IEM

Official PyTorch-based implementation of Paper "An Image-enhanced Molecular Graph Representation Learning Framework", IJCAI 2024.

[Paper] [Appendix] [Poster] [PPT]

News!

[2024/08/11] Uploaded Poster and PPT.

[2024/04/17] Accepted in IJCAI 2024 !!!

[2024/01/17] Repository installation completed.

TODO

• Publish pre-training dataset
• Publish supplementary material of IEM
• Publish downstream task data
• Release pre-trained teacher model

Environments

1. GPU environment

CUDA 11.6

Ubuntu 18.04

2. create conda environment

# create conda env
conda create -n IEM python=3.9
conda activate IEM

# install environment
pip install rdkit
pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu116 -i https://pypi.tuna.tsinghua.edu.cn/simple

pip install timm==0.6.12
pip install tensorboard
pip install scikit-learn
pip install setuptools==59.5.0
pip install pandas
pip install torch-cluster torch-scatter torch-sparse torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.13.1%2Bcu116.html
pip install torch-geometric==1.6.0
pip install dgl-cu116
pip install ogb

Pre-Training Teacher Model

1. Pre-training Dataset

The pre-trained datasets (2 million) can be accessed in following table

Name	Download link	Description
iem-200w.csv	OneDrive BaiduCloud	index and SMILES information for all molecules
image2d.tar.gz	OneDrive	2D images (RGB format)
image3d.tar.gz	BaiduCloud	multi-view 3D images (RGB format)
mol-basic-info.tar.gz	OneDrive	predefined knowledge, including atom, bound, geometry, attributes

Please download all data listed above and put it in datasets/pre-training/iem-200w/processed/ if you want to train your own teacher model from scratch.

2. ❄️Direct access to pre-trained teachers

The pre-trained teacher model can be accessed in following table.

Name	Download link	Description
Pre-trained teacher	IEM.pth	You can download the teacher and put it in the directory: resumes/pretrained-teachers/.

3. 🔥Train your own teacher model from scratch

If you want to pre-train your own teacher model, see the command below.

Usage:

usage: pretrain_teacher.py [-h] [--dataroot DATAROOT] [--dataset DATASET]
                           [--workers WORKERS] [--nodes NODES]
                           [--ngpus_per_node NGPUS_PER_NODE]
                           [--dist-url DIST_URL] [--node_rank NODE_RANK]
                           [--model_name MODEL_NAME]
                           [--warmup_rate WARMUP_RATE] [--lr LR]
                           [--momentum MOMENTUM] [--weight-decay WEIGHT_DECAY]
                           [--weighted_loss] [--runseed RUNSEED]
                           [--start_epoch START_EPOCH] [--epochs EPOCHS]
                           [--batch BATCH] [--imageSize IMAGESIZE]
                           [--temperature TEMPERATURE]
                           [--base_temperature BASE_TEMPERATURE]
                           [--resume RESUME] [--n_ckpt_save N_CKPT_SAVE]
                           [--log_dir LOG_DIR]

run command to pre-train teacher:

python pretrain_teacher.py \
	--nodes 1 \
	--ngpus_per_node 1 \
	--weighted_loss \
	--epochs 300 \
	--batch 128 \
	--lr 1e-2 \
	--ngpu 1 \
	--workers 4 \
	--dataroot ../datasets/pre-training/ \
	--dataset toy \
	--log_dir ./experiments/pretrain_teacher

Distillation Training on Downstream Tasks

All downstream task data is publicly accessible below:

8 classification tasks (RGB format):

Datasets	#Molecules	#Task	Links
Tox21	7,831	12	[OneDrive]
ToxCast	8,576	617	[OneDrive]
Sider	1,427	27	[OneDrive]
ClinTox	1,478	2	[OneDrive]
MUV	93,087	17	[OneDrive]
HIV	41,127	1	[OneDrive]
BBBP	2,039	1	[OneDrive]
BACE	1,513	1	[OneDrive]

4 regression tasks (BGR format):

