Research And Application Of MicroRNA Precursors Recognition Based On Multiple Kernel Learning

MicroRNAs are a class of single-stranded non-coding RNAs that are about 20-24 nucleotides.Experiments have shown that microRNAs can regulate about fifty percent of the protein coding process,and are involved in a variety of physiological processes such as growth and development of organisms,lipid metabolism and hormone secretion,as well as a variety of pathological processes such as leukemia and cancer.How to efficiently identify new microRNAs is a hot topic in current research.MicroRNA precursors(pre-microRNA)are a necessary stage for the formation of microRNAs.Therefore,the identification of new microRNAs can be started from the identification of new pre-microRNAs.However,existing pre-microRNA recognition models still have some problems,such as low recognition efficiency,single identified species,and poor representation of features.To address the above problems,the main research contents of this paper are as follows:(1)This paper summarized the research on microRNA recognition at home and abroad in recent years,and proposed new pre-microRNA features.On the basis of summarizing the 29pre-microRNA features proposed by the better mi Pred algorithm and the 32-dimensional triplet features proposed by triplet-SVM,this paper proposes 5 new features related to pre-microRNA sequences.Combined with the 11-dimensional triplet features selected by F-score,45 premicroRNA features were formed,which increased the richness of the existing pre-microRNA features.In terms of datasets,this paper used pre-microRNA datasets from multiple species,including 41 species of Arabidopsis thaliana,mosses,humans and Drosophila melanogaster,including plants,animals and viruses.(2)This paper proposes a cross-species pre-microRNA recognition model based on multiple kernel learning,and applies the localized multiple kernel learning model to the premicroRNA recognition for the first time,which improves the accuracy of the existing crossspecies pre-microRNA recognition methods and solves the problem of the low recognition rate of the existing cross-species pre-microRNA.Traditional multiple kernel learning methods have a single mapping method and cannot fully represent the characteristics of pre-microRNA datasets.Each kernel function in the multiple kernel learning method has its own characteristics,and the simple linear combination of them can make up for the shortcomings of the single kernel learning method.At the same time,a selection model is used to dynamically adjust the weight of each kernel function according to the input samples,and the weighted linear combination of the kernel functions is used to solve the problem of same weight of each kernel function in the existing multiple kernel learning methods.The experiments show that the localized multiple kernel learning model significantly improves the recognition rate of pre-microRNA compared with the existing pre-microRNA recognition methods.(3)In this paper,a pre-microRNA recognition model based on deep localized multiple kernel learning was proposed,and the idea of deep learning was applied to the classification of pre-microRNA,making full use of the shallow and deep characteristics of pre-microRNA.The existing multiple kernel learning model only optimizes the weight selection of the kernel function,but is still limited to the shallow kernel model,which cannot well map the deep features of the sample.In order to fully map the deep characteristics of pre-microRNA datasets,a deep synthetic kernel model was firstly studied.After the model was trained on the premicroRNA datasets,the deep synthetic kernel was preserved,and then the deep synthetic kernel was combined with other shallow base kernel functions through the localized multiple kernel learning model to form a deep localized multiple kernel learning model.In this way,after the deep localized multiple kernel learning model is trained on the same pre-microRNA datasets,the model can map the shallow and deep features of the pre-microRNA datasets and achieve better generalization effect.The experiments show that the proposed deep localized multiple kernel learning model can significantly improve the sensitivity,specificity and accuracy of cross-species pre-microRNA recognition model,and has a good comprehensive performance.