The three-dimensional (3D) organization of chromosomes in eukaryotic cells is key to understanding cellular functions such as DNA replication and gene transcription. 
Recent advances in (single-cell) high-throughput 3D genome mapping methods have ushered in a new era of unveiling spatial chromatin organization at an unprecedented resolution. However, there is a lack of algorithms that can effectively utilize these emerging data types and integrate them with other multimodal data. In this talk, I will introduce a series of new algorithms developed based on graph representation learning for the analysis of multiscale 3D genome organization. In particular, I will describe a generic hypergraph representation learning scheme called Hyper-SAGNN, which was one of the first methods that focus on hyperedge predictions for heterogeneous non-uniform hypergraphs. This framework was further extended to Higashi, an end-to-end machine learning solution for single-cell Hi-C analysis. I will demonstrate how Higashi, as the most systematic approach to date, reveals cell-to-cell variability of 3D genome features and the connections between 3D chromatin structure and cell-type specific gene regulation in complex brain tissues. Finally, I will highlight a more efficient and interpretable framework for the next generation single-cell 3D epigenome studies called Fast-Higashi based on scalable tensor decomposition. Together, these new methods have the potential to provide key insights into the structure and function connections of genome organization and will be of high value to a diverse group of biomedical researchers.