New preprint proposes that biomedical world models should be structurally constrained, auditable, and steerable — not merely larger predictive systems — and launches steerable.world as the project’s digital home.
BEIJING, May 7, 2026 — As investment accelerates around AI systems that aim to model cells, drugs, disease progression, and human biology, DeepoMe Limited today announced the release of a new preprint introducing SEWO — the Steerable Medicine World Model — a framework designed to shift the conversation from bigger biomedical predictors toward trustworthy, auditable, and steerable world models.
The manuscript, “World Models for Biomedicine: A Steerability Framework,” published on Preprints.org by MDPI, argues that the next generation of biomedical AI should not only predict biological trajectories, but also allow clinicians and researchers to guide them through explicit, inspectable, and biologically meaningful directional signals.
At the center of the work is a simple question that DeepoMe believes the field must answer before biomedical world models can be trusted in drug discovery, precision medicine, and clinical decision support:
Can you steer it?
“Everyone is building bigger engines. We are asking whether anyone has checked the steering system — or whether the vehicle even has one,” said Jianghui Xiong, lead author of the preprint. “A world model for medicine 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.”
From Black-Box Prediction to Steerable Guidance
The SEWO framework challenges the dominant assumption that biomedical AI progress should be measured primarily by model size, data scale, or benchmark accuracy. Instead, it proposes steerability as a foundational property of trustworthy biomedical world models.
In the SEWO context, steerability does not mean remote-controlling a model or manually adjusting every parameter. It means directional guidance under robustness constraints.
DeepoMe describes the concept through a metaphor: the rider and the horse. A rider does not micromanage every muscle of the horse. The rider provides directional signals through the reins, while the horse maintains balance, adapts to terrain, and moves forward with its own embodied robustness.
Likewise, a steerable medicine world model should allow human experts to provide stage-wise guidance — such as adding a therapeutic hypothesis, modifying a nutritional or environmental condition, or removing a confounding assumption — while the model maintains internal consistency in the presence of real-world noise, missing data, and distribution shifts.
“Medicine is not a straight road,” Xiong added. “It is uncertain terrain. A trustworthy biomedical world model should be less like a black-box oracle and more like a well-trained horse: robust on its own, but responsive to the right direction.”
Five Structural Constraint Points for Biomedical World Models
Rather than prescribing a single neural architecture, SEWO defines a set of structural constraint points that any candidate biomedical world model can be evaluated against, regardless of whether it is implemented through transformers, state-space models, graph systems, hybrid mechanistic models, or future AI architectures.
The preprint outlines five core constraint points:
1. State Representation
Biological states should be represented through modular and interpretable components, including modular Intrinsic Capability, or mIC, vectors that decompose biological function into auditable units.
2. Capability Quantification
SEWO introduces capability-centric measurement through the Capomics Index, expressed as CI = 1 − PAI, to estimate how far a biological system is from functional breakdown.
3. Input–Response Semantics
Environmental, nutritional, pharmacological, or therapeutic perturbations should be mapped into computationally tractable inputs with explicit biological meaning.
4. Counterfactual Transition Modeling
A valid biomedical world model should be able to simulate plausible “what-if” trajectories under different intervention scenarios, rather than only extrapolating from observed data.
5. Five-Gate Quality Control Loop
SEWO proposes a reasoning scaffold organized as State → Input → Response → ΔmIC → Phenotype, allowing each step of the model’s reasoning chain to be independently inspected, challenged, and falsified.
Together, these constraint points form what the authors call a deductive constraint framework: a structural chain of reasoning in which each link is explicit, auditable, and less vulnerable to the silent failures often associated with black-box predictors.
Why Steerability Matters Now
The release of the SEWO preprint comes at a moment of rapid expansion in AI-for-biology and AI-driven drug discovery. Recent advances in virtual cell modeling, biological foundation models, cellular simulation, and drug-response prediction have made “world models for biology” a major frontier for both academic and commercial research.
However, DeepoMe argues that the field still lacks a shared vocabulary for evaluating whether these systems are structurally valid, clinically interpretable, and safely guidable.
“Before billions of dollars are invested in biomedical world models for drug discovery and precision medicine, the field needs to clarify what it means for such a model to be valid,” said Xiong. “SEWO is our opening proposal for that conversation.”
The company positions SEWO not as a competing model architecture, but as a meta-level framework for assessing future biomedical world models. In this sense, SEWO can serve as a specification layer: a way to ask whether a system claiming to model biology is not only predictive, but also interpretable, constrained, counterfactual, and steerable.
Steering, Not Merely Predicting
A central claim of the SEWO preprint is that biomedical systems are not best understood as passive targets of prediction. They are dynamic, adaptive systems that respond to signals.
This idea mirrors many real biological interventions. A flavonoid, for example, may not simply “kill” a cancer cell directly. Instead, it may influence a signaling network, alter a protein–protein interaction subnetwork, or shift regulatory dynamics so that the cell’s own machinery responds differently.
SEWO extends this logic into biomedical AI: instead of asking AI to dictate outcomes from above, the framework asks whether a model can accept biologically meaningful directional input, recompute a coherent trajectory, and make its reasoning inspectable.
This principle is summarized by the project’s core phrase:
Steering, not predicting.
Launch of steerable.world
DeepoMe also announced the launch of steerable.world as the digital home for the SEWO project.
The website will serve as a hub for SEWO-related updates, preprint materials, future technical resources, community discussion, and the emerging steerability framework for biomedical world models.
Readers, researchers, clinicians, and AI-for-biology teams are invited to visit steerable.world to follow the project and participate in the broader conversation around steerable, auditable, and trustworthy biomedical AI.
Availability
Title: World Models for Biomedicine: A Steerability Framework
DOI: https://doi.org/10.20944/preprints202605.0366.v1
The preprint “World Models for Biomedicine: A Steerability Framework” is available on Preprints.org and open for community comment.
Important notice: This manuscript is a preprint and has not yet undergone peer review. The framework described should be interpreted as a research proposal and conceptual specification, not as a clinically validated medical device or approved diagnostic system.
About SEWO
SEWO, short for Steerable Medicine World Model, is a conceptual framework for biomedical world models that emphasizes steerability, robustness, certainty, counterfactual reasoning, and structural auditability. The framework is designed to help researchers and clinicians evaluate whether future AI systems for biomedicine can be guided through meaningful expert input while maintaining coherent and inspectable reasoning under uncertainty.
More information is available at steerable.world.
About DeepoMe Limited
DeepoMe Limited is a Beijing-based research company focused on structural, interpretable frameworks for biomedicine and trustworthy AI. Its work includes the Capomics platform for capability-centric biological analysis, the EvoSika initiative for evolutionary systems modeling, and the newly launched SEWO project for steerable medicine world models.
DeepoMe’s research emphasizes deductive rigor, structural auditability, and the development of AI systems that clinicians, researchers, and biomedical innovators can meaningfully inspect, guide, and trust.
Media Contact
DeepoMe Limited
Email: info@deepome.com
Web: steerable.world
Hashtags: #SEWO #SteerableWorldModel #BiomedicalAI #WorldModels #TrustworthyAI
