(141c) Nonlinear Dimensionality Reduction for Cellular Phenotyping From High Throughput Mass Cytometry Data

The advent of mass cytometry has enabled simultaneous, high-throughput measurements of ~40 proteins at the single cell level. Such unprecedented data makes possible, for the first time, resolution of cellular phenotypic diversity under normal conditions, and also systematic classification of alterations therein resulting from pathologies. Current techniques to interrogate cellular diversity use unsupervised clustering to classify individual cells into groups and produce visualizations in the form of a graph structure describing the relatedness of different cell types. Often, a major limitation in these methods is that number of clusters needs to be pre-specified by the user, and furthermore, the final visualization produced by the algorithm relies heavily on prior knowledge of existing cell types. We describe an alternative approach based on non-linear dimensionality reduction that overcomes these limitations, and demonstrate its advantages by applying it on two datasets. Our technique produces a two-dimensional embedding of cells based on their relative expression patterns and a simple kernel density estimator identifies unique subpopulations with significantly different expression patterns. We first analyzed published human bone marrow data and show that our method produces results consistent with the existing state-of-the-art clustering-based analysis tool. Furthermore, our unbiased density-based detection of cell types also identifies subsets not described under the existing taxonomy. Application of our method on CD8 T cells extracted from wild-type Black-6 mice of different ages, while demonstrating the stereotypy of cellular development in genetically identical mice, also identifies a distinct subpopulation of naïve cells that vary as a function of age.