DrugAI

Drug Efficacy Prediction using Graph Neural Network

We implemented 3 GCN model for efficacy prediction:

• DMPNN
• GCN
• GIN, see also GINEConv

Dependency

• numpy
• pandas
• python >= 3.7
• Pytorch >= 1.5
• Pytorch Geometric >= 1.7
• RDkit
• optuna: (optional) hyperparameter search

Usage

Input data

Need at least one file

1. a csv file with

• first column: SIMILES
• second to last column: float (regression), int (classification)

2. (optional) a pickle file with a tuple (train, test, val) of row indices (splits).

1. Train

python drug_gnn/train.py --data_path ${data} \
                        --task ${regression} \
                        --gnn_type dmpnn --log_dir checkpoints/dmpnn 

2. Predict

python drug_gnn/predict.py --data_path ${data} \
                        --task ${regression} \
                        --gnn_type dmpnn --log_dir checkpoints/dmpnn 

Hyperparameter tuning

python drug_gnn/hyperopt.py --data_path ${data} --task ${regression}  \
                            --gnn_type dmpnn \
                            --hyperopt_dir hyper_dmpnn

Drug Efficacy prediction

1. Train your model using LINCS 2020 Data

   • input data format:
     • shape： (num_smiles, num_landmark_genes)
     • first column are SMILE strings
     • the rest columns are expression values
     • columns names should be Entrez IDs
   • save best_model

2. Prediction step will generate two output file

   • Embeddings for each molecule: xxx.embeddings.npy
   • Predicted Landmark genes expression: xxx.pred.exprs.csv

3. Efficacy Score:

   • 1. Prepare a up- or down-regulated gene signatures (Entrez ID only): up.txt, down.txt
   • 2. Get transform matrix: GSE92743_Broad_OLS_WEIGHTS_n979x11350.gctx

        # convert to pandas DataFrame
        from cmapPy.pandasGEXpress.parse import parse
        weight = parse('GSE92743_Broad_OLS_WEIGHTS_n979x11350.gctx').data_df
        weight.to_csv("GSE92743_Broad_OLS_WEIGHTS_n979x11350.csv")

   • 3. Predicted output from step 2
   • 4. Run:

   python efficacy.py --weights GSE92743_Broad_OLS_WEIGHTS_n979x11350.csv \
                      --predicts xxx.pred.exprs.csv \
                      --up up.txt \
                      --down down.txt \
                      --output efficacy.csv

Results

1. Average pearson's correlation (AUC-like plot) shows GNN works pretty good for predicting transcriptional profiles
   

2. TSNE plot of drug's embeddings

3. Pearson's coefficiency distribution

Contact

Zhuoqing Fang: fangzq@stanford.edu

Others

This project is based on chemprop, and chiral_gnn