aurekaresearch/OpenDDE

OpenDDE – an all‑atom biomolecular foundation model

What it is – OpenDDE is an open‑source, large‑scale neural model that predicts the three‑dimensional structure of biomolecules (proteins, DNA, RNA, ligands, ions) by co‑folding all components together. It is positioned as a “foundation model” for drug‑discovery tasks such as structure prediction, design, and optimisation.

Key capabilities

• All‑atom co‑folding – the model can handle multiple chains (e.g., antibody‑antigen complexes) in a single forward pass.
• Conditional inference – with the --use_tfg_guidance flag you can steer the sampler using partial knowledge of an interface (contacts, epitope, movable chains) without retraining the weights.
• Scalable inference – supports single‑GPU, CPU, and a multi‑GPU “Fold‑CP” mode that distributes the computation across P GPUs (1 × P topology).
• Flexible preprocessing – optional MSA, template search, and RNA‑MSA pipelines can be toggled on/off; the latter require external tools (HMMER, Kalign) and large search databases.
• Model variants – a general‑purpose checkpoint (opendde.pt) and an antibody‑antigen‑tuned checkpoint (opendde_abag.pt).

How to get it

• Install via PyPI (opendde) using the uv tool, choosing the appropriate --torch-backend (CPU or CUDA 12.6) and optional [gpu] extra for cuEquivariance kernels.
• Or clone the repo and install from source (pip install -e .).
• Pre‑built Docker image (aurekaresearch/opendde:v1) for GPU inference.
• Model weights are hosted on Hugging Face; the repo provides a helper script to download the checkpoint and runtime assets into ~/.cache/opendde (or a custom OPENDDE_ROOT_DIR).

Running a prediction

1. Write a JSON job description (e.g., tiny.json with a short protein sequence).
2. Call the CLI:

   opendde pred -i tiny.json -o ./output -n opendde_v1 \
       --use_msa false --use_template false --use_rna_msa false \
       --sample 1 --step 200 --cycle 10
   

   The command produces a directory tree output/<job>/seed_<seed>/predictions/ containing the predicted coordinates.
3. For larger, multi‑GPU runs, launch with torchrun and the --foldcp_mode distributed flags as described in the README.

Documentation & resources

• Detailed installation guides (PyPI, source, Docker).
• JSON schema, MSA/template pipelines, kernel options, and a tutorial are all in the docs/ folder.
• Technical reports for the preview release (co‑folding) and the newer conditional‑inference feature are linked.
• Benchmarks (2026ARK‑AB, OpenBench2026) and a reproducibility guide for Fold‑CP are provided.

Status – The repository is labelled preview; CLI flags, JSON fields, and checkpoints may change, and predictions are not guaranteed to be reproducible across releases. It is not yet meant for production pipelines.

License – Apache‑2.0.

---

In short, OpenDDE is a research‑grade, open‑source foundation model for all‑atom biomolecular structure prediction, with a focus on co‑folding and conditional guidance, packaged as a Python CLI and Docker image.