SwinSite: Protein–Ligand Binding Site Prediction

SwinSite is a deep learning-based method for predicting protein–ligand binding sites using a hybrid of Swin Transformer and 3D CNN architectures. It processes protein structures in 3D and outputs likely binding regions.

---

Installation

1. Create conda environment and install PyTorch

conda create -n swinsite python=3.12
conda activate swinsite

# Install PyTorch with CUDA support
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

2. Install additional dependencies

pip install openbabel-wheel matplotlib py3dmol einops scikit-image scikit-learn h5py tqdm timm

---

Usage

Running the Prediction

Run the prediction script on a set of protein–ligand complexes:

python predict.py \
    -i ./example \
    -o ./output/example \
    -l ./logs/log.txt

Arguments:

• -i : Path to input directory (should contain subdirectories with protein.pdb and ligand.mol2)
• -f : File format (default: pdb, optional: mol2)
• -o : Output directory
• -l : Log file path

The predicted pockets and grids will be saved in the specified output directory.

---

Jupyter Notebook Example

To run both prediction and visualization in an interactive environment:

1. Launch Jupyter Lab:

   jupyter lab

2. Open the notebook:

   predict_and_visualize.ipynb
   

This notebook:

• Runs predict.py for inference
• Uses py3Dmol to visualize proteins, ligands, and predicted pockets

---

Example Folder Structure

example/
├── 1abc/
│   ├── protein.pdb
│   ├── ligand.mol2
├── 2xyz/
│   ├── protein.pdb
│   ├── ligand.mol2

Each subdirectory should contain a single protein–ligand pair.

---

Output Files

For each processed sample, the following files will be generated:

• pocket_*.mol2 : Atom-level predicted pockets
• grid_*.mol2 : Grid-based volumetric pockets
• log.txt : Log file containing progress and error messages

---