Research On MiRNA Target Gene Prediction Algorithms Based On Heterogeneous Network Embedding

MiRNAs are a kind of vital regulatory factors in humans.Accumulated studies have found that miRNAs are in charge of many complex diseases by modulating gene expression.Predicting miRNA-target interactions is beneficial for uncovering the crucial roles of miRNAs in regulating target genes and the progression of diseases,which is of great significance for exploring the pathogenesis of complex diseases and researching precision medical treatment.The rapid development of high-throughput sequencing technology as well as the emergence of large-scale genomic and biological data provides new opportunities for miRNA target identification,however,the effective mining of valuable information in these data still faces many challenges.Compared with conventional experimental methods,bioinformatics methods have become an indispensable complementary tool with the advantages of low cost and high efficiency.In this paper,the investigation of models for miRNA-target gene association prediction is studied based on biological network and deep lea rning techniques.The main work of our paper is outlined as follows:(1)Aiming at the problem of over-reliance on feature selection and underutilization of existed or validated data in previous miRNA-target gene association prediction methods,we designed an end-to-end miRNA target gene prediction model named MRMTI(Multi-Relation Mi RNA-Target Interaction)by considering both network structure information and sequence information.The MRMTI model first used multi-relational graph convolutional network to model the complex relationships between the nodes in heterogeneous network and learn high-quality representations for the nodes.Meanwhile,considering that sequence information often reflects the most essential features of biomolecules,the bidirectional long-short term memory network was introduced to mine the deeper features of gene sequences which were further fused with network topology embeddings to calculate the miRNA-target gene association scores.The experimental results showed that MRMTI model outperformed the comparison models under all evaluation criteria and had the ability to discover new miRNA-target gene associations.(2)In order to further alleviate the inherent sparsity problem of biological data and improve prediction accuracy of the model,we proposed a multi-feature fusion-based miRNA target gene prediction model named MSMTI(Multi-Source Mi RNA-Target Interaction).On the one hand,by introducing small molecule drug-related data,the model constructed a triple-layer heterogeneous network which contained rich information.Then,in order to fully explore the global and local structural information of the network and the potential structural relationships between nodes,the network embedding algorithm node2 vec was used to learn the pairwise topological features for miRNA-target gene node pairs.On the other hand,Doc2 vec algorithm was utilized to learn the sequential features of miRNAs.Meanwhile,the extraction of gene sequential features was further optimized using the BERT pre-trained language model in view of its significant advantages in text sequence feature extraction.Finally,the obtained features were fused and then combined with an attentional neural network for prediction.The well prediction performance of MSMTI and its ability to infer potential miRNA-target gene associations were validated by the comparisons with other prediction models and case studies.