DropGNN: Graph Neural Networks with DropNode for Drug Discovery

This repository contains an implementation of DropGNN, a graph neural network approach with DropNode regularization for drug discovery tasks, based on the research paper "DropGNN: Random Dropouts Increase the Expressiveness of Graph Neural Networks".

Table of Contents

• Overview
• Installation
• Usage
• Repository Structure
• Results
• Contributing
• Citation
• License

Overview

DropGNN introduces random dropouts of entire nodes during training to improve the expressiveness and performance of Graph Neural Networks (GNNs) on molecular property prediction tasks. Key features:

• Implements DropNode regularization for GNNs
• Improves model generalization
• Enhances expressiveness of message passing GNNs
• Particularly effective for drug discovery applications

The original paper demonstrates that DropGNN achieves state-of-the-art performance on several molecular property prediction benchmarks.

Installation

1. Clone this repository:

git clone https://github.com/JaanuNan/DropGNN.git
cd dropgnn.ipynb

2. Create and activate a virtual environment (recommended):

python -m venv venv
source venv/bin/activate  # On Windows use `venv\Scripts\activate`

Usage

Jupyter Notebook

The main implementation is provided in DropGNN.ipynb. Open it with Jupyter:

jupyter notebook dropgnn.ipynb

Python Script

Alternatively, you can use the standalone Python implementation:

python dropgnn.py

Configuration

Modify the hyperparameters in the notebook/script to experiment with different settings:

• Dropout rate
• Number of GNN layers
• Hidden layer dimensions
• Learning rate
• Training epochs

Repository Structure

dropgnn-implementation/
├── dropgnn.ipynb    # Main implementation notebook
├── dropgnn.py                     # Python implementation
├── documentation              # Implementation documentation
├── Research paper                # Paper Pdf
└── LICENSE

Results

Contributing

Contributions are welcome! Please open an issue or submit a pull request for any improvements.

Citation

If you use this implementation in your research, please cite the original paper:

@article{papp2021dropgnn,
  title={DropGNN: Random Dropouts Increase the Expressiveness of Graph Neural Networks},
  author={P{\'a}l Andras Papp and Roger Wattenhofer},
  journal={arXiv preprint arXiv:2111.06283},
  year={2021}
}

License

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