Prediction of drug–target interaction by label propagation with mutual interaction information derived from heterogeneous network

Abstract

The identification of potential drug-target interaction pairs is very important, which is useful not only for providing greater understanding of protein function, but also for enhancing drug research, especially for drug function repositioning. Recently, numerous machine learning-based algorithms (e.g. kernel-based, matrix factorization-based and network-based inference methods) have been developed for predicting drug-target interactions. All these methods implicitly utilize the assumption that similar drugs tend to target similar proteins and yield better results for predicting interactions between drugs and target proteins. To further improve the accuracy of prediction, a new method of network-based label propagation with mutual interaction information derived from heterogeneous networks, namely LPMIHN, is proposed to infer the potential drug-target interactions. LPMIHN separately performs label propagation on drug and target similarity networks, but the initial label information of the target (or drug) network comes from the drug (or target) label network and the known drug-target interaction bipartite network. The independent label propagation on each similarity network explores the cluster structure in its network, and the label information from the other network is used to capture mutual interactions (bicluster structures) between the nodes in each pair of the similarity networks. As compared to other recent state-of-the-art methods on the four popular benchmark datasets of binary drug-target interactions and two quantitative kinase bioactivity datasets, LPMIHN achieves the best results in terms of AUC and AUPR. In addition, many of the promising drug-target pairs predicted from LPMIHN are also confirmed on the latest publicly available drug-target databases such as ChEMBL, KEGG, SuperTarget and Drugbank. These results demonstrate the effectiveness of our LPMIHN method, indicating that LPMIHN has a great potential for predicting drug-target interactions.

A Machine Learning Approach for Drug‐target Interaction Prediction using Wrapper Feature Selection and Class Balancing

“…[1][2] Drug-Target Interaction (DTI) predictions can be performed using the experimental (in-vivo) methods or computational (in silico) methods. Computational methods are broadly classified into ligand-based approach, [3] molecular docking /structurebased approach, [4] text mining, [5] based on gene ontology, [6] chemo-genomic approach, [7] [20] network-based methods, [8,21,23] learning-based, [9][10][11][12][13][14][15][16][17]22] and others. [18][19] These available resources are data driven and are proved to be useful to address and develop a new predictor for DTIs.…”

Drug repositoning or target repositioning: a structural perspective of drug-target-indication relationship for available repurposed drugs

“…Those approaches are generally classified in feature vector-based machine learning and similarity-based machine learning. Similarity-based machine learning methods can be further grouped into three categories: kernel-based approaches, matrix factorization-based approaches and network-based approaches 17 . Compared with time consuming docking and information-demanding QSAR, machine learning methods can be faster and more efficient 18 .…”

Two distributed semi-supervised algorithms for mining of complex networks using Giraph

“…Discovering the disease-gene interaction (Hwang & Kuang, 2010), detecting drug-target interaction (X. Chen, Liu, & Yan, 2012;Yan, Zhang, & Zhang, 2016), and drug repositioning (Shahreza, Ghadiri, Mousavi, Varshosaz, & Green, 2017) are among the research works on this issue and have been appropriately used in biological subjects. In (Hwang & Kuang, 2010), an algorithm called MINProp and a regularization framework is introduced for label propagation over the subnetworks of a heterogeneous network.…”

“…In (Yan et al, 2016), an algorithm named LPMIHN is employed to discover the possible relations of drug-target using the heterogeneous network. In this algorithm, label propagation is done in each heterogeneous subnetwork separately, and interactions of heterogeneous subnetworks are used only as extra information to form similarity matrices.…”

“…Chen, Ding, & Wild, 2012;Tari & Patel, 2014;Zhu, Tao, Shen, & Chute, 2014), and 3) using network analysis (Alaimo, Pulvirenti, Giugno, & Ferro, 2013;H. Chen & Zhang, 2013;Hwang & Kuang, 2010;Li & Lu, 2012;Wang, Yang, & Li, 2013;Xia, Wu, Zhou, & Wong, 2010;Yan et al, 2016). Our proposed methods are also based on a heterogeneous network analysis approach and the distributed versions of MINProp and Heter-LP algorithms.…”