Integrated multi-similarity fusion and heterogeneous graph inference for drug-target interaction prediction

Drug-target interaction (DTI) prediction performs a crucial part in drug discovery and design. Although many computational approaches for such prediction have been proposed, current researches still generally adopt chemical similarities of drugs or the sequence similarities of targets. However, the valuable information of known interactions has not been noticed, and the existing noise and useless information reduce the accuracy of DTI prediction. In addition, many existing computational approaches ignore the behavior information between nodes of the DTI network. In this paper, we develop an ensemble computational approach called integrated multi-similarity fusion and heterogeneous graph inference. First, based on the known DTI network, the degree distribution of drug and target similarities are analyzed and the noise and useless information are removed to improve prediction accuracy. Second, based on drug and target similarities and known DTIs, a strategy of multi-similarity fusion is proposed to capture potential useful information from known interactions that is used for enhancing drug and target similarities. Third, the heterogeneous graph inference is used to predict the DTIs to capture the edge weight (closeness) and behavior information (diffusion) between nodes of a heterogeneous network. To assist the reproducibility of our work and its comparison to published results, we perform experiments on four benchmark datasets. Results show that our approach outperforms some existing approaches and can contribute to predicting potential DTIs.(c) 2022 Elsevier B.V. All rights reserved.