COVID-19 Multi-Targeted Drug Repurposing Using Few-Shot

This is a Pytorch implementation of the following paper:

Yang Liu, You Wu, Xiaoke Shen, Lei Xie. COVID-19 Multi-Targeted Drug Repurposing Using Few-Shot

Architecture

Installation

The following packages are used.

pytorch           1.5.0
rdkit             2020.09.4
tensorboardx      2.1  
torch-cluster     1.5.7       
torch-geometric   1.6.1       
torch-scatter     2.0.5         
torch-sparse      0.6.7        
torch-spline-conv 1.2.0         
torchvision       0.6.0       
tqdm              4.48.2    

Dataset download

All the data files can be downloaded from here dataset_MolGNN

It contains Zinc standard agent for contextPred node and edge pretraining, Chembl for fingerprint pretraining,, and jak1/2/3, mpro, ellinger, amu for covid-19 drug repurposing fintuning.

To use it, pleases unzip, and replace "dataset" argument with the saved location accordingly.

Training flow

1. Context prediction pretraining

   pretrain_contextpred.py --output_model_file OUTPUT_MODEL_PATH_CONTEXT
   

   This is the step1 in bash directory. It saves the result to OUTPUT_MODEL_PATH_CONTEXT to be further pretrained in step2.

2. Motif learning pretraining

   pretrain_fingerprint.py  --input_model_file OUTPUT_MODEL_PATH_CONTEXT \
   --output_model_file OUTPUT_MODEL_PATH_CONTEXT
   

   This is the step2 in bash directory. It is passed with the pretrained model from step1, then does Motif learning pretraining, and saves the reult to OUTPUT_MODEL_PATH_CONTEXT

3. Finetuning

   finetune.py --input_model_file OUTPUT_MODEL_PATH_MOTIF 
   

   This is the step3 in bash directory. It will fine-tune the pretrained model according to downstream task.

Reproducing results in the paper

Saved Trained model can be found in trained_model. To reproduce our result run

bash/step3_finetune_ours.sh