UC Irvine

The advancements in single-cell and spatial omics technologies have revolutionized our ability to dissect biological systems at unprecedented resolution, allowing us to explore the intricate internal mechanisms of cells, the fundamental units of life, and their complex spatial organization within tissues. However, the high-dimensional, heterogeneous, and often sparse nature of these datasets presents substantial computational challenges. This thesis addresses these challenges by developing a suite of innovative graph learning methodologies designed to systematically analyze single-cell and spatial omics data, spanning a comprehensive range of scales from subcellular genomic organization to multi-slice spatial tissue atlases.We first present our contributions to Genomic and Cellular Analysis. This section introduces scENCORE, a novel method leveraging single-cell epigenetic data and graph embedding to predict chromatin A/B compartment conformations, providing crucial insights into the genome's three-dimensional organization. Then, iHerd is presented as an integrative hierarchical graph representation learning framework, adept at quantifying network changes and prioritizing risk genes in disease contexts. We then introduce ExAD-GNN, an explainable graph neural network that utilizes single-cell data for robust disease state prediction and the identification of cell-type-specific biomarkers, deepening our comprehension of individual cellular complexities.Next, we detail our work on Microenvironmental and Intercellular Understanding within spatial omics. Here, Impeller is developed as a path-based heterogeneous graph learning method, specifically tailored for imputing missing gene expression values within individual cells from spatial transcriptomic data, thus enhancing data completeness. Building on this, iMiracle is introduced as an iterative multi-view graph neural network. This model precisely characterizes intercellular gene regulation from spatial transcriptomic data, elucidating the nuanced mechanisms of cell-to-cell interactions within their native contexts.Finally, our research extends to Large-Scale Spatial Omics Analysis. At the region-level, DISCO offers a diffusion model-based framework. By integrating graph neural networks, DISCO achieves large-scale spatial transcriptomics data completion, effectively addressing the challenges of macroscopic data sparsity. To enable comprehensive joint analysis of multi-slice spatial transcriptomic data and move towards holistic tissue representations, we present MUSE. This computational framework models each spatial slice as a graph and employs optimal transport for robust cross-slice alignment and the construction of integrated spatial atlases.Collectively, the methodologies developed in this thesis underscore the powerful capabilities of graph learning in overcoming the inherent complexities of single-cell and spatial omics data. They offer a cohesive computational paradigm for understanding biological systems across diverse scales, from microscopic genomic features to macroscopic tissue structures and functions, ultimately paving the way for the creation of comprehensive biological atlases.