The Research Of Disease Associated MiRNA Pairs Prediction Algorithm Based On Deep Tensor Factorization

MicroRNA(miRNA)regulate gene transcription and play a significant role in the emergence and advancing of complex diseases.As the biotechnology and computer technology developing rapidly,researchers have realized that the regulatory level of multiple miRNA is stronger than that of a single miRNA,and that complex diseases are often caused by the combination of miRNA.Therefore,exploiting miRNA combinations to treat complex diseases has become a promising strategy to improve the therapeutic effect,which provides great insights for exploring disease-associated miRNA pairs and comprehending their synergistic mechanism in complex diseases.The explosion of human genome data and massive amounts of biological information are valuable resources for computational methods to discover disease-related miRNA combinations.However,how to model the associations between miRNA combinations and diseases and apply computational methods to efficiently integrate multi-source data has become a major challenge.Hence,this paper takes miRNA combinations and diseases as research objects and proposes two end-to-end disease-miRNA combination association algorithms from the perspective of network representation learning and multi-source data aggregation by integrating tensor factorization technology.The main work is summarized as follows:(1)In view of the problem that existing researches on miRNA-disease binary association prediction and miRNA synergistic effects cannot apply to disease-miRNA pairs prediction tasks and fully exploit the topology information in association networks,this paper proposes a deep tensor factorization framework based on graph attention mechanism,named Graph TF,for predicting disease associated miRNA-miRNA pairs.The framework aggregates neighbor information in miRNA and disease similarity networks by learning weights of neighbor nodes adaptively,and reconstructs miRNA and disease features into a miRNA-miRNA-disease association tensor for association prediction.Extensive experiments under five-fold cross-validation of Graph TF show better performance than other algorithms.Robustness tests on different negative sample size and missing rate of associations also demonstrate the stability of Graph TF.Case studies and the enrichment analysis of miRNA pairs further verified the powerfulness of Graph TF to predict miRNA pairs of potential diseases.(2)Considering that Graph TF fails to fuse rich biological auxiliary data and heterogeneous association information,we propose a deep tensor factorization model MAGTF based on multi-source data aggregation.In this model,a simplified attention mechanism is applied to adaptively learn the node features in multi-source miRNA and disease similarity networks,and a multilayer perception is used for feature fusion.Then,MAGTF aggregates neighbor features in the miRNA-disease heterogeneous network to learn the features of miRNA and diseases.Finally,the learned global features are used for tensor reconstruction and prediction.Through five-fold cross validation experiments,MAGTF achieves good performance as other compared methods.Ablation study illustrates the validity of each module in the MAGTF model.In addition,case study further demonstrates the ability of MAGTF to discover potential disease-associated miRNA pairs in practical applications.