TGSA: Protein-Protein Association-Based Twin Graph Neural Networks for Drug Response Prediction with Similarity Augmentation

Here we provide an implementation of Twin Graph neural networks with Similarity Augmentation (TGSA) in Pytorch and PyTorch Geometric. The repository is organised as follows: Cancel changes

• data/ contains the necessary dataset files;
• models/ contains the implementation of TGDRP and SA;
• TGDRP_weights contains the trained weights of TGDRP;
• utils/ contains the necessary processing subroutines;
• preprocess_gene.py preprocessing for genetic profiles;
• smiles2graph.py construct molecular graphs based on SMILES;
• main.py main function for TGDRP (train or test);

• Please install the environment using anaconda3;
  conda create -n TGSA python=3.6
• Install the necessary packages.
  conda install -c rdkit rdkit
  pip install fitlog
  pip install torch (1.6.0)
  pip install torch-cluster (1.5.9) (https://pytorch-geometric.com/whl/)
  pip install torch-scatter (2.0.6) (https://pytorch-geometric.com/whl/)
  pip install torch-sparse (0.6.9) (https://pytorch-geometric.com/whl/)
  pip install torch-spline-conv (1.2.1) (https://pytorch-geometric.com/whl/)
  pip install torch-geometric (1.6.1)

• data/CellLines_DepMap/CCLE_580_18281/census_706/ - Raw genetic profiles from CCLE and the processed features. You can also preprocess your own data with preprocess_gene.py.

• data/similarity_augment/ - Directory edge contains edges of heterogeneous graphs; directory dict contains necessary data and dictionaries for mapping between drug data or cell line data.

• data/Drugs/drug_smiles.csv - SMILES for 170 drugs. You can generate pyg graph object with smiles2graph.py

• data/PANCANCER_IC_82833_580_170.csv - There are 82833 ln(IC50) values across 580 cel lines and 170 drugs.

• data/9606.protein.links.detailed.v11.0.txt and data/9606.protein.info.v11.0.txt - Extracted from https://stringdb-static.org/download/protein.links.detailed.v11.0/9606.protein.links.detailed.v11.0.txt.gz

• You can run python main.py --mode "train" to train TGDRP or run python main.py --mode "test" to test trained TGDRP.

• First, you can run heterogeneous_graph.py to generate edges of heterogeneous graphs.

• Then, you can run main_SA.py to generate node features of heterogeneous graphs using two GNNs from TGDRP/TGDRP_pre and to fine-tune sequentially the remained parameters from TGDRP/TGDRP_pre. To be specific, you can use the instruction python main_SA.py --mode "train"/"test" --pretrain 0/1 to fine-tune TGDRP/TGDRP_pre or to test fine-tuned SA/SA_pre.

MIT