💊MvDDI: A Multi-view Interaction Framework for Few-shot Drug-Drug Interaction

This is the code repository for the paper: MvDDI: A Multi-view Interaction Framework for Few-shot Drug-Drug Interaction (NLPCC 2025).

💻Environment Setting

cuda version: 11.7

python version: 3.8

Some required python packages:

networkx = 3.1
numpy = 1.19.5
rdkit = 2023.9.5
torch = 2.0.1
tqdm = 4.66.2
transformers = 4.39.1
pandas = 1.4.4
scikit-learn = 1.3.2

You can execute the following command to install the required packages:

pip install -r requirment.txt

📂Dataset

Few-shot Dataset

All few-shot datasets are in the data folder, Few-Drugbank is data/drugbank, Few-TWOSIDES is data/twosides, and Few-S1 and Few-S2 are in data/inductive_data.

The data in the specific dataset are as follows:

Transductive Setting

Class	Few-Drugbank	Few-TWOSIDES
Train	4,247	5,792
Valid	1,311	1,926
Test	37,577	790,298

Inductive Setting

Class	Few-S1	Few-S2
Train	5,213	5,213
Valid	677	1,293
Test	7,007	52,163

Data Format

Each data item in the Few-shot Dataset is a four-tuple (d1, d2, type, Neg samples). A specific example is:

DB01173,DB06144,15,DB00765$h

It indicates that DB01173 and DB06144 have a reaction type of category 15. DB00765$h indicates that DB00765 and DB06144 have no reaction type, which is represented as the corresponding Negative sample.

🔬Experiment

Transductive setting

python train_trans.py --data_dir data --dataset {drugbank, TWOSIDES} --fold {0, 1, 2} --smiles_model_path [SMILES_MODEL] --mol_model_path [MOL_MODEL] --dropout [DROPOUT] --device [DEVICE] --batch_size [BATCHSIZE] --model_path [MODEL_PATH] --test [TEST]

Argument

--data_dir  Few-shot dataset or alldata dataset. Choose from {data, alldata}.
--dataset Dataset to be trained on.Choose from {drugbank, twosides}
--fold fold of the dataset to be trained on. Choose from {0, 1, 2}
--smiles_model_path SMILES encoder. Default = model/MoLFormer-XL-both-10pct
--mol_model_path SMILES encoder. Default = model/grover_large.pt
--dropout dropout probability in [0-1]. Default=0.1.
--device Model train device. Default = cuda:1/cpu
--batch_size Batch size of DDI triplets. Default=32.
--model_path The path to use the trained model. Default=None.
--test Whether to perform testing only. Default=False.

Inductive setting

python train_induct.py --data_dir data --dataset inductive_data --fold {0, 1, 2} --smiles_model_path [SMILES_MODEL] --mol_model_path [MOL_MODEL] --dropout [DROPOUT] --device [DEVICE] --batch_size [BATCHSIZE] --s1 [S1_MODEL_PATH] --s2 [S2_MODEL_PATH] --test [TEST]

Argument

--data_dir  Few-shot dataset or alldata dataset. Choose from {data, alldata}.
--dataset Dataset to be trained on.
--fold fold of the dataset to be trained on. Choose from {0, 1, 2}
--smiles_model_path SMILES encoder. Default = model/MoLFormer-XL-both-10pct
--mol_model_path SMILES encoder. Default = model/grover_large.pt
--dropout dropout probability in [0-1]. Default=0.1.
--device Model train device. Default = cuda:1/cpu
--batch_size Batch size of DDI triplets. Default=32.
--s1 The path to use the trained s1 model. Default=None.
--s2 The path to use the trained s2 model. Default=None.
--test Whether to perform testing only. Default=False.

During the training process, the best model address is stored at:

model_save = f'./model/{args.dataset}/{datetime.now().strftime("%m_%d_%H_%M_%S")}/'

After the training is completed, the test set will be automatically called for testing, and the results will be saved in the save file.

result_save = f'./table/{args.dataset}/{datetime.now().strftime("%m_%d_%H_%M_%S")}.csv'

All training results will be saved in log files.

log_path = f'log/{args.dataset}/{datetime.now().strftime("%m_%d_%H_%M_%S")}.log'

📱Concat

Please create Github issues here or email Zihao Mao or Libo Qin if you have any questions or suggestions.