Research On Methods For Gene Regulatory Networks Construction Based On Multi-Omics Data

Gene regulatory networks are a powerful abstraction of biological systems,which control the activities of life by regulating the expression level of genes in cells.As the core of biological processes,gene regulatory networks control almost all cell activities and functions of organisms,and play a vital role in the biological life process.The regulation mechanism between genes is very complex,and it is difficult to explore the regulation rules by traditional wet-lab experiment.The construction of high quality gene regulatory networks has been a hot topic in systems biology.With the rapid development of computing technology and artificial intelligence theory,it has been greatly promoted the research of gene regulatory networks calculation methods.In recent decades,many approaches have been proposed to construct gene regulatory networks.Mining the regulation rules of genes can help us understand the growth and differentiation rules of organisms and understand the complex regulatory relationships among biomolecules.The accurate construction of gene regulatory networks is of great significance for revealing how transcription factors regulate the expression of target genes,control their own metabolic rates,and adapt to environmental changes.In the era of big data,a large number of omics data such as genomics,transcriptomics,proteomics and metabolomics have emerged,providing a solid data basis for the inference of gene regulatory networks.At present,there are many methods to construct gene regulatory networks based on gene expression data,however,a single type of biological data may not provide enough information to infer gene regulatory relationships accurately and efficiently.Different types of biological data can provide complementary information for the construction of gene regulatory networks,and the effective use of these multiomics data is the key to the construction of high-quality gene regulatory networks.In this thesis,the complex problem of gene regulation networks construction was transformed into a simple and easily solved binary classification problem.On this basis,the construction method of gene regulation network was studied based on multi-omics data.In summary,the main contents of the thesis are displayed as follows:(1)Numerous methods infer gene regulatory networks based on a single type biological gene expression data and a single model,ignore the information contribution of other multi-source data and the inherent bias of the model,and insufficient use of data exists,which leads to the problem of low accuracy and poor robustness.To solve the above problem,we proposed an integrative neural networks model,named Int NNs.Int NNs makes full use of gene expression data,protein sequence data,promoter sequence data and coding sequence data to extract different data features and feed them into multiple basic neural networks for training.Next,the predictions of these basic networks are integrated into an ensemble subnetwork to fuse different data features and complementary information of these basic networks to achieve more robust prediction of gene transcriptional regulatory network.Experimental results on real maize and human datasets show that the integration of multi-source data and multiple basic networks can further improve the accuracy for gene regulatory network construction.(2)Because of gene regulatory networks are dynamic,static gene expression data cannot truly reflect the dynamic information of the regulatory networks.In addition,considering the individual differences between different data sources,information fusion of multi-source data directly may override a unique and complementary genomic view of regulatory mechanism provided by each omics data,and most of the methods exist the problem of insufficient modeling of global data characteristics and individual differences in the use of multi-source data.To solve the above problem,we propose another new method,TDINNs,to inferring gene regulation Networks based on temporal dynamic integrated neural networks.TDINNs uses dynamic information from gene time series expression data,combine with other omics data such as protein sequence data to mine gene regulatory information,as well as use neural network to integrate individual information from different data sources and overall information from multi-source data to predict gene regulatory network.Experimental results on real maize biological datasets show that TDINNs performed better than other comparison methods in inferring gene regulatory networks.