Awesome

Target-Specific De Novo Peptide Binder Design with DiffPepBuilder

This is the official repository for the paper Target-Specific De Novo Peptide Binder Design with DiffPepBuilder.

For any questions, please open an issue or contact wangyuzhe_ccme@pku.edu.cn for more information.

News

• [2024/9/12] Our research article is now published in JCIM! Dive into the details by checking out the full paper here or on ArXiv.
• [2024/9/11] We released the PepPC-F and PepPC datasets for DiffPepBuilder on Zenodo. The training protocol has also been released. Please refer to the Training section for more details.
• [2024/7/22] We released the initial code, model weights, and a Colab demo for DiffPepBuilder.

Quick Start

We provide a colaboratory notebook to demonstrate the usage of DiffPepBuilder (in progress). Please click the following link to open the notebook in Google Colab:

Installation

We recommend using a conda environment to install the required packages. Please clone this repository and navigate to the root directory:

git clone https://github.com/YuzheWangPKU/DiffPepBuilder.git
cd DiffPepBuilder

Then run the following commands to create a new conda environment and install the required packages:

conda env create -f environment.yml
conda activate diffpepbuilder

Before running inference, please unzip the SSBLIB data in the SSbuilder directory:

cd SSbuilder
tar -xvf SSBLIB.tar.gz

The post-processing procedure requires Rosetta to be installed. Please download the latest version of Rosetta from the official website and follow the installation instructions.

Inference

To de novo generate peptide binders for a given target protein, please first download the model weights into experiments/checkpoints/ from Zenodo. You can use the following command to download the model weights:

wget https://zenodo.org/records/12794439/files/diffpepbuilder_v1.pth
mv diffpepbuilder_v1.pth experiments/checkpoints/

We provide an example of the target ALK1 (Activin Receptor-like Kinase 1, PDB ID: 6SF1) to demonstrate the procedures of generating peptide binders. Please note that the following pipeline can also be used to generate peptide binders for multiple targets simultaneously. The hotspots or binding motif of the target protein can be specified in JSON format, as showcased by the example file examples/receptor_data/de_novo_cases.json. To preprocess the receptor, run the experiments/process_receptor.py script:

python experiments/process_receptor.py --pdb_dir examples/receptor_data --write_dir data/receptor_data --peptide_info_path examples/receptor_data/de_novo_cases.json

This script will generate the receptor data in the data/receptor_data directory. To generate peptide binders for the target protein, please specify the root directory of DiffPepBuilder repository and then run the experiments/run_inference.py script (modify the nproc-per-node flag accordingly based on the number of GPUs available):

export BASE_PATH="your/path/to/DiffPepBuilder"
torchrun --nproc-per-node=8 experiments/run_inference.py data.val_csv_path=data/receptor_data/metadata_test.csv

The config file config/inference.yaml contains the hyperparameters for the inference process. Below is a brief explanation of the key hyperparameters:

You can modify these hyperparameters to customize the inference process. For more details on the hyperparameters, please refer to our paper.

After running the inference script, the generated peptide binders will be saved in the tests/inference/. To run the side chain assembly and energy minimization using Rosetta, please run the following script subsequently:

export BASE_PATH="your/path/to/DiffPepBuilder"
python experiments/run_redock.py --in_path tests/inference --ori_path examples/receptor_data --interface_analyzer_path your/path/to/rosetta/main/source/bin/rosetta_scripts.static.linuxgccrelease

Modify the interface_analyzer_path flag to the path of the Rosetta interface_analyzer executable. The script will generate the final peptide binders in the tests/inference/.../pdbs_redock/ directory and calculate the binding ddG values of the generated peptide binders. The results will be summarized in the tests/inference/redock_results.csv file.

Training

To train the DiffPepBuilder model from scratch, please download the training data from Zenodo and unzip the data in the data/ directory:

wget https://zenodo.org/records/13744959/files/PepPC-F_raw_data.tar.gz
mkdir data/PepPC-F_raw_data
tar -xvf PepPC-F_raw_data.tar.gz --strip-components=1 -C data/PepPC-F_raw_data

To preprocess the training data, run the experiments/process_dataset.py script:

python experiments/process_dataset.py --pdb_dir data/PepPC-F_raw_data --write_dir data/complex_dataset

This script will generate the training data in the data/complex_dataset directory. You can add max_batch_size flag to specify the maximum batch size for ESM embedding to avoid out-of-memory errors. Then split the data into training and validation sets:

python experiments/split_dataset.py --input_path data/complex_dataset/metadata.csv --output_path data/complex_dataset --num_val 200

You can modify the num_val flag to specify the number of validation samples. To train the DiffPepBuilder model, please specify the root directory of the DiffPepBuilder repository and then run the experiments/train.py script (modify the nproc-per-node flag accordingly based on the number of GPUs available):

export BASE_PATH="your/path/to/DiffPepBuilder"
torchrun --nproc-per-node=8 experiments/train.py

The config file config/base.yaml contains the hyperparameters for the training process. You can modify these hyperparameters to customize the training process. Checkpoints will be saved every 10,000 steps after validation in the tests/ckpt/ directory by default. Training logs will be saved every 2,500 steps.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Citation

Please cite the following paper if you use this code in your research:

@misc{wang2024targetspecificnovopeptidebinder,
      title={Target-Specific De Novo Peptide Binder Design with DiffPepBuilder}, 
      author={Fanhao Wang and Yuzhe Wang and Laiyi Feng and Changsheng Zhang and Luhua Lai},
      year={2024},
      eprint={2405.00128},
      archivePrefix={arXiv},
      primaryClass={q-bio.BM},
      url={https://arxiv.org/abs/2405.00128}, 
}

Acknowledgments

We would like to thank the authors of FrameDiff and OpenFold, whose codebases we used as references for our implementation.