Research On MiRNA Identification And Disease Related MiRNA Prediction Method Based On Deep Learning

MicroRNAs(miRNAs)are a type of endogenous non-coding RNA with a length of about 22 Nucleo Tides(NTs).Mi RNAs occupy an important position in many biological and physiological processes such as cell proliferation,differentiation,cell aging,and death,and are closely related to the occurrence and development of various human diseases.Accurately identifying miRNAs and studying the relationship between miRNAs and diseases can help to explore the pathogenic mechanism of diseases from the level of post-transcription and provide new ideas for disease prevention and treatment.In this thesis,miRNA-related data is used as the object of research.This thesis proposes corresponding solutions to the current problems in miRNA research,such as insufficient miRNA characterization,single information utilization,imbalanced categories and excessive dependence on node similarity information.The main work of this thesis are as follows:(1)In the current research on miRNA recognition,feature extraction depends on domain knowledge,lack of information fusion,and imbalance between positive and negative examples,which leads to low sensitivity of miRNA recognition.This thesis proposes a method for miRNA identification based on improved convolutional neural network(ICNN-MI).Based on the idea of deep learning,the primary sequence and secondary structure of miRNA are jointly;This thesis improves the traditional convolutional neural network structure to adapt to the input of miRNA,introduces the inception module to increase the adaptability to different scales,and replaces the fully connected layer with global pooling to reduce the number of parameters;Finally,a weighted loss function is used to solve the class imbalance problem.Experiments show that the results on the three benchmark datasets are superior to traditional machine learning methods.(2)To address the problem of inadequate use of information,excessive dependence on similarity information of nodes in the network and lack of negative samples in miRNA-disease association studies.This thesis proposes a Mi RNA-Disease Association prediction method based on Network Representation Learning(NRLMDA).This method constructs a miRNA-lnc RNA-disease heterogeneous network by introducing long-chain non-coding RNA(lnc RNA)and its association with diseases and miRNAs,enriching the biological information of the original network;The network representation learning node2 vec algorithm is used in the above-mentioned heterogeneous network to obtain the neighbor sequence of the node with a certain walking strategy,and deep learning is performed through the skip-gram model to obtain the low-dimensional feature vectors of the node;Finally,the association rule inference method based on miRNA-miRNA similarity predicts the association between miRNA and disease.This method can mine the topological structure characteristics of the global network,and does not need negative samples.NRLMDA’s experimental results on leave-one-out cross-validation,five-fold cross-validation and further case studies are superior to classical methods.