Jordan Hale: Can I just — I've been quiet about this for like two days and I need to say it out loud.
Ryan Castillo: That's slightly concerning. Go.
Jordan Hale: Your blood knows how old your neurons are. Not how old *you* are. How old your *neurons* are. Specifically. And also your astrocytes, and your muscle cells, and forty-something other cell types — all estimated from proteins floating around in your plasma. That's what this Nature Medicine paper is saying.
Ryan Castillo: The Wyss-Coray lab at Stanford. Seven thousand-plus proteins measured.
Jordan Hale: Seven thousand proteins. Across sixty thousand people. And you know what the study found — twenty to twenty-five percent of those people are aging faster than their chronological age in at least one cell type.
Ryan Castillo: Hold on — faster in *one* cell type. That distinction matters a lot.
Jordan Hale: No, totally — and that's exactly the weird part. Asynchronous aging. Your body isn't declining as one thing. It's more like... different parts are on completely different clocks. Neurons aging at seventy, immune cells aging at forty, same person.
Ryan Castillo: Which breaks the baseline assumption most people carry around. That aging is one unified process happening to one body.
Jordan Hale: Right — and chronological age, like, your birthday number, says nothing about any of this. Biological age is the functional molecular state of your cells, which can diverge substantially. That gap is what this whole study is built on.
Ryan Castillo: And validating that across three cohorts — including the UK Biobank at forty-four thousand people alone — that's not a fluke finding. That's a structural result.
Ryan Castillo: Walk through the actual machinery though, because I think people skip this step. You get a blood draw. One vial. And what the researchers are doing is measuring — okay, seven thousand proteins circulating in that plasma. Proteins that got there because cells shed them or secreted them into the bloodstream. Neurons, astrocytes, macrophages — none of those are in the blood, but they leave traces.
Jordan Hale: It's like reading a crime scene, kind of. The cell was never there, but it left something behind.
Ryan Castillo: That's exactly the move. And then machine learning models are trained to say — given this pattern of proteins, what's the functional state of, say, the astrocyte population? That output is what they're calling biological age for that cell type. You're never looking at an astrocyte. You're looking at what it left behind in your blood.
Jordan Hale: And they validated this across three completely separate groups of people, right? Give me the actual numbers because I want to hear them again.
Ryan Castillo: The Global Neurodegeneration Proteomics Consortium — fourteen thousand two hundred eighty-one people, SomaScan platform. UK Biobank — forty-four thousand four hundred fifty-eight people, Olink platform. And then the 1946 National Survey of Health and Development — one thousand eight hundred three individuals, back on SomaScan.
Jordan Hale: Wait, hold on. Two different platforms?
Ryan Castillo: SomaScan and Olink aren't interchangeable. They measure protein abundance differently — different antibodies, different detection chemistry. So when the UK Biobank cohort uses Olink and the other two use SomaScan, you're not just testing the biology across populations. You might be testing whether the model survives a measurement platform swap. And that's a live unresolved question.
Jordan Hale: Okay, yeah — I mean, Daisy Yi Ding posted a whole public thread on this, she's the Stanford co-author who's now at Khosla Ventures, and I read it twice and I still wasn't sure she fully put that question to bed. Like she summarized the findings clearly, but the cross-platform consistency thing, you know, she kind of... gestures at it?
Ryan Castillo: Right, but — and this is where I want to be precise — the Karolinska group's earlier foundational work on plasma proteomics aging, which this study builds on, also didn't fully resolve that. It's a field-wide tension, not just this paper.
Jordan Hale: Which makes me wonder — and this is maybe the uncomfortable version of your point — like, imagine a Tuesday morning, some forty-five-year-old woman gets a blood draw, the model spits out 'your astrocytes are aging at seventy.' Is that number robust? Or does it shift if the lab happens to run SomaScan instead of Olink? Because that's not a small question for her.
Ryan Castillo: And that's exactly what stays unresolved. You are never measuring the cell. You are measuring what it left behind. The age is a model output — change the platform, potentially change the number. The biology doesn't move. The inference does.
Jordan Hale: But okay, I need to — because we've been circling the measurement problem and I want to actually land on what breaks my brain about this. The APOE4 finding. Same person, same genome, right? Carriers show older astrocytes — which is the cell type tied to Alzheimer's risk — but younger macrophages. Like, their immune cells are aging backward while their brain cells age forward.
Ryan Castillo: That breaks the 'faster aging equals worse' story completely.
Jordan Hale: It does! And I don't know what to do with that emotionally. Like — you know, this person is walking around, she feels however she feels, and inside her body two completely opposite aging processes are running simultaneously. That's not a spectrum. That's a contradiction.
