DeepDDS: deep graph neural network with attention mechanism to predict synergistic drug combinations

Abstract

Motivation: Drug combination therapy becomes promising method in the treatment of cancer. However, the number of possible drug combinations toward cancer cell lines is too large, and it is challenging to screen synergistic drug combinations through wet-lab experiments. Therefore, the computational screening has become an important way to prioritize drug combinations. Graph attention network has recently shown strong performance in screening of compound-protein interactions, but it has not been applied to the screening of drug combinations. Results: In this paper, we proposed a deep learning model (DeepDDS) based on graph neural networks and attention mechanism to identify drug combinations that can effectively inhibit the viability of specific cancer cell line. The graph representation of drug molecule structure and gene expression profiles is taken as input to predict the synergistic effects of drug combinations. We compare DeepDDS with traditional machine learning methods (random forest, support vector machine) and other deep learning methods (DeepSynergy, DTF) on the same data set. Our experimental results show that DeepDDS achieved best performance by the AUC value 0.93. Also, on an independent test set released by AstraZeneca, DeepDDS is superior to other comparative methods by 12.2% higher than the suboptimal method. We believe that DeepDDS is a effective tool that can prioritize synergistic drug combinations.

SAFER: sub-hypergraph attention-based neural network for predicting effective responses to dose combinations

“…We observed that two deep neural networks, CCSynergy and DeepSynergy, performed better than the graph-based models, DeepDDs, TranSynergy, SAFER-C2, and SAFER-C3, achieving superior overall performance with 0.943±0.005, 0.935±0.15 of AUPRC, and 0.864±0.009, 0.845±0.013 of AUROC, which could be due to a larger number of hidden layer neurons between 2000 to 4096 that are able to capture more nuances in the data. However, such "deep" architecture is not suitable attention-based models because of graph smoothing problem (12).…”

“…DeepDDs (12) employed graph attention networks to emphasize the sub-structures of drugs' molecular graphs. While network biology helps to understand the dependence of genes and their roles in biological systems, genetic signatures may not truly reflect drug response phenotype without considering their cellular contexts.…”