Single-cell Methods and Spatial Analysis for Highly Multiplexed Tissue Images

Abstract

A transformation is currently underway in the study of tissue biology from one in which deep characterization of cell types and states is possible only following the dissociation of tissues into single cells to one in which spatial relationships can be preserved. Recent breakthroughs in highly multiplexed imaging enable the measurement of tens to hundreds of bio-molecules at single-cell resolution in tissues of human biopsies. These technologies extract molecular information about single cells in situ, thereby preserving their spatial characteristics within a tissue context. The data they produce are essential for understanding the organization of multi-cellular populations in health and diseases at multiple spatial scales ranging from single cells to groups of hundreds of thousands of cells. It is well established that tissue structure influences the underlying physiology of organisms and thus will assist in deciphering spatial signatures of disease. However, the high dimensional nature of single-cell, spatially resolved data poses difficulties for computational and statistical analyses. In this thesis I begin by providing a mathematical background of spatial point patterns and lattice data in the context of spatial biology. Embedded throughout is a survey of spatial statistical approaches to modeling the spatial dependence and heterogeneity of single cells in tissue. I follow this introduction with contributions to three key challenges inherent in the analysis of multiplexed data. First, I propose a robust segmentation framework to automatically extract single cells from gigapixel images in a manner that generalizes to diverse tissue types. Second, I present a novel approach for identifying cell phenotypes that capitalizes on protein spatial distributions and cellular morphology. I show that unsupervised deep clustering has the ability to extract cell states as well as neighborhoods of complex tissue motifs, with the potential to uncover rare, and previously unrecognized, cell phenotypes. Furthermore, I show that morphological information identifies molecular features of distinct histological classes of tumors. Third, I develop a statistical approach for performing differential spatial analysis between groups of cells in the same or different tissue that discriminates the most salient spatial features. In addition, I introduce a set of in-silico tissue generation models that are specified from a desired functional form of correlation structure. Realizations of these synthetic tissues are used to assess cell-cell interactions and overcome limitations in gene expression studies when there is no access to spatial information. Together these contributions will assist biologists and clinicians in understanding the key mechanisms and drivers of normal tissue structure and development as well as circumstances in which these go awry in human disease.