LncRNA-miRNA interaction prediction based on multi-source heterogeneous graph neural network and multi-level...

Identifying the interactions between long non-coding RNAs (lncRNAs) and microRNAs (miRNAs) has been demonstrated to unveil the mechanisms of biological processes, thus aiding in disease diagnosis and treatment. Traditional experimental methods are often laborious, costly, and time-consuming. Graph neural network-based approaches have made significant strides by effectively learning from graph data. However, the type selection of intermediate nodes in heterogeneous networks remains relatively simplistic and the varying importance of different types of information in different processes is frequently neglected in existing methods. In this study, we proposed a novel model for predicting potential LncRNA-MiRNA Interaction based on the Multi-source heterogeneous graph neural network and Multi-level attention mechanism (LMI-MM). In LMI-MM, a homogeneous network and a multi-source heterogeneous network of lncRNAs, miRNAs, diseases, small-molecule drugs, and mRNAs are constructed, graph neural network and graph representation learning method are employed to extract node features respectively. Besides that, the multi-level attention modules are introduced to incorporate adaptive aggregation that explicitly captures information importance through differentiated weighting. Compared with other models, LMI-MM exhibits superior performance across evaluation metrics such as AUC and AUPR. Further case studies demonstrate its effectiveness in identifying potential lncRNA-miRNA interactions.