foundry/models/mpnn

ProteinMPNN, LigandMPNN, and SolubleMPNN

Warning

Benchmarking: Please use the old repositories of ProteinMPNN, LigandMPNN, and SolubleMPNN for model benchmarking/comparison until the API and public weights stabilize. We are in the process of validating that the re-implementation (both a retrained version and the old weight loading option) is as performant as the original models.

Important

Issues: Please provide feedback on any issues you encounter with the ProteinMPNN/LigandMPNN/SolubleMPNN re-implementation. We are particularly interested in discrepancies between the original models and this re-implementation, issues with performance when loading the original weights from the old repositories, problems with inference hyperparameters/conditioning, and input/output bugs.

ProteinMPNN enables protein sequence design given a fixed backbone structure of a protein. LigandMPNN extends this functionality to enable fixed-backbone sequence design of proteins in the context of ligands (i.e. small molecules, ions, DNA/RNA, etc.). This module represents a re-implementation of the original ProteinMPNN and LigandMPNN models within the modelforge/atomworks framework.

For more information on the original models, please see:

• ProteinMPNN: Robust deep learning–based protein sequence design using ProteinMPNN | ProteinMPNN Original Github
• LigandMPNN: Atomic context-conditioned protein sequence design using LigandMPNN | LigandMPNN Original Github
• SolubleMPNN: Computational design of soluble and functional membrane protein analogues

This guide provides instructions on preparing inputs and running inference for ProteinMPNN/LigandMPNN, as well as training these models.

Installation

A. Installation using uv

git clone https://github.com/RosettaCommons/foundry.git \
  && cd foundry \
  && uv python install 3.12 \
  && uv venv --python 3.12 \
  && source .venv/bin/activate \
  && uv pip install -e ".[mpnn]"

B. Download Model Weights

ProteinMPNN

Please use the following settings with these ProteinMPNN weights:

• model_type: "protein_mpnn"
• is_legacy_weights: True

48 Nearest Neighbors, \sigma = 0.20 Å Gaussian noise during training:

wget https://files.ipd.uw.edu/pub/ligandmpnn/proteinmpnn_v_48_020.pt

Additional ProteinMPNN Weights

48 Nearest Neighbors, \sigma = 0.02 Å Gaussian noise during training:

wget https://files.ipd.uw.edu/pub/ligandmpnn/proteinmpnn_v_48_002.pt

48 Nearest Neighbors, \sigma = 0.10 Å Gaussian noise during training:

wget https://files.ipd.uw.edu/pub/ligandmpnn/proteinmpnn_v_48_010.pt

48 Nearest Neighbors, \sigma = 0.30 Å Gaussian noise during training:

wget https://files.ipd.uw.edu/pub/ligandmpnn/proteinmpnn_v_48_030.pt

LigandMPNN

Please use the following settings with these LigandMPNN weights:

• model_type: "ligand_mpnn"
• is_legacy_weights: True

32 Nearest Neighbors, \sigma = 0.10 Å of Gaussian noise during training, 25 ligand atom context:

wget https://files.ipd.uw.edu/pub/ligandmpnn/ligandmpnn_v_32_010_25.pt

Additional LigandMPNN Weights

32 Nearest Neighbors, \sigma = 0.05 Å of Gaussian noise during training, 25 ligand atom context:

wget https://files.ipd.uw.edu/pub/ligandmpnn/ligandmpnn_v_32_005_25.pt

32 Nearest Neighbors, \sigma = 0.20 Å of Gaussian noise during training, 25 ligand atom context:

wget https://files.ipd.uw.edu/pub/ligandmpnn/ligandmpnn_v_32_020_25.pt

32 Nearest Neighbors, \sigma = 0.30 Å of Gaussian noise during training, 25 ligand atom context:

wget https://files.ipd.uw.edu/pub/ligandmpnn/ligandmpnn_v_32_030_25.pt

SolubleMPNN

Please use the following settings with these SolubleMPNN weights:

• model_type: "protein_mpnn"
• is_legacy_weights: True

48 Nearest Neighbors, \sigma = 0.20 Å Gaussian noise during training:

wget https://files.ipd.uw.edu/pub/ligandmpnn/solublempnn_v_48_020.pt

Additional SolubleMPNN Weights

48 Nearest Neighbors, \sigma = 0.02 Å Gaussian noise during training:

wget https://files.ipd.uw.edu/pub/ligandmpnn/solublempnn_v_48_002.pt

48 Nearest Neighbors, \sigma = 0.10 Å Gaussian noise during training:

wget https://files.ipd.uw.edu/pub/ligandmpnn/solublempnn_v_48_010.pt

48 Nearest Neighbors, \sigma = 0.30 Å Gaussian noise during training:

wget https://files.ipd.uw.edu/pub/ligandmpnn/solublempnn_v_48_030.pt

Inference

Warning

Known Bug: There is currently an issue with loading MPNN user annotation (temperature, designed residues, etc.) from CIF/atom array annotations. Command line passing of these options works as expected, as does input_dict specificiation with MPNNInferenceEngine.

Important

API Instability: We are currently finalizing some cleanup work on the inference API. Please expect the API (including input formats and outputs) to stabilize in the upcoming weeks. Thank you for your patience!

Important

When using weights from the original ProteinMPNN/LigandMPNN/SolubleMPNN repositories, please ensure to set is_legacy_weights to True when running inference.

A. Command Line Inference

Detailed documentation coming soon!

B. JSON-based Inference

Detailed documentation coming soon!

C. Programmatic (Scripted) Inference

Detailed documentation coming soon!

Important

Currently, 'mpnn_bias' and 'mpnn_pair_bias' annotations cannot be saved to CIF files due to shape limitations. As a result, these annotations must be recreated (either directly with annotation on the atom array or via the input config dictionary) when reloading designed structures from CIF files.

Training

Instructions for training ProteinMPNN/LigandMPNN/SolubleMPNN models will be updated here shortly.

Important

Training Code and New Weights: We are working to release the dataframes used for retrianing the ProteinMPNN, LigandMPNN, and SolubleMPNN re-implementations. Also, we are finalizing the retraining runs and will release weights retrained within this repository shortly.