GBMdeconvoluteR accurately infers proportions of neoplastic and immune cell populations from bulk glioblastoma transcriptomics data

Abstract

Background: Characterizing and quantifying cell types within glioblastoma (GBM) tumors at scale will facilitate a better understanding of the association between the cellular landscape and tumor phenotypes or clinical correlates. We aimed to develop a tool that deconvolutes immune and neoplastic cells within the GBM tumor microenvironment from bulk RNA sequencing data.

Methods: We developed an IDH wild-type (IDHwt) GBM-specific single immune cell reference consisting of B cells, T-cells, NK-cells, microglia, tumor associated macrophages, monocytes, mast and DC cells. We used this alongside an existing neoplastic single cell-type reference for astrocyte-like, oligodendrocyte- and neuronal progenitor-like and mesenchymal GBM cancer cells to create both marker and gene signature matrix-based deconvolution tools. We applied single-cell resolution imaging mass cytometry (IMC) to ten IDHwt GBM samples, five paired primary and recurrent tumors, to determine which deconvolution approach performed best.

Results: Marker-based deconvolution using GBM-tissue specific markers was most accurate for both immune cells and cancer cells, so we packaged this approach as GBMdeconvoluteR. We applied GBMdeconvoluteR to bulk GBM RNAseq data from The Cancer Genome Atlas and recapitulated recent findings from multi-omics single cell studies with regards associations between mesenchymal GBM cancer cells and both lymphoid and myeloid cells. Furthermore, we expanded upon this to show that these associations are stronger in patients with worse prognosis.

Conclusions: GBMdeconvoluteR accurately quantifies immune and neoplastic cell proportions in IDHwt GBM bulk RNA sequencing data and is accessible here: https://gbmdeconvoluter.leeds.ac.uk.

Keywords: deconvolution; glioblastoma; immune; neoplastic; transcriptomics.

Figure 1.

(A) The process adopted to amalgamate several independent single cell GBM datasets and create a GBM-specific immune cell reference signature gene-expression matrix (for input to CIBERSORTx) or marker gene set (for input to MCPcounter). (B) The inherent batch effects in the amalgamated data are evident in dimensionality reduction plots where clusters initially separated by originating datasets (far left), but were removed by normalization (middle left and Supplementary Figure S1A). Initial clustering and cell type assignment of the normalized data was unable to resolve TAM and microglia, and T- and NK-cells (middle right) but further sub-clustering enabled these cell types to be further delineated (far right and Supplementary Figure S1B). (C) A dot plot showing the expression of chosen GBM-specific immune cell type markers (y-axis) in each cell type in the amalgamated single cell data (x-axis).

Figure 2.

(A) A schematic showing how patient samples were used for validation. Regions of formalin fixed tissue sections were annotated as high tumour cell content by a neuropathologist (circled) and were macro-dissected for RNA sequencing. At least three overlapping regions (squares) per sample were subjected to imaging mass cytometry (IMC) on a consecutive section. Brain schematic taken from Vecteezy.com (B) Left: A representative image from the IMC for GBM sample 64 with three of the chosen protein markers annotated. Right: The UMAP projection of cell types assigned according to the expression of cell type protein markers quantified by IMC. (C, D) Scatterplots of gold standard cell proportions quantified by IMC (y-axis) versus those predicted by gene expression based methods (annotated across the top) for immune (C) or neoplastic cancer (D) cell types indicated down the side. The Pearson’s correlation coefficient (r) is indicated. The dotted line is the line of best fit and the shaded area denotes the confidence interval. Marker genes for MCPcounter were either default (MCP_default), GBM-specific according to our research (MCP_GBM) or GBM-specific according to GBMap (MCP_GBMap) Neoplastic cells are astrocyte-like (AC), oligodendrocyte progenitor-like (OPC), neuronal progenitor-like (OPC) or mesenchymal (MES).

Figure 3.

MCP_GBM was used to score cell types in bulk GBM RNAseq data from The Cancer Genome Atlas (TCGA). (A) Heatmap of the correlations between cell type scores across all samples. (B) Boxplots showing distribution of cell type scores for patients with worse or better prognosis (determined by the lower and upper quartile of overall survival, respectively). (C) Heatmap of the correlations between cell type scores across samples from patients with worse (left) or better (right) prognosis. Significance is denoted by asterisks: *P < .05; ^**P < .01; ^***P < .001;
 ^****P < .0001; NS, not significant.