ODesign: A World Model for Biomolecular Interaction Design

1 Introduction

• Existing generative models have advanced within individual molecular domains, but do not perform the cross-modality reasoning needed to program complex biological functions.Structure-prediction models jointly represent multiple modalities, yet their objectives remain limited to predicting existing structures rather than generating molecules with desired functions.Source: §1 · p. 2
• Protein-centric entity-level models do not capture the multimodal nature of biological systems or provide hierarchical control across molecular, motif, and atomic levels.Atomic representations can miss chemical and biological priors encoded in higher-order units and may not ensure the intended molecular modality.Source: §1 · p. 2
• ODesign uses a unified token space and task-oriented masking for cross-modality generation with molecular-, motif-, and atomic-level conditional control.Rigid and flexible conditioning supports fixing target structures or co-generating target–binder interfaces.Source: §1 · p. 2
• Across eleven benchmark tasks spanning proteins, small molecules, RNA, and DNA, ODesign consistently outperforms modality-specific models and improves design throughput.The authors present this unified framework as a step toward a general-purpose molecular world model.Source: §1 · p. 2

2 Architecture and training of ODesign

• ODesign abstracts chemical building blocks into unified generative tokens, then decodes them through an all-atom diffusion module for multimodal generation.The paper adopts a two-stage design paradigm, generating a backbone before designing its sequence.Source: §2 · p. 4
• Built on an AlphaFold3-like architecture, ODesign uses Pairformer to model token–target interactions and conditional diffusion to decode constrained three-dimensional coordinates.Upsampling and downsampling pathways support cross-modality generation at full-atom resolution.Source: §2 · p. 4
• Structural information is injected into the embedding layer, Pairformer, and diffusion decoder; rigid mode fixes the target, while flexible mode co-generates binder and receptor conformations.Generation can also be conditioned on specified target epitopes, which are randomly selected during training.Source: §2 · p. 4
• Training uses monomeric and complex structures from the PDB, with sampling probabilities and masking strategies adjusted for downstream design tasks.The five modules comprise embedding, target-structure conditioning, Pairformer interaction modeling, all-atom coordinate diffusion, and inverse folding for type generation.Source: §2 · p. 4

3 Protein Designability

• Protein design spans entity-, motif-, and atomic-level control and targets proteins, small molecules, and nucleic acids, but existing models address only subsets of these tasks.The paper identifies low computational throughput as a practical bottleneck: a single-GPU run may yield too few candidates to fill a 96-well screening plate.Source: §3 · p. 4

3.1 Protein-binding Protein Binder Design

• ODesign generated an average of 2,672 successful protein binders per GPU-day across targets, versus 555 for RFDiffusion, an order-of-magnitude throughput improvement.ODesign outperformed all compared methods on each target; the flexible variant slightly exceeded the rigid variant.Source: Figure 2 · p. 5, §3.1 · p. 6

3.2 Ligand-binding Protein Binder Design

• ODesign outperformed existing methods by 12.1-fold over RFDiffusion-AA and 57.8-fold over BoltzDesign1 in ligand-binding protein design.ODesign-Flex surpassed ODesign-Rigid across all cases; success required iPAE<10 and complex_pLDDT>0.7.Source: Figure 2 · p. 5, §3.2 · p. 6

3.3 Motif Scaffolding

• ODesign achieved an 18.4% average success rate versus 14.3% for RFDiffusion-AA, with successful designs in 27 of 30 MotifBench cases.It also exceeded the second-best method by 38.6% on MotifScore.Source: Figure 2 · p. 5, §3.3 · p. 6

3.4 Interface Design

• ODesign produced lower mean pocket-aligned RMSD by 2.24 Å and lower mean min-iPAE by 0.68 than PocketGen across four interface-design cases.The authors interpret these results as indicating more stable pockets and better alignment with intended binding geometry.Source: Figure 2 · p. 5, §3.4 · p. 7

3.5 Atomic Motif Scaffolding

• ODesign achieved a 37.5-fold average efficiency improvement in generating filter-passing designs on the AME atomic-motif benchmark.The task asks a model to reproduce the spatial geometry of a ligand and key residues represented at atomic level.Source: Figure 2 · p. 5, §3.5 · p. 7

4 Nucleic Acid Designability

• Most existing nucleic-acid design models focus on single molecules rather than application-oriented target-binding aptamers.Aptamers can bind proteins, small molecules, other nucleic acids, and whole cells.Source: §4 · p. 7

