saCNN

The saCNN is a protein-ligand prediction tool based on 3D convolutional neural network with spatial attention mechanisms, to encourage spatial feature learning. It can focus more on the voxels near interaction centers. You can quickly get started with the saCNN tool according to the following instructions.

Installation

Create virtual environment and install packages:

conda create -n saCNN htmd=2.0.6 -c acellera -c conda-forge
conda activate saCNN
pip install torch

Quick start

Git clone

Clone this repository by:

git clone https://github.com/xfcui/saCNN.git

Data preparation

After creating a virtual environment, you need to prepare data and trained model. We provide a sample data in the data/dataset/3jvr directory, which contains the files of protein (3jvr_protein.pdb) and ligand (3jvr_ligand.mol2). We also provide the trained model under the checkpoint/model.pkl.

Data processing

Run the following command to complete the characterization of protein and ligand. The file path of protein, ligand and feature generation are set in data.sh file.

bash src/data.sh

Model inference

Run the following command to complete the affinity prediction of protein and ligand. The file path of feature and model checkpoint are set in inference.sh file.

bash src/inference.sh

Usage

If you want to run our model on your own data, you need to provide the protein (.pdb) file and ligands (.mol2) file.

Authors: Yuxiao Wang, Zongzhao Qiu, Qihong Jiao, Cheng Chen, Zhaoxu Meng and Xuefeng Cui*
Contact: xfcui@email.sdu.edu.cn