Research On The Construction Algorithm Of Gene Regulatory Network Based On Deep Neural Network

There is a complex set of gene regulatory mechanisms in organisms,which control the growth and development of organisms,researchers define this regulatory mechanism as gene regulatory networks(GRNs),gene regulatory networks are important tools for understanding biological systems.With the rapid development of single-cell sequencing technology,the scale of single-cell transcriptomic data is increasing,and accurate inference of gene regulatory networks from large-scale transcriptomic data is crucial for understanding biological regulatory processes.In recent years,researchers have proposed various algorithms to reconstruct gene regulatory networks.Existing algorithms usually decompose the regulatory network reconstruction problem into multiple sub-problems,and use machine learning methods to mine data.These methods to some extent still suffer from problems such as computational complexity,computational accuracy,and so on.Aiming to address these issues,the main work of this research is as follows:This thesis first presents BiRGRN,a novel method for inferring GRNs from gene expression data with pseudo-time series.BiRGRN utilizes a bidirectional Recurrent Neural Network(BiRNN)to infer gene regulatory networks,which is a complex deep neural network capable of capturing complex,nonlinear,and dynamic relationships between variables.The model maps neurons to genes,and maps connections between neural network layers to regulatory relationships between genes.Based on the deep network,we transform the reconstruction problem of GRNs into a regression problem.The algorithm uses the gene expression data of multiple previous time nodes to predict the gene expression data of the next time node.Furthermore,the algorithm utilizes a bidirectional structure to integrate forward and reverse inference results and exploits a set of incomplete prior knowledge to filter out some candidate edges.BiRGRN is biologically explainable and mathematically flexible.To verify the accuracy of the proposed algorithm,we apply BiRGRN to four simulated datasets and three real sc RNA-seq datasets.Compared with other state-of-the-art algorithms,the experimental results demonstrate that BiRGRN can accurately infer GRNs from time-series sc RNA-seq data.Furthermore,this paper proposes GraConGRN,which is a brand-new gene regulatory network reconstruction algorithm for steady-state gene expression data.GraConGRN utilizes Graph Convolutional Neural Networks(GCN)and Convolutional Neural Networks(CNN)to infer gene regulatory networks.The algorithm first uses GCN to extract the features of gene expression data and obtain the low-dimensional embedded expression of genes,which is used to reconstruct the gene-relation-gene triples from the obtained low-dimensional embedded expression.Then,we utilizes convolutional neural networks to predict the triples.GCN is a convolutional network structure specially developed for non-Euclidean graphs,which can effectively extract the interaction relationship between genes.Based on the encoder-decoder structure,the algorithm transforms the reconstruction problem of GRNs into a classification problem and determine whether the current triplet is a legal triplet by scoring the triplet.GraConGRN also has biological proximity and structural flexibility.To verify the accuracy of the proposed algorithm,we apply GraConGRN to nine simulated datasets and three real sc RNA-seq datasets.Compared with competing algorithms,the experimental results show that GraConGRN can accurately reconstruct GRN from steady-state gene expression data.