Here is the contradiction most readers will miss: the impressive part of ALADYNOULLI is not that it predicts disease better than the calculators we already use. It is that the same machinery driving that accuracy is precisely what keeps the model out of the exam room.
The two things are the same thing, and one of them works today while the other doesn't yet mean anything a doctor can prescribe against.
Start with the gap it closes—the one no risk calculator could. For decades a physician could tell you that your Alzheimer's risk was elevated, or that your coronary risk was high. What no tool could tell you was when.
Eric Topol names the stakes precisely: we could estimate a patient's Alzheimer's risk, but "could not temporally place when it might occur. Age 65 or 95? It makes a huge difference."
We could estimate a patient's Alzheimer's risk, but could not temporally place when it might occur. Age 65 or 95? It makes a huge difference.
It makes a huge difference because it is the number that actually changes behavior. A risk that resolves at 95 is a footnote. A risk that resolves at 65 reorganizes a life—when you screen, when you intervene, how aggressively you treat.
ALADYNOULLI is the machine built to answer "when." Published in Nature by Sarah Urbut and colleagues at Mass General Brigham and Harvard, it was trained across three biobanks—the UK Biobank, Mass General Brigham, and All of Us—totaling more than 683,000 individuals, 348 diseases, and up to 52 years of follow-up.
The model is a Bayesian generative framework that jointly reads longitudinal diagnoses, age, and polygenic risk. Topol offers the cleanest metaphor for what it does: risk shifts continually "like a GPS navigation when you make a wrong turn and you are re-routed."
That dynamic re-routing is the first real capability. This is not a one-time score at your annual physical. As a life unfolds and new diagnoses arrive, the projected trajectory updates—and the paper reports that signatures were strongly predictive of future disease risk, particularly when predictions were updated over time as new patient information became available.
The second capability reorders how we think about a diagnosis at all. ALADYNOULLI reduces 348 diseases into 21 latent signatures—hidden patterns that capture the biological processes running underneath the labels.
And a single diagnostic label can decompose into distinct signatures. Breast cancer, for instance, could reportedly load onto signature 8 or onto signature 7—the latter tied to inflammation and metabolic abnormalities, per Topol's account.
Urbut's framing of why this matters is worth quoting in full: "two patients with the same diagnosis are not the same patient: this model shows they often have different underlying signature profiles, which can translate into different progression patterns and different responses to the same treatment."
That last clause is where the promise gets sharp. The same signature machinery can reportedly flag likely medication failure—per Topol's account, predicting when SSRI treatment for depression is likely to fail.
Now the accuracy, which is genuine and decisive. Against the Pooled Cohort Equation for one-year coronary risk, ALADYNOULLI reportedly posted 0.89 versus 0.68 for one-year coronary risk, per Topol's account (metric not specified as AUC). Against the Gail model for breast cancer, reportedly 0.783 versus 0.54, per Topol's account.
Those are not incremental margins.
The signatures themselves are not statistical mirages. Across the three cohorts, the model recovered 21 replicable signatures with a median 80% cross-cohort composition preservation, and the biological subtypes within diagnostic categories separated at Cohen's d up to 4.25, with P ≤ 1 × 10⁻⁸ for 95% of comparisons.
They are also biologically concordant with what we already know. Carriers of familial hypercholesterolemia enrich in the cardiovascular signature; carriers of clonal hematopoiesis of indeterminate potential enrich in the inflammation signature. A signature-based genome-wide association study reportedly surfaced 151 significant loci, including cardiovascular associations that single-trait analyses missed.
So the discovery engine is real. The prediction is real. The temporal placement is real.
Here is the honest pivot.
"Signature loading" is not a clinical variable anyone prescribes against. No clinician today knows what to do with the fact that a patient's breast cancer loads onto signature 7 rather than signature 8.
The signatures are validated as statistically replicable and biologically coherent. That is a claim about the model's internal integrity and its concordance with known biology. It is not a claim that acting on a signature changes an outcome.
Those are different orders of evidence, and the distance between them is the whole ballgame.
Consider what "clinical utility" would actually require. It is not enough that signature 7 correlates with inflammation and metabolic abnormality. Someone has to show that knowing a patient sits in signature 7 leads to a different decision—and that the different decision produces a better result than the decision made without it.
None of that has been tested. The evidence establishes replicability and biological concordance; it does not establish that any signature-informed action defers, prevents, or better treats a single disease.
Topol gestures at the missing piece himself when he raises "counterfactuals, what-if scenarios for the individuals to reduce the risk, markedly defer a condition's appearance, or even fully prevent it." That is the aspiration. It is not a finding.
The specific missing experiment is a prospective trial. Take a population, generate signature-informed predictions and the "what-if" interventions those predictions imply, act on them in one arm, withhold them in another, and measure whether the predicted disease arrives later, or not at all.
Until that trial exists, ALADYNOULLI is an extraordinary description of how disease unfolds—and an unproven guide to changing it. Predicting when Alzheimer's arrives is not the same as pushing it from 65 toward 95. The model can do the first. Nobody has yet shown it can drive the second.
This is the same tension running through this issue's open thread on aging clocks: an early biological-clock movement is a measurement, not yet a validated surrogate endpoint, because no one has tied the movement to a deferred outcome in a prospective trial. Different tools, identical gap. Both are validated as measurements and unproven as levers.
The distinction is not pedantic, because the failure mode is specific. A model this fluent at telling a compelling story about a patient's future invites clinicians and patients to act on the story before anyone has shown that acting on it helps. Signature 7 versus signature 8 sounds like a treatment decision. Right now it is a research finding wearing the costume of one.
The value here is unusually clear even under that constraint. ALADYNOULLI has given us a validated map of how 348 diseases co-evolve across half a century, a dynamic risk estimate that re-routes as a life changes, and a way to see that the same diagnosis hides different biologies underneath.
What it has not given us—what nothing in the evidence claims it has given us—is a demonstration that any of this changes what happens to a patient. The clock is now on the risk. The trial that would put that clock in the exam room has not been run.