Nuclear morphology from H&E slides shows a small, composition- confounded association with hypoxia transcription in glioma: a cross-cohort benchmark

Abstract Hypoxia is a hallmark of glioblastoma but is not measured in routine diagnosis, whereas hematoxylin and eosin (H&E) slides are universally available. We asked whether 31 interpretable nuclear and tissue-architecture features from H&E slides of TCGA-GBM, CPTAC-GBM and TCGA-LGG are associated with a bulk Hallmark hypoxia transcription score, and how far composition, IDH status, signature choice and technical confounding explain it. Hypotheses were defined post hoc, and only later-processed CPTAC patients (n = 60–61) formed a temporal hold-out. No feature was significant in TCGA-GBM alone (0/31), whereas 9/31 were significant in the pooled analysis. Three selected features had pooled ρ ≈ 0.19 but were non-significant in the hold-out (ρ = 0.05–0.14; about 34% power). The score correlated strongly with immune, stromal and myeloid scores (ρ = 0.35–0.71), and composition predicted it better than morphology. Associations vanished after adjustment for IDH and grade in TCGA-LGG and for sequencing depth in Visium, and no independent prognostic value was found. In HCMI-CMDC, the only further open cohort with paired slides and RNA (n = 35; low power), the features did not replicate (ρ = −0.15 to − 0.20; 0/31 BH), and the pooled estimate fell to ρ ≈ 0.15. The findings are hypothesis-generating.