%0 Generic
%A Max, Frank
%C Heidelberg
%D 2024
%F heidok:34849
%R 10.11588/heidok.00034849
%T Modeling epigenetic heterogeneity across time and genome in single-cell multi-omics experiments
%U https://archiv.ub.uni-heidelberg.de/volltextserver/34849/
%X The genomic sequence of an organism is nearly identical in all its cells and over its lifetime. Epigenomic marks, however, such as DNA methylation and chromatin accessibility, are subject to drastic changes across different tissues and throughout organism development. Recent advancements, notably the development of multi-omics single-cell technologies, allow for simultaneous interrogation of DNA methylation, chromatin accessibility, and transcriptomes within individual cells. This offers unique opportunities to gain insight into mechanisms by which the epigenome shapes gene expression and influences cell fate. However, analyzing these datasets poses major challenges: Typically, smaller numbers of cells can be assayed per experiment than conventional single-cell RNAseq with lower coverage due to small amounts of input material. This means that classical statistical methods are underpowered to detect subtle changes in DNA methylation and chromatin accessibility. Furthermore, current tests can only detect differences between discrete and pre-defined cell populations, whereas single-cell approaches allow for studying continuous processes in organismal lineage development.    To address this, I propose computational methods for decomposing single-cell epigenetic heterogeneity across developmental time and genomic loci. This thesis introduces new concepts, leveraging pseudotemporal ordering of cells to conduct statistical inferences upon epigenetic changes. At the core of these developments is GPmeth, a Gaussian process framework designed to model highly sparse single-cell methylation and accessibility information by enforcing smooth variation across pseudotime and genomic coordinates and thus effectively sharing information between cells and genomic positions. Importantly, this model does not rely on averaging methylation signals across fixed genomic windows but can identify differentially methylated/accessible regions in a data-driven way. Testing GPmeth against other models without dynamic aggregation of methylation data revealed increased sensitivity to detect even subtle epigenetic changes.    Application of GPmeth to scNMT-seq data from mouse embryonic stem cells undergoing gastrulation revealed over 3000 enhancer elements that exhibited dynamic changes  in chromatin accessibility or DNA methylation rates during germ layer formation. The detailed spatiotemporal model allowed for a precise definition of differentially methylated regions, validated by transcription factor binding motif analysis. Furthermore, the clustering of temporal epigenetic patterns identified lineage-specific enhancers in an unsupervised manner.    I expect GPmeth to be a valuable tool for studying time-resolved epigenetic regulation in several emerging multimodal single-cell datasets.