A recurrent interferon, stress, and survival axis identifies a rare malignant programme across glioblastoma single-cell, spatial, and longitudinal cohorts

Glioblastoma exhibits profound intratumoral heterogeneity and rapid adaptation under stress, complicating durable therapeutic control. Here, a biologically grounded transcriptional axis is defined by three complementary modules capturing type I interferon signalling, immediate-early stress response, and pro-survival signalling enriched in unfolded protein response components. Using this equal-weighted axis, a rare axis-high programme was identified in a glioblastoma training cohort (GSE173278; 75,075 cells after filtering; 5.95% prevalence) and mapped to external cohorts through marker-anchored cohort-specific clusters. To enable reproducible detection, an interpretable 40-gene regularised logistic regression classifier was trained using fold-internal panel selection and Platt calibration fitted strictly on out-of-fold predictions. The conservative out-of-fold training evaluation showed informative discrimination and good probability calibration (AUC 0.767, AP 0.116, Brier 0.055, ECE 0.011). When the final model was applied to marker-anchored operational labels, ranking performance was high in an independent glioblastoma cohort (GSE131928; AUC 0.950, AP 0.351) and in an exploratory pan-cancer transfer setting (GSE203612; AUC 0.944, AP 0.343), although calibration slopes indicated dataset-dependent probability-scale shift. Longitudinal CARE glioblastoma samples showed that recalibration on T1 samples improved probability-scale reliability on later T2/T3 samples without changing rank discrimination. Spatial stratification based on available sample annotations linked higher detector probabilities to necrotic, bulk, and tumour-region samples. Irradiation feature-count contrasts supported positive shifts of the 40-gene panel and axis-related modules, while cell-level treated/control summaries were interpreted as context-specific perturbational evidence. Together, these results define a rare axis-high glioblastoma programme and provide a recalibratable gene-panel detector for probability-guided experimental triage and mechanistic follow-up.