Abstract
Cells within a single tumor are known to display extensive phenotypic and functional heterogeneity. Many life-threatening features of cancer, including drug resistance, metastasis and relapse, are facets of intratumor heterogeneity. With emerging single-cell measurement technologies, the field is poised to make important strides in understanding and controlling this heterogeneity. However, these technologies require coordinated advances in analytical methods to interpret the complex data they produce.
Acute myeloid leukemia (AML) is an aggressive bone marrow malignancy in which the importance of cellular heterogeneity has been well characterized. However, previous studies have only scraped the surface of the heterogeneity in this disease. Using mass cytometry, which measures single cells in ~40 simultaneous proteomic features, we developed novel methods for analyzing phenotypic heterogeneity in cancer. Our approach provides an extensive compendium of surface-marker and signaling phenotypes in AML that extends current boundaries of knowledge.
The heart of our approach is a graph-based representation of the single-cell samples. In this representation, each cell is modeled by a node connected to its neighbors—the cells most phenotypically similar to it. Constructed by local rules connecting cells, the graph as a whole represents the phenotypic structure of the sample. The graph can be partitioned into subsets of densely interconnected nodes, called communities, which represent distinct phenotypic subpopulations. Unlike parametric methods such as mixture models, this method makes no assumption about the size, distribution, or number of subpopulations.
Using our graph-based approach, we deconstructed several AML samples into discrete phenotypes. Comparing phenotypes across patients, we found a striking degree of order. Every phenotype identifiable phenotype was discoverable in multiple (but not all) patients, implying a constraint on the space of allowable AML phenotypes. For each phenotype we also identified cognate healthy cell types at different stages of bone marrow maturation, indicating a constraint that is linked to normal developmental programs.
Our data contain measurements of under various environmental perturbations and we designed a method to statistically quantify evoked signaling responses, producing high-dimensional signaling phenotypes for each subpopulation, which we regard as a representation of cellular functional potential. We found a tight coupling between surface and signaling phenotypes in healthy cells that is disrupted in AML. We identified a primitive signaling phenotype, derived from healthy stem and progenitor cells, which was not correlated with the primitive surface marker profile typically used to define primitive cells in AML. Using single-cell frequencies to deconvolve existing bulk gene expression data, we identified genes associated with this primitive signaling phenotype. These genes are enriched for primitive hematopoietic annotations and produce a clinically predictive signature that is more powerful than genes associated with the primitive surface profile, validating the utility of our approach and indicating novel regulators of the primitive hematopoietic cell state.
Citation Format: Jacob H. Levine, Erin F. Simonds, Sean C. Bendall, Garry P. Nolan, Dana Pe'er. Computational dissection of phenotypic and functional heterogeneity in acute myeloid leukemia. [abstract]. In: Proceedings of the AACR Special Conference on Hematologic Malignancies: Translating Discoveries to Novel Therapies; Sep 20-23, 2014; Philadelphia, PA. Philadelphia (PA): AACR; Clin Cancer Res 2015;21(17 Suppl):Abstract nr A32.