Background: Glioblastoma (GBM) is the most aggressive primary brain tumor with extremely poor prognosis. Conventional diagnostic and prognostic approaches remain inadequate, highlighting the need for integrative strategies to improve patient outcomes. Methods: We analyzed ligand–receptor (L–R) interactions in TCGA-GBM transcriptomes using BulkSignaL-R, and validated their spatial expression patterns with single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics datasets. Prognostic histopathological features were extracted from hematoxylin and eosin (H&E)-stained sections through omics-guided feature identification, followed by classification using machine learning algorithms. Results: We identified four pivotal L–R pairs (LTB–CD40, VEGFA–ITGB1, FN1–COL13A1, and TGM2–ITGB1) to construct a risk model, which served as an independent prognostic factor for overall survival. The multivariate Cox regression analyses revealed that the risk score was significantly associated with Overall Survival (OS) (HR = 1.67, 95% CI: 1.25–2.25, p < 0.001). High-risk patients exhibited distinct molecular signatures, including CALN1 mutations, specific CNV patterns, and enriched Notch/interferon-γ signalings. scRNA-seq and spatial transcriptomics revealed that these L–R pairs were predominantly expressed in gMES-like glioma cells, OPC-like cells, and pericytes. Finally, our deep learning model successfully stratified risk groups based on histological features, identifying specific tumor regions (Clusters 0, 2, 4, and 5) as critical determinants of prognosis (AUC = 0.750 by Logistic Regression). Conclusions: We developed a novel multi-modal framework integrating L–R interactomics and deep learning-based pathomics. This approach not only elucidates the molecular and spatial landscape of glioma intercellular communication but also provides a methodological framework for risk stratification.
Gao et al. (Thu,) studied this question.