Machine-learning assisted subclassification of glioblastoma by developing an endoplasmic reticulum stress-related methylation signature.

Introduction: Glioblastoma (GBM) is a highly aggressive brain tumor with significant heterogeneity, leading to poor prognosis and limited treatment options. Developing innovative molecular subtyping approaches is important for gaining deeper insights into disease pathogenesis and optimizing treatment strategies. DNA methylation has been implicated in the regulation of endoplasmic reticulum stress (ERS), which disrupts protein folding and activates the unfolded protein response (UPR), ultimately determining cellular survival or apoptotic outcomes. Methods: ERS-related DNA methylation profiles were integrated with non-negative matrix factorization (NMF) to establish a molecular classification framework for GBM. An ERS-based signature was further developed using recursive feature elimination with cross-validation (RFECV), and a random forest (RF) model was constructed for subtype prediction. The model was then applied to an external TCGA cohort for validation and downstream characterization. Results: The NMF-based framework stratified GBM patients into four distinct subtypes. The RF model achieved an accuracy of 92.4% in the independent test set. Application of the model to the TCGA cohort revealed distinct molecular and clinical characteristics across subtypes. In particular, Subtype 2 was associated with an immune-inflamed phenotype, lower tumor purity, and poorer prognosis. Connectivity Map (CMap) analysis further identified MEK inhibitors as preliminary candidate compounds for specific subtypes. Discussion: These findings support an association between ERS-related epigenetic modifications and GBM heterogeneity, and provide an epigenetic framework for refined molecular stratification and further exploration of subtype-related therapeutic strategies.