Detection of Drug-Drug Interactions Through Knowledge Graph Integrating Multi-attention with Capsule Network

Drug-drug interaction (DDI) prediction is a challenging problem in drug development and disease treatment. Current computational studies mainly solve this problem by designing features and extracting features from multi-sources. They have limitations in accuracy, knowledge and universality. In this paper, a universal computational method for DDI prediction is proposed. The model is made up of several identical layers. Each layer is composed by a multi-attention unit and a capsule unit. Multi-attention is responsibility for extracting neighborhood features and capturing representation under various semantic relation. Capsule unit is responsibility for features aggregating. Each layer can be viewed as first-order representation. High-order representation is obtained by increasing layers. The structure of whole method is easily understood and implemented. Experiment results show that proposed method can achieve a promising predicting performance and stands out when compared with several popular network embedding methods. Proposed method is a reliable model for DDI prediction and is suitable for other interaction predictions.