Creating processed datasets

Experimental Setup

• Setup1: Leave-responses-out (LRO) cross validation
• Setup2: Leave-cell lines-out (LCO) cross validation (Unseen cell lines with 1/5th cell lines for each cancer type hold out during testing)
• Below, we provide detailed instructions and commands on how to reproduce the provided processed and cleaned data files from source files.
• Each setup requires different type of data split as discussed below.

Get gene expression data

• CCLE: Download CCLE expression data from here and save it as CCLE/CCLE_expression.csv.
• Combined: Download the combined_rnaseq_data file with gene expression data for cell lines renamed for CTRPv2 from here and save it in Combined/.

Preprocess data from source files

• To create the processed data files, first download the source datasets as mentioned below.
• For CTRPv2 with adjusted AUCs, we download the following files from here:
  • For drug information: drug_info
  • For adjusted AUCs: combined_single_response_agg
• For PRISM, we download the following files from here
  • For drug information and AUCs: secondary-screen-dose-response-curve-parameters.csv
• Run the commands provided in commands_ctrpv2.log and commands_prism.log to clean the files within respective directories for the following steps.

For LRO Setup

mkdir -p ctrpv2/LRO/
python scripts/create_cv.py --data_dir /content/DrugRank/data/Combined/ --save_dir ctrpv2/LRO/

The above script creates the list of AUCs in aucs.txt and splits in splits.txt. Run this command similarly for PRISM.

To create setup for pLETORg:

python scripts/create_setup_LRO.py --data_dir ctrpv2/LRO/ --save_dir ctrpv2/LRO/pletorg/ --genexp_file Combined/combined_rnaseq_data_combat 

• The above script does the following:
  • Creates the full cell line - drug sensitivity matrix and saves as auc_matrix.txt. M x N matrix
  • Saves the list of cell IDs as cells.txt. # cells = M
  • Saves the list of drug IDs as drugs.txt. # drugs = N
  • Note that the order of cells across rows and the order of drugs across columns must be consistent with the respective ordering in the files cells.txt and drugs.txt.
• Rerun the script accordingly for PRISM.

To select informative genes using Elastic Net

python scripts/lasso.py --data_dir ctrpv2/LRO/ --l1_ratio 0.5
python scripts/lasso.py --data_dir prism/LRO/ --l1_ratio 0.1

• This runs Elastic Net on the drug response matrix to identify informative genes for each cell line, following the standard setup in pLETORg.

To compute pairwise cell line similarity using the selected genes as features

• To compute cosine similarities:

python scripts/cell_similarity.py --data_dir ctrpv2/LRO/ --genes ctrpv2/LRO/genes.txt --sim cos
python scripts/cell_similarity.py --data_dir prism/LRO/ --genes prism/LRO/genes.txt --sim cos

• To compute simliarities using RBF kernel:

python scripts/cell_similarity.py --data_dir ctrpv2/LRO/ --genes ctrpv2/LRO/genes.txt --sim rbf
python scripts/cell_similarity.py --data_dir prism/LRO/ --genes prism/LRO/genes.txt --sim rbf

For LCO Setup

python scripts/create_setup_LCO.py --data_dir ctrpv2/LRO/ --save_dir ctrpv2/LCO/pletorg/
python scripts/create_setup_LCO.py --data_dir prism/LRO/ --save_dir prism/LCO/pletorg/

• Creates the auc matrices separately for training, val and test in each fold
• Saves the list of cell IDs as *_cells.txt and the union of cells in ../cells.txt, and corresponding pairwise similarity matrix
• Note that the order of cells across rows is consistent with the respective ordering in the similarity matrix file.