Official implementation of MolMFD model in the paper: Molecular representation learning via multimodal fusion and decoupling (Information Fusion 2025) [MolMFD]

The code of MolMFD is built upon [UniMol] documentation, please refer to its dependencies.

Download the pre-training dataset from [Pretrain_data] and place it in the ./data/pretrain directory.

You can pretrain the model by

bash script_pretrain.sh

Download the downstream dataset from [MPP_data] and place it in the ./data directory.

You can finetune the model for classification tasks by

bash script_classification.sh

You can finetune the model for regression tasks by

bash script_regression.sh

Download the downstream dataset from [PLDPP_data] and place it in the ./data directory. Download the pretrained pocket weights from [Pocket_weights] and place it in the ./weights directory, naming it 'pocket_checkpoint.pt'.

You can finetune the model for protein-ligand docking pose prediction tasks by

bash script_docking.sh
bash infer_docking.sh
bash docking.sh

Please kindly cite this paper as follows. Thank you.

@article{zang2025molecular,
  title={Molecular representation learning via multimodal fusion and decoupling},
  author={Zang, Xuan and Zhang, Junjie and Tang, Buzhou},
  journal={Information Fusion},
  pages={103493},
  year={2025},
  publisher={Elsevier}
}