Automatic Classification of Cellular Expression by Nonlinear Stochastic Embedding (ACCENSE)

Abstract

Mass cytometry enables an unprecedented number of parameters to be measured in individual cells at a high throughput, but the large dimensionality of the resulting data severely limits approaches relying on manual "gating." Clustering cells based on phenotypic similarity comes at a loss of single-cell resolution and often the number of subpopulations is unknown a priori. Here we describe ACCENSE, a tool that combines nonlinear dimensionality reduction with density-based partitioning, and displays multivariate cellular phenotypes on a 2D plot. We apply ACCENSE to 35-parameter mass cytometry data from CD8 + T cells derived from specific pathogen-free and germ-free mice, and stratify cells into phenotypic subpopulations. Our results show significant heterogeneity within the known CD8 + T-cell subpopulations, and of particular note is that we find a large novel subpopulation in both specific pathogen-free and germ-free mice that has not been described previously. This subpopulation possesses a phenotypic signature that is distinct from conventional naive and memory subpopulations when analyzed by ACCENSE, but is not distinguishable on a biaxial plot of standard markers. We are able to automatically identify cellular subpopulations based on all proteins analyzed, thus aiding the full utilization of powerful new single-cell technologies such as mass cytometry.immunophenotyping | machine learning | class discovery | CyTOF | FACS

A distinct subpopulation of CD25−T-follicular regulatory cells localizes in the germinal centers

“…Vaccination was required to induce substantial fractions of CD25 − Tfr cells in the spleen and LNs, whereas the Peyer's patches had a constant population of these cells unaffected by vaccination. To supplement this manual gating, we also used Automatic Classification of Cellular Expression by Nonlinear Stochastic Embedding (ACCENSE), a tool that uses a dimensionality reduction technique, t-distributed stochastic neighbor embedding, which allows the visualization of multiple parameters in two dimensions and is useful in the identification of novel subpopulations (16). Here, we took total CD4 T cells from the Peyer's patches and mapped them by expression of PD1, BCL6, CXCR5, Foxp3, and CD25.…”

Data-driven identification of prognostic tumor subpopulations using spatially mapped t-SNE of mass spectrometry imaging data

“…The application of ACCENSE density-based partitioning (12) to the t-SNE map led to the detection of a large number of clusters, but the Kaplan-Maier curves lacked statistical significance (SI Appendix, Fig. S2B).…”

Immunohistochemistry and mass spectrometry for highly multiplexed cellular molecular imaging

“…More sophisticated, but still purely molecular (non-spatial), approaches to deal with this scale of data currently include the SPADE tools on Cytobank 40 and the ACCENSE method. 41 The former emphasizes hierarchical or developmental connections between cell populations, whereas the latter can detect population clusters without hierarchical constraints. To give an idea of the scale of the challenge: CyTOF data processed through ACCENSE provided cell-by-cell high-dimensional information highlighting the probable existence of at least 24 subclasses of CD8 þ T cells.…”

“…To give an idea of the scale of the challenge: CyTOF data processed through ACCENSE provided cell-by-cell high-dimensional information highlighting the probable existence of at least 24 subclasses of CD8 þ T cells. 41 Now, imagine combining such per-cell molecular complexity with geographic distribution tools. These could be used to characterize the spatial distributions of highly refined cell subclasses, and add information on distances and possible interactions between populations, such as immune cell types, and distinct tumor subregions.…”