DrugComb Cross dataset prediction model

Cross-study prediction of drug combination treatment response

for the dataset used in this study, please refer to DrugComb

Dependencies

• python (3.6.5)
• LightGBM (2.3.2)

Obtaining training dataset

Download the drug combination screening dataset from DrugComb data portal: https://drugcomb.org/download/ and put it under a new directory ./dataset

QC

QC.ipynb

Split training and test set by study name

python split_train_test.py

Construct feature set

python preprocess_feature.py

Build model

You can start training model simply by executing the following bash file:

sh bash.sh

which will train 20 different models with different feature combinations.

You can also refer to ./master and run python main.py -h

usage: main.py [-h] [-f FEATURES [FEATURES ...]]

Build Drugcomb drug combination prediction machine learning models across studies.

optional arguments:
  -h, --help            show this help message and exit
  -f FEATURES [FEATURES ...], --features FEATURES [FEATURES ...]
                         Features selected for model, including:
                                drug_categorical;
                                cell_line_categorical;
                                cancer_gene_expression;
                                chemical_structure;
                                monotherapy_ri;
                                monotherapy_ic50;
                                drc_baseline;
                                drc_intp_linear;
                                drc_intp_lagrange;
                                drc_intp_4PL;
                                (default = ['drug_categorical', 'cell_line_categorical']

this will generate results, save in a new folder ./results

Performance visualization

demo_results.ipynb

Reference

Zhang, H., Wang, Z., Nan, Y. et al. Harmonizing across datasets to improve the transferability of drug combination prediction. Commun Biol 6, 397 (2023). https://doi.org/10.1038/s42003-023-04783-5