Multi-source Data-driven Drug Repositioning based on Deep Self-attention Network

Drug-target interaction (DTI) prediction is the core of drug repositioning, which can promote rapid screening of candidate drugs and shorten drug development time. Methods based on graph neural networks (GNN) have attracted widespread attention and shown certain advantages in DTI prediction. However, extracting deep features from graph structure information in heterogeneous bioinformatic network (HBIN) still faces challenges. In this work, we propose a novel deep self-attention model called BCHNDTI. First, the model integrates the biological information of drugs and targets to construct HBIN, and preprocesses graph information by using the path integral thermonuclear (PIHK) method. Then, we use Graph Convolutional Network (GCN) and Multi-Head Attention to obtain representations of HBIN features and input them into a novel self-attention network, which can be used to learn deep feature representations of drugs and targets. Finally, we use the XGBoost classifier to predict DTIs. Compared with several recent DTI prediction methods, the results indicated that BCHNDTI performs better on the benchmark dataset for identifying DTIs. The case study further confirmed its ability to predict potential drug-target interactions.