Quantifying Cellular Heterogeneity in Cancer and the Microenvironment

Abstract

In spite of recent advances in therapy, cancer remains a leading cause of death worldwide. Therapy response is often unpredictable and relapse frequently occurs. In many cases, this therapy resistance is attributed to subsets of therapy resistant cancer cells and surrounding stromal cells that support a resistant phenotype. A better understanding of cellular heterogeneity in cancer is therefore crucial in order to develop novel therapeutic strategies and improve patient outcomes. Experimental technologies like mass cytometry (CyTOF) allow for high-content, multi-parametric single-cell analysis of human tumor samples. However, analytical tools and workflows are still needed to standardize and automate the process of identifying and quantitatively describing cell populations in the resulting data. This dissertation presents a novel workflow for automated discovery and characterization of novel and rare cell subsets, quantification of cellular heterogeneity, and characterization of cells based on population-specific feature enrichment. First, a modular workflow is described that combines biaxial gating, dimensionality reduction, clustering, and hierarchically clustered heatmaps to maximize rare population discovery and to create an interpretable visualization of cell population characteristics. Next, a novel method is introduced for quantifying cellular heterogeneity based on two-dimensional mapping of cells in phenotypic space using tSNE analysis. Finally, an algorithmic method termed Marker Enrichment Modeling (MEM) is introduced that automatically quantifies population-specific feature enrichment and generates descriptive labels for cell populations based on their feature enrichment scores. MEM analysis is shown to identify features important to cell identity across multiple datasets, and MEM labels are effectively used to compare populations of cells across tissue types, experiments, institutions, and platforms. Going forward, the tools presented here lay the groundwork for novel computational methods for machine learning of cell identity and registering cell populations across studies or clinical endpoints. Automated methods for identifying and describing cell populations will enable rapid discovery of biologically and clinically relevant cells and contribute to the development of novel diagnostic, prognostic, and therapeutic approaches to cancer and other diseases.