Pathway-guided deep neural network toward interpretable and predictive modeling of drug sensitivity

Abstract

Motivation: To efficiently save cost and reduce risk in drug research and development, there is a pressing demand to develop in-silico methods to predict drug sensitivity to cancer cells. With the exponentially increasing number of multi-omics data derived from high-throughput techniques, machine learning-based methods have been applied to the prediction of drug sensitivities. However, these methods have drawbacks either in the interpretability of mechanism of drug action or limited performance in modeling drug sensitivity. Results: In this paper, we presented a pathway-guided deep neural network model, referred to as pathDNN, to predict the drug sensitivity to cancer cells. Biological pathways describe a group of molecules in a cell that collaborates to control various biological functions like cell proliferation and death, thereby abnormal function of pathways can result in disease. To make advantage of both the excellent predictive ability of deep neural network and the biological knowledge of pathways, we reshape the canonical DNN structure by incorporating a layer of pathway nodes and their connections to input gene nodes, which makes the DNN model more interpretable and predictive compared to canonical DNN. We have conducted extensive performance evaluations on multiple independent drug sensitivity data sets, and demonstrate that pathDNN significantly outperformed canonical DNN model and seven other classical regression models. Most importantly, we observed remarkable activity decreases of disease-related pathway nodes during forward propagation upon inputs of drug targets, which implicitly corresponds to the inhibition effect of disease-related pathways induced by drug treatment on cancer cells. Our empirical experiments show that pathDNN achieves pharmacological interpretability and predictive ability in modeling drug sensitivity to cancer cells. Availability: The web server, as well as the processed data sets and source codes for reproducing our work, is available at

Looking at the BiG picture: Incorporating bipartite graphs in drug response prediction

“…We slightly modified NRL2DRP to predict continuous values instead of binary values (so that it can be applied to our data). PathDNN [7] is another deep learning method that proposes to add some level of explainability to the drug response prediction problem by constraining the neural network connectivity using a pathway mask. This method uses drug targets and gene expressions, both of which should be in any of the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways [26].…”

“…To evaluate the performance of BiG-DRP and BiG-DRP+, we used 5-fold cross validation and compared these results across different baselines and other drug response prediction approaches, namely NRL2DRP [17] and PathDNN [7]. We tested on two data-splitting methods, leave-pairs-out and leave-cell lines-out, which represent two possible scenarios of data availability.…”

“…Due to the similarity in molecular and chemical structure of different drugs and their mechanisms of action, machine learning (ML) methods that can take advantage of these similarities are of great interest. Instead of training a different ML model for each drug, one can formulate the drug response prediction as a paired prediction problem, such that a model takes in a (drug, CCL) pair as input and trains a single model for all drugs and CCLs [7][8][9]. This increases the number of samples, and enables information sharing across many drugs and drug families.…”