Statistical Differential Analyses of Hi-C Contact Maps
- Author(s): Liu, Huiling
- Advisor(s): Ma, Wenxiu
- et al.
Recent advances in Hi-C techniques have allowed us to map genome-wide chromatin interactions and uncover higher-order chromatin structures, thereby shedding light on the principles of genome architecture and functions. However, statistical methods for detecting changes in chromatin organizations are still in the early stage. In this dissertation, we proposed two statistical methods, namely DiffGR and scHiCDiff, for differential analysis in Hi-C contact maps.
The first method DiffGR detects differentially interacting genomic regions at the scale of topologically-associating domains (TADs) between two Hi-C contact maps. Specifically, we utilized the stratum-adjusted correlation coefficient (SCC) to measure similarity of local TAD regions. We then developed a non-parametric approach to identify statistically significant changes of genomic interacting regions. Through simulation studies, we demonstrated that DiffGR can robustly and effectively discover differential genomic regions under various conditions. Furthermore, we successfully revealed cell type-specific changes in genomic interacting regions using real Hi-C datasets.
The second method scHiCDiff focuses on detecting differential chromatin interactions (DCIs) in single-cell Hi-C data. The three-dimensional genome organization constructed from the conventional bulk Hi-C protocol represents an ensemble based on thousands to millions of nuclei, but not the actual genome organizations in individual cells. Unlike bulk Hi-C, single-cell Hi-C enables the exploration cell-specific chromosomal structures. However, interpretation and analysis of single-cell Hi-C data is at very early stage. To characterize the significant changes between different cells at the single-cell level, we built scHiCDiff which applied non-parametric tests and parametric models in distinguishing DCIs from single-cell Hi-C data. Our evaluation proved that these methods, especially the zero-inflated negative binomial (ZINB) and negative binomial hurdle(NBH) models, can effectively detect reliable and consistent DCIs of single cells between different conditions, which better capture cell type-specific variations of chromosomal structures.