The Research Of MiRNA-disease Associations Prediction Algorithm Based On Deep Network Fusion

miRNA is an important post-transcriptional regulator and play a crucial role in the occurrence and development of diseases.Discovering disease-related miRNAs is a fundamental and critical biomedical task.Since it is time-consuming and laborious to identify miRNA-disease associations through traditional biological experiments,computational methods have become an indispensable auxiliary method.Most of the existing computational methods are affected by the sparsity and noise of biological data result in their poor prediction performance.With the prevalence of high-throughput experimental technology,an abundance of biological networks has emerged dramatically.How to efficiently integrate multiple network information and make full use of the information complementarity between biological networks to further improve the prediction performance has also become one of the challenges of the computational method.From the perspective of multi-network fusion,this paper proposes two network fusion methods based on deep learning for disease related miRNAs identification.The main work is summarized as follows:(1)Considering that most existing computational methods are difficult to effectively integrate multi-omics information and capture nonlinear network structure information,a framework deep MD based on multi-modal autoencoders is proposed.The framework utilizes a multi-modal autoencoder to fuse the structural information of multiple miRNA similarity networks and disease similarity networks respectively,and exploits graph convolutional neural networks to further extract high-quality node features to infer potential miRNA disease associations.In the five-fold cross-validation experiment,the deep MD algorithm outperforms comparison algorithm in multiple metrics.In addition,in the experiment of predicting associations of novel diseases and novel miRNAs,deep MD also achieves superior performance,which demonstrate the practicality of the model.(2)As the deep MD algorithm tends to ignore high sparse networks and cannot integrate heterogeneous network information,a miRNA-disease association prediction model MNGNN based on graph neural network is developed.This model exploits the multi-graph convolutional encoder and the bipartite graph convolutional encoder to extract the information of both the homogeneous network and the heterogeneous network respectively.In addition,the model also uses a discriminator to capture the dependence between global features and local features.We performed a five-fold cross-validation experiment and compared our model with state-of-art methods for verifying the performance of MNGNN.The experimental results suggest that MNGNN is significantly better than the comparison algorithm in most of metrics.At the same time,ablation experiments and case study respectively verified the effectiveness of each sub-module of MNGNN and the ability to predict new potential associations of the model.