Fast Non-autoregressive Inverse Folding with Discrete Diffusion (NeurIPS MLSB 2023)

This repository offers an implementation of discrete diffusion for inverse protein folding. It has pre-trained models, training routines, and inference scripts, ensuring a streamlined experience for protein folding predictions and experiments.

• Provide code of designability metric for proper reproducibility.
• Configure PMPNN ARM sampling temperature correctly.

Clone the repository, navigate to its root directory, and create a conda environment using the provided YAML file. Activate the environment as follows:

conda env create -f environment.yml
conda activate your-env-name

Within the activated environment and the root directory of the repository, execute:

pip install -e .

For discrete diffusion inference with purity sampling, run

python experiments/inference_diff.py --sampling_type purity_sample

Refer to configs/clean/inference_diff.yaml for a complete description of inference args.

To compute designability numbers, run

python scripts/run_esmfold_csv.py --csv_path your-csv-path ...

passing in your CSV path generated by inference_diff.py.

We provide pretrained weights for ProteinMPNN trained on the CATH 4.2 dataset under the weights directory. Both ARM and Discrete Diffusion weights are available.

To train an ARM model from scratch, run

python experiments/train_arm.py ...

To train a discrete diffusion model from scratch, run

python experiments/train_diff.py ...

Please reach out to [email protected] with any questions or concerns.

This project is endorsed under the MIT License - refer to the LICENSE.md file for details.