Using representer point selection for interpretable drug response prediction (OpenDrug)

About

This directory contains the code and resources of the following paper:

"Using representer point selection for interpretable drug response prediction " Under review.

Overview of the Model

OpenDrug is a two step interpretable deep learning model for drug response prediction. OpenDrug is a kind of post-hoc interpretable deep learning model methods which identify important feature and training example by reanalysis model after whole training process.

Step 1. Train OpenDrug to predict drug response.

For each drug, we trained a multi-layer neural network model to predict the drug response based on the molecular features (gene expression and mutation feature) of each cell

Step 2. Interpreting OpenDrug and identify important training example and feature.

For each drug and each testing cell line, the importance of training data and feature can be interpreted by reanalysis the neural network.

For further details, see Online Methods of our paper.

Sub-directories

• [src] contains implementation of OpenDrug used for the peer review.
• [example_data] contains one example drug data. The drug response data is save as pickle file.
• [baseline_methods] contains the implementation of baseline methods including elastic net, linear model, and random forest used in comparison.
• [evaluation] contains the implementation of GCN used in evaluating important training example identified by OpenDrug
• [ipynb] contains the tutorial for processing data

Data

• We constructed dataset based on Genomics of Drug Sensitivity in Cancer (GDSC) project(Iorio et al., 2016), which measured the drug responses elicited by a panel of 265 drugs on 1,001 tumor cell lines, which can be found and downloaded at the GDSC website (https://www.cancerrxgene.org/gdsc1000/GDSC1000_WebResources)

• We selected genes from the “mini cancer genome panel” and further removed gene expression features for which the standard deviations fell into the lowest 20% over all the training samples, and gene mutation features with less than 5 somatic mutations across all the training samples.

• We also provide a process tutorial in ipynb/GDSC_data_preprocessing.ipynb

Dependency

• python 3.7
• pytorch 1.7.0+cu101
• ogb 1.2.4+cu101
• scikit-learn 0.23.2
• lime 0.2.0.1
• ax-platform 0.1.19

Code Usage

1. Train OpenDrug to predict drug response.

   • Benchmark drug response performance for OpenDrug. Notice,most drug response auc is higher than 0.7, therefore, we unsampled data which has lower drug response AUC.

     cd src
     python train.py --drug_id <drug_id> --data_dir <data_dir> 
     

   • Leave-one-out cross-validation implementation which can be used in further interpreting model.

     cd src
     python train.py --drug_id <drug_id> --data_dir <data_dir> --model_dir <model_dir>
     

   • Baseline method implementations used in our paper.

     cd baseline_methods
     python run_{elastic_net,linear_model,rf}.py  --drug_id <drug_id> --data_dir <processed data path> --model_dir <path to save model> 
     

2. Interpreting OpenDrug and identify important training example and feature.

   cd src
   python generate_representer_matrix_index.py --drug_id <drug_id> --data_dir <processed data path> --index_list <test_index_list> 
   --model_dir <model path> --output_dir <path to save interprate result> 
   

   The key idea of OpenDrug is that it can interpret drug response prediction model in a post-hoc way. For each test tumor sample, this script can be used to identify the most important training sample for the prediction (relate to the manuscript Eq.5) and sample-specific feature(gene) importance for the prediction (related to manuscript Eq.6).

   Besides precessed data file and model file trained in step 1, the input files also including a test_index_list file which denotes the sample index that need to be interpreted.

   This script generate two numpy format file, feature importance and sample importance. "_feature_importance.npy" denotes to feature importance interprated by OpenDrug, the matrix weight (W_ij) represent importance of feature j on training data. And "_sample_importance.npy" denotes to the sample importance interprated by

3. Evaluating important training example and feature

   Evaluation 1: OpenDrug perform outperform other methods by using identified important training feature.

   cd evaluation
   python Open_Drug_feature_selection.py --drug_id <drug_id> --data_dir <processed data path> --index_list 
   <test_index_list> --model_dir <saved training model> --output_dir <path to save performance result> 
   --representer_value_matrix_dir_prefix <path of representer graph calculated in step 2> 
   

   Similarly, other feature interpreting baseline can be evaluated by

   cd evaluation
   python random_feature_selection.py --drug_id <drug_id> --data_dir <processed data path> --index_list 
   <test_index_list> --output_dir <path to save performance result> 
   python lime_feature_selection.py --drug_id <drug_id> --data_dir <processed data path> --index_list 
   <test_index_list> --output_dir <path to save performance resault> 
   

   The output result is saved in pickle file including predicted drug response and test data index.

   Evaluation 2: Verifying the representers using graph convolutional networks (GCN).

   cd evaluation
   python train_gcn.py --data_path <path save processed data> --type <random/OpenDrug> 
   --representer_value_matrix_dir_prefix < only used for OpenDrug, path of representer graph calculated in 
   step 2>
   

   The output result is saved in pickle file including person correlation with different percentage of top feature selected.

   ---

   If you have any question, please feel free to contact to me. Email: sht18@mails.tsinghua.edu.cn, majianzhu@gmail.com

   ---

License

OpenDrug is licensed under the Apache License, Version 2.0: http://www.apache.org/licenses/LICENSE-2.0