Comparative analysis of transcriptomic profile, histology, and IDH mutation for classification of gliomas

Abstract

Gliomas are currently classified through integration of histology and mutation information, with new developments in DNA methylation classification. However, discrepancies exist amongst the major classification methods. This study sought to compare transcriptome-based classification to the established methods. RNAseq and microarray data were obtained for 1032 gliomas from the TCGA and 395 gliomas from REMBRANDT. Data were analyzed using unsupervised and supervised learning and other statistical methods. Global transcriptomic profiles defined four transcriptomic glioma subgroups with 91.4% concordance with the WHO-defined mutation subtypes. Using these subgroups, 168 genes were selected for the development of 1000 linear support vector classifiers (LSVC). Based on plurality voting of 1000 LSVC, the final ensemble classifier confidently classified all but 17 TCGA gliomas to one of the four transcriptomic profile (TP) groups. The classifier was validated using a gene expression microarray dataset. TP1 cases include IDHwt, glioblastoma high immune infiltration and cellular proliferation and poor survival prognosis. TP2a is characterized as IDHmut-codel, oligodendrogliomas with high tumor purity. TP2b tissue is mostly composed of neurons and few infiltrating malignant cells. TP3 exhibit increased NOTCH signaling, are astrocytoma and IDHmut-non-codel. TP groups are highly concordant with both WHO integrated histology and mutation classification as well as methylation-based classification of gliomas. Transcriptomic profiling provides a robust and objective method to classify gliomas with high agreement to the current WHO guidelines and may provide additional survival prediction to the current methods.

Retrospective Validation of a 168-Gene Expression Signature for Glioma Classification on a Single Molecule Counting Platform

“…Histologically the brain tumors were classified as astrocytoma ( n = 10), glioblastoma ( n = 27), oligoastrocytoma ( n = 4), and oligodendroglioma ( n = 7) ( Table 1 ). We applied our transcriptomic classification algorithm [ 13 ] to classify these 48 FFPE tissues into TP1 ( n = 30), TP2A ( n = 4), TP2B ( n = 4), and TP3 ( n = 1).…”

“…We previously reported on our ensemble algorithm which combines 1000 linear support vector classifiers trained from TCGA RNASeq data clustered using our combined UMAP and density-based clustering algorithm [ 13 ]. We applied this algorithm to our AU cohort data after batch normalization, z-score transformation, and using Empirical Bayes to combine these data with the TCGA data.…”

“…There is a 14% disagreement rate between histology and molecular testing [ 13 ]. The inter-observer disagreement with histology diagnosis is also high [ 12 ].…”

“…We recently reported a transcriptomics-based classification scheme for gliomas [ 13 ]. The Cancer Genome Atlas (TCGA) brain cancer cohort [ 8 ] was classified into four-subtypes: transcriptome profile one (TP1), T2a, TP2b and TP3.…”

“…We present our results on the technical reproducibility of the technology. Our study demonstrates validation of 168 gene classifier [ 13 ] developed for classification of gliomas on nCounter platform using retrospectively collected brain cancer samples.…”