EPCO-18. UNSUPERVISED MACHINE LEARNING ON TUMOR IMMUNE TRANSCRIPTOMIC DATA REVEALS DISTINCT IMMUNOLOGIC SUBTYPES OF GLIOBLASTOMA

Abstract

Abstract Glioblastoma (GBM) has demonstrated a limited therapeutic response to single-agent immunotherapies, highlighting the need for a better understanding of GBM tumor immune phenotypes to inform targeted immunologic treatments. To elucidate GBM immunologic phenotypes we performed unsupervised machine learning on microarray transcriptomic data from 815 newly diagnosed GBM patients. We utilized immune genes only and performed k-means clustering for k=2-10 with within-cluster sum of squares and silhouette methods used to identify the optimal number of clusters. Identified subtypes were further characterized using xCell to estimate the tumor infiltration of 64 immune and stromal cell types. Two primary immunologic subtypes were identified: subtype 1 (423 patients) and subtype 2 (392 patients). There was no difference in patient age (mean age 58 years) or gender between subtypes. Subtype 1 was primarily classical subtype GBMs (50.5%, p< 0.001), while subtype 2 was composed of primarily mesenchymal subtype GBMs (74.2%). Relative to subtype 1, subtype 2 was characterized by high expression of immune checkpoint and co-stimulatory proteins, including CD86, OX40L, IDO1, and LAG3. Subtype 2 also had increased activation of immune genes associated with cytolytic activity and antigen processing. In addition, subtype 2 had increased infiltration of M2 macrophages, monocytes, dendritic cells, and neutrophils while subtype 1 had higher infiltration of CD8 naïve T cells. Multivariate logistic regression confirmed that subtype 2 was characterized by monocytes, M1 and M2 macrophages, neutrophils, and dendritic cells while subtype 1 was characterized by CD8 naïve T cells, regulatory T cells, and naive B cells. Subtype 2 had reduced progression-free survival, but there was no difference in overall survival between the two subtypes. In summary, unsupervised machine learning of GBM immune transcriptomic data identified two immune phenotypes, one of which was more immunologically active than the other, potentially informing future personalized immunotherapy treatments.