4.1 RNA aptamer design

• ODesign achieved nearly twice the success rate of RNAFrameFlow on RNA monomer design for sequences 10–150 nucleotides long.Success meant RMSD<5 Å or TMScore>0.45 after refolding.Source: §4.1 · p. 7, Figure 3 · p. 8
• In a zero-shot protein-binding RNA task on unseen targets, ODesign achieved a 77.95% success rate at a protein-aligned RMSD threshold of 5 Å.The study generated and refolded 400 RNA aptamers for each of ten protein targets curated from post-January 13, 2023 PDB complexes.Source: §4.1 · pp. 7, 9

4.2 DNA aptamer design

• ODesign was extended to single-stranded DNA aptamers; their designability declined more rapidly with length than RNA, possibly because single-stranded DNA was less represented in training data.For targeted DNA design, four protein–DNA complexes were benchmarked, with 400 generated aptamers per target; examples localized near electrostatically complementary protein regions.Source: §4.2 · p. 9

5 Small Molecule Designability

• Small molecules have complex topologies, and more than 75% of FDA-approved drugs are small molecules.Their structural diversity raises demands on model learning, while many existing design models are trained specifically on small-molecule data.Source: §5 · p. 9

5.1 Protein-binding Small Molecule Design

• Across six targets, ODesign generated the most valid and successful molecules per GPU per day; versus TargetDiff, counts were 3.9-fold and 9.0-fold higher, respectively.It also exceeded SurfGen by 14.0-fold in valid and 48.4-fold in successful molecule count.Source: §5.1 · pp. 9, 11, Figure 4 · p. 10

5.2 DNA-binding Small Molecule Design

• Mean min-iPAE 1.29 and mean iPTM 0.55 accompanied high-confidence DNA interfaces; a DB1476 case reproduced minor-groove binding with refinements fitting groove contours.The study notes fewer than 1% as many nucleic-acid–small-molecule complexes as protein–small-molecule complexes in PDBBind and no specialized models for DNA-binding design.Source: §5.2 · p. 11

5.3 RNA-binding Small Molecule Design

• Mean min-iPAE 2.0 and mean iPTM 0.57 accompanied valid ligands across three RNA targets, with a riboswitch case preserving native binding interactions in flexible RNA conformations.The generated ligands also formed additional interfacial interactions.Source: §5.3 · p. 11

• Across eleven benchmark tasks, ODesign consistently outperforms modality-specific models, with up to two to four orders of magnitude higher design throughput.The conclusion describes the model as supporting controllable generation across proteins, nucleic acids, and small molecules.Source: §6 · p. 11
• ODesign supports specifying epitopes on arbitrary targets and controllable design across molecular modalities.The authors describe a unified foundation model that learns cross-modality state representations and supports generation under multiple biological conditions.Source: §6 · p. 11
• Performance declines under some out-of-distribution conditions, with DNA aptamer designability decreasing more rapidly with sequence length than RNA aptamer designability.Fine-grained semantic control remains unavailable for properties such as protein secondary-structure composition and small-molecule aromaticity.Source: §6 · pp. 11–12
• The authors propose advancing ODesign toward a self-evolving world model through in silico evolution, physics-informed constraints, drug-likeness priors, and LLM-based reasoning agents.The proposed direction includes closed-loop design–evaluation–update cycles.Source: §6 · p. 12

• Odin Zhang conceived the research idea and led the project; Odin Zhang and Xujun Zhang designed the initial model architecture.Shuangjia Zheng contributed to the conceptual idea and provided foundational resources and strategic guidance.Source: Author Contributions Statement · p. 12
• The machine-learning team developed ODesign’s main architecture, inverse-folding module, benchmarking framework, and task-specific protein, small-molecule, and nucleic-acid design components.Peicong Lin managed large-scale training and integration of the universal model.Source: Author Contributions Statement · p. 12
• The experimental and applied-design team conducted in silico design efforts, real-world design campaigns, and experimental validations.Gang Du led the in silico design efforts.Source: Author Contributions Statement · p. 12

• The listed authors include Odin Zhang, Xujun Zhang, Haitao Lin, Cheng Tan, Qinghan Wang, Yuanle Mo, and Qiantai Feng.The roster also names collaborators including Gang Du, Yuntao Yu, and Shuangjia Zheng.Source: Authors and Affiliations · p. 12
• Affiliations include Lingang Laboratory, Shanghai Jiao Tong University, Zhejiang University, Fudan University, and the Chinese University of Hong Kong.Source: Authors and Affiliations · p. 12
• The affiliation list also includes McGill University, MIT, Harvard University, the University of Washington, and Zhejiang University.It additionally lists Shanghai Artificial Intelligence Laboratory and Shanghai Jiao Tong University School of Medicine.Source: Authors and Affiliations · pp. 12–13

• The authors declare that there are no conflicts of interest.Source: Competing Interests Statement · p. 13

