Cytofast is a visualization tool used to analyze output from clustering. Cytofast can be used to compare two clustering methods: FlowSOM and Cytosplore. Cytofast can rapidly generate a quantitative and qualitative overview of mass cytometry data and highlight the main differences between different clustering algorithms.
The complexity of data generated by mass cytometry has necessitated new tools to rapidly visualize analytic outcomes. Clustering methods like Cytosplore or FlowSOM are used for the visualization and identification of cell clusters. For downstream analysis, a newly developed R package, Cytofast, can generate a rapid visualization of results from clustering methods. Cytofast takes into account the phenotypic characterization of cell clusters, calculates the cell cluster abundance, then quantitatively compares groups. This protocol explains the applications of Cytofast to the use of mass cytometry data based on modulation of the immune system in the tumor microenvironment (i.e., the natural killer [NK] cell response) upon tumor challenge followed by immunotherapy (PD-L1 blockade). Demonstration of the usefulness of Cytofast with FlowSOM and Cytosplore is shown. Cytofast rapidly generates visual representations of group-related immune cell clusters and correlations with immune system composition. Differences are observed in the clustering analysis, but separation between groups are visible with both clustering methods. Cytofast visually shows the patterns induced by PD-L1 treatment that include a higher abundance of activated NK cell subsets, expressing a higher intensity of activation markers (i.e., CD54 or CD11c).
Mass cytometry (cytometry by time-of-flight, or CyTOF) allows detection of a wide range of intracellular or extracellular biomarkers in millions of single cells. The high-dimensional nature of mass cytometry data necessitates certain analysis tools, such as cell clustering techniques like SPADE1, FlowMaps2, FlowSOM3, Phenograph4, VorteX5, and scaffold maps6. In addition, various dimensionality reduction-based techniques have been developed (i.e., principal component analysis [PCA]7, t-distributed stochastic neighbor embedding [t-SNE]8, hierarchical stochastic neighbor embedding [HSNE]9, uniform manifold approximation and projection [UMAP]10, and diffusion maps11) to improve the speed, interpretation and visualization of high-dimensional datasets.
Downstream analysis of high-dimensional flow and mass cytometric data often lacks automatic processes to perform statistical tests on cluster frequency and links with clinical outcomes. Previously, we developed an R-based workflow known as Cytofast12, which allows for visual and quantitative downstream analyses of clustering techniques by Cytosplore or FlowSOM.
The protocol described here clarifies the use of Cytofast in R and shows how to generate quantitative and qualitative heatmaps and graphs. Furthermore, it facilitates the determination of connections between observed immune phenotypes and clinical outcomes. This report also describes the analysis of a specific mass cytometry dataset using two different clustering procedures: FlowSOM and Cytosplore. By using Cytofast with both clustering methods, it is correspondingly shown that the activation phenotype of NK cells is influenced by PD-L1 immune checkpoint blockade.
Cytofast is a rapid computational tool that provides a quick and global exploration of cytometric data by highlighting and quantifying treatment-specific cellular subsets. The protocol described aims to further process clustering analyses with Cytosplore or FlowSOM. Other clustering analysis tools are suitable for Cytofast, but this requires the use of Cytofast to assign each cell to a subset. Cytofast, however, is not a clustering method, and therefore requires clustering procedures before use.
The analysis performed here showed that certain CD161+ NK cell subsets in the tumor microenvironment were sensitive to a PD-L1 blockade. This was evidenced by changes in their phenotype and abundance, which were observed using both Cytosplore and FlowSOM as clustering methods. Both methods distinguished the main NK cell cluster (CD11b+ NKG2A+) with slightly different frequencies (15%–20% for Cytosplore, 30%–40% for FlowSOM). The differences in abundance and this approximation did not affect the global pattern, because both dendrograms displayed in the right panels of Figure 2 and Figure 3 showed similar results. By using Cytofast, it is thus possible (independent from the clustering method chosen) to segregate PD-L1-treated and untreated mice based on analyses of NK cell cluster phenotype and abundance.
Depending on the recorded parameters, modifications to the protocol are needed. Specifically, certain parameters such as time and background must be removed while performing the clustering analysis. In addition, it is important that each cell is assigned to a subset. The cfData function will simply add the raw cell counts per cluster per sample into the cfList. From this step, the cytoheatmap can be built as explained in section 3.
Cytofast has been successfully used as a visualization and quantification tool to compare different clustering methods13. This R package is also compatible with advanced features, such as the globaltest14, which can test associations between groups of clusters using clinical variables. In the future, the globaltest tool and other algorithms can be integrated with Cytofast for more in-depth visualization and quantification.
The authors have nothing to disclose.
We acknowledge funding from the European Commission of a H2020 MSCA award under proposal number 675743 (ISPIC). We thank Tetje van der Sluis and Iris Pardieck for testing the protocol.