Research On Small Molecule-miRNA Association Prediction Based On Heterogeneous Graph Neural Network

As a new therapeutic target of small molecule drugs,miRNA can affect various diseases including cancer by regulating gene expression.Therefore,identifying the association of small molecule drugs and miRNA is of great significance for disease diagnosis,treatment and clinical application of drugs.Since traditional biological experiments are time-consuming and laborious,many computational models have been proposed to predict the association between small molecule drugs and miRNA.However,most of the association prediction models proposed at present have problems,such as inability to deal with isolated nodes,insufficient use of multi-source information,and ignoring the impact of different types of nodes,so the prediction effect is poor.In view of the above problems,this paper proposes two deep learning-based models to predict the small molecule-miRNA association on the basis of the association network and similarity network of various biomolecules:(1)Heterogeneous network embedding for small molecule-miRNA association prediction(HNESMA).Firstly,HNESMA integrates the associations among various biomolecules in the data set into three kinds of association networks,namely: small molecule-miRNA association network,small molecule-disease association network,and miRNA-disease association network.Then,multiple embeddings of each biomolecule node are obtained through structural deep network embedding.In the structural deep network embedding,the first-order similarity and second-order similarity of the network are jointly optimized to make the embedded representation of nodes more robust.Finally,multiple embeddings are fused to obtain the feature vector of each small molecule-miRNA association pair,and this feature vector is used to predict the potential small molecule-miRNA association.(2)Heterogeneous graph neural network for small molecule-miRNA association prediction(HGNNSMA).Firstly,HGNNSMA effectively integrates small molecule drug similarity,miRNA similarity,disease similarity and the three association networks mentioned above into a complete heterogeneous network.HGNNSMA uses the improved restart random walk algorithm on the heterogeneous network to sample various types of strongly correlated neighbors of fixed size for each node.And the SDNE method in HNESMA is used instead of the deepwalk algorithm of the original model to obtain the structural initial features of each node.Then,the initial features of the same type of nodes are aggregated by Bi-LSTM to obtain general embeddings for each type of node.Finally,the attention mechanism is used to calculate the weight of each embedding type,and the final embedding of each node is obtained by aggregating the three types of embedding and the embedding of the node itself according to the weight,and then input into the corresponding classifier to predict the score of the small molecule-miRNA association relationship.Both the different computational models have achieved better results in the performance evaluation.The results of cross validation show that the AUC of the two proposed models HNESMA and HGNNSMA reach 0.9132 and 0.9525 respectively,which are superior to some existing computational methods.Furthermore,the results of the case studies of the two models also demonstrate the practicality of both models.