SAKE-PP: Spatial Attention Kinetic GNN for Protein–Protein Interactions

Overview

SAKE-PP is a spatially equivariant graph neural network for scoring protein–protein interaction (PPI) decoys via interface RMSD regression. The model combines Laplacian eigenvector-based orientation, physics-inspired attention, and geometric message passing to capture both local and global features of protein interfaces. It supports training and evaluation workflows with batch I/O using .pt tensors.

We are currently migrating the code from the HPC environment and may encounter some minor bugs along the way. We plan to have completed most of the testing and system integration in one week

Quick Start

Installation

Clone the repository

git clone https://github.com/yxnyu/SakePP.git
cd SakePP

Option 1: Create conda environment and install dependencies

./install.sh

Option 2: PIP Install package

pip install -r requirements.txt

or, alternatively,

pip install -e .

Launch training

To re-construct by using default settings, at the very first please make sure .h5dataset and .ptcheckpoint already exist in the given directory:

SakePP/
  |- weights/
    |- PPI_pdbbin_irmsd.h5
    |- processed_pdb.pt

You can download these files from Google Drive: (https://drive.google.com/drive/folders/1SemY9YfQb-4r21VLN9_uMEFWAJfcl0Fh)

And in the github, we also provide a trainning using in the antibody-antigen test checkpoint.

Then, run the main script:

python main.py

To customise your training settings, i.e. dataset or pretrained weight, or using a different path importing your dataset, please edit ./config/config.yaml, and put your files in the correct directory correspondingly.

To customise the model, i.e. hyper-parameters of model, please edit ./config/models/models.yaml.

Development Roadmap

Because the current workflow uses HDF5-based batch I/O for .pt tensors in both training and testing, we are actively developing the following features to enhance usability and scalability:

Native .pdb and .mmCIF support for end-to-end preprocessing
Streaming data loaders for large-scale docking decoys
Integration with AlphaFold3, ZDOCK, and other docking pipelines
Visualization tools for inspecting spatial attention maps

Citation

If you use SAKE-PP in your research or development, please cite:

@article{sakepp2024,
  title={Quantifying Protein-Protein Interaction with a Spatial Attention Kinetic Graph Neural Network},
  author={Yuzhi Xu et al.},
  journal={},
  year={2025}
}

License

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

SakePP

Install / Use

README