Research On MiRNA-target Gene Association Prediction Algorithm Based On Network Embedding

As one of the most important small molecules in humans,MicroRNA regulates the expression of target genes in a variety of ways and is associated with many complex diseases.The development of next-generation sequencing technology and the accumulation of massive biomedical data have facilitated the understanding of the regulatory mechanism of miRNAs in biological networks and the pathogenesis of complex diseases.Bioinformatics algorithm has greatly accelerated the progress of the relationship prediction between miRNA and target genes.However,the existing methods need to be improved in the aspect of effectively using the known data to extract features.We take the miRNA and target gene as the research object,used the multi-source data of miRNA and target gene,and combined the calculation method of network representation learning and deep learning to design models that can predict potential miRNA-target gene interactions,and speed up biological experiments.process.The main work of our paper was as follows:(1)In view of the existing miRNA target gene prediction methods that do not make efficient use of the existing data,feature selection ignores the relationship between the data and so on.Based on the neighbor information in the network,we designed the miRNA-target gene association prediction algorithm Multi-MTI(Multi-Information MiRNA-Target Interaction).On the one hand,this method starts from the network association information,constructs a miRNA-target gene bipartite network fused with neighbor information,and uses the network representation learning method Deepwalk to learn the relationship characteristic between miRNAs and target genes.On the other hand,starting from the sequence characteristics of miRNAs and target genes,the Role2 vec method is used to learn the sequence characteristics of miRNA and target genes.Splice the learned data features and relationship features,and we used an attention network to predict the relationship between miRNAs and target genes.Through 5-fold cross-validation,Multi-MTI has achieved better results in multiple evaluation indicators compared with other comparison methods.The ablation experiment and parameter analysis show that every aspect of the Multi-MTI model has a better effect on the results.The case analysis of the prediction results of unknown samples,further demonstrates that Multi-MTI has the ability to predict potential miRNA-target gene associations.(2)To further explore the characteristics of the relationship between pairs of miRNAs and targets in the network,an algorithm MDCNN(Metapath-based Deep Convolutional Neural Network)is proposed.On the one hand,in the form of meta-path,the relationship of network node pairs is linked to improving the interpretability of the algorithm.On the other hand,the effectiveness of the learned features plays an important role for miRNA-target genes.The model is designed as an end-to-end framework,and the parameters can be updated during the training process to adaptively learn the best Feature representation.Then,through 5-fold cross-validation,MDCNN obtained higher values in each evaluation index compared with Multi-MTI and other comparison methods.It also confirmed that the model is indeed greatly improved compared to Multi-MTI,and the improvement of the model is very effective.The ablation experiment showed that the use of multi-source information enables the model to capture more node information.The parameter sensitivity analysis shows that the influence of parameters on the model is small and it has strong robustness.Finally,the case study of hsa-mir-26b-5p and CDKN1 A showed that MDCNN has the ability to predict new potential miRNA target gene associations.