A new inter-cluster similarity measure for high-dimensional data can facilitate analysis of heterogeneous mass-cytometry data

Mass cytometry (CyTOF) is a high-dimensional, high-throughput technology to analyze quantities of 30–40 proteins simultaneously in single cells and is widely used to investigate heterogeneity in tumour tissue samples. Also, by detecting phospho-proteins, (phospho-)CyTOF can be used to investigate intracellular signalling-pathway activity. Analysis of CyTOF data usually involves clustering and visualization of the high-dimensional data, which are typically performed using independent methods and are strongly influenced by user-controlled input parameters. To reduce user influence on CyTOF analysis results and harmonize the clustering and visualization processes, I developed ASTRICS, a new measure of similarity between clusters of high-dimensional data points based on local dimensionality reduction and triangulation of alpha shapes. I propose a multi-stage clustering strategy called CytoClue in which ASTRICS is used to automatically generate a weighted graph from an initial set of fine-grained clusters, which are obtained by an existing algorithm (e.g. FlowSOM) and are represented by graph nodes. The graph is visualized by force-directed layout and used for further clustering by a graph-based algorithm. This poster introduces ASTRICS and CytoClue and presents results of applying them to phospho-CyTOF experiments that were conducted to investigate heterogeneity between and within diffuse large B-cell lymphoma (DLBCL) cell lines.