Table 1.
Patient characteristics.
Fig 1.
Overall workflow for generating multiscale, multiparametric data, extraction of various features and/or conversion to higher scales and analysis approaches to differentiate phenotypes.
Multiparametric MRI images (Panels A & B) were segmented for ROIs and various image features to characterize tumor and subregions (necrosis, enhancing and edema) within the tumor. Multiplexed immunofluorescence tissue analysis (MxIF) (Panel C) provided (left-to-right) a virtual H&E (vH&E), which is a pseudo-colored DAPI and AF image, and corresponding overlays of 46 markers (examples shown are for proliferation and angiogenesis markers). Single cell data were generated for every multiplexed marker and intensities were binned into levels for each cell (low, medium or high using the 33rd and 67th quantiles as the thresholds). The molecular “state” of each cell was computed using the ordinal level of each protein. For visualization purposes, the molecular “state” of a cell was overlaid on the vH&E image (Panel D). Genomics data (Panel E & F), including IDH1 mutation status, were summarized into pathways, cancer hallmarks, and enrichments for each tumor. Cell-level biomarker and MRI feature data were clustered across all glioma patients and by IDH 1 status (Panel G) and molecular and spatial heterogeneity were analyzed relative to IDH1 mutation status or tumor grade (Panel H).
Fig 2.
Distribution and clustering of cells based on protein expression from all treatment-naïve patients.
Unsupervised clustering of cell biomarker data revealed 7 distinct subsets (clusters) of cells. Cluster 2 is dominated by IDH1wt and Cluster 6 is dominated by IDH1mt cases. Clusters 1, 4 and 7 (which were less diverse patient groups) show higher staining intensities of most markers (and cancer hallmarks) compared to Clusters 2, 5 and 6. Iron metabolism marker expression was generally high in Cluster 2, but low in Cluster 6.
Fig 3.
Biomarker images and lollipop plots for cluster 2 and cluster 6.
Staining images for representative cases in Cluster 2 (A) and Cluster 6 (D), including a vH&E image (top left), segmented image (top middle) showing individual cells, an image with cluster assignment to individual cells (top right) and a number of single marker or multi-marker overlays representing expression of different hallmark proteins ((a): DNA breaks, gH2AX. (b): Iron metabolism; FTL, FTH1; (c): Cell Death, Cleaved Caspase 3; (d): Proliferation, EGFR, pERK, Ki67; (e): Immune MHC1, PDL1; (f): Stemness, Nestin, SOX2; (g): Angiogenesis, VEGFR2, SMA, S100A4, CD31; (h) Metabolism, FASN overlaid on DAPI (i) Metabolism, GSK3b, PKM2, CA9; (j) Invasion, GFAP, Collagen IV; (k) Vimentin, Cofilin & NCad.Panel B and Panel E show the protein expression profiles of individual clusters (2 & 6, respectively); “lollipop” lines originate at the average expression of proteins in all cells measured from all cases. Lines to the left of the average vertical axis show lower than average expression, while to the right show higher than average expression. Cluster 2 (Panel C) and cluster 6 (Panel F) trend towards separating cases by IDH1 mutation status. Specifically, Cluster 6, which shows a lower than average expression of most hallmark proteins, is significantly positively correlated to IDH1 mutation (Panel F); Cluster 2 cells with higher iron metabolism (FTL, FTH1) show a trend towards lower representation in IDH1 mutant samples (Panel C).
Fig 4.
Cell cluster composition and Oncoprint of treatment naive gliomas.
For each glioma case, Panel A portrays the fractional distribution of its cells within each of the 7 clusters. Panel B depicts the genomic profile of each glioma case.
Fig 5.
IDH1 mutation status drives cell phenotype at both the gene and the protein level.
