Download PDFOpen PDF in browserMixed Expert Model for Drug-Target Interaction PredictionEasyChair Preprint no. 891412 pages•Date: October 3, 2022AbstractComplex interactions between biology entities (drugs, diseases, side-effects, etc.), have posed difficulties for drug discovery and treatment. Despite the significant efforts that have been invested in drug-target interaction prediction, existing methods are still afflicted by the highly sparse datasets for drug-target interaction prediction and ignore the impact of interactions between different types of biological entities when constructing heterogeneous biology networks. To address the issue, we develop a framework based on the mixed expert model, named MEDTI, which captures the intratype relationships of interactions between the same type of biological entities and inter-type relationships of interactions among different biological entities for drug-target interaction prediction. The MEDTI consists of three main components: the edge representation extractor, the type-prior information extractor, and the mixed expert discriminator. The edge representation extractor is responsible for constructing a heterogeneous biology network from numerous types of biology networks, such as drug-drug interaction, drug-target interaction, drug-disease association, and drugside-effect association networks. Then the edge representation extractor maps the representation of different types of networks into their type spaces. The type-prior information extractor exploits prior information of different types of networks by using a type gate to aggregate information of each type spaces in heterogeneous biology networks. Cooperating with the type-prior information extractor, the mixed expert discriminator resists the implications of different types of biology entities and predicts the unobserved drug-target interaction. Experiments on real-world heterogeneous biology network datasets show that the MEDTI can outperform the state-of-the-art methods and predict different drug-target interactions accurately. Keyphrases: Drug-target interaction prediction, mixed expert model, type gate
|