OpenDDE

OpenDDE (Open Drug Discovery Engine) is an open-source, all-atom biomolecular co-folding foundation model developed by Aureka, an AI TechBio company. It predicts the joint three-dimensional structure of biomolecular complexes — proteins bound to other proteins, small-molecule ligands, nucleic acids, ions, and antibody-antigen pairs — from the sequences and chemical definitions of their components. The authors position the model as a shared structural reasoning layer that can serve as the entry point to a scalable, AI-driven drug discovery pipeline rather than as a stand-alone structure predictor.

The release addresses a gap in the co-folding landscape. Systems that reach comparable accuracy on protein complexes and protein-ligand binding are frequently proprietary or restrict their weights to non-commercial use. OpenDDE instead publishes its training code, inference pipelines, model checkpoints, and evaluation benchmarks under the Apache-2.0 license, making one of the larger openly available structural reasoning models accessible to academic groups, startups, and pharmaceutical teams alike.

OpenDDE is distributed as two checkpoints: a general-purpose model (opendde.pt) covering the full range of complex types, and an antibody-antigen-tuned variant (opendde_abag.pt) optimized for immune recognition interfaces. It is an early-stage preview release; the authors note that command-line flags, input and output formats, and the checkpoints themselves may change between versions, and that predictions are not yet guaranteed to be reproducible across releases.

#Key Features

• All-atom co-folding: Models complexes spanning proteins, small-molecule ligands, DNA, RNA, and ions within a single unified all-atom framework, resolving binding interfaces at atomic resolution.
• Fully open release: Training code, inference pipelines, weights, and benchmarks are published under Apache-2.0, with checkpoints hosted on Hugging Face and code on GitHub.
• Antibody-antigen specialization: The dedicated opendde_abag.pt checkpoint improves prediction of CDR-epitope contacts and antibody-antigen binding interfaces.
• Atomic latent reasoning: Combines a diffusion-based structure generator with atomic latent reasoning and inference-time optimization to handle large, heterogeneous assemblies.
• Scaling-law analysis: The work identifies two scaling-law directions for co-folding models, mapping practical routes to improvement through data, model, inference, and training scaling.

#Technical Details

OpenDDE couples a pairformer-style representation trunk with a diffusion module that denoises atomic coordinates, augmented by atomic latent reasoning and inference optimization for large-scale complexes. The released model has roughly 655 million trainable parameters and was trained using approximately 414,000 GPU-hours. Antibody-antigen co-folding accuracy is reported across three benchmarks under top-ranked selection (with oracle selection in parentheses): 51.0% success on PXMeter-AB (65.9%), 70.0% on FoldBench-AB (81.9%), and 66.4% on the accompanying 2026ARK-AB benchmark (80.1%). The 2026ARK-AB set is released alongside the model to support standardized antibody-antigen evaluation.

#Applications

OpenDDE targets structure-based drug discovery workflows: predicting how small molecules dock into protein pockets, modeling protein-protein assemblies, and resolving antibody-antigen interfaces to guide biologics design. Its downstream framing extends to de novo design, binding-affinity estimation, and structure-conditioned optimization of therapeutic candidates. Because the full training and inference stack is open, academic laboratories, biotech startups, and pharmaceutical teams can run the model locally, fine-tune it, or integrate its structural predictions into larger design loops. Aureka frames the model as one component of a broader dry-wet discovery system that pairs computational modeling with high-throughput automated wet-lab screening for complex antibody modalities.

#Impact

By releasing a competitive co-folding model together with its training code, checkpoints, and benchmarks under a permissive license, OpenDDE lowers the barrier to structural reasoning for groups that cannot access proprietary systems, and its scaling-law analysis offers the community a roadmap for extending co-folding accuracy. The most significant caveat is maturity: as a preview release, its interfaces and weights are explicitly subject to change and predictions are not yet stable across versions, so results should be treated as an evolving early-stage capability rather than a fixed reference. Its emphasis on antibody-antigen prediction, an interaction class that remains difficult for general co-folding models, makes it particularly relevant to open biologics discovery.