GMPNN-CS

Source code for "Drug-drug Interaction Prediction with Learnable Size-Adaptive Molecular Substructures"

Required packages

• Python == 3.7
• PyTorch == 1.6
• PyTorch Geometry == 1.6
• rdkit == 2020.09.1 (for data preprocessing)

Dataset

Download

Datasets can be downloaded using the python library Therapeutics Data Commons.

Data Preprocessing

Molecular graph representation and negative samples generation. All saved in data/preprocessed.

python data_preprocessing.py -d {drugbank,twosides} [-n NEG_ENT] [-s SEED] -o {all,generate_triplets,drug_data,split} [-t_r TEST_RATIO] [-n_f N_FOLDS]

Arguments:

-d  Dataset to be preprocessed. Choose from {drugbank, twosides}.
-n Number of negative samples per positive sample. Default=1
-s Seed for the random number generator. Default=0
-o Preprocessing operation to perform. Choose from {all, generate_triplets, drug_data, split}.
    1. generate_triplets: generates triplets + negative samples.
    2. drug_data: transform drugs to graphs.
    3. split: stratified splitting of the dataset in -n_f number of folds.
    4. all: does all the above at once.
-t_r test set ratio [0-1]. Default=0.2
-n_f number of folds. Default=3

Training

python train_on_fold.py -d {drugbank, twosides} -fold FOLD -n_iter N_ITER [-drop DROPOUT] [-b BATCH_SIZE]

Arguments:

-d  Dataset to be trained on. Choose from {drugbank, twosides}.
-fold fold of the dataset to be trained on.
-n_iter Number of iterations for message passing.
-drop dropout probability in [0-1]. Default=0.
-b Batch size of DDI triplets. Default=512.