OpenDDE open-source engine: co-folding biomolecular model

Aureka AI OpenDDE project introduced the Open Drug Discovery Engine (OpenDDE) in a paper submitted to arXiv on 4 Jul 2026 (arXiv:2607.03787). The project presents an open-source, all-atom biomolecular foundation model that treats co-folding as its entry point and releases training code, inference pipelines, checkpoints, and benchmarks alongside the paper.

What is OpenDDE?

OpenDDE is an open-source all-atom biomolecular foundation model built around co-folding, designed to act as a shared structural reasoning layer for modeling sequence-structure-function relationships across biomolecular complexes. The paper frames co-folding not as a terminal output but as the foundation for de novo design, affinity estimation, structure-conditioned optimization, and other downstream tasks.

The authors label OpenDDE a platform rather than a standalone predictor: it combines co-folding with atomic latent reasoning to support structure prediction today and provide a basis for design and scoring workflows in the future.

How does OpenDDE work and what did the paper release?

OpenDDE integrates advances in all-atom architecture, atomic latent reasoning, inference optimization, and large-scale data processing, and the paper claims it achieves "IsoDDE-level co-folding accuracy" within an openly accessible framework. The model uses co-folding as its core method and positions that capability as a shared reasoning layer across sequence, structure and function.

Alongside the manuscript, the project releases training code, inference pipelines, checkpoints, and benchmarks. The arXiv submission is listed as arXiv:2607.03787 and was submitted on 4 Jul 2026; the submission metadata shows a 15,794 KB file in the record. The paper also identifies two scaling-law directions for co-folding models and describes practical routes to improve performance via data, model, inference, and training scaling.

Why it matters

OpenDDE shifts the framing of structure prediction from an end task to an engineering layer that can be reused across downstream drug discovery problems. By publishing code, checkpoints, and benchmarks, the project aims to "democratize access to frontier biomolecular intelligence" and "accelerate global collaboration," language the authors use to describe the release.

That open stance matters because it makes the components needed for replication and community-driven improvement publicly available: training pipelines, inference code, and benchmarks provide the base material for other groups to validate claims, iterate on architectures, or pursue domain-specific optimization such as affinity estimation and structure-conditioned design.

What to watch

Watch for community validation of the claimed "IsoDDE-level co-folding accuracy" and for follow-up work exploiting the paper's two identified scaling-law directions for co-folding models. Also track adoption of the released artifacts: successful reuse of the training code, inference pipelines, checkpoints, and benchmarks will be the clearest signal that OpenDDE is serving as a shared structural reasoning layer beyond the original team.

The paper's central promise is that an open, all-atom, co-folding foundation model can move work from predicting molecular structures toward designing, scoring, and optimizing therapeutic candidates for human health; the next milestones will be reproducible benchmarks and community-driven extensions that test those claims.