Everyone is building bigger AI models for biology. But here's a question nobody seems to be asking:
Can you steer it?
The Problem with Bigger Predictors
The field of AI for biomedicine is exploding. Virtual cell models, drug-response predictors, biological foundation models — billions of dollars are flowing into systems that aim to model cells, drugs, disease progression, and human biology.
But almost all of these systems share a critical limitation: they predict, but they cannot be steered.
A biomedical world model should not merely forecast what may happen next. It should allow a clinician or researcher to ask:
"What if we move in this direction instead?"
And then provide a reliable, auditable answer.
Introducing SEWO: Steerable Medicine World Model
DeepoMe Limited has released a new preprint: "World Models for Biomedicine: A Steerability Framework", introducing SEWO — a conceptual framework that proposes steerability as a foundational property for trustworthy biomedical AI.
📄 Preprint: https://doi.org/10.20944/preprints202605.0366.v1
SEWO is not another neural architecture. It's a meta-level framework — a specification layer that helps evaluate whether any biomedical world model (transformer, graph network, state-space model, or future architecture) is not only predictive, but also interpretable, constrained, counterfactual, and steerable.
The Rider and the Horse
Think of it this way. A rider doesn't micromanage every muscle of the horse. The rider provides directional signals through the reins. The horse maintains balance, adapts to terrain, and moves with its own embodied robustness.
Likewise, a steerable medicine world model should:
- Accept directional guidance from human experts (add a therapeutic hypothesis, modify a nutritional condition, remove a confounding assumption)
- Maintain internal consistency despite noise, missing data, and distribution shifts
- Make its reasoning inspectable at every step
Five Structural Constraint Points
SEWO defines five constraint points that any biomedical world model should satisfy:
1. State Representation
Biological states should be decomposed into modular, interpretable components — specifically, modular Intrinsic Capability (mIC) vectors that break biological function into auditable units.
2. Capability Quantification
How far is a biological system from functional breakdown? SEWO introduces the Capomics Index: CI = 1 − PAI — a single metric to quantify system resilience.
3. Input–Response Semantics
Every perturbation (drug, nutrient, environmental factor) should map to computationally tractable inputs with explicit biological meaning — not just latent vectors.
4. Counterfactual Transition Modeling
A valid biomedical world model must simulate plausible "what-if" trajectories: What happens if we intervene here? What if we remove this assumption?
5. Five-Gate Quality Control Loop
Every reasoning chain follows: State → Input → Response → ΔmIC → Phenotype
Each gate can be independently inspected, challenged, and falsified. No black boxes.
Why This Matters for AI Engineers
If you're building AI systems for biomedicine, SEWO offers a practical checklist:
- [ ] Can your model's state representation be decomposed into interpretable modules?
- [ ] Can you quantify how close a system is to failure?
- [ ] Do inputs map to biologically meaningful perturbations?
- [ ] Can you simulate counterfactual intervention scenarios?
- [ ] Can each step of your model's reasoning be independently audited?
If the answer is no to any of these, you may have a powerful predictor — but not a steerable world model.
Steering, Not Predicting
A flavonoid doesn't simply "kill" a cancer cell. It influences a signaling network, alters protein–protein interactions, shifts regulatory dynamics — and the cell's own machinery responds.
SEWO extends this logic to AI: instead of asking AI to dictate outcomes from above, we should build systems that accept biologically meaningful directional input, recompute coherent trajectories, and make their reasoning transparent.
Steering, not predicting.
Get Involved
The SEWO project is open for community discussion:
- 🌐 Project home: steerable.world
- 📄 Preprint: doi.org/10.20944/preprints202605.0366.v1
- 📧 Contact: info@deepome.com
Important note: This manuscript is a preprint and has not yet undergone peer review. The framework is a research proposal and conceptual specification, not a clinically validated system.
Hashtags: #SEWO #SteerableWorldModel #BiomedicalAI #WorldModels #TrustworthyAI #MachineLearning #AI
Top comments (0)