DIMDRP: A Double Iteration Method for Drug Response Prediction

Jianxing Ouyang,Minzhu Xie,Xinqiu Liu

doi:10.1109/access.2019.2942217

Jianxing Ouyang, Minzhu Xie + Show 1 more

Open Access

https://doi.org/10.1109/access.2019.2942217

Copy DOI

Abstract

Using molecular profiles to predict the drug response is of practical importance in precision medicine and have been extensively studied. Due to the complexity of molecular information, the existing methods couldn't capture enough information and their prediction performances are not satisfying. In this study, we propose a method called DIMDRP (double iteration method for drug response prediction) which improves a lot in the prediction accuracy. DIMDRP integrates several important molecular information including miRNA expression, drug chemical structure, target interaction, drug-target interaction and cell line-drug response, and constructs a heterogeneous network. Then an improved information flow iteration algorithm is used to calculate association scores of cell line-drug responses, and we prioritize the cell lines for each query drug. The cross-validation experiments show that the average area under curve (AUC) of DIMDRP is as high as 0.8953, and two other measurement metrics are applied to assess the performance. When compared to other approaches, our method shows a significant advantage for all metrics. We also consider the specific tissue condition into our competition from a practical aspect. Last, a case is studied to predict novel cell line-drug responses, some of which can be evidenced by previous experiments.

Highlights

The high complicated mechanism of cancers makes it difficult to find effective medicines for personal treatment, and different individuals respond differently to the same medicine
In this study, we proposed DIMDRP to predict drug responses which was based on a heterogeneous network integrating the drug similarity information, target similarity information and interaction information between cell line, drug and target
To evaluate the effectiveness of our method, we conducted leave-oneout cross-validation (LOOCV) experiments on the entire database and tissue-specific database downloaded from Genomics of Drug Sensitivity in Cancer (GDSC) and Cell Line Encyclopedia (CCLE)

Summary

Introduction

The high complicated mechanism of cancers makes it difficult to find effective medicines for personal treatment, and different individuals respond differently to the same medicine. Conventional methods of discovering new drug responses are always based on many biological experiments that cost lots of money and time. With the great development in biological technologies, more and more molecular information about human genome, transcriptome and proteome has been accumulated, and it provides researchers a novel way to study the association between drugs and cell lines. There exists a trend to conduct the drug response prediction analysis with these accessible data [1]. Plenty of large-scale studies have been conducted to collect the corresponding biological data. Two famous of them are Genomics of Drug Sensitivity in Cancer (GDSC) and Cancer Cell Line Encyclopedia (CCLE) which provide public access

Methods

Results

Conclusion