The Open Gate: OpenDDE Brings IsoDDE-Class Co-Folding to Apache-Licensed Drug Discovery

For three years, the most consequential shift in computational drug discovery has been structural: co-folding models that predict proteins, nucleic acids, and small-molecule ligands in a single pass, rather than treating each component as a separate prediction problem. AlphaFold 3 opened that door in 2024. Isomorphic Labs' IsoDDE pushed further on protein–ligand generalization, antibody–antigen interfaces, and binding affinity — but kept the engine closed.

On July 4, 2026, the Aureka AI Research OpenDDE project posted a manuscript to arXiv claiming the first Apache-licensed all-atom generative foundation model to reach IsoDDE-level co-folding accuracy. Checkpoints, inference pipelines, and benchmarks followed on GitHub and Hugging Face under Apache 2.0. The release does not settle every argument about open versus closed biotech AI. It does, however, give academic labs, biotech startups, and national programs something they have lacked: a reproducible baseline at the frontier of antibody–antigen structure prediction.

Why co-folding became the engine, not the endpoint

Structure prediction used to be a destination. You folded a protein, docked a ligand, scored the pose, and hoped the geometry survived contact with a wet lab.

Co-folding reframed the workflow. Models like AlphaFold 3 and the open Protenix and OpenFold3 lineages jointly model biomolecular complexes — proteins with DNA, RNA, ions, modified residues, and drug-like small molecules — at atomic resolution. That joint view matters because binding is rarely a static fit; it is an interface problem spanning chemistry, flexibility, and cross-molecular context.

IsoDDE, Isomorphic Labs' unified computational drug-design system, demonstrated that co-folding could anchor a broader engine: pocket identification, affinity estimation, and design modules conditioned on predicted structure. The catch, repeated across the OpenDDE paper and community discussion, is reproducibility. Closed frontier systems resist independent validation, slow community extension, and leave policy debates about biosecurity and access unresolved because outsiders cannot inspect what insiders claim.

OpenDDE enters that gap explicitly. Its authors describe co-folding not as a scoreboard metric but as a "shared structural reasoning layer" — a substrate for sequence–structure–function modeling that can, in principle, support de novo design, affinity prediction, and structure-conditioned optimization downstream.

What OpenDDE actually ships

The project released as OpenDDE-Preview on July 3, 2026, with a candid stability warning: CLI flags, JSON schemas, and checkpoints may change; predictions are not guaranteed reproducible across releases; the stack is not yet production-ready. That honesty matters in a field where "preview" often means demo-grade.

What researchers can run today:

• Two checkpoints on Hugging Face: a general-purpose opendde.pt and an antibody–antigen-tuned opendde_abag.pt.
• A Docker image (aurekaresearch/opendde:v1) and a command-line toolkit covering inference, MSA preprocessing, and structure conversion.
• Training code and benchmarks under Apache 2.0, with the arXiv manuscript documenting architecture, scaling laws, and test-time compute behavior.

Architecturally, OpenDDE is a 655-million-parameter all-atom model built on advances in atomic latent reasoning — internal refinement of local geometry, chemical context, and cross-molecular interfaces before structure generation — plus inference optimizations including context-parallel execution for large complexes. Training stages progressively increase crop size and atom budgets, with a late-stage de novo design objective mixed into the loss at small probability (0.1–0.2), signaling where the project intends to go even while today's release emphasizes structure prediction.

The paper's framing is worth taking seriously: OpenDDE unifies structure prediction and conditional design as masked generation under different partitions of known versus unknown molecular components. Fixed chains, motifs, or binding contexts are provided as conditions; the model denoises the rest. That is not marketing language — it is an architectural bet that the same diffusion backbone can serve discovery pipelines end to end.

The antibody–antigen numbers that matter

Therapeutic antibodies are where co-folding models earn or lose credibility. Interfaces are flexible, chemically heterogeneous, and expensive to get wrong in the clinic.

OpenDDE reports DockQ success rates — the fraction of predicted interfaces crossing quality thresholds — on three antibody–antigen benchmark collections, compared against AlphaFold 3, Chai-1, Boltz-1, OpenFold3, Protenix-v1, and Biohub's ESMFold2:

The 2026ARK-AB benchmark is itself a contribution: 164 PDB antibody–antigen complexes, 159 unique interface clusters, filtered for recent structures with low homology to training-era data. On that set, OpenDDE's 66.4% success rate compares to ESMFold2 at 51.0%, Protenix-v1 at 50.0%, OpenFold3 at 21.9%, Boltz-1 at 25.0%, and Chai-1 at 16.7%. The authors argue the gains hold in medium- and high-DockQ regimes, not only at permissive cutoffs — meaning improved binding geometry, not marginal acceptable poses.

Broader FoldBench performance reinforces balance: protein monomer mean LDDT of 0.890, protein–protein DockQ success 0.769, and antigen–antibody 0.700 — with the paper noting the largest improvement on interaction-heavy systems relative to ESMFold2 (0.581), AlphaFold 3 (0.488), and Protenix-v1 (0.485) on that task class.

Oracle performance exposes a ranking problem — and an opportunity

Raw accuracy tells half the story. OpenDDE also reports oracle selection: pick the best structure among multiple stochastic samples using ground-truth DockQ, approximating upper-bound model capacity when ranking is perfect.

