A multi-source similarity fusion method based on hypergraph convolutional networks and graph transformers for predicting miRNA-disease associations.

Identifying the potential miRNA-disease association (MDA) has a greater impetus to the development of drug prevention, treatment and other fields. For the traditional prediction methods, they are characterized by long time and large cost. With the continuous development of bioinformatics, the prediction of MDA using computational methods can better advance the human development in the field of MDA. The existing computational methods, on the other hand, usually have the disadvantages of low prediction accuracy, weak generalization ability and single data used. In order to better develop the association information and mine the problem of association information between multiple nodes, we propose a new method: HGC-GraphT. Firstly, for solving the problem of single data, the method introduces new multi-source similarity information by introducing new similarity information from multiple sources, such as drug-disease association information, protein information, circRNA information and LncRNA information. At the same time, in order to better mine the problem of association information between multiple nodes, this method helps to mine more edge information by constructing a hypergraph, and uses hypergraph convolutional neural network to fully obtain the attribute features within the hypergraph. Then for the problem that a single model extracts information in a sparse way and the reliability of the information is questionable, we add a biological entity graph to supplement the problem of the model acquiring a single feature. Finally, we use a multilayer perceptron to predict unknown miRNA-disease correlation scores. To validate the effectiveness of the method, we conducted a series of experiments using the human microRNA disease database HMDD v3.2. The results of the experiments show that the prediction performance of HGC-GraphT outperforms the other five compared methods in terms of AUC, AUPR, ACC, F1 score, recall and precision. In addition, case studies have shown that HGC-GraphT can accurately predict miRNAs associated with colon cancer and gastrointestinal tumors. In conclusion, the HGC-GraphT method proposed in this paper can effectively predict miRNA-disease associations and has a reliable auxiliary function for biomedical research.