Source code for our IJCAI 2021 paper Learning Attributed Graph Representation with Communicative Message Passing Transformer
The code was built based on Molecule Attention Transformer and The Annotated Transformer. Thanks a lot for their code sharing!
We have reimplemented the CoMPT model by using pytorch, which have been finished all todo step in this readme, please consider this code firstly, thank you for your attention to our work!
- cuda >= 9.0
- cudnn >= 7.0
- RDKit == 2020.03.4
- torch >= 1.4.0 (please upgrade your torch version in order to reduce the training time)
- numpy == 1.19.1
- scikit-learn == 0.23.2
- tqdm == 4.52.0
Tips: Using code conda install -c conda-forge rdkit can help you install package RDKit quickly.
| Dataset | Tasks | Type | Molecule | Metric |
|---|---|---|---|---|
| bbbp | 1 | Graph Classification | 2,035 | ROC-AUC |
| tox21 | 12 | Graph Classification | 7,821 | ROC-AUC |
| sider | 27 | Graph Classification | 1,379 | ROC-AUC |
| clintox | 2 | Graph Classification | 1,468 | ROC-AUC |
| esol | 1 | Graph Regression | 1,128 | RMSE |
| freesolv | 1 | Graph Regression | 642 | RMSE |
| lipophilicity | 1 | Graph Regression | 4,198 | RMSE |
| 1H-NMR | 1 | Node Regression | 12,800 | MAE |
| 13C-NMR | 1 | Node Regression | 26,859 | MAE |
For the Graph-level task (Graph classification, Graph Regression), you can download the source dataset from Molecule-Net.
For the Node-level task (Node Regression), you can download the source dataset from NMRShiftDB2, or use a preprocess dataset cleaned by nmr-mpnn, thanks a lot for their code sharing!
In the folder ./Data, we have preprocessed every mentioned dataset by the corresponding jupyter notebook. All source datasets can be refered in the ./Data/<dataset>/source/, and all preprocess files can be refered in the ./Data/<dataset>/preprocess/.
You can also run the corresponding jupyter notebook in the path ./Data/<dataset>/preprocessing.ipynb to generate the <dataset>.pickle files.
To train a graph-level task, run:
python train_graph.py --seed <seed> --gpu <gpu> --fold 5 --dataset <dataset> --split <split>
where <seed> is the seed number, <gpu> is the gpu index number, <dataset> is the graph-level dataset name (bbbp, tox21, sider, clintox, esol, freesolv, lipophilicity), <split> is the split method that mentioned by Molecule-Net (random, scaffold, cv).
To train a node-level task, run:
python train_node.py --seed <seed> --gpu <gpu> --dataset nmrshiftdb --element <element>
where <seed> is the seed number, <gpu> is the gpu index number, <element> is the element name(1H for 1H-NMR, 13C for 13C-NMR).
All hyperparameters can be tuned in the utils.py
- Clean the unuse function and write more comments.
- Replace the unnoticed Chinese comments in English.
- Generate the split-fold files in
.csvformat, rewrite the code and then make a bash script to train all folds in parallel. - Make a suitable padding way to adapt the molecules with more than 100 atoms, which will be used in the protein (long period).
- Try our best to reduce the training time and the using memory, especially for the large dataset (long period).
Please cite the following paper if you use this code in your work.
@inproceedings{ijcai2021-309,
title = {Learning Attributed Graph Representation with Communicative Message Passing Transformer},
author = {Chen, Jianwen and Zheng, Shuangjia and Song, Ying and Rao, Jiahua and Yang, Yuedong},
booktitle = {Proceedings of the Thirtieth International Joint Conference on
Artificial Intelligence, {IJCAI-21}},
publisher = {International Joint Conferences on Artificial Intelligence Organization},
editor = {Zhi-Hua Zhou},
pages = {2242--2248},
year = {2021},
month = {8},
note = {Main Track}
doi = {10.24963/ijcai.2021/309},
url = {https://doi.org/10.24963/ijcai.2021/309},
}