ODesign: A World Model for Biomolecular Interaction Design

ODesign Team

ODesign: A World Model for Biomolecular Interaction Design

Odin Zhang^1,3,5,*,†, Xujun Zhang^1,3,*, Haitao Lin^1,5,*, Cheng Tan^5,6,*, Qinghan Wang^1,3,*, Yuanle Mo^1,5,*, Qiantai Feng^4,6,*, Gang Du^1,3, Yuntao Yu^1,3, Zichang Jin^1,3, Ziyi You^1,3, Peicong Lin¹, Yijie Zhang⁷, Yuyang Tao¹, Shicheng Chen³, Jack Xiaoyu Chen⁸, Chenqing Hua⁷, Weibo Zhao⁵, Runze Ma^1,2, Yunpeng Xia¹, Kejun Ying¹, Jun Li⁹, Yundian Zeng³, Lijun Lang⁵, Peichen Pan³, Hanqun Cao⁵, Zihao Song¹⁰, Bo Qiang¹⁰, Jiaqi Wang¹⁰, Pengfei Ji¹¹, Lei Bai⁶, Jian Zhang¹², Chang-yu Hsieh³, Pheng Ann Heng^5,†, Siqi Sun^6,†, Tingjun Hou^3,†, Shuangjia Zheng^2,1,†

¹Lingang Laboratory, Shanghai, China
²Global Institute of Future Technology, Shanghai Jiao Tong University, Shanghai, China
³College of Pharmaceutical Sciences, Zhejiang University, Zhejiang, China
⁴Research Institute of Intelligent Complex Systems, Fudan University, Shanghai, China
⁵Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong, China
⁶Shanghai Artificial Intelligence Laboratory, Shanghai, China
⁷McGill University, Montreal, Canada
⁸Institute for Medical Engineering Science and Department of Biological Engineering, Massachusetts Institute of Technology, Cambridge, MA, USA
⁹T. H. Chan School of Public Health, Harvard University, Boston, MA, USA
¹⁰Department of Biochemistry, University of Washington, Seattle, WA, USA
¹¹Institute of Catalysis, Department of Chemistry, Zhejiang University, Hangzhou, China
¹²School of Medicine, Shanghai Jiao Tong University, Shanghai, China
^*Equal contribution ^†Corresponding authors

Web Server Technical Report

Abstract

Biomolecular interactions underpin almost all biological processes, and their rational design is central to programming new biological functions. Generative AI models have emerged as powerful tools for molecular design, yet most remain specialized for individual molecular types and lack fine-grained control over interaction details. Here we present ODesign, an all-atom generative world model for all-to-all biomolecular interaction design. ODesign allows scientists to specify epitopes on arbitrary targets and generate diverse classes of binding partners with fine-grained control. Across entity-, token-, and atom-level benchmarks in the protein modality, ODesign demonstrates superior controllability and performance to modality-specific baselines. Extending beyond proteins, it generalizes to nucleic acid and small-molecule design, enabling interaction types such as protein-binding RNA/DNA and RNA/DNA-binding ligands that were previously inaccessible. By unifying multimodal biomolecular interactions within a single generative framework, ODesign moves toward a general-purpose molecular world model capable of programmable design.

Highlights

First-of-its-kind all-to-all world model: Unified framework enables cross-modality molecular generation—proteins, nucleic acids, and small molecules—within a single architecture built upon AlphaFold3-like structure prediction.
Multi-level controllable generation: Task-oriented masking mechanism provides fine-grained conditional control at entity-, token-, and atom-levels, supporting binder design, motif scaffolding, and atomic motif engineering.
Flexible and rigid design modes: Supports both fixed-target (rigid receptor) and co-design (flexible receptor) strategies, enabling epitope-specific generation through hotspot residue guidance.
Superior performance and throughput: Consistently outperforms modality-specific baselines across 11 benchmark tasks, achieving 2-4 orders of magnitude improvement in design throughput with minutes per sample.
Previously inaccessible interactions: Enables novel design capabilities including protein-binding RNA/DNA aptamers and nucleic acid-binding small molecules that were computationally infeasible before.

ODesign main results showcasing target and designed biomolecular complexes

ODesign is an All-to-All design framework that simultaneously models targets (grey) and designed structures (violet) across proteins, nucleic acids, and ligand complexes.

Design cases

Showcasing the dynamic design process of protein, cyclic peptides, nucleic acids and ligand molecules.

Architecture

1. Embedding Module

Unified generative tokens abstract minimal chemical units across modalities. Target 3D structures are incorporated as initial coordinates and distance constraints.

2. Conditional Module

Dual control mechanism: initial coordinate "guess" and distance-based interaction modeling. Hotspot residue features enable epitope-specific design.

3. OInvFold Module

Unified inverse folding assigns sequences to protein/nucleic acid backbones and atom types to ligand scaffolds with modality-specific decoders.

In-silico Performance

ODesign consistently outperforms modality-specific baselines across protein, nucleic acid, and ligand design benchmarks.

Protein Design

Performance on protein-centric benchmarks: binder design, ligand-binding protein design, motif scaffolding, interface design, and atomic motif scaffolding