Ratios of the average staining intensities for 21 MxIF markers in clusters 6 and 2 were calculated (Panel A). Following deconvolution of the transcriptomes using CellDistinguisher, RNA expression counts (FPKM) for the mRNAs were used to distinguish “class types” (n = 3) across the bulk sequenced specimens, then ratios of the expression values for the same 21 genes compared between Class 2 and Class 3 (Panel B). Fractional composition of each patient case by Cluster 2 or Cluster 6 (Panel C) or according to Class 2 or Class 3 (Panel D) was determined. Cases dominated by cells belonging to protein cluster 2 were more likely to be found in IDH1 wild-type tumors, while cases for which cells from cluster 6 dominated were mostly IDH1 mutated tumors (Panel C). Similarly, the fractional composition of glioma cases comprised of gene expression class 3 were present in higher proportions in IDH1 wild type samples, while class 2 cell types were more abundant in the IDH1 mutant ones (Panel D). The distinguisher genes of class 3 were enriched in genes related to cancer hallmarks of “inducing angiogenesis”, “enabling replicative immortality” and “evading growth suppression” (see S4 Fig and S2 Table).
Fig 6.
Computed molecular and spatial heterogeneity metrics using the multi-omics heterogeneity analysis (MOHA) tool.
The method first converts the continuous marker intensity measures of each segmented cell into an ordinal value representing either a high, medium, or low state. Panel (a) presents an example for the Sustaining Proliferative Signaling cancer hallmark. This gene set is composed of three markers: EGFR, Ki67, Nestin. The state of each of these markers can either be high (2), medium (1), or low (0). Therefore, the three-marker gene set has 27 possible molecular states presented in the color-coded legend (far left). The scatter plot (center) presents the spatial and molecular heterogeneity of treatment naïve gliomas and recurrent GBM samples. Images of tissues from four treatment naïve gliomas (A-D) and four recurrent GBM (E-H) are presented with each segmented cell colored by their expressed molecular state. The spatial state distributions of these eight samples are presented above the scatter plot. For the 4-gene set “inducing angiogenesis” (SMA [ACTA2], VEGFR2 [KDR], CD31 [PECAM1], and S100A4) hallmark, IDH1 mutation status discriminates those cases with relatively lower molecular heterogeneity and relatively higher spatial heterogeneity in grade III treatment-naïve glioma or recurrent glioblastoma. Mann-Whitney test p-values are presented for each nonpaired comparison. (panel b).
Fig 7.
MRI-derived features appear to differentiate patients that carry an IDH1 mutation (IDH1mt) and those that are wild type (IDH1wt), regardless if subjects are treatment naïve or have recurring GBM.
T1w post contrast MRI for IDH1wt subjects (A,E), and IDH1mt (B,F). The white outlines show the extent of the tumor as delineated by the expert neuroradiologist (LW). (C,G) Across the two cohorts, a similar trend may be notice when comparing the mean T1 post-contrast intensity signal in the peritumoral edema region, suggesting an increase in enhancement in the IDH1mt in the peritumoral edema region when compared to the IDH1wt (C and G). An opposite trend is observed when comparing the normalized enhancing core volume across IDH1wt and IDH1mt (D and H), indicating that IDHmt patients have limited to no enhancement. None of these comparisons reach statistical significance after multiple comparison correction using false discovery rate.
Fig 8.
Comprehensive rendering of multiscale measurements in gliomas.
Multiscale modalities depicted include: 1) clinical information (red), 2) IDH1 mutational status (blue), 3) MRI derived variables (green), 4) RNA expression level of genes involved in the “inducing angiogenesis” hallmark (black), and 5) MxIF angiogenesis markers or cell clusters (magenta). The data is binned in low, medium and high categories. Across the treatment-naïve gliomas (a) and the recurrent (post-treatment) glioblastoma* (b) cohorts, it can be observed that IDHmt patients have low angiogenesis according to RNA expression levels and expression of S100A4 and VEGRF. Those subjects also have high fraction of cells in clusters 1 and 2 (low angiogenesis markers), and low fraction of cells in cluster 4 and 5 (high angiogenesis markers (S8 Fig, cluster profiles of angiogenesis clusters). Moreover, MR Images for the same subjects have lower normalized enhancing cores volumes and measure higher intensities on T1 post contrast. *Recurrent GBM (5 subjects are not shown since they were missing MxIF).