An AI agent proposed and refined aging modules under a frozen protocol. Of 7 concepts, 2 reach nominal significance (ECM stiffening +0.031, TRM efferocytosis +0.021); none survive FDR. Replay is public: steeramed.com/en/bench/… I’m rarely on LinkedIn — email: jianghui@deepome.com
Our claim: 332 capability modules are the right unit for steering human health. Any such claim has to lose somewhere. Next three posts: where a 117-axis panel beats the atlas, where we drop to chance, and which axes have no handle.> I’m rarely on LinkedIn — email: jianghui@deepome.com
Put a human on Mars, Europa, Kepler-22b — which capability reserves get remodeled first? A pure test of deduction: no new data, only whether the reasoning survives published biology. Which reserve would you bet breaks first? I’m rarely on LinkedIn — email: jianghui@deepome.com
Which representation ranks repurposing drugs best? It depends on the disease: Hallmarks won cancer (recall@20 0.900), food-as-medicine won depression, TCM won sedatives; the full atlas was strictly best in only 9 of 23: https://www.preprints.org/manuscript/202608.0998/v1
Everyone is building virtual cells. We’re after the missing layer between cell and patient — emergence, pathway compensation, and whether this person responds to this intervention. The figure contrasts the two: https://steeramed.com/en/
We built SteeraMed Bench to answer one question: can an AI-generated “aging module” actually predict drug effects? The setup: 332 modules × 1,916 repurposing drugs × 23 disease categories, one frozen protocol for everyone. The diagram shows the instrument: https://steeramed.com/en/bench/
Our new preprint maps human health as 332 capability modules — 117 nutrient, 105 food-medicine, 72 aging-hallmark, 38 syndrome axes. Figure 1 shows how the atlas and its frozen evaluation protocol fit together: https://www.preprints.org/manuscript/202608.0998/v1
The AI virtual cell is fashionable. But a patient is not a bigger cell — between them sit emergent transitions, pathway compensation, heterogeneity. Missing is not a larger model but another scale of map, plus action semantics and a retest loop a virtual cell doesn’t have.
Biomedical knowledge lives in three forms: travelogues (reviews), museums (databases), maps (computable knowledge). Astronomy has star catalogs, chemistry the periodic table. Biology has the largest data output in science — and still no coordinate system of its own.
Our starting axiom: a body is a set of capabilities for adapting to its environment. Not a gene list, not a network diagram. Everything downstream — capability axes, response mapping, module panels — is deduced from that one sentence, then tested.
Systems biology tells us how a living system runs. A world model tells us how it moves under an action. A steerable model asks a third question: how does the rider get there without breaking the horse? Not a rebrand — a different task.
Two hundred years of biomedicine ran on induction: collect data, find patterns, explain. AI is usually deployed to make that loop faster. We are trying the other direction — deduce from first principles, then let data refute it: https://steerable.world
SteeraMed Bench: 332 module axes, 1,916 DrugBank small molecules, 5 disease tasks, 23 ATC-derived categories. Versioned snapshots, published failure modes, open nominations: steeramed.com/en/bench/
Why modules: one gene is too small a unit to steer a body, a whole transcriptome too large to act on. Our atlas has 332 axes — 117 nutrient targets, 105 food-medicine compounds, 72 extended aging hallmarks, 38 syndrome proxies from Chinese medicine.
Most AI-for-biology work asks what will happen. We ask a narrower question: which knob, turned how far, moves a person — and what measurement would show it. Prediction without a handle is not yet medicine.
New blog: notes on steerable biology — what can actually be changed in a human body, and how we would know it moved. Method notes, failure boundaries, and open problems from building SteeraMed Bench: steeramed.com/en/bench/