Kernelized-Rank-Learning (KRL)

He* X, Folkman* L & Borgwardt K (2018): Kernelized rank learning for personalized drug recommendation
Bioinformatics, 34(16), 2808–2816
*equal contributions

Dependencies

Python (2.7.12); for the KRL method: numpy (1.13.1), scipy (0.19.1), scikit-learn (0.18.1), joblib (0.11), cvxopt (1.1.9); for reproducing our experiments: PyYAML (3.12), openpyxl (2.4.8); for analysing the results: Jupyter (4.3.0), pandas (0.20.3), matplotlib (2.2.2), seaborn (0.9.0). All of the packages are available via pip (9.0.1). Tested on Ubuntu (16.04).

KRL usage

The implementation of KRL is in KRL/KRL.py:

Y_pred = KRL(X_train, Y_train, X_test, k, lambda, gamma, njobs)

Input:

X_train: n1 x p matrix of n1 samples and p feaures

Y_train: n1 x m matrix of n1 samples and m drugs

X_test:  n2 x p matrix of n2 samples and p feaures

k:       an integer indicating the evaluation parameter k in NDCG@k used during training

lambda:  a float indicating the regularization parameter

gamma:   a float indicating the width of the RBF kernel or 'linear' for linear kernel

njobs:   an integer indicating the number of cores that will be used for parallelization

Output:

Y_pred:  n2 x m matrix of n2 samples and m drugs

Scripts for reproducing our experiments

Download and preprocess the GDSC dataset:

python dataset.py URLs_to_GDSC_datasets.txt

Edit the configuration file to specify which experiments you wish to run:

vim config.yaml

Split the data for training and testing, create folds for cross-validation:

python prepare.py config.yaml

Run experiments for KRL, LKRL, KRR, EN and RF on the GDSC datasets:

python exp.py config.yaml

Run experiments for KBMTL on the GDSC datasets. The code below was tested using Matlab R2016b. Execute this only if KBMTL was specified in the configuration file:

bash prepare_KBMTL.sh
matlab -nodisplay -nosplash -nodesktop -r "run_KBMTL"

Evaluate results on the GDSC datasets:

python results.py config.yaml > results.txt

Data, experiments, and results for KRL on the clinical trials datasets:

This analysis was based on the full training dataset and thus, training is computationally demanding. To test the code in a quick fashion, set the --few_drugs argument to limit the training to only a few drugs of interest.

python create_clin_trial_data.py
python clinical_KRL.py data/KRL_data_for_docetaxel.npz --Lambda 1e-05 --gamma 7.6e-05 --prefix results/KRL_docetaxel_
python clinical_KRL.py data/KRL_data_for_bortezomib.npz --Lambda 1e-05 --gamma 1.8e-08 --prefix results/KRL_bortezomib_
python results_clin_trial.py data/KRL_data_for_docetaxel.npz --drug docetaxel --prefix results/KRL_docetaxel_ --lambdas 1e-05 --gamma 7.6e-05
python results_clin_trial.py data/KRL_data_for_bortezomib.npz --drug bortezomib --prefix results/KRL_bortezomib_ --lambdas 1e-05 --gamma 1.8e-08

Data, experiments, and results for KRL on the TCGA breast cancer (BRCA) cohort:

This analysis was based on the full training dataset and thus, training is computationally demanding. To test the code in a quick fashion, set the --few_drugs argument to limit the training to only a few drugs of interest.

python create_TCGA_data.py
python clinical_KRL.py data/KRL_data_for_TCGA_BRCA.npz --Lambda 1e-05 --gamma 3.1e-05 --prefix results/KRL_TCGA_BRCA_
python results_TCGA.py data/KRL_data_for_TCGA_BRCA.npz --prefix results/KRL_TCGA_BRCA_ --lambdas 1e-05 --gamma 3.1e-05

Plot the results:

jupyter notebook plots.ipynb

Analysis of the published results

All results and configuration files for our paper are located in the folder paper/. To re-analyze these results, a Jupyter notebook is provided:

jupyter notebook plots.ipynb

Installation tests

Run tests and compare the results (takes approx. one hour on a desktop computer):

bash test/run_tests.sh
diff KEEPK_results.txt test/KEEPK_results.txt
diff SAMPLE_results.txt test/SAMPLE_results.txt 

Contact

Any questions can be directed to:

• Xiao He: xiao.he [at] bsse.ethz.ch
• Lukas Folkman: lukas.folkman [at] bsse.ethz.ch