of caloric restriction, she gets a new blood draw, clock drops three years, clinic celebrates. But nobody ran her cardiac output. Nobody checked inflammatory markers. The score moved. Whether her biology moved is — no one knows.
Iris Holm: And that's the TranslAGE finding, right? They evaluated 18 different clocks and found meaningful inconsistencies across technical replicates — same blood sample, run twice, different answer.
Cyrus Reed: Which means the Oslo runner's three-year drop might partially be measurement noise. Not biology. And none of those intervention studies — not the semaglutide work, not caloric restriction — demonstrated clock reduction alongside improvements in tissue function or survival. Biomarker movement, full stop.
Iris Holm: We're lowering a number we don't trust to stay put when you retest it.
Cyrus Reed: And the deeper problem — the one we haven't touched yet — is whether moving the needle actually touches the underlying mechanism at all. Horvath has a framing for this, face versus clockwork, that I think reframes everything we just said, but we'll get there.
Iris Holm: Face versus clockwork. Horvath's own language — and it's the most honest thing said about this field. Resetting the face of a clock doesn't move the hands of time underneath it.
Cyrus Reed: Right — but wait, what would it even look like if you did move the clockwork? Because that's the actual test nobody's run. Like, imagine a researcher — she drops her GrimAge score by two years, which maps to roughly 20% reduced mortality association in some cohorts. Is there any tissue-level evidence her biology changed? Inflammatory load, organ function, anything?
Iris Holm: That's exactly the gap. Longitudinal outcome data — did people whose clock scores dropped actually live longer, get sick less, show measurable tissue-level improvement. That evidence doesn't exist yet.
Cyrus Reed: No way — none of it?
Iris Holm: The 2025 Lancet Healthy Longevity meta-analysis — it strengthened the predictive case. EAA links to frailty, multisystem physiological decline, across cross-sectional and longitudinal data, the InCHIANTI cohort included. But prediction is not mechanism. Better correlation is not a causal chain.
Cyrus Reed: So frailty as the downstream validator — that's real. The clock reading connects to actual clinical decline, not just a number. But connecting A to C doesn't tell you B is the lever you pull.
Iris Holm: GrimAge predicting 10% higher mortality per year of acceleration — that's powerful. It is not proof that reversing the acceleration reverses the risk. Those are — frankly, those are different sentences.
Cyrus Reed: Which means someone celebrating a DunedinPACE drop after six months of caloric restriction is — I mean, is she celebrating slowing aging, or celebrating moving a surrogate biomarker that tracks aging but might not be aging itself?
Iris Holm: That's the question the field hasn't closed. And until it does, we're intervening on the face.
Cyrus Reed: And that's — wait, that's actually the thing I want to land before we stop. Because there's a gap nobody names out loud: population-level tools versus individual clinical translation. GrimAge is genuinely powerful at the cohort level, predicting who trends toward earlier death across thousands of people. But the commercial test you take Tuesday and retest Friday? That individual number has accuracy problems the population average papers over. And nobody selling you a kit mentions that.
Iris Holm: The question is: what would it actually take to close that loop. Not a clock score moving. Longitudinal outcome data — did the person age more slowly. Mechanistic evidence that the biology shifted, not the readout.
Cyrus Reed: Which nobody has yet. And honestly — sitting here — I think that's the right place to land. The clocks are real. GrimAge, DunedinPACE, Horvath's original 353 sites, all of it. Real signal. The mechanism is just... still open.
Iris Holm: Someone handed you a number. Your body, three years younger. We started there.
Cyrus Reed: And the honest answer is: still depends on what it measured, who trained it, and whether anyone tracked what happened next. Which is more than one question.
