A novel approach for drug response prediction in cancer cell lines via network representation learning

Prediction of cancer patient's response to therapeutic agent is important for personalized treatment. Because experimental verification of reactions between large cohort of patients and drugs is time-intensive, expensive and impractical, preclinical prediction model based on large-scale pharmacogenomic of cancer cell line is highly expected. However, most of the existing computational studies are primarily based on genomic profiles of cancer cell lines while ignoring relationships among genes and failing to capture functional similarity of cell lines.

In this study, we present a novel approach named NRL2DRP, which integrates protein-protein interactions and captures similarity of cell lines' functional contexts, to predict drug responses. Through integrating genomic aberrations and drug responses information with protein-protein interactions, we construct a large response-related network, where the neighborhood structure of cell line provides a functional context to its therapeutic responses. Representation vectors of cell lines are extracted through network representation learning method, which could preserve vertices' neighborhood similarity and serve as features to build predictor for drug responses. The predictive performance of NRL2DRP is verified by cross-validation on GDSC dataset and methods comparison, where NRL2DRP achieves AUC > 79% for half drugs and outperforms previous methods. The validity of NRL2DRP is also supported by its effectiveness on uncovering accurate novel relationships between cell lines and drugs. Lots of newly predicted drug responses are confirmed by reported experimental evidences.

The code and documentation are available on https://github.com/USTC-HIlab/NRL2DRP.

Supplementary data are available at Bioinformatics online.