Research On Single-cell Gene Regulatory Network Inference Based On Graph Neural Network

Gene regulation is the core control system of life activities and plays an important role in understanding the mechanism of living organisms reacting to surrounding environments from genotypes to phenotypes.Gene regulatory network(GRN)abstracts the interactions between transcription factors(TFs)and target genes,where nodes are TFs or genes and links are their regulatory relationships.Increasing evidence has suggested that altered transcriptional regulation is a prominent mechanism of common human diseases.GRN provides abundant information on gene interactions,which contributes to demonstrating pathology,predicting clinical outcomes,and identifying drug targets.In the past few decades,high-throughput sequencing technology has revolutionized the biological field entirely.The opportunity to study transcriptomes in great detail using RNA-seq has powered many essential discoveries and is now a conventional method in biomedical research.However,previous RNA-seq data is typically presented in "bulk",which shows an average of gene expression patterns across thousands to millions of cells;this might obscure biologically relevant differences between various cell types.Since the first discovery of singlecell RNA sequencing(scRNA-seq)in 2009,it signified an approach to overcome the problem of bulk data.By isolating single cells,capturing their transcripts,and generating sequencing libraries in which the transcripts are mapped to individual cells,scRNA-seq allows the assessment of basic biological properties of cell populations and systems with unprecedented resolution and provides an opportunity to reconstruct cell-type-specific GRN.Although many methods have been proposed to infer GRNs,the available algorithms for reconstructing GRNs are commonly designed for bulk RNA-seq data.Few of them are applicable to analyze scRNA-seq data under the circumstance of low signal-to-noise ratio and dropout.Therefore,it is imperative to develop new methods to analyze scRNA-seq data and illustrate the cellular mechanisms of gene regulations.In this thesis,we present GENELink(GENE regulatory network via Link prediction)to infer potential interactions between TFs and target genes by using graph attention network(GAT)regarding GRN inference as a graph-based link prediction problem.Given a set of genes with some of their observed interactions,we apply GAT to project gene expression profiles with the knowledge-based interaction matrix to low-dimensional space.The goal of GENELink is to optimize this embedding space to learn high-quality gene representations serving for downstream similarity measurement or causal inference of pairwise genes.Compared to eight state-of-the-art GRN reconstruction methods,GENELink achieves better performance in terms of AUROC and AUPRC on seven benchmark scRNA-seq datasets with four types of groundtruth networks.We further apply GENELink on scRNA-seq datasets of human breast cancer lung metastasis and reveal regulatory heterogeneity of Notch and Wnt signaling pathways between primary tumor and lung metastasis.Moreover,the ontology enrichment results of unique lung metastasis GRN indicate that mitochondrial oxidative phosphorylation(OXPHOS)is functionally important during the seeding step of the cancer metastatic cascade,which is validated by oligomycin pharmacological assays.