ZeroDDI: A Zero-Shot Drug-Drug Interaction Event Prediction Method with Semantic Enhanced Learning and Dual-Modal Uniform Alignment
Accepted by IJCAI2024
The authors are Ziyan Wang, Zhankun Xiong, Feng Huang, Xuan Liu, Wen Zhang.
create a conda virtual env:
conda create -n name python=3.8
The required libraries are:
numpy
torch 1.11.0+cu113
addict
yapf
sklearn
pandas
torch_geometric 2.1.0
rdkit
tensorflow
deepchem
networkx
transformers>=4.26
matplotlib
sacremoses
bs4
lxml
Due to space limitations, we compressed the dataset. You can unzip all xxx.zip data in its fold.
There are three folds in /data/DrugBank5.1.9/
For example, the data of fold2 is in zsl2/ and gzsl2/
python main.py --config configs/zeroddi.py
or python main.py --config configs/zeroddi_fold2.py
You can also create our own config python file for different datasets or models.
After training, the parameters of models are saved in ./work_dirs/
Then, you can test the model by:
python main.py --config configs/zeroddi.py --zsl_para work_dirs/zeroddi/model_parameter/zsl_model_best_epoch100_seed42.pkl --gzsl_para work_dirs/zeroddi/model_parameter/gzsl_model_best_epoch100_seed42.pkl