OpenDDE: An Open-Source All-Atom Foundation Model for Biomolecular Structure Prediction and Drug Discovery
Aureka Research recently released OpenDDE (Open Drug Discovery Engine), an open-source, all-atom foundation model for biomolecular structure prediction, built as the intended structural core of a broader AI drug-discovery platform spanning design, optimization, and affinity estimation. The current release is strongest as a structure-prediction model, particularly for antibody-antigen complexes, where it outperforms AlphaFold3 and ESMFold2.
The premise is that structure prediction shouldn't be the endpoint. It should be the structural reasoning layer underneath design, scoring, and optimization, and that's the direction Aureka Research says it wants to take OpenDDE.
Why release this now
This lands in the middle of a lot of excitement (and some frustration) around Isomorphic Labs' proprietary IsoDDE, which pushed further toward an operational drug-design engine, with reported advances in protein-ligand generalization, affinity prediction, and binding-pocket identification substantial enough that Mohammed AlQuraishi described IsoDDE as being "on the scale of an AlphaFold 4." Folks outside Isomorphic Labs have mostly been left guessing at how it was done, and much of that capability is presumed to trace back to internal data from Isomorphic's pharma partnerships, data the rest of us don't have access to.
OpenDDE is not yet an open-source equivalent of IsoDDE. But it's a real step toward one, and, as I'd flagged as a plausible direction earlier this year, we're getting closer to having an open-source counterpart to IsoDDE than I expected this soon. OpenDDE makes that possibility feel a lot less speculative.
Reasoning at the atomic level
A conventional residue token compresses an entire amino acid or nucleotide into a single representation. OpenDDE argues that's too coarse, since different parts of a residue play different structural roles, so it expands residues into seven categories of structural tokens (protein backbone, protein side chain, DNA/RNA backbone, DNA/RNA base, ligand and individual atom tokens), then runs a Structural Refiner over them using pair-conditioned attention and triangular pair updates. The idea is a coarse-to-fine path: residue-level relations, then structural-token relations, then atom coordinates, letting the model reason explicitly about things like how a side chain should pack across a protein-protein interface or how an antibody loop should face an antigen surface, rather than jumping straight from residue identity to atom positions.
One diffusion framework for prediction and design
Coordinate losses alone can place two chains roughly near each other while still leaving steric clashes, unrealistic interface gaps, or poorly packed side chains. OpenDDE adds a differentiable shape-complementarity objective, supervising whether interacting surfaces actually face one another, whether their normals are oppositely oriented, and whether the interface gap and side-chain packing look physically reasonable, at the pair, token, and global level.
Structure prediction and conditional design are also framed as the same denoising problem: a "known atoms" mask fixes structural context and a "target atoms" mask marks what the model must generate. With nothing fixed, that's ordinary structure prediction; with part of the system (an antigen, a pocket, a motif) held fixed, that's conditional design. It's an elegant unification, and it means OpenDDE doesn't need a separate design network. But the current report only provides structure-prediction benchmarks, not design benchmarks, so this unified formulation is a demonstrated architectural capability rather than a validated design capability, at least for now.
How well did it work?
OpenDDE is a 655-million-parameter all-atom diffusion model, trained for roughly 414,000 GPU-hours. On antibody-antigen co-folding across three benchmarks (PXMeter-AB, FoldBench-AB, and a new temporally-held-out benchmark the authors call 2026ARK-AB), it posts the highest reported ranked DockQ success rate among tested models: 51.0%, 70.0%, and 66.4% respectively, with oracle (best-of-sample) success reaching 65.9%, 81.9%, and 80.1%. On the broader FoldBench suite it's tied or best on protein monomer, protein-protein, RNA, and protein-RNA prediction, though it's competitive rather than best on protein-ligand pose prediction (60.1% vs. AlphaFold3's 64.9%), a reminder that antibody-antigen prediction is where the real advantage currently lives.
The model generates better answers than it can rank
The gap between ranked and oracle performance is the most interesting result in the report. Increasing the number of diffusion seeds keeps improving oracle success (reaching roughly 90% with hundreds of samples), which means OpenDDE frequently samples a high-quality structure, but its confidence model doesn't reliably recognize it as the best one. That's a generation-vs-ranking bottleneck, not a generation-quality problem, and it suggests that better confidence calibration, clustering, or external rescoring could unlock real gains without retraining the underlying structure generator. On the systems side, OpenDDE also ships Fold-CP, a four-GPU context-parallel scheme for the pair representation, letting it handle a 2,000-token complex that runs out of memory on a single GPU.
Worth knowing before you get too excited
The design capability isn't benchmarked convincingly in this release; there's no reported binder-design success rate, sequence recovery, or wet-lab validation, so it should be treated as a promising framework, not an established design platform. The 2026ARK-AB benchmark, while useful for testing temporal generalization, was assembled and evaluated by the OpenDDE authors themselves, so independent replication would strengthen the generalization claim. The scaling-law plots comparing OpenDDE against other models are also correlational rather than a controlled scaling study, since the compared models differ in architecture, data, and training objectives, not just compute. And the developers explicitly label this a preview release: the technical report and the public repository disagree on inference settings (four cycles/20 diffusion steps vs. ten cycles/200 steps), so exact reproduction will need version-pinned configs.
Put together: OpenDDE is a strong open antibody-antigen and biomolecular co-folding foundation with a promising architecture for future design, not yet a fully validated end-to-end drug-discovery engine. But as an open, Apache-licensed, all-atom co-folding model that beats AlphaFold3 and ESMFold2 where it's strongest, it's a meaningful data point in how quickly the open-source side is catching up to what closed labs are claiming.
For similar tools, see the BioMoDes Biomolecular Structure Prediction page.
References
- OpenDDE Project, Aureka AI Research. Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine. Technical Report (2026).
- OpenDDE code (GitHub).
- Project homepage.
- Model weights (HuggingFace).
- Nature. On IsoDDE and the “AlphaFold 4” characterization.
This post is an AI-reworded, expanded version, in my own voice, of a summary I originally posted on LinkedIn.
I try my best to make the information on this website as accurate as possible.
If you find any errors in the contents of this page or any other page on this website,
I would greatly appreciate that you kindly get in touch with me at
contact[at]abeebyekeen[dot]com.
If you are interested in joining my free weekly “BioModes and Top Reads” newsletter, please subscribe below.