Research On MiRNA Target Prediction And Functional Identification Algorithms

microRNAs(miRNAs)are a class of 19~22nt non-coding RNAs that play fundamental roles in multiple biological processes,including cell differentiation,proliferation and apoptosis as well as disease process.The application of computational methods plays an important role in miRNA target prediction and miRNA function identification,promoting the development of relevant research.In this thesis,according to the biological characteristics of miRNA,the related issues such as miRNA target prediction,miRNA regulatory module identification,miRNA-disease association prediction and biomarker mi RNAs prediction are further investigated.The creative work mainly consists of the following four parts:(1)A novel prediction method based on convolutional neural network(CNN)is proposed for predicting miRNA target genesmiRNAs affect the protein biosynthesis and further the occurrence of disease by regulating gene expression.Thus,the general premise of investigating miRNA-disease association is aware of miRNA target genes.In this study,we proposed miRTDL based on CNN according to miRNA target mechanism.In the respect of target feature selection,the complementary features,accessible features and conservative features of miRNA-gene interaction are explored based on the miRNA secondary structure.As a result,20 features are selected for target prediction.In the respect of dataset construction,according to the targeting rules obtained from biological experiments,constraint relaxation method is applied to constructing balanced training data,smoothing out the impact of unbalanced training set on prediction.In the respect of selecting machine learning method,due to the incomprehension about exactly target mechanism,the CNN is selected based on that CNN model extracts features by itself and thus can not be effected by input data.This step avoids the impact of expert domain knowledge on prediction,and improves the prediction accuracy.The miRTDL is applied to 1606 human miRNA-gene pairs and achieves high accuracy.The contributions of target feature selection,balanced dataset construction and machine learning model selection to the final prediction are validated.Compared with the existing methods,miRTDL achieves higher prediction performance and the significance of targeting features is analyzed based on the prediction results.The conclusion is got that the complementary feature is the most significant feature in miRNA-gene interaction,and conservative feature is more important than accessible feature,which can be used to guide biological experiments.(2)A method based on topic model is proposed for identifying miRNA regulatory moduleThe researchers comprehend miRNA-miRNA collaborative regulatory mechanism and miRNA-gene target mechanism by identifying miRNA regulatory module,and further grasp the complex process of participation in diseases.In this study,CCRM model is proposed to identify miRNA regulatory module based on the spatio temporal feature of miRNA function and miRNA and gene expression profile is used as training data.According to the biological premise that the miRNAs with similar expressions always have similar function,miRNA regulatory module is identified.The function of miRNAs and mi RNA-gene target relationship is inferred by the known miRNAs and genes within the same regulatory module.Due to the various functions,miRNAs and their target genes appear in modules repeatedly,and the associations are built based on these miRNAs and genes.Furthermore,the miRNA regulatory network is constructed which shows the dynamic evolution characteristics of miRNA function.11 experiments are performed,through which the miRNA-miRNA collaborative regulatory mechanism,miRNA-gene target mechanism and associations between modules are investigated.As a result,the reliability of CCRM model is validated.(3)A prediction method based on hidden conditional random field for predicting miRNA-disease associationThe abnormal function of miRNA is one of the most important reasons which lead to disease,therefore,the research on miRNA-disease association is significant for clinical diagnosis.In this study,hidden conditional random field is used to predict disease associated miRNAs.In the respect of sample selection,the biological knowledge used by the previous methods which based on network functional similarity network is insufficient,it is hard to obtain reliable training model based on small amount of input data.Besides that,the previous methods are established on static regulatory network,which can not reflect the miRNA dynamic function.Thus,miRNA expression profile is used here.In the respect of constructing training dataset,the network functional similarity is used to detect direct miRNA-disease associations and indirect mi RNA-disease associations.After that,decision fusion method is used to assign reliable class label to each mi RNA in expression profile.In the respect of selecting machine learning model,hidden conditional random field is chosen which can detect the meaningful sub-sequence of miRNA expression values.ROC curve is used to describe prediction results,comparing with previous methods,HCRF obtains the highest AUC value.Compared with HMM and CRF models,the results shows that HCRF outperforms the other two models obviously.Besides that,the performance of decision fusion method is investigated.The method proposed here are useful in providing reliable disease-associated miRNA candidates for biological validation experiments,and demonstrating the extensive application prospect of expression profile in miRNA-disease association identification.(4)An approach based on locally linear embedding and clustering for identifying miRNAs as biomarkerBiomarker miRNAs play an important role in early diagnosis of cancers.However,there are few research based on computational methods in identifying biomarker miRNAs at present.In this study,based on the spatio temporal feature of miRNA expression,locally linear embedding method is used to reduce the dimension which reserves the spatial topological structure of miRNAs.Based on the biological premise that miRNAs differentially expressed under various biological samples,and co-expressed miRNAs distribute in cluster,the density-based clustering method is applied to identify co-regulated mi RNAs.As a result,the biomarker miRNAs are predicted according to the clustering frequency and the miRNA expression pattern under various cancers are analyzed.Furthermore,according to the general rule of disease development,the common biomarkers of different cancers are detected.The reliability of identified biomarkers are validated by comparing with published literatures and database.The machine learning methods provide new tools for predicting biomarkers.