AAAI 2025 paper: MORE: Molecule Pretraining with Multi-Level Pretext Task

We used the following packages under Python 3.7.

pytorch 1.13.1
torch-cluster 1.6.1
torch-geometric 2.6.0
torch-scatter 2.1.1
torch-sparse 0.6.17
rdkit 2022.9.5

The pre-training and downstream datasets used in our experiments are referred to the paper Strategies for Pre-training Graph Neural Networks. You can download the biology and chemistry datasets from their repository.

• To run the codes successfully, the downloaded datasets should be placed in /dataset_conf and /dataset_info for pre-training

(If you're using 3D-level pretext task, you'll need to use the /dataset_conf)

(If you are not using 3D-level pretext task, you'll need to use the /dataset_info)

• To run the codes successfully, the downloaded datasets should be placed in /dataset for fine-tuning

We use Pretrain/dataset_conf/zinc_2m_MD and Pretrain/dataset_info/zinc_2m_MD

(Preprocessed data from zinc_standard_agent dataset, you can get here

Please run pretraining.py for downstream adaptations.

The pre-trained models we use follow the training steps of the paper Strategies for Pre-training Graph Neural Networks and GraphMAE

Please run finetune.py for downstream adaptations.

We provide pretrained MORE (Finetune/pre-train/MORE.pth)

1. Pretraining settings

2. Fine-tuning settings

For pretraining, Pretrain/example.ipynb

For Fine-tuning, Finetune/example.ipynb