The official implementation of our ICLR 2025 Spotlight paper "Boltzmann-Aligned Inverse Folding Model as a Predictor of Mutational Effects on Protein-Protein Interactions",
which establishes a bidirectional connection between log-likelihood in protein inverse folding models and
git clone https://github.com/aim-uofa/BA-DDG
cd BA-DDG
conda env create -f env.yml
conda activate BA-DDG
Dataset | Download Script |
---|---|
SKEMPI v2 | data/get_skempi_v2.sh |
Download the trained weights from Google Driver and put them into the ./ckpt folder.
Load the pre-trained ProteinMPNN model and make inference in an unsupervised setting.
cd training
python train_skempi.py --config_path ../config/inference_ba-cycle_skempi.json
You can choose different pre-trained ProteinMPNN weights by modifying the ckpt_path
parameter in the config file. Available full protein backbone models include: vanilla_model_weights/v_48_002.pt, v_48_010.pt, v_48_020.pt, v_48_030.pt, and soluble_model_weights/v_48_002.pt, v_48_010.pt, v_48_020.pt, v_48_030.pt.
Load the pre-trained BA-DDG model and make inference.
cd training
python train_skempi.py --config_path ../config/inference_ba-ddg_skempi.json
You can set the wandb flag to use Weights & Biases.
cd training
python train_skempi.py --config_path ../config/train_ba-ddg_skempi.json --use_wandb
@article{jiao2024boltzmannaligned,
title = {Boltzmann-Aligned Inverse Folding Model as a Predictor of Mutational Effects on Protein-Protein Interactions},
author = {Xiaoran Jiao and Weian Mao and Wengong Jin and Peiyuan Yang and Hao Chen and Chunhua Shen},
year = {2024},
journal = {arXiv preprint arXiv: 2410.09543},
url = {https://arxiv.org/abs/2410.09543v1},
pdf = {https://arxiv.org/pdf/2410.09543.pdf}
}
If you have any issue about this work, please feel free to contact Xiaoran Jiao.
For non-commercial academic use, this project is licensed under the 2-clause BSD License. For commercial use, please contact Chunhua Shen.