Learning size-adaptive molecular substructures for explainable drug–drug interaction prediction by substructure-aware graph neural network

Note

If you want to use the code in the cold start scenario, you should add dropout to MLP layers and use weight decay.

Requirements

numpy==1.18.1
tqdm==4.42.1
pandas==1.0.1
rdkit==2009.Q1-1
scikit_learn==1.0.2
torch==1.11.0
torch_geometric==2.0.4
torch_scatter==2.0.9

Step-by-step running:

1. DrugBank

• First, cd SA-DDI/drugbank, and run data_preprocessing.py using
  python data_preprocessing.py -d drugbank -o all
  Running data_preprocessing.py convert the raw data into graph format.
  Create a directory using
  mkdir save

• Second, run train.py using python train.py --fold 0 --save_model

  to train SA-DDI. The training record can be found in save/ folder.

  Explanation of parameters

  • --n_iter: number of iterations
  • --fold: {0, 1, 2}
  • --epochs: number of epochs
  • --weight_decay: weight decay
  • --batch_size: batch size
  • --save_model: whether save the model or not, for example, 'python train.py' will not save the model and 'python train.py --save_model' will save the model.
  • --lr: learning rate

2. TWOSIDES

• First, cd SA-DDI/drugbank, and run data_preprocessing.py using
  python data_preprocessing.py -d twosides -o all
  Running data_preprocessing.py convert the raw data into graph format. Create a directory using
  mkdir save

• Second, run train.py using python train.py --fold 0 --save_model

  to train SA-DDI. The training record can be found in save/ folder.