Predicting MiRNA-disease Association Based On Graph Neural Network

Since it was discovered,micro RNA has attracted the attention of an increasing number of researchers.In particular,the discovery of its regulatory role in cellular activity has led to the discovery that it is inextricably linked to many diseases.Uncovering the association between micro RNAs and diseases is of great importance to the study of disease onset,development and treatment.However,due to the long time cycle and high cost of resource consumption in biological experiments,the prediction of miRNA-disease associations using computational methods has now become an important tool to guide traditional biological experiments,greatly improving the efficiency of discovering disease-associated miRNAs.Although the proposed computational methods have achieved good performance,there are still some problems.First,most of the currently proposed methods for predicting miRNAdisease associations are based on the assumption that similar miRNAs are associated with the same diseases,and that similar diseases are associated with the same micro RNAs.This assumption helps to predict miRNA-disease associations to a certain extent,but the introduction of a priori knowledge of "similarity",which varies between models,leads to inaccurate prediction results.Secondly,the lack and inaccuracy of the similarity data used in the computational approach has an impact on the prediction results.Therefore,there is still a need to improve the methods for predicting miRNA-disease associations.In this paper,we propose a graph neural network prediction algorithm HLGNN-MDA(Heuristic Learning based on Graph neural networks for miRNA-disease association prediction).HLGNN-MDA learns the association patterns between miRNAs and diseases from the graph structural features of existing miRNA-disease association networks through graph neural networks for the purpose of predicting miRNA-disease association.First,HLGNN-MDA extracts enclosing subgraphs around miRNA-disease node pairs.Then to represent the structure of the enclosing subgraphs,the nodes of the subgraphs were annotated.Then,the association between miRNA-disease pairs was taken as the supervision information,and the labeled subgraphs were input into the neural network for training.Finally,the trained graph neural network is used to predict the potential miRNA-disease association.Meantime,this paper improves the graph neural network so that it can simultaneously learn information between miRNA and disease nodes and topological relationships between homogeneous nodes,compensating for the similarity information between homogeneous nodes and improving the applicability of graph neural networks in bipartite graph networks.In the experimental section,this paper first compares the proposed HLGNN-MDA with MKRMDA,BNPMDA,IMCMDA,BLHARMDA and LFEMDA in a ten-fold crossvalidation.HLGNN-MDA obtains maximum values in five of the six metrics.The paper then discusses the effect of the number of hops of the enclosing subgraph and the number of nodes of the graph pooling layer unification on the graph neural network.The effect of improvements to the graph convolution layer of the graph neural network is also highlighted.HLGNN-MDA uses existing association data to make predictions of potential association.Forty-nine of the top 50 miRNAs from the predictions could be found in the validation database;97 of the top 100 miRNAs could be found in the validation database.Then,this paper also conducted case studies on breast cancer,hepatocellular carcinoma and renal cell carcinoma,and 50,50 and 47 of the top 50 miRNAs of the prediction results were validated respectively.It proves that the HLGNN-MDA proposed in this paper can effectively predict the association between miRNA and disease.