Implementation of "Learning Multi-View Molecular Representations with Structured and Unstructured Knowledge"
Warning: This repository is still under development. If you encounter any problem implementing our model, please open an issue or concat [email protected]
The model architecture is implemented in open_biomed/models/multimodal/mvmol.py
.
The model checkpoint is currently available at Google Drive. Install the checkpoint and put it under ./ckpts/fusion_ckpts/mvmol
.
For evaluating MV-Mol on downstream tasks, run the following scripts:
# cross-modal retrieval on PCDes
bash scripts/multimodal/mtr/run_pcdes.sh mvmol [device id]
# molecule captioning on CheBI-20
bash scripts/multimodal/molcap/train.sh mvmol [device id]
# text-based molecule generation on CheBI-20
bash scripts/multimodal/text2smi/train.sh mvmol [device id]
# We will be incorporating the molecular property prediction task soon