proteinmpnn-cli 0.2.0

Structure-conditioned protein sequence design using message passing neural networks

proteinmpnn - A cli adaptation of Kuhlman lab's fork

[!WARNING] This is a work-in-progress.

This repo contains a clean-up of Kuhlman lab's fork of ProteinMPNN, converting it into an easy-to-use cli.

This modernization includes

• using uv for dependency and package management.
• using typer to construct a CLI with plenty of flavor.

Clone this repo and run

uv run proteinmpnn --help

Current features

• Running inference for a single pdb using proteinmpnn run-single. Use --help to get a look into the optional arguments. This is meant to replace the single-protein analyses.
• Computing conditional/unconditional probabilities of amino-acids per location. Check proteinmpnn compute-probs --help for more context.

Other improvements on Kuhlman's fork

• The usual two-step sequence with generate_json.py and then running it is no longer necessary.
• Unit testing using pytest, as well as backwards compatibility test (making sure that we don't deviate from the original behavior).
• Linting using ruff to make the code more developer-friendly.

Original readme

This repo includes the Kuhlman Lab fork of ProteinMPNN. It includes all the functionality of the original ProteinMPNN repo (linked here), with the following additions:

• Improved input parsing for custom design runs
• Multi-state design support
• Additional utilities to provide integration with EvoPro

Read ProteinMPNN paper.

Installation:

git clone [e-mail]:Kuhlman-Lab/proteinmpnn.git
cd proteinmpnn
mamba create env -f setup/proteinmpnn.yml

NOTE (July 2025):

ProteinMPNN uses CUDA 11.3, which is too old for the new H100 GPUs (CUDA 11.8+). This means it may hang if run from the default mpnn environment.

To fix this, we can generate a CUDA 12.4 environment as follows:

# Install original env without torch/cuda dependencies
mamba env create -f setup/proteinmpnn_cu12.4.yml -n mpnn_cu12.4

# Install torch/cuda 12.4 dependencies
pip install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 --index-url https://download.pytorch.org/whl/cu124

To use this, simply replace conda activate mpnn with conda activate mpnn_cu2.4 wherever present.

Usage Guidelines:

General Usage

The different input arguments available for each script can be viewed by adding -h to your python call (e.g., python generate_json.py -h).

ProteinMPNN accepts PDB files as input and produces FASTA files as output.

Unlike the original repo, our ProteinMPNN organizes the different input options (aka arguments) into .flag files:

• json.flags is used to specify design constraints, like fixed residues and symmetry
• proteinmpnn.flags is used to specify prediction flags, like which sampling temperature and model variant to use.

In general, there are two steps to running ProteinMPNN:

1. Run the generate_json.py script and pass it the json.flags file.

• This makes a new file called proteinmpnn_res_specs.json containing parsed design information.

2. Run the run_protein_mpnn.py script and pass it proteinmpnn.flags and proteinmpnn_res_specs.json to obtain the actual ProteinMPNN prediction.

Useful Flags

Used in json.flags:

--default_design_setting: this is an optional filter to allow/disallow certain residue types during design. By default, it is set to all, which allows all 20 amino acids. Possible settings include: all-hydphob: exclude hydrophobic residues (CDEHKNPQRSTX) all-hydphil: exclude hydrophilic residues (ACFGILMPVWYX) all-CLD: exclude specific amino acids (in this case, Cys, Leu, and Asp) L+polar: mix-and-match amino acids and categories (in this case, allow all polar amino acids and also Leu)

Used in proteinmpnn.flags: --model_name: specifies which ProteinMPNN model checkpoint to use. Possible options include: v_48_002: vanilla (default) model with k=48 neighbors and 0.02A noise s_48_010: soluble protein model with k=48 neighbors and 0.1A noise

--sampling_temp: specifies the sampling temperature, which changes how diverse the generated sequences will be. Ranges from 0 to 1, inclusive. A temperature of 0 returns the "best" prediction every time (zero diversity), while a temperature of 1 will return completely random samples. Recommended range is 0.0 - 0.3 or so.

--dump_probs: if included, ProteinMPNN will save the predicted sequence probability table for each scaffold. This will be a numpy array of shape [L, 21], for a protein of length L. If multiple sequences are generated per scaffold, probabilities will be averaged before saving. A helper script for visualizing these tables is included at run/helper_scripts/other_tools/view_probs.py.

Example Cases

Example input and expected output files, as well as jobscripts and flag files, for many different design tasks are included in examples/. For a summary and explanation of each example, see examples/EXAMPLES.md. Currently supported protocols include:

1. Monomer Design (with user-friendly parsing of designable residues)
2. Binder Design
3. Oligomer Design (with support for abitrary symmetries in homooligomers)
4. Multi-state Design (with support for multiple complex design constraints)

---

Unit Testing

TODO

Code organization:

• run/run_protein_mpnn.py - the main script to initialialize and run the model.
• run/generate_json.py - function to automatically generate json of design constraints.
• run/helper_scripts/ - helper functions to parse PDBs, assign which chains to design, which residues to fix, adding AA bias, tying residues etc.
• examples/ - simple example inputs/outputs and runscripts for different tasks.
• model_weights/ - trained proteinmpnn model weights.
  • v_48_... - vanilla proteinmpnn models trained at different noise levels.
  • s_48_... - solublempnn models trained at different noise levels.
  • ca_48_... - Ca-only models trained at different noise levels.

License

ProteinMPNN is distributed under an MIT license, which can be found at proteinmpnn/LICENSE. See license file for more details.