Biotech · global
OpenDDE Preprint Releases Open-Source All-Atom Model, Moving Drug AI Toward a Testable Structural Foundation
Aureka AI has released the code, weights, and evaluation workflow for its co-folding model together, so molecular structure reasoning in drug discovery no longer remains confined to closed systems; but this is still an early preprint result, and there remains a path of validation before it can truly screen drugs, estimate affinity, or design therapies.
One of the most costly mistakes in drug development is misjudging how a molecule will approach, interfere with, or stabilize a biological target before it even enters the laboratory. The OpenDDE preprint published by Aureka AI on arXiv targets exactly this early decision-making stage: using an open-source all-atom biomolecular foundation model to help infer the three-dimensional relationships among proteins, antibodies, nucleic acids, and other molecules in complexes.
OpenDDE stands for Open Drug Discovery Engine. The research team positions it as a structural reasoning layer centered on “co-folding,” meaning it does not only predict the shape of a single protein, but attempts to handle the conformations that arise when multiple molecules bind to one another at the same time. The paper says the system currently focuses mainly on complex structure prediction, and will only later be extended to de novo molecule design, affinity estimation, and structure-conditioned candidate drug optimization.
The signal in this release is not only the model itself, but also the degree of openness. The preprint states that OpenDDE releases training code, inference workflows, model checkpoints, and benchmarks, under an Apache-2.0 license. For biomedical AI, this means external researchers have a chance to rerun the workflow, examine data processing and evaluation settings, rather than only reading result summaries from a closed model.
Technically, OpenDDE uses a coarse-to-fine structure generation architecture: it first builds residue-level relational representations, then unfolds them into finer structural tokens, such as protein backbones, side chains, nucleic acid backbones, bases, and ligand or atomic units, before a diffusion-based coordinate generator produces all-atom positions. The team also added a training objective for interface shape complementarity, hoping the model will not merely place atoms in nearby positions, but better reflect the orientation, spacing, and local packing of binding surfaces.
The main validation presented in the preprint focuses on antibody-antigen structure prediction. The research team says OpenDDE achieved DockQ success rates of 51.0%, 70.0%, and 66.4% on the PXMeter-AB, FoldBench-AB, and 2026ARK-AB datasets, respectively, and outperformed models including AlphaFold3, Chai-1, Boltz-1, OpenFold3, Protenix-v1, and ESMFold2 in several comparisons. These tasks have practical significance because early antibody therapy design often requires judging whether an antibody binds an antigen epitope in the correct pose; but these figures still come from evaluations reported by the team in the preprint, and await more independent reproduction.
Background Context
In recent years, biomolecular AI has gradually moved from single-protein folding toward complexes, antibodies, ligands, and design tasks. AlphaFold3 brought atom-level modeling of interactions among multiple types of molecules into focus, while subsequent open-source and commercial systems have tried to push structure prediction into drug discovery workflows. OpenDDE differs in that it places an “accessible open-source foundation” at the center of its argument, responding to the problem that closed high-performance models are difficult to audit, compare, and modify.
The limitations are equally clear. The paper’s authors acknowledge that OpenDDE currently focuses mainly on protein-protein and antibody-antigen interfaces, and cannot directly claim to already be suitable for small-molecule ligand docking, virtual screening, or affinity ranking; these tasks may require more chemistry-oriented data curation, post-training, and specialized evaluation. The team also notes that comparisons with closed systems such as IsoDDE are limited by insufficient public information, making it difficult to determine whether differences ultimately come from architecture, data, training strategy, or inference engineering.
Therefore, OpenDDE is more like an open structural infrastructure layer than a mature automated drug discovery machine. It may lower the barrier for academic and small teams to enter advanced biomolecular modeling, and may also allow different laboratories to use the same set of tools to examine antibody interfaces, candidate binding proteins, or new design hypotheses; whether it can truly shorten the drug development cycle will still depend on whether subsequent independent benchmarks, wet-lab feedback, and preclinical decisions can connect with the structural reasoning provided by this model.