Geometry of the Gene Expression Space of Individual Cells

Abstract

There is a revolution in the ability to analyze gene expression of single cells in a tissue. To understand this data we must comprehend how cells are distributed in a high-dimensional gene expression space. One open question is whether cell types form discrete clusters or whether gene expression forms a continuum of states. If such a continuum exists, what is its geometry? Recent theory on evolutionary trade-offs suggests that cells that need to perform multiple tasks are arranged in a polygon or polyhedron (line, triangle, tetrahedron and so on, generally called polytopes) in gene expression space, whose vertices are the expression profiles optimal for each task. Here, we analyze single-cell data from human and mouse tissues profiled using a variety of single-cell technologies. We fit the data to shapes with different numbers of vertices, compute their statistical significance, and infer their tasks. We find cases in which single cells fill out a continuum of expression states within a polyhedron. This occurs in intestinal progenitor cells, which fill out a tetrahedron in gene expression space. The four vertices of this tetrahedron are each enriched with genes for a specific task related to stemness and early differentiation. A polyhedral continuum of states is also found in spleen dendritic cells, known to perform multiple immune tasks: cells fill out a tetrahedron whose vertices correspond to key tasks related to maturation, pathogen sensing and communication with lymphocytes. A mixture of continuum-like distributions and discrete clusters is found in other cell types, including bone marrow and differentiated intestinal crypt cells. This approach can be used to understand the geometry and biological tasks of a wide range of single-cell datasets. The present results suggest that the concept of cell type may be expanded. In addition to discreet clusters in gene-expression space, we suggest a new possibility: a continuum of states within a polyhedron, in which the vertices represent specialists at key tasks.

Assessing rejection-related disease in kidney transplant biopsies based on archetypal analysis of molecular phenotypes

“…It is an unsupervised method, since the algorithm is given only the raw data, not any descriptive or outcome-related information (e.g., diagnoses). It has been used in a wide variety of fields, including sports analytics (33), astrophysics (34), marketing (35), and bioinformatics/medicine (16,36,37). Although it is common to show archetype results on a PCA plot as we do here, AA is conceptually very different from PCA.…”

Geometric analysis of pathways dynamics: Application to versatility of TGF-β receptors

“…Our analysis of such models also showed that branches and bifurcations of states occur naturally in the context of cell cycle modelling (Noel et al, 2012). In a more general context, geometric analysis of single-cell expression data from human and mouse tissues showed that gene expression is structured in clusters but also in continua of states within polyhedra whose vertices can be understood as specialized key tasks (Korem et al, 2015). These findings were interpreted in terms of multi-objective optimization solutions (Korem et al, 2015), but could also suggest transient behaviour between specialized states.…”

Take ACTION to characterize the functional identity of single cells

“…The functional identity of each cell is closely associated with its underlying type [3] and a number of methods have been proposed to directly identify cell types from the transcriptional profiles of single cells [4][5][6][7][8][9]. A majority of these methods rely on classic measures of distance between transcriptional profiles to establish cell types and their relationships.…”

“…To validate cell types identified using ACTION, we compare our method to four recently proposed methods: SCUBA [5], SNNCliq [7], single-cell ParTI [8,13], and TSCAN [9] (see Supplementary Text 1 for a brief description of these methods) to predict annotated cell types on the same four datasets (see Methods, Component 2). All methods (including ACTION) have no user-configurable parameters -except for SNNCliq.…”

“…In another direction, Korem et al [8], adopted a previously developed method, called Pareto task inference (ParTI) [13], and…”