A.1.1 Dataset

• Training combines weighted PDB crystal structures with structures distilled from AlphaFold2 and OpenFold predictions.PDB structures were quality-filtered, including temporal, sequence, resolution, completeness, and structural-plausibility checks; evaluation uses low-homology recent structures and PoseBusters V2, with its 308 targets excluded from training.Source: §A.1.1 · p. 19

A.1.2 Input Features

• ODesign extends AlphaFold3 input features with hotspot-residue and masked-token indicators.Source: §A.1.2 · p. 19

A.1.3 MaskGenerator

• MaskGenerator masks at All, Entity, Token, or Atom granularity, supporting free generation, binder design, motif scaffolding, and atomic-motif scaffolding.Entity masking generates new entities from partial structures; Token masking selects contiguous segments, while Atom masking retains condition atoms in selected tokens. Symmetry-aware masking and ligand-specific handling further constrain masking.Source: §A.1.3 · p. 20

A.1.4 Cropping Methods

• ODesign uses stochastic contiguous, spatial, and spatial-interface cropping to train within GPU memory limits while preserving molecular integrity and interaction regions.Spatial-interface cropping centers on contacting tokens within 15Å; small-molecule samples use this strategy, retaining only the target ligand when multiple ligands are selected.Source: §A.1.4 · p. 21

A.2 Architecture

• The architecture combines target conditioning and Pairformer interaction modeling with conditional diffusion for all-atom coordinates, followed by OInvFold sequence or atom-type design.OInvFold decodes polymer residues autoregressively, while its ligand pathway predicts atom types and chemical properties in parallel.Source: §A.2 · pp. 21–22, §A.2.3 · p. 23, §A.2.4 · p. 23

A.2.1 Unified Generative Token Mechanism

• A unified generative token represents each modality’s minimal chemical unit, with modality-specific suffixes selecting proteins, nucleic acids, or ligands.Source: §A.2.1 · p. 22

A.2.2 Condition Module with 3D Target Structure

• ODesign conditions generation on target 3D structure through coordinate and distance features, supporting both fixed rigid targets and flexible-target co-generation.Hotspot residues are annotated using intermolecular distance and entity identity, then sampled during training to expose the model to varying hotspot counts and positions.Source: §A.2.2 · pp. 22–23

A.2.5 OInvFold Module

• OInvFold uses local geometric frames and an SE(3)-equivariant graph encoder, decoding protein and nucleic-acid sequences autoregressively and ligand atom types in parallel.The ligand pathway treats atoms as units because small molecules have graph-like bond topologies rather than linear sequence patterns.Source: §A.2.5 · p. 24

A.3.1 Composition

• ODesign’s structure objective combines distance, coordinate, bond, local-distance, and ligand-bond reconstruction losses, with rigid-body alignment before loss calculation.Source: §A.3.1 · p. 25

A.3.2 Global Align and Partial Align

• Global alignment yields significantly lower validation loss than Partial Align when training from a pretrained checkpoint.This result supports allowing the model to explore coordinate-frame distributions rather than fixing the frame from conditional atoms.Source: §A.3.2 · p. 25

A.4 Modalities

• ODesign reports its lowest loss for ligands, higher loss for proteins, and highest loss for nucleic acids, particularly RNA.The paper relates these differences to training-data scale and quality, atom counts, alignment, and nucleic-acid flexibility.Source: §A.4 · pp. 25–26

B.2.1 Hotspot Center

• The hotspot-center method starts sampling at the geometric center of specified hotspot residues for interface-focused design tasks.Source: §B.2 · p. 27, §B.2.1 · p. 28

B.2.2 Global Center

• The global-center method centers sampling on the geometric center of the reference structure for conditional designs using global information.Source: §B.2.2 · p. 28

B.2.3 User-Provided Center

• The user-provided center method initializes sampling at user-specified 3D coordinates when prior knowledge indicates the desired binding pocket.Source: §B.2.3 · p. 28

B.2.4 User-Provided Residue Center

• The user-provided residue-center method initializes sampling at the geometric center of selected conditional residues when users have specific design preferences or prior knowledge.Source: §B.2.4 · p. 28

B.3 Partial Diffusion

• Partial diffusion modifies an existing binder by adding noise in a back stage and denoising in a forth stage.Smaller SNR values increase divergence and diversity but make function preservation more challenging; larger values reduce diversity while maintaining greater similarity to the starting molecule.Source: §B.3 · pp. 28–29

B.4 Inference Speed

• Each ODesign design iteration costs one fold-model pass, compared with 200 fold-model executions for RFdiffusion and RFdiffusion-AA.On ligand-binding protein design, sampling-time benchmarks also report ODesign as more efficient than RFdiffusion-AA and Boltz-1.Source: §B.4 · p. 29

