Research On Prediction Method Of MiRNA-disease Associations Based On Tensor Decomposition Model

Predicting the MiRNA-disease interactions may reveal the etiology of various diseases.In recent years,related scholars have proposed many intelligent computerbased algorithms to solve the sparse miRNA-disease interactions prediction problem.However,most methods only focus on the binary data associations for miRNAs and diseases,ignoring the complexity and diversity within the associations.To better solve the above problems,this paper proposes two tensor decomposition model-based research methods to predict the potential multiple types of miRNA-disease associations.The main works are arranged as follows:(1)A multi-type miRNA-disease interactions prediction method,named DFMDT,is proposed based on multi-source data fusion and tensor decomposition.Considering the sparse nature of existing association data,DFMDT fuses multiple data in a tensor model to make full use of the information between miRNAs and diseases.First,the method calculates the fusion similarity for multiple miRNAs and diseases respectively,constructs the three-dimensional tensor of miRNA-disease-type based on existing association data and uses the calculated similarities for neighbor learning.Then,the method performs tensor decomposition on the learned tensor together with the similarity matrix to obtain the factor matrices for miRNA,disease and association type.Finally,the method predicts all disease-related miRNAs and their association types based on tensor reconstruction.This method is optimized using the alternating direction multiplier method and the conjugate gradient algorithm.During the experiment,to ensure the balance of positive and negative samples,this method also performs random sampling based on K-means clustering for positive and negative samples.The experimental results show that DFMDT can achieve 2.51%-52.39%improvement in prediction performance compared with existing methods,and can identify the association information for new miRNAs and new diseases.(2)A multi-type miRNA-disease interactions prediction method,MFMDT,is proposed based on model fusion.The method fuses XGBoost classifier and tensor decomposition model,which solves the problem of type-dimension data imbalance after original tensor neighbor learning in DFMDT.MFMDT first performed neighbor learning on miRNA-disease two-dimensional associations to obtain new interactions.Then,the method uses a classifier to classify the newly added interactions by type,resulting in an enhanced miRNA-disease-type three-dimensional tensor.Finally,the method performs tensor decomposition and reconstruction using the enhanced tensor together with the fusion similarity for miRNA and disease to predict potential miRNAdisease interactions and their types.The experimental results show that MFMDT can achieve 1.14%-72.85% improvement in prediction performance,and can identify the associated information of new diseases.And the prediction performance under different experimental settings is also more stable.