How a Nobel‑Lab‑Born Team Uses a World Model to Break Modal Islands in AI‑Driven Molecular Design

Recent breakthroughs such as AlphaFold have shown that AI can predict protein structures with near‑experimental accuracy, but the next challenge is moving from seeing to designing molecules that act in the right place and way to modulate pathways or reconstruct biological functions.

The Modal‑Island Problem

Current AI‑driven molecular design tools treat proteins, small molecules, and nucleic acids as separate tasks, creating “modal islands” that prevent a single model from learning the continuous interactions across these biological modalities. This fragmentation arises from divergent representations, training data, and task objectives, even though real biological systems operate without such boundaries.

ODesign: A Full‑Modal World Model

ODesign, an open‑source research project led by Zhang Haotian and colleagues from the Baker Lab, proposes a unified generative framework that places proteins, DNA, RNA, small molecules, and ions into a single modeling space. The core idea is to identify a “Minimal Common Generative Unit” (MCGU) that can be shared across modalities, abstracted as unified tokens (Modality Token and Unit Token). During inference, the model first generates shared chemical primitives and then fills in modality‑specific atomic details.

The architecture uses a Pairformer to learn interactions between these tokens and a full‑atom diffusion module to produce three‑dimensional structures that satisfy spatial constraints.

Cross‑Modal Transfer and Benchmarks

Experimental results demonstrate that ODesign’s value lies not merely in handling more tasks but in exhibiting genuine cross‑modal transfer:

On protein design, ODesign achieves roughly a 10× increase in computational throughput over RFDiffusion and a 20× increase in candidate flux on the AME benchmark compared to RFDiffusion2.

For nucleic acids, where data are scarce, ODesign’s RNA monomer design success rate is about twice that of RNAFrameFlow, and zero‑shot protein‑RNA complex design reaches an average success rate of 77.9%.

In small‑molecule design, ODesign outperforms SurfGen by more than 40× and extends capability to DNA/RNA‑bound small‑molecule design, covering regions where traditional models fail.

Beyond computational metrics, wet‑lab validation shows that ODesign has produced candidate molecules with nanomolar to picomolar affinity on eight targets, surpassing methods such as RFDiffusion, BindCraft, BoltzGen, and PXDesign by several‑fold to hundreds‑fold in binding affinity.

Team Background and Future Roadmap

The core team—Zhang Haotian, Ying Kejun, and Wang Jiaqí—are alumni of David Baker’s Nobel‑winning laboratory, combining expertise in physics, pharmacy, medicine, and computer science. While ODesign proves the feasibility of unified generation and cross‑modal transfer, the authors acknowledge that it remains a prototype and is not yet a scalable, fault‑tolerant industrial tool.

To bridge this gap, they outline a three‑stage roadmap:

AI4Bio : Commercialization through real‑world biomolecular design tasks for pharma and research institutions, addressing multi‑constraint optimization (binding, toxicity, permeability, immunogenicity, synthesizability).

AI4AI : A self‑iterating scientific intelligence framework that organizes unstructured literature, experimental data, and model outputs into a navigable knowledge graph, enabling hypothesis generation, attribution assessment, and result‑driven model refinement.

AI4Phy : An autonomous experimental loop where generated candidates are synthesized, tested, and fed back into the model, allowing continuous calibration of the model against physical reality.

This vision shifts AI4Bio from merely accelerating candidate discovery to providing a programmable pipeline where hypotheses are generated, experimentally validated, and used to refine the underlying world model of molecular interactions.