Update: The training code is now open source and updated in the ProteinDecoy-main.zip file （ We keep an example file for all the files: .pdb, .xyz, DNA_feature( HMM, PSSM, SS) The core training and model files are train_esm_mix.py / sparseconvunet_inference.py

CrossBind

Official Pytorch implementation of CrossBind: Collaborative Cross-Modal Identification of Protein Nucleic-Acid-Binding Residues.

Getting Started

Setup

To set up the environment for CrossBind, follow these steps:

1. Create Environment:

   Use conda to create a new environment with the dependencies listed in environment.yaml.

   conda env create -f environment.yaml
   conda activate Spn_3.7

2. Compile SparseConvNet operations:

   Navigate to the lib/ directory and compile the SparseConvNet operations.

   cd lib/
   python setup.py develop

Data Preparation

To prepare your data for CrossBind, perform the following:

1. Download Dataset:

   The dataset containing DNA/RNA PDB files can be downloaded from the following sources:

   • GraphBind: CSBio
   • GraphSite: GitHub - biomed-AI/GraphSite

2. Prepare XYZ Files:

   To convert original PDB files into XYZ format, you will need to use LIG_TOOL.

   git clone https://github.com/realbigws/PDB_Tool.git

   After cloning the repository, modify the file paths in datasets/prepare_pdb_to_xyz.py to match your local setup, then run the script:

   cd datasets/
   python prepare_pdb_to_xyz.py

Load ESM2 Representation:

For details on loading the ESM2 representation, refer to the documentation available at GitHub - facebookresearch/esm.

Training

To fine-tune the CrossBind model, you can customize the model settings in the configuration files located in cfgs/*.yaml. Select the appropriate configuration file for your needs.

• Run the full version of CrossBind:

  python train_esm_mix.py --log_dir SparseConv_default --cfg_file cfgs/SparseConv-Cath-Decoys-Clf-Only.yaml --gpu 0

  if you want pre-train the point encoder with a self-supervised way, use train_pointsite_contrastive.py first, and load the pre-trained 'pkl' model in train_esm_mix.py.

Visualization Case

For visual case studies of the results: