PLCMPNN

Reimplementation of the CMPNN model by using dgl and PyTorch Lightning.

For the full version, please use Baidu Netdisk to view, the access link is as follows：

链接：https://pan.baidu.com/s/18YHY3_zpG0Mp7dfdBJn2gg?pwd=l1t7 提取码：l1t7

The original implementation of CMPNN could be referred at CMPNN and CMPNN-dgl. Thanks a lot for their working!

The IJCAI 2020 paper could be referred at Communicative Representation Learning on Attributed Molecular Graphs.

Notes: Since the limitation of dgl, we could not implement the reverse edge in the model. The original calculation could be found in the paper Algorithm 1 line 8, so we delete the minus of reverse edge as the substitution and find it will not cause performance decrease a lot.

Dependencies

• cuda == 11.1
• dgl-cuda11.1 == 0.7.2
• numpy == 1.23.2
• ogb == 1.3.5
• pandas == 1.4.3
• python == 3.9.13
• pytorch-lightning
• RDKit == 2022.9.3
• scikit-learn == 1.1.3
• torch == 1.9.1+cu111
• tqdm == 4.64.1

Overview

We report the different models and their corresponding results.

1. PLCMPNN

These result could be refered from AttentiveFP, SMILES Transformer and GROVER. Thanks a lot for their working!

Dataset	BACE	BBBP	ClinTox	HIV	SIDER	Tox21	ToxCast	MUV	ESOL	FreeSolv	Lipophilicity	QM7	QM8	QM9
Molecules	1513	2039	1478	41127	1427	7831	8577	93087	1128	642	4200	6834	21786	133885
Task	1 GC	1 GC	2 GC	1 GC	27 GC	12 GC	617 GC	17 GC	1 GR	1 GR	1 GR	1 GR	12 GR	12 GR
Metrics	AUC-ROC	AUC-ROC	AUC-ROC	AUC-ROC	AUC-ROC	AUC-ROC	AUC-ROC	AUC-PRC	RMSE	RMSE	RMSE	MAE	MAE	MAE
GraphConv	0.854±0.011	0.877±0.036	0.845±0.051	-	0.593±0.035	0.772±0.041	0.650±0.025	-	1.068±0.050	2.900±0.135	0.712±0.049	118.9±20.2	0.021±0.001	-
AttentiveFP	0.863±0.015	0.908±0.050	0.933±0.020	0.832±0.021	0.605±0.060	0.807±0.020	0.579±0.001	-	0.853±0.060	2.030±0.420	0.650±0.030	126.7±4.0	0.0282±0.001	-
Smile Transformer	0.719±0.023	0.900±0.053	0.963±0.064	0.683	0.559±0.017	0.706±0.021	-	0.009	1.144±0.118	2.246±0.237	1.169±0.031	-	-	-
GROVER	0.894±0.028	0.940±0.019	0.944±0.021	-	0.658±0.023	0.831±0.025	0.737±0.010	-	0.831±0.120	1.544±0.397	0.560±0.035	72.6±3.8	0.0125±0.002	-
MPNN	0.815±0.044	0.913±0.041	0.879±0.054	-	0.595±0.030	0.808±0.024	0.691±0.013	-	1.167±0.430	2.185±0.952	0.672±0.051	113.0±17.2	0.015±0.002	-
DMPNN	0.852±0.053	0.919±0.030	0.897±0.040	0.794±0.016	0.632±0.023	0.826±0.023	0.718±0.011	0.045±0.011	0.980±0.258	2.177±0.914	0.653±0.046	105.8±13.2	0.0143±0.002	3.451±0.174
PLCMPNN	0.874±0.014	0.962±0.004	0.952±0.013	0.835±0.001	0.631±0.010	0.843±0.001	0.743±0.009	0.045±0.007	0.557±0.032	1.391±0.119	0.693±0.017	60.380±1.338	0.012±0.000	6.767±0.405

2. PLCMPNN-OGB

These result could be refered from GCN and GIN. Thanks a lot for their working!

Dataset	ogbg-molbace	ogbg-molbbbp	ogbg-molclintox	ogbg-molhiv	ogbg-molsider	ogbg-moltox21	ogbg-moltoxcast	ogbg-molmuv	ogbg-molesol	ogbg-mollipo
Molecules	1513	2039	1477	41127	1427	7831	8576	93087	1128	4200
Task	1 GC	1 GC	2 GC	1 GC	27 GC	12 GC	617 GC	17 GC	1 GR	1 GR
Metrics	AUC-ROC	AUC-ROC	AUC-ROC	AUC-ROC	AUC-ROC	AUC-ROC	AUC-ROC	AUC-PRC	RMSE	RMSE
GCN	0.689±0.070	0.678±0.024	0.886±0.021	0.760±0.012	0.598±0.015	0.775±0.009	0.667±0.005	0.110±0.029	1.015±0.025	0.771±0.025
GIN	0.735±0.052	0.697±0.019	0.841±0.038	0.771±0.015	0.576±0.016	0.776±0.006	0.661±0.005	0.098±0.027	0.998±0.025	0.704±0.025
PLCMPNN-OGB	0.856±0.007	0.703±0.004	0.924±0.019	0.787±0.008	0.646±0.005	0.757±0.005	0.625±0.016	0.106±0.035	0.853±0.007	0.725±0.008

