Characterizing cell subsets using marker enrichment modeling

Abstract

Learning cell identity from high-content single-cell data presently relies on human experts. We present marker enrichment modeling (MEM), an algorithm that objectively describes cells by quantifying contextual feature enrichment and reporting a human- and machine-readable text label. MEM outperforms traditional metrics in describing immune and cancer cell subsets from fluorescence and mass cytometry. MEM provides a quantitative language to communicate characteristics of new and established cytotypes observed in complex tissues.

Figure 1. Marker enrichment modeling (MEM) automatically labels human blood cell populations in Dataset A

a) Cells from normal human blood were previously grouped into 7 canonical populations using viSNE analysis and expert review of 25D mass cytometry data. b) MEM labels were computationally generated for each canonical cell subset using the other six populations as reference. The population labeled by immunologists as “CD4⁺ T cells” was labeled by MEM as ▲CD4⁺⁶ CD3⁺⁵ ▼CD8a⁻⁴ CD16⁻³ and comprised 48.72% of PBMC in this sample. In contrast, the MEM label ▲CD16⁺⁹ CD56⁺² CD11c⁺²▼CD4⁻⁷ CD3⁻⁴ CD44⁻³ was generated for the population gated as “NK cells”. Heatmaps show protein enrichment values used to generate MEM labels and the median protein expression values for each protein on each cell subset. Variability in protein expression across the 7 canonical cell populations is shown below to highlight proteins that were expressed homogeneously (low variability, e.g. CD45) and those that were expressed heterogeneously (high variability, e.g. CD8a, CD4). c) Graphs show decreasing f-measure (clustering accuracy) as markers were excluded from k-means cluster analysis based on high to low absolute MEM or median values, compared to random exclusion.

Figure 2. Hierarchical clustering based solely on MEM label groups T cells and B cells measured in diverse studies using different cytometry platforms

A) MEM label values were compared for each of 80 populations (CD4+ T cells and B cells) from 3 human tissues representing 6 mass cytometry studies and 1 fluorescence flow cytometry study. The normalized RMSD (i.e. similarity) for two populations was 100% when MEM label exponents were identical for all of the shared proteins. Populations are shown clustered according to MEM label percent similarity. Tissue type, source study (numbered 1-7 and referenced in online methods), and individual sample IDs are indicated to the right. *indicates samples stimulated by bacterial superantigen Staphlococcus enterotoxin B (SEB). B) Representative MEM labels for CD4+ T cells (top) and B cells (bottom) from SEB-stimulated normal human blood (1.4, top, mass cytometry), normal human bone marrow (5, mass cytometry), normal human tonsil (2.5, mass cytometry), SEB-stimulated normal human blood (1.4, bottom, fluorescence flow cytometry), and normal human blood (6.1, mass cytometry).

Figure 3. MEM correctly grouped immune and cancer cell populations from glioma tumors using nine proteins expressed on cancer cells in Dataset D

(A) A heatmap of MEM enrichment scores is shown for 52 populations of cells identified in tumors from 4 glioblastoma patients (G-08, G-10, G-11, G22) in an unsupervised manner using viSNE. MEM scores were then calculated based only on the nine measured proteins expected to be expressed on cancer cells (S100B, TJF1, GFAP, Nestin, MET, PGFRα, HLA-DR, and CD44). (B) Each population was annotated for a cell type based on review of the MEM label and classified as tumor infiltrating APCs (blue), tumor infiltrating T cells (green), or non-immune tumor cells (red). (C) A heatmap of median intensity values is shown for the 13 measured proteins from each of the 52 tumor cell populations. Expression of CD45, CD3, and CD64 was used to assess the respective identity of leukocytes, T cells, and antigen presenting cells.