HGCN_MDA: Hierarchical Heterogeneous GraphConvolutional Neural Network for PredictingmiRNA-Disease Associations

Abstract miRNAs are non-coding, single-stranded RNA molecules approximately 22 nucleotides in length. They play pivotal roles in numerous biological regulatory processes and have close associations with a variety of diseases, including cancer, cardiovascular diseases, and neurodegenerative disorders. miRNAs have become new candidate molecular targets for disease diagnosis and treatment. Therefore, the swift and effective identification of associations between miRNAs and diseases holds significant clinical significance. Presently, the available methods for predicting miRNA-disease associations remain quite limited. Moreover, there is a scarcity of approaches that specifically address the varying importance of different nodes in miRNA and disease interactions, and then model these variations. Consequently, there is an urgent and challenging need to develop computational models for forecasting potential miRNA-disease associations. In this paper, we propose a hierarchical heterogeneous graph convolutional network named HGCN_MDA for the prediction of miRNA-disease associations. Our approach encompasses several key steps: First, we gather data on the associations between miRNAs and diseases, miRNAs and miRNAs, and diseases and diseases. Subsequently, we compute the importance scores of nodes in the graph, which serve as indicators for stratifying the hierarchical network. We then construct a multi-layer network to extract features based on their respective importance. Our approach combines global and local features, which are integrated by the attention mechanism. Finally, a fully connected layer is used to build a prediction model. Comprehensive experimental comparisons demonstrate that our proposed HGCN_MDA has better ability to predict positive samples and comprehensive performance.