TAG-CP

Target based Attentive Graph neural network & Combination Prediction (TAG-CP).

Introduction

TAG-CP offers a novel computational model for synergistic drug combination through integrating drug-target relationship to represent small molecules with the framework of attentive graph neural network.

• To begin with, please get the codes with git clone https://github.com/ZJunBio/TAG-CP.git or download the .zip file with magnet https://github.com/ZJunBio/TAG-CP/archive/refs/heads/master.zip, and run the following scripts or commands in the tag-cp directory;
• The NN directory saves the code used to build the deep learning model;
• The data_preprocess folder saves the codes for handling training or testing data;
• The test folder saves the test combinations and python code for predicting drug combinations.

Environment Requirement

• The code has been tested running under Python 3.9.12. The key packages are as follows:

  • pytorch == 1.13.0

  • torch-geometric == 2.3.1

  • rdkit == 2023.3.1

  • pandas == 2.0.1

  • numpy == 1.24.3

  • scikit-learn == 1.3.0

• You can prepare the environment with conda, please try again if you failed to create the environment :

  $ conda env create -n tag_cp -f requirement.yml
  $ conda activate tag_cp

  • If you have not installed conda，please refer to Installing Miniconda

Obtaining the graph embedding of drugs

For now, this graph attention network (GAT) is a transductive learning model and allowed embedding learning for 1362 drugs or compounds. We offer a dictionary-structured file serialized with python pickle module and you can directly use them for further study , or you can run the GAT model to generate the low-dimensional graph embedding.

• With offered file, you can directly use the representation of compounds:

  > import pickle
  > with open("data/drugs/graph_re.pickle", 'rb') as file: 
      cid_repre = pickle.load(file)
      #The key of a cid_repre record is the pubchem CID of a compound;
      #The value of a cid_repre record is the embedding of a compound;
  > #cid_repre[5311104] = array([0, 0, 0, 0.0024659 , 0...]

• Enter the jupyter-lab at terminal under conda environment, and run the GAT model within the notebook named GAT_model.ipynb located in NN directory, and the PubChem CIDs of drugs are saved in drug_list.txt

  $ jupyter-lab

Predicting the drug combination

1. The input format of csv file.

   drug_row,drug_col,cell_line_name
   PubChem CID 1,PubChem CID 2,Cell's name in Cell Model Passports database
   PubChem CID 3,PubChem CID 4,Cell's name in Cell Model Passports database
   

2. You can generate the synergy probability of prepared drug combinations with follow commands, the predict_pytorch.py and results are saved under test directory.

   $ python data_preprocess/preprocess.py
   $ cd test
   $ python predict_pytorch.py lung_test.csv lung_prediction.csv
   $ # where the lung_prediction.csv is a user specified file including output.