Ligandmpnn Noise010

Model Details

LigandMPNN sequence-design model with ligand/atom context features.

Converted LigandMPNN sequence-design checkpoint with ligand and non-protein atom context.

Model Provenance

• Upstream Project: LigandMPNN
• Primary Paper: Atomic context-conditioned protein sequence design using LigandMPNN
• Upstream License: MIT for upstream LigandMPNN code and model parameters

Model Specification

[!NOTE] Checkpoint metadata: k_neighbors=32, atom_context_num=25; the noiseXXX suffix identifies the training-noise variant.

Links

• Hub repository: oteam/ligandmpnn-noise010
• Upstream paper: Atomic context-conditioned protein sequence design using LigandMPNN
• Upstream repository: LigandMPNN
• Code: ofoldx/pipelines/design.py
• Project repository: https://github.com/OTeam-AI4S/OFoldX
• Issues: https://github.com/OTeam-AI4S/OFoldX/issues

Usage

The artifact depends on the ofoldx library. Install it with pip:

pip install ofoldx

Pipeline Usage

Load the artifact from oteam/ligandmpnn-noise010 with the OFoldX task pipeline. Use AutoModel or AutoProcessor only when you need lower-level control:

from ofoldx.pipelines import Pipeline

pipeline = Pipeline.from_pretrained("oteam/ligandmpnn-noise010")

When a matching processor is available, load it with AutoProcessor.from_pretrained(...) and pass the processed batch to the model.

Interface

• Task: design_generation
• Artifact kind: pipeline
• Architecture: ligandmpnn
• Runtime files: manifest.json, config.json, and model.safetensors when present

Training Details

OFoldX did not train these weights. This repository contains a converted checkpoint and OFoldX runtime metadata for loading it.

Training Data

LigandMPNN was trained on PDB assemblies as of 2022-12-16 from X-ray crystallography or cryo-EM structures better than 3.5 A and shorter than 6,000 residues, with 30% sequence-identity clustering and non-protein atomic context such as ligands, nucleic acids, and metals. OFoldX does not redistribute the training set.

Training Procedure

Upstream LigandMPNN uses categorical cross-entropy sequence-design training with Adam, mixed precision/checkpointing, and released noisy-backbone checkpoint variants. OFoldX converts released LigandMPNN checkpoints into model.safetensors; it does not run LigandMPNN training.

Evaluation

OFoldX conversion reports and contract tests validate artifact structure and checkpoint loading. Task-level scientific evaluation should be checked against the corresponding upstream model release or paper.

Limitations

• This artifact is distributed for research use.
• Inputs must match the model-specific processor and expected biomolecular representation.
• OFoldX is pre-alpha, so APIs and artifact metadata may still change before a stable release.

Citation

Please cite the upstream LigandMPNN work for the source checkpoint. If OFoldX supports your work, please also cite or link the OFoldX project repository.

@article{dauparas2025atomic,
  author = {Dauparas, Justas and Lee, Gyu Rie and Pecoraro, Robert and An, Linna and Anishchenko, Ivan and Glasscock, Cameron and Baker, David},
  title = {Atomic context-conditioned protein sequence design using LigandMPNN},
  journal = {Nature Methods},
  year = {2025},
  doi = {10.1038/s41592-025-02626-1}
}

Contact

Please use OFoldX GitHub issues for questions or comments about this model card.

License

The Hub license metadata, when present, reflects the source checkpoint or upstream project license. The OFoldX project license is not yet finalized. The source checkpoint is associated with the upstream license noted above: MIT for upstream LigandMPNN code and model parameters. Review both OFoldX and upstream terms before redistribution or production use.