README.md

On the Expressive Power of Sparse Geometric MPNNs

Overview

This repo contains code for the experimental part of the paper "On the Expressive Power of Sparse Geometric MPNNs".

Driven by applications in chemistry and various scientific fields, we investigate the expressive capabilities of message-passing neural networks (MPNNs) for geometric graphs. We demonstrate which graphs invariant and equivariant MPNNs can separate. Our findings indicate that this separation is primarily influenced by graph structure rather than the underlying point cloud. This repository provides implementations for sparse separation, power graph experiments, and chemical property prediction.

Installation

To set up the project environment and install all necessary dependencies, follow these steps:

git clone https://github.com/yonatansverdlov/E-GenNet.git
cd E-GenNet
conda env create -f egenet_environment.yml
conda activate egenet

Usage

This project includes different types of experiments: Sanity Check, Hard Example, Power-Graph & k-Chain Experiments, and Chemical Property Prediction.

Sanity Check

To verify the correctness of our model, run:

python sanity_check.py

Hard Example

To demonstrate a pair that no I-GWL can separate, but our model can, we train our model on such a pair:

python hard_example.py

Power-Graph & k-Chain Experiments

In this section, we showcase our ability to separate sparsest graphs, demonstrating that our model can outperform I-GGNN. Our experiments validate that the k-chain pairs can be separated across ten different seeds. To run:

python k_chain.py

For baselines, refer to the Colab notebook illustrating our tasks, results, and baselines:

Chemical Property Prediction

We present our results on predicting chemical properties along with train/test/val accuracy.

Data

Please download the Kraken, Drugs, and BDE datasets from MARCEL GitHub and place them in the /dataset directory.

Options

• dataset_name options: Drugs, Kraken, BDE
• task options:
  • Drugs: ip, ea, chi
  • Kraken: B5, L, burB5, burL
  • BDE: BindingEnergy

Training

To start training, use:

python chemical_property_exps.py --dataset_name <dataset_name> --task <task>

License

This project is licensed under the MIT License.

Contact

For any questions or feedback, reach out to me at yonatans@campus.technion.ac.il.

Authors

• Yonatan Sverdlov - yonatans@campus.technion.ac.il
• Nadav Dym - nadavdym@technion.ac.il

Citation

@inproceedings{sverdlov2025expressive,
  title={On the Expressive Power of Sparse Geometric MPNNs},
  author={Yonatan Sverdlov and Nadav Dym},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2025},
  eprint={2407.02025},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2407.02025}
}