Technical Summary: OpenDDE – An Open-Source All-Atom Biomolecular Foundation Model

Problem Statement

Accurately modeling biomolecular interactions remains a central bottleneck in biology and therapeutic discovery. While recent advances like AlphaFold3 and the proprietary IsoDDE system have demonstrated the potential of co-folding (jointly predicting structures of protein complexes, nucleic acids, and ligands) to achieve experimental-level accuracy, the most capable systems remain closed-source. This limits reproducibility, independent validation, and community-driven extension. Furthermore, incremental improvements in co-folding accuracy are slowing, raising questions about whether current tokenization, data curation, and architectures are approaching practical bottlenecks. There is a critical need for an open, scalable, all-atom foundation model that can serve as a shared structural reasoning layer for drug discovery, moving beyond isolated structure prediction toward de novo design and affinity estimation.

Methodology

OpenDDE (Open Drug Discovery Engine) is introduced as an open-source, Apache-2.0 licensed, all-atom biomolecular foundation model. It utilizes a diffusion-based framework to generate atomic coordinates, unifying structure prediction and conditional design.

Architecture and Reasoning

OpenDDE employs a coarse-to-fine reasoning path that operates on biomolecular tokens before generating coordinates:

1. Residue-Token Trunk: The model first builds single and pair representations using a Pairformer-style trunk to capture global relational context across chains, residues, and ligands.
2. Structural-Token Refinement: Residue-level states are expanded into chemically meaningful structural tokens (backbone, side-chain, base, ligand/atom). A dedicated Structural Refiner performs fine-grained relational reasoning over these tokens, refining local geometry, chemical context, and cross-molecular interfaces.
3. Atomic Latent Reasoning: Unlike direct sequence-to-coordinate mapping, OpenDDE incorporates latent reasoning to refine representations of local geometry and interfaces before the final all-atom structure generation.
4. Diffusion Generation: A diffusion-based coordinate generator denoises atom positions conditioned on the refined structural token representations.

Training Strategy

The model is trained on 655M parameters using NVIDIA Ampere and Hopper GPUs. The training follows a multi-stage curriculum designed to move from precision to breadth and back to precision:

• Data Curation: Training data is filtered with a cutoff of September 2021 to prevent leakage. It includes a mix of weighted PDB data, AFDB multimers, reproduced Teddymer datasets, MGnify monomers, Swiss-Prot, disordered motifs, and SAbDab antibody structures.
• Curriculum Learning:
  • Early Stages: Focus on large-scale, experimentally validated structures to learn accurate folding patterns and local geometry.
  • Middle Stages: Increased use of distilled datasets to broaden coverage over sequence and structure space.
  • Later Stages: Shift back to high-quality experimental and task-specific data (including SAbDab) to improve atomic accuracy, interface fitting, and confidence calibration.
• Loss Functions: The training incorporates shape-complementarity losses to supervise interface fitting (rewarding compatible contacts, penalizing steric clashes) alongside standard Smooth LDDT losses.
• Unified Framework: Structure prediction and de novo design are treated as instances of the same conditional denoising problem. By masking specific atoms (fixing them as context), the model can perform conditional generation around provided structures without requiring a separate architecture.

Scaling Laws

The authors investigate scaling laws along two axes: estimated training tokens (steps × GPUs × crop tokens) and estimated training cost (tokens × trainable parameters). They observe a clear, sublinear positive relationship between training scale and antibody-antigen (Ab-Ag) performance, suggesting that biomolecular modeling is entering a scaling regime similar to large language models.

Key Contributions

1. Atomic Latent Reasoning: The integration of latent reasoning into biomolecular modeling, allowing the model to refine local geometry and interface representations before coordinate generation.
2. Extensible Foundation: A unified architecture designed not just for structure prediction, but as a shared structural reasoning layer for future modules including de novo design, affinity prediction, and structure-conditioned optimization.
3. Scaling and Distillation: Empirical identification of scaling directions for co-folding models and revised distillation strategies for monomeric and multimeric structures to improve robustness and generalization.
4. Open Release: The release of training code, inference pipelines, checkpoints, and benchmarks under the Apache-2.0 license, aiming to democratize access to frontier biomolecular intelligence.

Results

OpenDDE was benchmarked against state-of-the-art closed models (AlphaFold3, Chai-1) and open models (Boltz-1, OpenFold3, Protenix-v1, ESMFold2) on antibody-antigen structure prediction tasks.

• Benchmark Performance: OpenDDE achieved the highest success rates across three major benchmarks (PXMeter-AB, FoldBench-AB, and the newly curated 2026ARK-AB).
  • On FoldBench-AB, OpenDDE reached a 70.0% success rate (DockQ > 0.23) under ranking-based selection, significantly outperforming the next best open model (Protenix-v1 at 48.5%) and closed models like AlphaFold3 (48.8%).
  • On 2026ARK-AB (low-homology, newly released systems), OpenDDE achieved 66.4% success, outperforming ESMFold2 (51.0%) and Protenix-v1 (50.0%).
• High-Quality Predictions: The model showed particular strength in the medium- and high-quality regimes (DockQ > 0.49 and > 0.8), indicating improved modeling of binding geometry rather than just marginal recovery.
• Oracle Performance: Under an oracle selection setting (selecting the best candidate from multiple samples), OpenDDE achieved up to 81.9% success on FoldBench-AB, suggesting strong latent sampling capacity.
• Test-Time Scaling: The model exhibits consistent performance gains with increased inference-time sampling (more seeds), demonstrating that additional compute can systematically improve prediction accuracy.
• Case Studies: Qualitative analysis on new biological systems (e.g., PDB 8Z8V, 8ZP2, 8JYR) showed OpenDDE accurately recovering binding geometries and epitope-facing poses, even for membrane proteins and recent cryo-EM structures.

Significance and Claims

The paper positions OpenDDE as the first Apache-licensed all-atom generative foundation model reported to reach IsoDDE-level co-folding accuracy. Its primary significance lies in:

• Democratization: Providing an open, reproducible foundation for the community to explore, refine, and deploy biological intelligence at scale.
• Scalability: Demonstrating that co-folding models can scale effectively with data and compute, narrowing the gap with proprietary systems.
• Foundation for Drug Discovery: Establishing a structural core that can evolve from predicting molecular structures to designing, scoring, and optimizing therapeutic candidates.

The authors explicitly note limitations, acknowledging that comparisons with IsoDDE are constrained by the lack of public details on the latter's training recipe. They also clarify that the current model is optimized for macromolecular interfaces (protein-protein, antibody-antigen) and is not yet fully optimized for protein-small molecule docking or affinity ranking, which will require future data curation and chemistry-aware post-training. OpenDDE is presented as the "folding-centered foundation" of a larger, extensible drug discovery engine, rather than a complete, standalone system.