SoupFold: soup-opendde-pack

SoupFold (paper) folds a complex from the pair representations of several co-folding models. This repository holds one checkpoint that makes one such combination a single model, SoupOpenDDE: OpenDDE as the anchor, ESMFold2 and Protenix-v1 as peers, and the two networks that map the peers into OpenDDE's representation space. The peers' trunks and OpenDDE's trunk run in one process, the representations are mixed in memory with equal weights, and OpenDDE goes on to diffusion and confidence. No representation is written to disk.

This is a redistribution of third-party parameters. See NOTICE.md for each part's license, source and changes.

file	what it is
soup_opendde_pack.pt	the checkpoint, 18.2 GB (SHA256SUMS has its hash)
NOTICE.md	licenses, upstream revisions with file hashes, and every change from the originals
LICENSES/	the upstream license files

Use

git clone -b soup-opendde-pack https://github.com/hsjang0/soupfold.git && cd soupfold
uv venv --python 3.12 && uv pip install -e .
.venv/bin/soupfold-pack --input examples/9mnb/9mnb.json

The first run downloads this checkpoint and the CCD data (from aurekaresearch/OpenDDE).

The checkpoint is a pickle, and the code loads it with weights_only=False, then unpickles the bundled ccd.pkl. Loading a pickle can run code, so check the hash in SHA256SUMS and only load a file you trust.

Limitations

• The transfer networks are not trained on RNA or DNA and may not generalize to complexes with nucleic acids.
• Outputs are computational predictions and need scientific validation.

Citation

Please cite SoupFold and the models it bundles; the BibTeX entries are in the SoupFold README.

@article{jang2026soupfold,
  title   = {Co-folding with a Soup of Representations},
  author  = {Jang, Hyosoon and Kim, Taewon and Ahn, Sungsoo},
  journal = {arXiv preprint arXiv:2609.15552},
  year    = {2026}
}