Ben Okonkwo: Marcus, did you sleep okay this week? I ask because it's — relevant. Sort of.
Marcus Vale: Terribly. Why is it relevant?
Ben Okonkwo: Because your DNA methylation pattern shifted while you didn't sleep, and there are people who would now charge you two hundred dollars to tell you how old your cells think you are. Which is — that's today's episode.
Marcus Vale: Huh. Okay. Epigenetic clocks.
Ben Okonkwo: Epigenetic clocks. And the thing I want to sit with before we get into the commercial layer is just — the foundational disorientation of the premise. Which is: the calendar is a genuinely bad instrument for measuring aging. Steve Horvath demonstrated that in 2013 using 353 CpG sites — methylation markers — that could estimate biological age across all tissues from a single model. What fell out of that was the finding that epigenetic age acceleration, the gap between your DNA methylation age and your chronological age, predicts mortality more strongly than birth year alone. A 40-year-old with accelerated epigenetic aging can have a higher mortality risk than a 60-year-old who doesn't.
Marcus Vale: That's the number that stops people cold. The 40 and 60 thing.
Ben Okonkwo: Right — but here's what doesn't follow automatically from that. Predicting who dies first is not the same as understanding why. The Horvath clock is optimized for chronological age, not mortality. The second-generation clocks — GrimAge, PhenoAge — those are the ones actually trained on health outcomes. And now I'm already in it, which means we should probably just go.
Marcus Vale: Let's go. Because the move from measuring age to predicting mortality — that's not a refinement. That's a completely different scientific claim. And I don't think the field paused long enough to acknowledge that.
Ben Okonkwo: Right — but before mortality prediction, we need the plain mechanism, because I think people hear 'epigenetic clock' and imagine something exotic. It's actually almost mundane once you see it. Your DNA has millions of specific sites — CpG sites — where a methyl group can attach to a cytosine base. Think of it like rust accumulating on a metal beam. Predictably, over decades, certain spots rust and others don't. Horvath's insight was: that rust pattern is a calendar.
Marcus Vale: And the number that gets me — 353 sites. Out of how many potential CpG sites in the genome?
Ben Okonkwo: Millions. The Illumina MethylationEPIC array reads something like 850,000 of them. Horvath used a penalized-regression model — basically a very disciplined filter — and landed on 353 that did the job across every tissue type he tested.
Marcus Vale: Every tissue. That's the actual shock, right? Your liver and your neurons are — I mean, those are almost completely different biological environments — and the same 353 sites read the same age.
Ben Okonkwo: That universality is what made the 2013 paper land the way it did. Nobody expected one model to generalize across cell types. The working assumption had been that different tissues age differently at the epigenetic level. And they do in some ways — but the shared signal was strong enough that a single statistical model could cut through all of it.
Marcus Vale: Okay, so the surface story is: methylation drift is a clock because it accumulates predictably. Clean. Intuitive. Now what's wrong with it.
Ben Okonkwo: The complication is — and this is the one that actually stops me — predictable accumulation doesn't tell you whether the accumulation is causing anything. A clock tells you time is passing. It doesn't tell you why time passes. There's a paper in Nature Aging showing that you can build an accurate aging clock from purely simulated data. Random, meaningless methylation drift. Stochastic noise. And the model still works.
Marcus Vale: Wait — you can train on fake drift and still predict age accurately?
Ben Okonkwo: That's what the paper showed. Which means — okay, now I have to say this carefully — it doesn't prove the clock is meaningless. But it does mean that prediction and mechanism can come completely apart. Roughly 28% of the genome shows methylation drift with age. That's an enormous signal. But some of it, maybe a lot of it, could just be noise accumulating at scale.
Marcus Vale: That's — frankly that changes the whole downstream claim. Because if the drift is partly stochastic, then you're not reading biological programming. You might be reading biological static.
Ben Okonkwo: Exactly the problem. And the field kind of acknowledged this and then — shipped the product anyway. GrimAge, PhenoAge, the whole second generation optimized for mortality. They moved the target from 'estimate birth year' to 'predict who dies first.' Which is commercially smarter, scientifically more defensible. But the underlying question — is methylation a cause, a symptom, or just a shadow — that question didn't get answered first.
Marcus Vale: A shadow of aging, not aging itself. That's the bet the whole consumer market is currently making without knowing it's a bet.
Ben Okonkwo: And that shadow framing is exactly where I want to slow down, because — okay, so the shadow gets more complicated when you realize the field didn't produce one better instrument. It produced three different instruments that are all being sold as 'your biological age.'
Marcus Vale: Three different clocks, three different outputs.
Ben Okonkwo: Right. And they're solving three genuinely different problems. Horvath's original clock — that's answering 'what year were you born,' basically. GrimAge and PhenoAge, those are retrained on mortality data, health outcomes. They're answering 'when do you die.' And then DunedinPACE — which came after — that's not a cumulative estimate at all. It's a rate. It's asking: how fast are you aging right now.
Marcus Vale: A speedometer versus an odometer.
Ben Okonkwo: Exactly that. And the reason that distinction matters — actually, let me make it concrete. Imagine a 52-year-old accountant. Normal blood pressure, exercises three times a week. She takes all three tests. One comes back and says her biological age is 48. Another says her mortality risk profile looks like a 55-year-old. And DunedinPACE says she's aging at 1.2 years per calendar year.
Marcus Vale: Those are not the same finding.
Ben Okonkwo: Not remotely. And nobody sent her a guide for which one to act on. The first one says she's younger than her birth certificate. The second says she's at higher mortality risk than she looks. The third says the trajectory is accelerating. Those can all be simultaneously true — or they can be pointing at three different biological stories. We actually don't know.
