atlasfold-m-260725

AtlasFold-M-260725

AtlasFold-M-260725 predicts multimeric protein complex structures directly from one sequence per chain without requiring a multiple sequence alignment (MSA).

Installation

pip install "atlasfold[fold]"

Optional cuEquivariance kernels can be installed with:

pip install "atlasfold[fold,cuequiv]"

Usage

The AtlasFold-M and AtlasLM weights are downloaded automatically from Hugging Face:

from atlasfold.pretrained import load_model
from atlasfold.runner_multimer import MultimerFoldingRunner

model = load_model("atlasfold-m-260725", device="cuda")
runner = MultimerFoldingRunner(model)
result = runner.fold(
    "example_complex",
    ["MKTAYIAKQRQISFVKSHFS", "GGHVDHGKSTTTGHLIYK"],
)
print(result.best.iptm)

Command-line inference also downloads the weights automatically:

python run_atlasfold.py \
    --model multimer \
    --input-fasta multimers.fasta \
    --out-dir predictions/multimers

Use --cache-dir PATH to select a cache location. --model-path PATH remains available as an optional local checkpoint override.

See the AtlasFold repository for complete CLI and Python API documentation.

Files

• weights/atlasfold-m-260725.pth: AtlasFold-M multimer state dict, exported from the atlas-m-v2-25k.pth checkpoint.
• SHA256: 210e3e9cfaac92d276ae44910d9b66d0673ecd5c225261e64dccf58c8dfe123a

License

The source code and model weights are released under the MIT License.

Citation

If you use this model in your research, please cite:

@article{seo2026atlasfold,
  author = {Seo, Seonghwan and Kim, Hyeongwoo and Moon, Seokhyun and Kim, Woo Youn and {Team KAIST}},
  title = {AtlasFold: Protein structure prediction with metagenomic-scale language models},
  year = {2026},
  doi = {10.64898/2026.09.04.749352},
  URL = {https://www.biorxiv.org/content/10.64898/2026.09.04.749352v2},
  journal = {bioRxiv}
}