AI-driven transcriptomic classification of glioblastoma: Associations with survival and tumor microenvironment.

2042 Background: Glioblastoma (GBM) is the most lethal primary brain tumor in adults, with a median survival of ~15 months despite current therapies (surgery, radiation, temozolomide). Advances like immune checkpoint inhibitors, anti-angiogenic agents, and tumor vaccines have shown suboptimal results. The 2021 WHO classification highlights molecular markers (e.g., IDH, MGMT) for better stratification, but these fail to fully capture tumor microenvironmental dynamics. Using an AI-driven transcriptomic approach, we identified novel prognostic subtypes in IDH-wildtype GBM, aiming to refine stratification, enhance understanding of tumor biology, and guide personalized therapeutic strategies. Methods: We accessed microarray data from The Cancer Genome Atlas (TCGA) (n=353 newly diagnosed, IDH-WT GBM) for a training set and RNA-seq data from the Chinese Glioma Genome Atlas (CGGA) (n=170 primary and n=106 recurrent tumors) for validation. A proprietary SphereBio machine learning–based algorithm was used to derive transcriptomic signatures with prognostic relevance. Subtypes were assessed via Kaplan–Meier analyses in the training cohort and tested in both primary and recurrent validation cohorts. Immune/stromal infiltration was quantified using a tumor deconvolution tool (DA_505), and pathway enrichment (GAGE) was performed on the validation sets. Results: AI-driven clustering revealed three transcriptomic subtypes with significant survival differences in both the training (p<0.0001) and primary validation (p=0.0004) cohorts. In the recurrent cohort, a similar survival trend by subtype was observed, though significance was diminished (p=0.12), likely due to limited sample size and therapy-related changes. Immune/stromal deconvolution showed distinct infiltration patterns: subtypes enriched for CD4+ and CD8+ T cells correlated with prolonged survival. Pathway enrichment analysis in both primary and recurrent tumors highlighted potential targets involving embryogenesis, immune modulation, cell cycle, and stress response. The persistence of a consistent survival trend and comparable microenvironment and pathway patterns suggest that these transcriptomic subtypes remain biologically relevant even after standard treatment. Conclusions: Our integrated transcriptomic and microenvironment-focused approach identified three prognostically distinct GBM subtypes, validated across independent cohorts. These findings underscore the utility of AI-driven transcriptomic signatures for personalized stratification, with the potential to guide targeted therapeutic strategies and inform clinical trial design in GBM.