Marcus Vale: The GrimAge and PhenoAge pivot though — I mean, that's the move that actually interests me commercially. Retrain on mortality instead of birth year, and suddenly you're not in the 'novelty biomarker' category. You're in clinical risk prediction. That's a completely different market.
Ben Okonkwo: It is — but the load-bearing assumption is whether they outperform what a cardiologist already has. Because a risk calculator built on age, smoking, cholesterol — that's not exotic. If GrimAge doesn't beat that on actual patient outcomes, what's the methylation array adding?
Marcus Vale: That's the proof point I'd want. Specific study, head-to-head, real endpoints.
Ben Okonkwo: DunedinPACE sidesteps that comparison slightly, because it's not claiming to predict mortality. It's claiming to measure rate. Which is almost harder to validate, because you'd need decades of follow-up to know if the speedometer reading at 52 actually predicted the trajectory.
Marcus Vale: Huh. So the most intuitive framing — 'are you aging faster or slower' — might be the hardest one to actually test.
Ben Okonkwo: And that tension gets worse before it gets better — because there are two findings sitting right at the center of the field right now that point in completely opposite directions, and we haven't touched either of them yet.
Marcus Vale: Those two findings — and they don't resolve. That's the thing. One of them is a paper in Nature Aging showing you can build an accurate aging clock from purely simulated, random methylation noise. No biology. Just stochastic drift accumulating. And the model still predicts age.
Ben Okonkwo: Stochastic noise. Meaning the signal the entire field is reading might be — okay, I want to say this carefully — it might be mathematically real and biologically empty at the same time.
Marcus Vale: Which breaks DunedinPACE specifically. Because the speedometer framing only works if there's a real biological rate underneath it. If the signal is noise accumulating at a predictable pace — the speedometer is reading static.
Ben Okonkwo: That's — yeah, that's the one that actually unsettles me. Because DunedinPACE was designed to be the most actionable clock. The rate instrument. And if the rate it's measuring is partly stochastic drift, then — I mean, what experiment would even distinguish those two things?
Marcus Vale: Now here's the countervailing fact. Semaglutide.
Ben Okonkwo: The GLP-1 trial.
Marcus Vale: Randomized clinical trial. HIV-associated lipohypertrophy patients. First high-profile RCT showing a pharmacological agent — semaglutide specifically — actually shifts the epigenetic clock reading. Measurably slows it.
Ben Okonkwo: And that's the cut — because reversibility is supposed to be the proof. If you can move the reading with a drug, the reading must mean something real. Except — actually, no, that logic doesn't hold. Reversibility could just as easily mean the clock is measuring something downstream from the actual damage. You shifted the shadow, not the thing casting it.
Marcus Vale: So semaglutide moves the number, and we can't tell if we moved aging or just the meter.
Ben Okonkwo: Right. And the trial population is narrow — HIV-associated lipohypertrophy. That is not a generalizable gerotherapeutic claim yet. The large-scale trial that would tell us whether semaglutide slows aging in a healthy 55-year-old has not been run.
Marcus Vale: Frankly that's the commercial problem in one sentence. The longevity investor community is already pricing in the GLP-1 aging thesis off a single narrow RCT.
Ben Okonkwo: And here's what doesn't get said enough — the stochastic paper and the semaglutide RCT don't cancel each other out. They can both be true. The clock could be partially noise and still be sensitive enough to register a drug effect. Those findings just leave you with a much murkier claim than either camp wants to make.
Marcus Vale: Which is — basically the whole field in one tension. Accurate prediction, uncertain mechanism, one drug that moves the needle in a specific population. That's what's actually for sale right now.
Ben Okonkwo: And the clock can be right for the wrong reasons. That's the sentence nobody puts in the consumer brochure.
Marcus Vale: And that's actually where the infrastructure problem sits. Because we're not talking about a few boutique tests anymore — this is proliferating. Consumer epigenetic age testing is scaling commercially right now, before the scientific community has any consensus on what actionable decision a DNAm age readout should actually drive.
Ben Okonkwo: Right — and the population generalizability piece makes it worse. The iCAS-DNAmAge clock was developed specifically for Chinese cohorts because the clocks trained predominantly on European populations perform differently in other ancestry groups. That's not a minor caveat. We're about to sell biological age measurement globally and we haven't systematically validated whether the methylation signatures even replicate across populations.
Marcus Vale: A global product built on a European training set. That's — I mean, that's the generalizability gap dressed up as personalization.
Ben Okonkwo: And what lands at the end of all of it — the unglamorous truth — is that clinicians looking at a DNAm age readout still don't have a clearer path to extending healthspan than exercise, diet, sleep, stress management. That's what the interventions keep pointing back to. The measurement infrastructure is genuinely sophisticated. The actionable output is... not.
Marcus Vale: The basics. Which nobody needed a methylation array to discover.
Ben Okonkwo: No. And the thing I keep turning over — actually, I don't know if I've fully resolved this — is the mirror image. We may have built the most sophisticated molecular mirror ever pointed at human aging. And what it reflects back might be noise dressed as signal. Or signal we don't yet know how to read. I genuinely can't tell which. And I'm not sure the field can either, right now.
Marcus Vale: You started by asking if I slept okay. And whether my methylation shifted because of it.
Ben Okonkwo: I did. And I think the honest answer we've landed on is — it probably did shift, something in that pattern changed, and we have no idea what to tell you to do about it beyond the thing your grandmother already knew.
Marcus Vale: Sleep better. Which I'm going to try. Genuinely appreciate the thinking today.
