LigandMPNN

This repository contains oteam/ligandmpnn, converted from ligandmpnn_v_32_010_25.pt.

LigandMPNN designs protein sequences on a fixed backbone using nearby non-protein atoms as structural context.

Disclaimer

This is an unofficial OFoldX implementation of LigandMPNN. The upstream authors did not write, review, or endorse this implementation or model card.

Model Details

LigandMPNN is an autoregressive inverse-folding model for protein sequence design with explicit atomic context. In addition to the protein backbone, it embeds nearby ligand, nucleic-acid, metal, and other non-protein atoms, allowing sequence generation to respond to the local chemical environment around an interface or binding site.

Variants

ligandmpnn_v_32_005_25.pt: trained with coordinate-noise level 0.05.
ligandmpnn_v_32_010_25.pt: trained with coordinate-noise level 0.10.
ligandmpnn_v_32_020_25.pt: trained with coordinate-noise level 0.20.
ligandmpnn_v_32_030_25.pt: trained with coordinate-noise level 0.30.

Released weight variants differ in the coordinate perturbations used during training but use the same architecture and public operations. The Pipeline provides sequence design and four probability routes: sequence-conditioned and sequence-free autoregressive evaluation, target-last conditional evaluation, and target-first unconditional evaluation. Side-chain packing is provided separately by MPNN Packer, with LigandMpnnDesignAndPackPipeline as the explicit two-model composition.

Its prepare stage returns a LigandMpnnDesignAndPackPipelineInput containing the original structure, design input and observation policy. A filter_by_occupancy=False choice during preparation also applies to subsequent packing; forward consumes that retained choice.

Model Specification

Hidden Size	Encoder Blocks	Atom-Context Blocks	Context Blocks	Decoder Blocks	Neighbors	Context Atoms	Num Parameters (M)
128	3	2	2	3	32	25	2.619

Usage

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

pip install ofoldx

Sequence Design

Load the concrete design Pipeline and design on a read structure:

import ofoldx
from ofoldx.io import read_structure

structure, report = read_structure("complex.cif")
pipeline = ofoldx.pipeline("design", model="oteam/ligandmpnn")
output = pipeline.design(structure)

Staged Execution and Scoring

The callable design Pipeline uses the same preparation, computation and decoding as its explicit stages:

prepared = pipeline.prepare(structure)
raw = pipeline.forward(prepared.input, seed=7)
result = pipeline.postprocess(prepared, raw)
result = pipeline(structure, seed=7)

Scoring has its own task Pipeline and reuses the same converted weights:

scorer = ofoldx.pipeline("score", model="oteam/ligandmpnn")
score = scorer(structure, num_decoding_orders=8, seed=7)

For an in-memory model, construct LigandMpnnScorePipeline with the designer's model and processor. Its prepare, forward and postprocess stages evaluate the structure's sequence without sampling new identities.

Training Details

Training Data

LigandMPNN was trained on protein structures with local atomic context from ligands, nucleic acids, metals, and other non-protein entities.

Training Procedure

LigandMPNN uses autoregressive sequence recovery with coordinate perturbations for robustness. OFoldX exposes a differentiable teacher-forced forward path and a typed criterion for training and fine-tuning on caller-supplied structures. This does not reproduce the upstream data curation, sampling, or training procedure; callers remain responsible for selecting structures and for custom training behavior beyond the provided model and typed objective.

Citation

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

If OFoldX supports your research, also cite the OFoldX software:

@software{chen2026ofoldx,
  author = {Chen, Zhiyuan},
  title  = {OFoldX},
  year   = {2026},
  url    = {https://github.com/OTeam-AI4S/OFoldX},
}

Contact

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

License

The original model parameters are distributed under the MIT license, retained in LICENSE.

The OFoldX implementation is separately licensed under AGPL-3.0-or-later; see license.md and license-faq.md.

These converted parameters use native OFoldX tensor names and the canonical vocabulary. Conversion changes the parameter representation and vocabulary axes; it does not change the upstream weight license.

Upstream source revision: 26ec57ac976ade5379920dbd43c7f97a91cf82de.

Downloads last month: -; Downloads are not tracked for this model. How to track

Safetensors

Model size

2.63M params

Tensor type

F32

Inference Providers NEW

Other

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

oteam
/

ligandmpnn