Molecular Representation Learning via Heterogeneous Motif Graph Neural Networks

This is an official PyTorch implementation of the experiments in the following paper:
Molecular Representation Learning via Heterogeneous Motif Graph Neural Networks (ICML 2022)
Zhaoning Yu, Hongyang Gao

Requirements

pytorch                       1.9.0
rdkit-pypi                    2021.9.2
ogb                           1.3.1
dgl                           0.6.1
networkx

Part 1: Heterogeneous Motif Graph Construction

Run python preprocess.py to construct HM-graph for TUDataset.
Change the parameter of drop_node() function in the ops.py to drop noises in the motif dictionary.
Run python preprocess_hiv.py and python preprocess_pcba.py to construct HM-graph for ogbg-molhiv and ogbg-pcba dataset.
For ogbg-pcba dataset, because there are 11 graphs do not have motifs, you need to substract 11 from self.num_cliques.

Part 2: Training and evaluation

Run python main.py for TUDataset.
Run python main_ogbg_molhiv.py for ogbg-molhiv.
Run python main_molpcba.py for ogbg-pcba.

Cite

If you find this repo or paper to be useful, please cite our paper.

@inproceedings{yu2022molecular,
  title={Molecular Representation Learning via Heterogeneous Motif Graph Neural Networks},
  author={Yu, Zhaoning and Gao, Hongyang},
  booktitle={International Conference on Machine Learning},
  pages={25581--25594},
  year={2022},
  organization={PMLR}
}