Oracle success rates reach 65.9% (PXMeter-AB), 81.9% (FoldBench-AB), and 80.1% (2026ARK-AB) — well above ranked performance. Test-time scaling amplifies the pattern: on FoldBench-AB, oracle success rises from roughly 66% with one seed to above 90% with hundreds of seeds, while ranked high-quality (DockQ > 0.8) success moves more modestly from about 28% to 34%.

That gap is scientifically informative. OpenDDE frequently generates high-quality antibody–antigen conformations that its confidence head does not always surface as the top-ranked prediction. For drug-discovery teams, the implication is operational: inference budget is not waste if candidate pools are scored and filtered downstream. For model developers, calibration and selection are the next bottleneck — a familiar story from language-model scaling, now transplanted to biomolecular diffusion.

Scaling laws suggest co-folding is entering the LLM regime

The manuscript plots antibody–antigen performance against estimated training tokens and training cost (tokens × parameters). OpenDDE sits at the high-compute, high-performance end of a smooth, sublinear curve spanning Protenix-v1, AlphaFold 3, Protenix-v2, SeedFold, and ESMFold2, approaching but not fully closing the gap to IsoDDE as a reference point.

For OpenDDE specifically, the paper estimates roughly 2.04 × 10¹⁰ training tokens and 1.33 × 10¹⁹ training-cost units (tokens × 655M parameters), trained across multi-stage schedules on hundreds of GPUs. The authors' conclusion — that biomolecular structure modeling is entering the scaling regime that transformed language models — is bold but backed by a plotted trend, not a slogan.

Practical takeaway: incremental architectural tweaks alone may yield diminishing returns; data curation, distillation strategy, inference-time sampling, and compute scale appear jointly responsible for where each model lands on the curve.

Open versus closed at the therapeutic frontier

July 2026 is a crowded month for biomolecular AI. Biohub's MIT-licensed ESMC / ESMFold2 "world model" stack, posted to bioRxiv in June, pushes open complex prediction and binder design from evolutionary-scale sequence modeling. OpenDDE attacks the same frontier from the co-folding lineage rooted in AlphaFold 3–style all-atom diffusion, explicitly benchmarking against ESMFold2 and claiming IsoDDE parity within an open license.

Neither release ends the commercial advantage of closed systems. IsoDDE remains unavailable for independent replication. AlphaFold 3 access is gated. OpenDDE's preview status means production drug programs will still treat it as a research instrument, not a validated GxP toolchain.

What changes is the reference plane. When a startup pitches a novel antibody epitope model, reviewers can ask how it performs against a public OpenDDE checkpoint on 2026ARK-AB. When a university consortium negotiates cloud credits, scaling-law plots offer a compute roadmap rather than a vendor slide. When regulators and funders debate open science in biosecurity contexts, Apache-licensed weights give a concrete artifact to stress-test — unlike press releases about closed superiority.

What remains unproven

Verification discipline requires naming limits clearly:

• Peer review: The arXiv manuscript is preprint science. Benchmarks are author-reported; independent replication on 2026ARK-AB will matter.
• IsoDDE parity: "IsoDDE-level" is the authors' framing against a closed system outsiders cannot rerun head-to-head under identical conditions.
• Design and affinity modules: Today's release centers structure prediction; de novo design training appears in late stages but is not the headline validated capability.
• Production readiness: GitHub marks the release preview-grade, with explicit non-reproducibility caveats across versions.
• Therapeutic translation: Accurate antibody–antigen co-folding does not, by itself, produce developable biologics — manufacturability, immunogenicity, and in vivo behavior remain downstream.

OpenDDE also builds on the AlphaFold 3 ecosystem — Protenix, OpenFold, ColabFold — and the paper cites IsoDDE as the performance north star. Credit and lineage are shared; the novelty claim is openness at the frontier, not a from-scratch physics breakthrough.

The structural reasoning layer thesis

Drug discovery organizations already speak in pipelines: hit identification, lead optimization, ADMET, tox, CMC. OpenDDE's authors argue for an upstream layer beneath those stages — a co-folding engine that turns sequence inputs into consistent complex geometry, then feeds design, scoring, and optimization modules that share the same representation stack.

That thesis aligns with how July's broader AI-biology moment is evolving. OpenAI's GPT-Rosalind and LifeSciBench push agentic tool use across literature, genomics, and assay planning. Google DeepMind's Co-Scientist, published in Nature, targets hypothesis generation in life sciences. Anthropic's biology agents work highlights deterministic retrieval layers atop messy public databases. OpenDDE is the structural counterpart: less conversational, more geometric, but equally aimed at making virtual experiments precede expensive wet-lab ones.

The open gate is not merely a license footer. It is a bet that biomolecular intelligence compounds faster when folding, reasoning, and scaling can be inspected, challenged, and extended in public — starting with the hardest interface class in modern therapeutics, and with the humility of a preview stamp on the repository README.

### Sources

• arXiv — Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine (July 4, 2026)
• GitHub — aurekaresearch/OpenDDE (July 3, 2026 preview release)
• Hugging Face — aurekaresearch/OpenDDE model weights (2026)
• bioRxiv — Language Modeling Materializes a World Model of Protein Biology (June 4, 2026)
• OpenAI — Introducing GPT-Rosalind for life sciences research (2026)
• Google DeepMind — Co-Scientist: A multi-agent AI partner to accelerate research (2026)