Multi-Modal Foundation · aurekaresearch
OpenDDE
An all-atom biomolecular foundation model using co-folding for structure prediction and drug discovery.
What OpenDDE is
OpenDDE is an open-source, all-atom biomolecular foundation model developed by aurekaresearch. It models interactions across multiple molecular classes, turning biomolecular co-folding—the simultaneous structural prediction of interacting molecular entities—into a computational engine for structure prediction, molecular design, and lead optimization in drug discovery.
The system supports multimolecular assemblies encompassing protein chains, DNA, RNA, ligands, and ions, including explicit covalent bonds between entities. It builds upon concepts and components from the AlphaFold 3 ecosystem, including AlphaFold 3, Protenix, OpenFold, and ColabFold, offering researchers configurable workflows for structural biology and chemistry.
How it works
OpenDDE employs an all-atom diffusion architecture to predict the coordinates of complex biomolecular assemblies. Structure generation is configured through sampling steps, recycling iterations, and random seeds, executing in FP32 precision with automated kernel dispatch for triangle operations.
The model incorporates multiple sequence alignments (MSAs) for protein and RNA sequences, along with structural template searches that rely on external tools such as HMMER and Kalign. For larger inputs, inference can be distributed across a four-GPU context-parallel mesh using Fold-CP.
Key facts
- Release
- 2026-07-03 (Preview)
- Architecture
- All-atom diffusion co-folding model
- Supported entities
- Proteins, DNA, RNA, ligands, ions, covalent bonds
- Checkpoints
- opendde.pt (general-purpose), opendde_abag.pt (antibody-antigen tuned)
- Distributed execution
- 4-GPU Fold-CP context parallelism (2x2 mesh)
What you can use it for
- Biomolecular complex prediction
- Predicting 3D structures of assemblies containing combinations of protein chains, nucleic acids, small-molecule ligands, and ions.
- Antibody-antigen modeling
- Using the specialised opendde_abag.pt checkpoint to resolve structural interfaces between antibodies and target antigens.
- Drug lead optimization
- Co-folding small-molecule ligands with target macromolecular complexes to evaluate binding poses and structural interactions.
- Alignment-free rapid screening
- Executing rapid predictions without multiple sequence alignments or template searches, using only the model weights and runtime files.
- Context-parallel distributed folding
- Splitting large molecular structures across four GPUs using the Fold-CP context-parallel mode.
Run it here
We list this model but cannot run it yet: unsupported task: other.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Processes diverse molecular entities within a unified all-atom co-folding framework, including proteins, nucleic acids, ligands, and ions.
- Includes a specialised checkpoint fine-tuned specifically for antibody-antigen complex prediction.
- Supports multiple execution modes ranging from single-GPU and CPU to 4-GPU context-parallel distributed inference.
- Allows preprocessing components such as protein MSAs, RNA MSAs, and template searches to be selectively bypassed.
Watch out for
- Outputs are not guaranteed to be reproducible across releases during the preview phase.
- Command-line flags, checkpoint formats, and JSON input schemas may change between versions.
- Full preprocessing requires large external sequence databases, network access, and binaries such as HMMER and Kalign.
- Inference on CPU is slow without GPU-accelerated triangle kernels.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · apache-2.0
The apache-2.0 licence carries terms worth reading before you depend on it.
Questions
Can I use OpenDDE commercially?
Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.
What molecular inputs does OpenDDE accept?
OpenDDE takes JSON input files specifying protein chains, DNA sequences, RNA sequences, ligands, ions, and optional covalent bonds.
Which checkpoint should be selected for antibody studies?
The dedicated opendde_abag.pt checkpoint is tuned specifically for antibody-antigen complexes, whereas opendde.pt serves general-purpose co-folding.
Can OpenDDE run without downloading large search databases?
Yes, setting options to disable protein MSAs, template searches, and RNA MSAs allows inference using only the checkpoint and minimal runtime files.
How does Fold-CP multi-GPU inference work?
Fold-CP runs across four GPUs in a 2-by-2 context-parallel mesh launched via torchrun, enabling the processing of larger molecular assemblies.
Written from OpenDDE's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.