This repository is the official implementation of the paper "BridgeNet: A Dual-Curvature Product Manifold Approach for Molecular Property Prediction".
BridgeNet is a novel architecture based on Graph Neural Networks (GNNs) designed to enhance the performance of molecular dynamics simulations and molecular property prediction. By introducing a dual-curvature product manifold, the model can more accurately capture the complex underlying geometric and topological structures within molecular graph data, achieving exceptional representation capabilities across multiple downstream tasks.
It is recommended to use Anaconda or Miniconda to manage your virtual environment. The main dependencies are as follows:
- python >= 3.9.0
- pytorch >= 2.0.0
- torch_geometric >= 2.3.0
- rdkit >= 2022.09.1
- numpy == 1.23.5
- pandas == 1.5.3
- scipy == 1.11.4
Our model was evaluated on 9 fundamental datasets from MoleculeNet (BBBP, BACE, ClinTox, Tox21, SIDER, HIV, ESOL, FreeSolv, and Lipophilicity), as well as the quantum chemistry dataset QM9.
- Classification Tasks: Evaluated using ROC-AUC.
- Regression Tasks: Evaluated using RMSE (for MoleculeNet) or MAE (for QM9).
All raw data will be automatically downloaded and processed via PyTorch Geometric's dataset classes upon the first run. The preprocessed .pt files will be saved in the specified output directory.
- Clone this repository and install the dependencies.
- Run the main script to start training and evaluation. The script will automatically trigger data preprocessing if it's the first run.