Graph Neural Network Models For Inferring Gene Regulation From Single-Cell Transcriptome Data

Gene regulatory networks(GRNs)are the regulatory interaction structures that are formed by the interactions between genes and genes within a cell or within a particular genome,especially those based on gene regulation.GRNs play an important role in the regulation of intracellular expression and can be used to depict and understand the dynamic process of cell differentiation.However,discovering tissue-or cell typespecific GRNs is a key challenge.The development of single-cell RNA sequencing(sc RNA-seq)technology provides an opportunity to solve this issue.Compared with bulk RAN-seq technology,sc RNA-seq data can reveal the heterogeneity and dynamic differentiation process of cells.Meanwhile,inferring GRNs from sc RNA-seq data remains challenging due to its high dimensional and noisy characteristics.In recent years,dozens of computational algorithms have been developed to infer GRNs from sc RNA-seq data.Among them,the unsupervised learning strategy that constructing the correlation-based co-expression networks of TFs and their target genes has been widely adopted to infer the GRNs.But their performance on single-cell data from real organisms is limited.In addition,deep learning models based on supervised learning,such as the utilization of convolutional neural networks and variational autoencoders to explore nonlinear regulatory laws,have made significant progress.However,these methods often only focus on the regulatory relationship between gene pairs,fail to leverage the intrinsic global regulatory structure of a GRN that is crucial to explore the regulatory modes between TFs and their target genes in the complex biological systems.Therefore,this dissertation is aimed to infer cell-type-specific GRNs from sc RNAseq data,focusing on the regulatory structure among multiple genes,causality,explicit association modeling and architecture adaptive.The specific work of this dissertation is as follows:1.To reduce the influence of technical noise in single cell data and to explore the regulatory structure of genes,this dissertation developed a graph-based deep learning model,Deep RIG,to reconstruct the GRNs from single-cell transcriptomics.To learn the global regulatory structure,Deep RIG takes the trick of calculating the correlation coefficients for each pair of genes from gene expression profiles of sc RNA-seq data,and builds a prior regulatory graph based on the co-expression mode which is robust to noise.Then it utilizes a graph autoencoder model to embed the global regulatory information contained in the graph into gene latent embeddings and to reconstruct the gene regulatory network.Extensive benchmarking results demonstrate that Deep RIG can accurately reconstruct the GRNs and outperform existing methods on multiple simulated networks and real-cell regulatory networks.Additionally,this dissertation applied Deep RIG to the samples of human peripheral blood mononuclear cells and triple-negative breast cancer,and presented that Deep RIG can provide accurate celltype-specific GRNs inference and identify novel regulators of progression and inhibition.2.To enhance the model’s ability of inferring causal relationships between genes,this dissertation introduces GRACE,a novel graph-based causal autoencoder framework that combines a structural causal model with graph neural networks to enable GRN inference and gene causal reasoning from sc RNA-seq data.By explicitly modeling causal relationships between genes,GRACE facilitates the learning of regulatory context and gene embeddings.With the learned gene signals,this model successfully decodes the causal structures and alleviates the accurate determination of multiple key attributes of gene regulation that are important to determine the regulatory levels.Through extensive evaluations on seven popular benchmarks,this dissertation demonstrates that GRACE outperforms 11 state-of-the-art GRN inference methods,with the incorporation of causal mechanisms significantly enhancing the accuracy of GRN and gene causality inference.Furthermore,the application to human peripheral blood mononuclear cell samples reveals cell type-specific regulators in monocyte phagocytosis and immune regulation,validated through network analysis and functional enrichment analysis.3.Given the diverse distribution of sc RNA-seq data,this dissertation proposes Auto GERN,a GNN framework specifically designed for inferring GRNs from sc RNAseq gene expression data.To leverage and organize the complex edge information in the prior regulatory graph,Auto GERN explicitly models and learns the edge representations within the message passing space.The learned edge embeddings are further utilized to infer regulatory interactions between pairs of genes through a multilayer perceptron.To enhance its power and flexibility,this dissertation simultaneously considers two different spaces designed for message passing,intra-layer and inter-layer spaces.Furthermore,this dissertation employs a robust search algorithm from existing Auto GNNs to automatically adapt the GNN architecture to different sc RNA-seq datasets.Extensive experiments on seven real-cell sc RNA-seq datasets show that Auto GERN consistently yields good performances on seven benchmarks and outperforms state-of-the-art algorithms for GRNs inference.Additionally,this model exhibites remarkable effectiveness and robustness on real sc RNA-seq datasets with unbalanced samples,as demonstrated through its application to multiple unbalanced scenarios of negative samples.These results highlight the significant accuracy enhancement achieved through explicit edge modeling within GNNs for GRN inference.