Awesome

Few-shot Graph Learning for Molecular Property Prediction

Introduction

This is the source code and dataset for the following paper:

Few-shot Graph Learning for Molecular Property Prediction. In WWW 2021.

Contact Zhichun Guo (zguo5@nd.edu), if you have any questions.

Datasets

The datasets uploaded can be downloaded to train our model directly.

The original datasets are downloaded from Data. We utilize Original_datasets/splitdata.py to split the datasets according to the molecular properties and save them in different files in the Original_datasets/[DatasetName]/new. Then run main.py, the datasets will be automatically preprocessed by loader.py and the preprocessed results will be saved in the Original_datasets/[DatasetName]/new/[PropertyNumber]/propcessed.

Usage

Installation

We used the following Python packages for the development by python 3.6.

- torch = 1.4.0
- torch-geometric = 1.6.1
- torch-scatter = 2.0.4
- torch-sparse = 0.6.1
- scikit-learn = 0.23.2
- tqdm = 4.50.0
- rdkit

Run code

Datasets and k (for k-shot) can be changed in the last line of main.py.

python main.py

Performance

The performance of meta-learning is not stable for some properties. We report two times results and the number of the iteration where we obtain the best results here for your reference.

Acknowledgements

The code is implemented based on Strategies for Pre-training Graph Neural Networks.

Reference

@article{guo2021few,
  title={Few-Shot Graph Learning for Molecular Property Prediction},
  author={Guo, Zhichun and Zhang, Chuxu and Yu, Wenhao and Herr, John and Wiest, Olaf and Jiang, Meng and Chawla, Nitesh V},
  journal={arXiv preprint arXiv:2102.07916},
  year={2021}
}