Repository logo

Gene Regulatory Networks at Single-Cell Resolution: an approach to exploring the impact of genomic regulation on cellular heterogeneity



Change log


Xu, Zhihan 


The computational analysis of single-cell RNA sequencing data provides a great opportunity to infer gene regulatory associations in different tissues, cell types and cells. However, there are many challenges still to be overcome. In this thesis, I discuss why this opportunity is fascinating but challenging from a computational point of view, and I build a computational method to demonstrate the plausibility of inferring one gene regulatory network (GRN) for each single cell, also evaluations are conducted from multiple perspectives.

In Chapter 2, I investigate data-fitting models in existing computational methods which infer GRNs. To fully take advantage of the single-cell resolution in the input sequencing data, leads to the use of machine learning approaches, such as instance-wise feature selection models. I assess the eligibility at a methodological level, implement an instance-wise feature selection model, and aim to generate GRNs which learn cell-to-cell variations from single-cell sequencing. Based on the implementation above, I build a new method to infer GRN at single-cell resolution or single-cell specific GRN (scGRN). The inferred scGRN can be used to explore GRNs across cell types. Since the true underlying biological mechanism in scGRN cannot be known, my method is benchmarked based on available biological network databases which do not reveal cell-to-cell variation on GRN.

Analysis of scRNA-seq data can provide biological insights into cellular heterogeneity because of the single-cell resolution; since scGRNs also contain cell-resolution information, it implies that scGRN can also be analysed about its corresponding cell-related properties. Since scGRN is generated from scRNA-seq data, the inherent cell-related patterns in scGRN shall not deviate substantially from the results in the existing analysis for scRNA-seq. A resulting scGRN suggesting very different cellular information may harm its reliability even before conducting GRN validation, so cellular information provides a powerful angle to evaluate scGRN. In Chapter 3, I build a pipeline to analyse scGRNs and I refer to three cell-related properties for comparison - cell types, cell-type trajectories and cell-type specific marker genes. The results demonstrate the difference in the implication of derived cell-related properties from scGRN and scRNA-seq.

As this thesis’s ultimate goal, the analysis for resulting scGRN endows an unprecedented opportunity to explore changes in regulatory patterns along cell types or tissues. Thanks to the single-cell resolution in biologically interpretable scGRN, it is flexible to conduct various analyses such as implications from cell type specific GRNs. Chapter 4 focuses on the exploration of interactions between regulatory edges and cell subpopulations from scGRN. In other words, scGRNs are analysed to explore the changes in GRN edges across cell types, cellular lineages or organs. Based on the existing biological understanding, some expected patterns are summarised to evaluate scGRN. Unlike a quantitative score indicating the performance of a method, the evaluation of patterns is still sufficient to differentiate unreasonable scGRN candidates, and this work provides a novel perspective to evaluate GRNs. Besides the utility for evaluation, the observed patterns also suggest some meaningful biological implications about the impact of genomic regulation on cellular differentiation.

I hope the methodology developed in this thesis is helpful to inspire more method developers to pursue scGRN, and I hope multi-angle evaluations of scGRN demonstrated here can facilitate more biological insights into the regulatory mechanisms driving cellular heterogeneity.





Teichmann, Sarah


deep learning method, gene regulatory network, single-cell RNA sequence


Doctor of Philosophy (PhD)

Awarding Institution

University of Cambridge