This repo is the implementation of the dual graph transformer for molecular property prediction. This model integrates atom and bond graphs to encode the comprehensive molecular information, including atom and bond features, graph topology and structure, and 3D spatial information if available for enhanced molecular property prediction performance.
To install, run the following commands in sequence
conda create -n dgt python=3.10.16
conda activate dgt
conda install mamba
mamba install graph-tool==2.45
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.1.0+cu121.html
pip install torch-geometric==2.0.4
pip install torchmetrics==1.2.0
pip install ogb==1.3.6
pip install tensorboardX==2.6.2.2
pip install wandb==0.18.7
pip install rdkit==2025.9.1
pip install libauc==1.1.0
Since molecular SMILES is not provided along with the download link from the torch_geometric package for the QM9 dataset. We provide that in the datasets/QM9_split folder, together with the scaffold splitting results. After downloading the QM dataset, please copy files under datasets/QM9_split to datasets/QM9/raw.
Config files for reproducing our results are provided in the configs folder.
To train from scratch, run python main.py --cfg config_file_path --repeat 1 seed 0 wandb.use False.
This implementation is developed from graphgps. For more information, please check https://github.com/rampasek/GraphGPS.git