Methods For Interrogating Global and Single-Cell 3D Structures

The three-dimensional (3D) organization of the human genome plays a fundamental role in gene regulation, DNA replication, and cellular differentiation. Alterations in chromatin architecture, including chromatin loops and topologically associating domain (TAD) boundaries, are associated with cancer and developmental disorders. Chromatin conformation capture technologies such as Hi-C have enabled genome-wide characterization of these structures but require large numbers of cells and deep sequencing, limiting their ability to investigate cellular heterogeneity. To address this limitation, this dissertation presents scHiCcompare, a computational framework for differential analysis of single-cell Hi-C data. The method addresses extreme data sparsity through global LOESS normalization, Random Forest-based imputation with a pooling strategy, and Gaussian Mixture Model clustering to identify differential chromatin interactions between biological conditions. The framework demonstrates robust performance across simulated and experimental datasets. At a larger structural scale, this dissertation introduces EpiTADformer, a Transformer-based deep learning model for high-resolution TAD boundary prediction. Rather than relying solely on Hi-C contact maps, EpiTADformer leverages epigenomic signals from neighboring genomic regions to accurately identify TAD boundaries, outperforming existing computational approaches. Together, scHiCcompare and EpiTADformer provide complementary computational methods for investigating genome organization at the levels of chromatin interactions and domain architecture. Both methods are implemented as publicly available R and Python software packages with comprehensive documentation, enabling reproducible analyses. Their strong performance across diverse datasets and meaningful associations with biological and epigenetic features provide powerful tools for advancing the study of chromatin architecture, gene regulation, and disease mechanisms.