C.1.1 Free Generation; C.1.2 Protein-Binding Protein Design; C.1.3 Ligand-Binding Protein Design

• Protein benchmarks cover unconditioned backbone generation, protein-target binders, and ligand-target binders, using task-specific designability or interaction filters.Free generation compares classical models across 100–800-residue backbones and defines designability using pLDDT > 70 and backbone-aligned RMSD < 2.0Å. Protein-binding protein design uses 11 targets and filters on ipAE, pLDDT, ipTM, and complex RMSD; ligand-binding protein design tests IAI, OQO, FAD, and SAM with complex pLDDT and iPAE criteria.Source: §C.1.1 · p. 29, §C.1.2 · pp. 29–30, §C.1.3 · p. 30

C.1.4 Motif Scaffolding; C.1.5 Atomic Motif Scaffolding; C.1.6 Interface Design

• MotifBench evaluates 30 targets using N, Cα, C-aligned motif RMSD < 1.0Å and Cα-aligned Cα RMSD < 2.0Å; AME tests 41 systems, requiring catalytic-residue heavy-atom RMSD < 1.5Å without clashes.AME RMSD is assessed after backbone alignment on catalytic residues in at least one Chai-1-predicted LigandMPNN sequence.Source: §C.1.4 · p. 31, §C.1.5 · p. 31

C.2.1 DNA Free Generation; C.2.3 RNA Free Generation

• Free DNA and RNA generation are evaluated across 10–150-nucleotide lengths using refolding-based RMSD or TM-score designability criteria.DNA has no established benchmark, so ODesign samples 50 backbones per length at 10-nucleotide intervals; RNA is compared with RNAFrameFlow over the same length range. For both, a backbone is designable when at least one designed sequence meets C4’ RMSD < 5Å or TM-score > 0.45.Source: §C.2.1 · p. 32, §C.2.3 · p. 32

C.2.2 Protein-Binding DNA Design; C.2.4 Protein-Binding RNA Design

• Protein-binding DNA and RNA benchmarks test aptamer design on post-cutoff protein–nucleic acid complexes using phosphate-atom RMSD success criteria.The DNA test set contains four complexes, and the RNA test set contains ten; both are selected from structures released after January 13, 2023, to assess generalization. Successful designs require Cα-aligned DNA or RNA phosphate-atom RMSD < 5Å, respectively.Source: §C.2.2 · p. 32, §C.2.4 · pp. 32–33

C.3.1 Protein-Binding Ligand Design; C.3.2 DNA-Binding Ligand Design; C.3.3 RNA-Binding Ligand Design

• Small-molecule generation was benchmarked against protein, DNA, and RNA targets, using protein validity and success criteria and AlphaFold3 interface-confidence and ligand-structure-fidelity metrics for nucleic-acid targets.Protein success required QED > 0.25 and Vina Dock score < -8.18. DNA and RNA evaluations each used three PDB complexes released after January 13, 2023, and measured minimum iPAE, iPTM, refolded ligand PTM, and DNA- or RNA-C1′-aligned ligand RMSD.Source: §C.3.1 · p. 33, §C.3.2 · p. 33, §C.3.3 · p. 34

C.3.3 RNA-Binding Ligand Design

• Free-backbone generation masks standard tokens, then symmetrizes the token mask and maps it to atoms.The all-mask strategy removes side chains during atom mapping.Source: §C.3.3 · p. 40
• Binder design masks one entity when there is only one, or randomly selects between one and N−1 entities when there are multiple.The resulting mask is symmetrized and mapped to atoms.Source: §C.3.3 · p. 40
• Motif scaffolding masks random token segments, while atomic motif scaffolding conditions on all standard tokens and randomly masks 20% of them at the atom level.Both strategies apply symmetry-aware postprocessing; atomic conditioning selects atoms in a sampled chemical-bond neighborhood.Source: §C.3.3 · pp. 41–43
• Inference embeds masked pairwise distances as distogram constraints, processes them with a Pairformer, then applies conditional diffusion to generate coordinates.The diffusion initialization preserves coordinates marked as fixed, and noise is zeroed for those coordinates during sampling.Source: §C.3.3 · pp. 44–45
• JSON S1–S4 list configurations for protein-binding-protein, ligand-binding-protein, motif-scaffolding, and atom-scaffolding tasks.Source: §C.3.3 · p. 46
• JSON S5–S6 list configurations for RNA free generation and protein-binding-RNA tasks.Source: §C.3.3 · pp. 46–47
• JSON S7 is labeled as a configuration for ligand generation.Source: §C.3.3 · p. 47