One of the most important problem in personalized medicine research is to precisely predict the drug response for each patient. Due to relationships between drugs, recent machine learning-based methods have solved this problem using multi-task learning models. However, chemical relationships between drugs have not been considered. In addition, using very high dimensions of -omics data (e.g., genetic variant and gene expression) also limits the prediction power. A recent dual-layer network-based method was proposed to overcome these limitations by embedding gene expression features into a cell line similarity network and drug relationships in a chemical structure-based drug similarity network. However, this method only considered neighbors of a query drug and a cell line. Previous studies also reported that genetic variants are less informative to predict an outcome than gene expression. Here, we develop a novel network-based method, named GloNetDRP, to overcome these limitations. Besides gene expression, we used the genetic variant to build another cell line similarity network. First, we constructed a heterogeneous network of drugs and cell lines by connecting a drug similarity network and a cell line similarity network by known drug-cell line responses. Then, we proposed a method to predict the responses by exploiting not only the neighbors but also other drugs and cell lines in the heterogeneous network. Experimental results on two large-scale cell line data sets show that prediction performance of GloNetDRP on gene expression and genetic variant data is comparable. In addition, GloNetDRP outperformed dual-layer network- and typical multi-task learning-based methods.
Keywords: drug similarity network; gene expression-based cell line similarity network; genetic variant-based cell line similarity network; global drug response prediction; heterogeneous network of drugs and cell lines.
Copyright © 2018 Elsevier Ltd. All rights reserved.