Datasets	#Molecules	#Task	Links
ESOL	1,128	1	[OneDrive]
Lipo	4,200	1	[OneDrive]
Malaria	9,999	1	[OneDrive]
CEP	29,978	1	[OneDrive]

Usage:

usage: distillation_training.py [-h] [--dataroot DATAROOT] [--dataset DATASET]
                                [--graph_feat {min,all}]
                                [--label_column_name LABEL_COLUMN_NAME]
                                [--image_dir_name IMAGE_DIR_NAME]
                                [--img_type {RGB,BGR}] [--gpu GPU]
                                [--ngpu NGPU] [--workers WORKERS]
                                [--num_layers NUM_LAYERS]
                                [--feat_dim FEAT_DIM]
                                [--JK {concat,last,max,sum}]
                                [--t_dropout T_DROPOUT]
                                [--gnn_type {gin,gcn,gat,graphsage}]
                                [--resume_teacher RESUME_TEACHER]
                                [--resume_teacher_name RESUME_TEACHER_NAME]
                                [--lr LR] [--weight_t WEIGHT_T]
                                [--weight_te WEIGHT_TE]
                                [--weight_ke WEIGHT_KE] [--seed SEED]
                                [--runseed RUNSEED] [--split {scaffold}]
                                [--epochs EPOCHS] [--start_epoch START_EPOCH]
                                [--batch BATCH] [--resume RESUME]
                                [--pretrain_gnn_path PRETRAIN_GNN_PATH]
                                [--model_name MODEL_NAME]
                                [--task_type {classification,regression}]
                                [--save_finetune_ckpt {0,1}]
                                [--log_dir LOG_DIR]

For examples, you can run the following code to distillation training:

python distillation_training.py \
	--resume_teacher ../resumes/pretrained-teachers/image3d_teacher.pth \
	--resume_teacher_name image3d_teacher \
	--image_dir_name image3d \
    --dataroot ../datasets/fine-tuning/ \
	--dataset esol \
	--task_type regression \
	--feat_dim 512 \
	--batch 32 \
	--epochs 100 \
	--lr 0.001 \
	--split scaffold \
	--weight_te 0.01 \
	--weight_ke 0.001 \
	--log_dir ./experiments/image3d_teacher/esol/rs0/ \
	--runseed 0 \
    --gnn_type gin \
	--pretrain_gnn_path ../resumes/pretrained-gnns/GraphMVP.pth

Note: The multi-view 3D images used in the pre-training stage are in RGB format, and the multi-view 3D images used in the distillation stage are RGB format (8 classification) and BGR format (4 regression), respectively. Since there is data augmentation of grayscale images in the pre-training stage, we do not distinguish between RGB and BGR in distillation stage (See the link for details). If you want to convert the BGR format to RBG format for downstream tasks, please use the parameter --img_type BGR.

🌟Tips: Although the paper has provided detailed experimental descriptions, in order to accelerate your reproduction, please focus on the following points and parameters:

1. Use --gnn_type and --pretrain_gnn_path to specify different GNN methods and corresponding initialization pre-training weights;

2. Perform grid search for --weight_te and --weight_ke in [0.001, 0.01, 0.1, 1, 5];

3. For specific --weight_te and --weight_ke values, set --runseed from 0 to 9 and calculate the mean and variance.

Reference

If our paper or code is helpful to you, please do not hesitate to point a star for our repository and cite the following content.

@inproceedings{ijcai2024p675,
  title     = {An Image-enhanced Molecular Graph Representation Learning Framework},
  author    = {Xiang, Hongxin and Jin, Shuting and Xia, Jun and Zhou, Man and Wang, Jianmin and Zeng, Li and Zeng, Xiangxiang},
  booktitle = {Proceedings of the Thirty-Third International Joint Conference on
               Artificial Intelligence, {IJCAI-24}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  editor    = {Kate Larson},
  pages     = {6107--6115},
  year      = {2024},
  month     = {8},
  note      = {Main Track},
  doi       = {10.24963/ijcai.2024/675},
  url       = {https://doi.org/10.24963/ijcai.2024/675},
}