Unsupervised Machine Learning Analysis of Bulk Rna-sequencing Data Reveals Tumor Immune Microenvironment Subtypes of Idh-mutant Gliomas

Introduction: Recent literature suggests that IDH-mutant glioma tumor immunology differs significantly from IDH-wildtype gliomas and is more immunosuppressed, highlighting the need for IDH-mutant specific immunologic analyses. In addition, the possibility of different IDH-mutant glioma immunologic subtypes has not yet been investigated, and such subtypes could play an important role in decision-making surrounding the use of immunotherapies to treat these tumors. We aim to utilize unsupervised machine learning methods to identify distinct subtypes of IDH-mutant gliomas based on immune bulk RNA-sequencing transcriptomic data.

Methods: IDH-mutant glioma samples in The Cancer Genome Atlas (TCGA) were identified and selected. K-means unsupervised clustering analysis was performed on the immune-related gene expression data. Identified clusters were validated using self-organizing map (SOM) clustering. The xCell software was then used to estimate the tumor infiltration of 64 immune and stromal cell types.

Results: A total of 439 patients with IDH-mutant gliomas were included in the study. Unsupervised machine learning analysis trained on immune gene expression data yielded two distinct clusters. There were no significant differences between Cluster 1 (n=394) and Cluster 2 (n=45) in mean age at diagnosis (40.8 years), gender (54.9% male), or race. While both clusters shared high expression of certain immune genes including CLU, SPARC, FTL, and FTH1, Cluster 1 IDH-mutant gliomas were distinctly characterized by high expression of HSPA8, DDR1, and NCAM1. In contrast, Cluster 2 IDH-mutant gliomas had high expression of B2M, CD74, and CD81. Multivariate logistic regression analysis revealed B-cells (p=0.041), CD8+ naive T-cells (p=0.004), Tregs (p=0.006), Monocytes (p=0.027), and plasma cells (p=0.038) to be independently associated with Cluster 1 gliomas, while M1 (p < 0.001) and M2 (p < 0.001) macrophages were associated with Cluster 2 gliomas.

Conclusion : Unsupervised machine learning categorized IDH-mutant gliomas into lymphocyte-enriched and macrophage-enriched immunologic subtypes. Additional research validating these findings and their underlying mechanisms is warranted.