Multiple Dimensional Scaling For MiRNA-Disease Association Prediction Based On Katz Model

The development of miRNA disease association prediction has brought enormous opportunities for the diagnosis and treatment of human diseases.A series of computational methods have been developed to predict potential miRNA disease associations.We found that the processing methods for miRNA and disease similarity information are different,and the results of association prediction are also significantly different Therefore,this thesis applies the Katz model to the heterogeneous network composed of miRNA similarity and disease similarity,reconstructs the heterogeneous network,and finally performs multidimensional scaling embedding on the reconstructed heterogeneous network to predict miRNA-disease association.The main work and innovations of the thesis are as follows:(1)Build similarity heterogeneous networks.Firstly,we combine the sequence similarity and Gaussian interaction kernel similarity of miRNA to obtain the miRNA similarity matrix ISm.Combining semantic similarity and Gaussian interaction kernel similarity of diseases,we obtain disease similarity matrix ISd.Then integrate ISm、ISd D and miRNA-disease adjacency matrix A,and construct similarity heterogeneous network S.(2)Reconstruct similar heterogeneous networks.Apply the Katz model on the similarity heterogeneous network S to obtain the correlation prediction score for each pair of nodes,expressed in the form of a block matrix.Then use LOOCV to adjust the model parameter k,when the Kmd that in the similarity prediction score matrix K AUC value reaches its optimal value,extract the miRNA calculated the similarity prediction score matrix Kmm,and diseases calculation similarity prediction score matrix Kdd.Finally,replace ISm with Kmm and ISd with Kdd in heterogeneous network S to obtain the reconstructed heterogeneous network G.(3)Perform multidimensional scaling embedding(MDS)on heterogeneous network G.Firstly,obtain the inner product matrix B of the multidimensional scaled objective matrix Z,and perform eigenvalue decomposition on B to obtain the objective matrix Z.Then establish a linear mapping from the miRNA space/disease space to the miRNAdisease feature space,and obtain vector representations of miRNA and diseases.and then convert the correlation prediction between miRNA and disease into similarity calculation between vectors.Use the sigmoid function to obtain the feature based similarity score,and obtain the correlation prediction score of miRNA-disease.(4)Perform result analysis on the correlation prediction of miRNA-disease.We draw a heat map of nine diseases and all miRNA association scores,intuitively and comprehensively display the prediction results,and analyze the miRNAs related to breast cancer and lymphatic cancer.Next,we compare K-MDS with classic methods Katz,RKNNMD A,and WBSMD A,proving that our method has indeed improved performance in miRNA disease prediction.