Ryan Castillo: Hold on — because here's where I get stuck. The astrocytes look older. But are they actually aging faster, or are they shedding differently because early Alzheimer's pathology is already underway? If amyloid is accumulating, astrocytes activate, they leak different proteins into plasma. The model reads that shedding pattern as 'older.' That's not aging — that's disease announcing itself.
Jordan Hale: That's — yeah. That's the causation problem.
Ryan Castillo: And vesriram — the immunologist on X — flagged exactly this. Confounding by cell-type shedding dynamics, vascular permeability changes. If the blood-brain barrier is compromised, you get different protein leakage. The model can't distinguish 'this cell is aging' from 'this cell is sick and venting.'
Jordan Hale: No, I know — but wait, actually, the signal appears fifteen years before diagnosis. Fifteen. If disease is already underway fifteen years out, fine, call it announcing. It's still useful.
Ryan Castillo: Is it useful if you can't intervene on the mechanism you don't understand?
Jordan Hale: I mean — yes? Like, cardiovascular risk scores don't tell you the exact causal chain either. Inflammation markers are model outputs. Stress doesn't have a direct biological address. We use proxies all the time in medicine and we act on them.
Ryan Castillo: IntegralAnswers made the sharper version of that point though — not that proxies are bad, but that no cell age was ever directly measured here. The biological age of an astrocyte in this study is entirely a model output. Change the training data, change the architecture, the number shifts. That's not a property of the cell. That's a property of the algorithm.
Jordan Hale: Okay, but does that make it meaningless? Across three independent cohorts, predicting ALS, predicting Alzheimer's, predicting mortality — that held.
Ryan Castillo: It makes it a strong predictor. I'm not disputing that. What I'm saying is prediction and causation are doing very different work, and conflating them is how you end up with a treatment that targets the wrong thing.
Jordan Hale: And we just — we don't resolve that here. Because you've got one person, APOE4 carrier, old astrocytes, young macrophages, and we genuinely can't tell you if her brain is aging toward Alzheimer's or if Alzheimer's is already aging her brain. That distinction is everything. And the study doesn't close it.
Ryan Castillo: Let's put a real person in the room. Forty-three-year-old, no symptoms, gets a blood draw through whatever consumer longevity panel is running on SomaScan this year. Report comes back — neurons aging at seventy, muscles at fifty, immune cells at forty. She walks into her GP's office with that printout. What does the doctor do?
Jordan Hale: Nothing. Like — I mean, nothing validated. There's no clinical protocol. No intervention that's been established in response to these signatures. So the doctor is holding a number that predicted Alzheimer's fifteen years out in a population study and has absolutely no FDA-cleared playbook for what comes next.
Ryan Castillo: That's the gap. And it's not a small one.
Jordan Hale: It's huge. And you know what makes it weirder — that number, 'your neurons are aging at seventy,' that is not a measurement of her neurons. The npj Aging commentary on this whole field says biological age, quote, exists only as the output of the algorithm claiming to measure it. It's defined by the training data and the model architecture. Her neurons were never in the tube.
Ryan Castillo: So she's being handed a self-referential definition and told to be scared of it.
Jordan Hale: Or reassured by it. Which is almost worse, I think?
Ryan Castillo: The number that matters here is actually the extreme tail. One to three percent of the sixty thousand people in this study showed accelerated aging in ten or more cell types simultaneously. Not one cell type. Ten. That's a different category.
Jordan Hale: Wait — ten or more? I knew the number but I keep forgetting how tight that cluster is.
Ryan Castillo: Exactly. One to three percent. So the multi-system accelerated ager is actually rare — which weirdly makes the whole thing harder to use clinically, not easier. Because most people in that twenty to twenty-five percent bracket are accelerating in one system. One. And we have no idea what to tell them.
Jordan Hale: And the Knight Initiative for Brain Resilience — Wyss-Coray's group at Stanford, this is all sitting under that umbrella — they're not claiming clinical utility yet. Like that's not what the paper says. But the moment this kind of data hits a consumer platform, that distinction evaporates.
Ryan Castillo: Which is exactly the problem. Because the study is a population-level predictor. It says — in aggregate, this signature correlates with disease over fifteen years. It does not say: this specific woman's neurons are going to fail. Those are not the same claim.
Jordan Hale: And she doesn't experience aggregate. She experiences Tuesday morning, sitting across from a doctor who's — you know, who's never seen this printout before, who has no intervention to offer, and who is now trying to explain that the number predicting her risk is, technically, not a measurement of anything that was ever directly observed.
Ryan Castillo: No validated diagnostic test. No established intervention. And the 2019 Nature Medicine work — the earlier Wyss-Coray work on non-linear plasma proteome changes across the lifespan — that already showed aging isn't a smooth curve. It's undulating, it spikes at certain ages. We've known the biology is messy. This study makes that messiness cell-type specific. Which is more precise and somehow more actionable information in research and less actionable information in a clinic.
Jordan Hale: That's — yeah. That tension is real.