Note:

(1) GC=Graph Classification, GR=Graph Regression

(2) The FreeSolv dataset may need to more works on tuning the hyper-parameters since there are only 642 molecules in this dataset.

(3) All split types follow the ratio of 0.8/0.1/0.1 in train/valid/test.

Running

1. PLCMPNN

To reproduce all the results, run firstly:

cd dataloader

python dataset.py

it will generate a pickle file in the data/preprocess with the same dataset name, this pickle file contain 4 objects:

• smiles_list: All SMILES strings in the dataset.
• mols_dict: Unique SMILES strings -> RDKit mol object.
• graphs_dict: Unique SMILES strings -> dgl graph object.
• labels_dict: Unique SMILES strings -> label list.

Then run:

python main.py --gpu <gpu num> --data_name <dataset> --split_type <split>

• <gpu num> is the number of gpu used by the Trainer.

• <dataset> is the dataset name, we provide 14 datasets that mentioned in the overview, more datasets and their results will be updated.

• <split> is the split type name, we provide [scaffold, random] in the code.

Others parameters could be refered in the main.py.

After running the code, it will create a folder with the format <args.data_name>_split_<args.split_type> in the ./result/ folder. Meanwhile, a folder with the format plcmpnn_seed_<args.seed>_batch_<args.batch_size>_lr_<args.learning_rate> in the folder with the format <args.data_name>_split_<args.split_type>.

If choose scaffold split or random split, the folder will contain:

├── result
│   ├── bace_split_random
│   │   ├── plcmpnn_seed_666_batch_64_lr_0.0001
│   │   │   ├── checkpoints
│   │   │   ├── results.txt
│   │   ├── plcmpnn_seed_666_batch_64_lr_0.0001_logs
│   │   ├── test.pickle
│   │   ├── train.pickle
│   │   └── valid.pickle
│   ├── bace_split_scaffold
│   │   ├── plcmpnn_seed_666_batch_64_lr_0.0001
│   │   │   ├── checkpoints
│   │   │   ├── results.txt
│   │   ├── plcmpnn_seed_666_batch_64_lr_0.0001_logs
│   │   ├── test.pickle
│   │   ├── train.pickle
│   │   └── valid.pickle

2. PLCMPNN-OGB

To reproduce all the results, run firstly:

cd dataloader

Then run:

python main.py --gpu <gpu num> --data_name <dataset>

• <gpu num> is the number of gpu used by the Trainer.

• <dataset> is the dataset name, we provide 10 datasets that mentioned in the overview, more datasets and their results will be updated.

Others parameters could be refered in the main.py.

After running the code, it will create a folder with the format <args.data_name> in the ./result/ folder. Meanwhile, a folder with the format ogbg_seed_<args.seed>_batch_<args.batch_size>_lr_<args.learning_rate> in the folder with the format <args.data_name>.

The folder will contain:

├── result
│   ├── ogbg_molbace
│   │   ├── ogbg_seed_666_batch_64_lr_0.0001
│   │   │   ├── checkpoints
│   │   │   ├── results.txt
│   │   ├── ogbg_seed_666_batch_64_lr_0.0001_logs
│   │   ├── mapping
│   │   ├── processed
│   │   ├── raw
│   │   ├── split
│   │   └── RELEASE_v1.txt

Note:

Split type is the defult type used by package ogb, the split type name could be seen in the ./result/<args.data_name>/split/ folder.

Citation:

Please cite the following paper if you use this code in your work.

@inproceedings{ijcai2020-392,
  title     = {Communicative Representation Learning on Attributed Molecular Graphs},
  author    = {Song, Ying and Zheng, Shuangjia and Niu, Zhangming and Fu, Zhang-hua and Lu, Yutong and Yang, Yuedong},
  booktitle = {Proceedings of the Twenty-Ninth International Joint Conference on
               Artificial Intelligence, {IJCAI-20}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},             
  editor    = {Christian Bessiere}	
  pages     = {2831--2838},
  year      = {2020},
  month     = {7},
  note      = {Main track}
  doi       = {10.24963/ijcai.2020/392},
  url       = {https://doi.org/10.24963/ijcai.2020/392},
}