Enhanced heterogeneous network model with ensemble similarities for identifying protein targets of drugs

Abstract

Currently, computational identification of drug target proteins is widely used to help saving cost and time for drug discovery and development. One of the most efficient approaches is a prediction of drug-target interactions based on similarity scores between drugs and target proteins. Despite various data about drugs and targets extensively available, only chemical structures and protein sequences are mostly used to compute drug-drug and target-target similarity scores, respectively. In this thesis, the Forward similarity integration (FSI) Framework is proposed for systematically integrating multiple similarity measures to construct a heterogeneous network propagation model with a suitable similarity integration. Seven drug-similarity measures, nine target-similarity measures, and four similarity integration methods were formulated and used in the FSI framework. Thus, the suitable heterogeneous network model combines three drug-similarity measures integrated by using similarity network fusion (SNF) and one target-similarity measure based on protein sequences. The model selected by FSI reached an accuracy of 99.8% and significantly outperformed the models with random, full similarity integration, and the models without similarity integration. Also, the case studies of newly discovered drug-target interactions demonstrate the practicality of the proposed method for drug-target interaction prediction.