A machine learning analysis of a "normal-like" IDH-WT diffuse glioma transcriptomic subgroup associated with prolonged survival reveals novel immune and neurotransmitter-related actionable targets

Abstract

Background: Classification of primary central nervous system tumors according to the World Health Organization guidelines follows the integration of histologic interpretation with molecular information and aims at providing the most precise prognosis and optimal patient management. According to the cIMPACT-NOW update 3, diffuse isocitrate dehydrogenase-wild type (IDH-WT) gliomas should be graded as grade IV glioblastomas (GBM) if they possess one or more of the following molecular markers that predict aggressive clinical course: EGFR amplification, TERT promoter mutation, and whole-chromosome 7 gain combined with chromosome 10 loss.

Methods: The Cancer Genome Atlas (TCGA) glioma expression datasets were reanalyzed in order to identify novel tumor subcategories which would be considered as GBM-equivalents with the current diagnostic algorithm. Unsupervised clustering allowed the identification of previously unrecognized transcriptomic subcategories. A supervised machine learning algorithm (k-nearest neighbor model) was also used to identify gene signatures specific to some of these subcategories.

Results: We identified 14 IDH-WT infiltrating gliomas displaying a "normal-like" (NL) transcriptomic profile associated with a longer survival. Genes such as C5AR1 (complement receptor), SLC32A1 (vesicular gamma-aminobutyric acid transporter), MSR1 (or CD204, scavenger receptor A), and SYT5 (synaptotagmin 5) were differentially expressed and comprised in gene signatures specific to NL IDH-WT gliomas which were validated further using the Chinese Glioma Genome Atlas datasets. These gene signatures showed high discriminative power and correlation with survival.

Conclusion: NL IDH-WT gliomas represent an infiltrating glioma subcategory with a superior prognosis which can only be detected using genome-wide analysis. Differential expression of genes potentially involved in immune checkpoint and amino acid signaling pathways is providing insight into mechanisms of gliomagenesis and could pave the way to novel treatment targets for infiltrating gliomas.

Keywords: Amino acid neurotransmission; Biomarkers; C5AR1; Glioma; IDH-WT; MSR1; SLC32A1; SYT5; Transcriptomic; Tumor immune checkpoints.

Fig. 1

Unsupervised clustering of TCGA expression data associated with 243 samples. The 243 samples are composed of 238 IDH-WT gliomas and 5 healthy samples. We used 3903 genes differentially expressed between normal and tumor samples in this analysis, and the clustering ordering was performed using a Pearson correlation

Fig. 2

Comparison analysis with TCGA, Ceccarelli, and Aibaidula studies. a Comparison of the identified clusters with tumor subgroups identified in TCGA and Ceccarelli studies. The color code for the different clusters is provided at the bottom of the figure. b Comparison of the IDH-WT NL gliomas with previously published molecular glioma subgroups associated with longer survival only shows partial overlap (in green, uncommon IDH-WT gliomas from TCGA, 2015; in blue, pilocytic astrocytoma-like subgroup from Ceccarelli et al. 2016; in pink, molecularly low-grade from Aibaidula et al., 2017)

Fig. 3

Comparison of tumor purity estimation for NL and OT IDH-WT tumors. The tumor purity was estimated with the R package ESTIMATE. Statistical analysis was performed with the Wilcoxon signed-rank test

Fig. 4

Kaplan-Meier survival curves for NL vs OT tumors identified in TCGA transcriptomic dataset. NL IDH-WT tumors are associated with a better survival profile (p value < 0.05 (*); p value = 0.0052). Statistical analysis was performed with the log-rank test

Fig. 5

Histological distribution in NL vs OT IDH-WT gliomas. Distribution of tumor grade (a) and tumor histology (b) in NL gliomas (n = 14) vs other IDH-WT gliomas (n = 224). Cramer’s V with the chi-square test was used to measure the association between tumor type (NL/OT) and tumor grade/histology. Morphological oligoastrocytomas are included in TCGA dataset, as per the 2007 WHO classification of CNS tumors. Those included in this cohort would be classified as astrocytomas in the 2016 classification scheme, based on the absence of IDH1/2 and 1p/19q alterations

Fig. 6

Mutation and genomic alteration profile in NL vs OT tumors. Mutation burden associated with NL vs OT tumors. The log₁₀ of the number of single nucleotide variations (SNVs) obtained from the Ceccarelli study was counted for each type. Mann-Whitney U tests were used for the statistical analysis

Fig. 7

Expression levels of the C5AR1, SLC32A1, MSR1, and SYT5 genes. The boxplots show the C5AR1, SLC32A1, MSR1, and SYT5 gene expression level related to the tumor type (a) or the tumor grade (b). The NT type (in gray) corresponds to normal tissues. Statistical analysis was calculated with Mann-Whitney U tests (*p value < 0.05; **p value < 0.01; ***p value < 0.001)

Fig. 8

Kaplan-Meier survival curves for NL vs OT tumors identified in independent datasets. Kaplan-Meier survival curves for NL vs OT tumors identified in the first (a) and the second (b) CGGA datasets using the SLC32A1/MSR1 and C5AR1/SYT5 gene signatures. Statistical analysis was performed using a log-rank test

Fig. 9

Immune cell type distribution in NL vs OT tumors. Distribution of the immune cell types in NL and OT tumors estimated with Cibersort (a), quanTIseq (b), and Epic (c) software. Statistical analysis was calculated with Mann-Whitney U tests (d), and the p values which are significant (< 0.05) are highlighted in pink