Graph-based modeling of multiparametric MRI deciphers molecular states of high-grade glioma invasion with prognostic implications

AbstractThe infiltrative, non-enhancing margin of IDH wildtype high grade glioma (IDHwt HGG) harbors distinct molecular programs that drive invasion and therapeutic resistance, yet remains largely unevaluable by conventional tissue sampling approaches and by conventional imaging. Here we show that this invasive architecture is encoded within multiparametric MRI (mpMRI) feature relationships and can be decoded using a graph-based framework trained on multiregional image-localized biopsies. Across 134 spatially matched biopsy-imaging pairs from 35 patients with primary IDHwt HGG (29 glioblastomas (GBM) and 6 non-glioblastoma HGGs), unsupervised graph community detection identifies two imaging-defined clusters that localize to invasive tumor regions without molecular supervision. Transcriptomic profiling associates these clusters with neuronal (NEU) and glycolytic-plurimetabolic (GPM) molecular programs. Building on this framework, a graph convolutional network (GCN) accurately predicts NEU and GPM transcriptional states in independent training and validation cohorts and significantly outperforms conventional convolutional neural networks. Applied to whole-tumor mpMRI volumes, the trained GCN generates spatially resolved probability maps that quantify the distribution and relative burden of NEU and GPM programs across both MRI contrast-enhancing and non-enhancing invasive regions. These imaging-derived molecular maps stratify patients by overall survival. Increased GPM burden is associated with poorer survival, consistent with the aggressive behavior associated with mesenchymal-like transcriptional programs in IDHwt HGG. In contrast, increased NEU burden is associated with improved survival, identifying a previously unrecognized imaging-derived prognostic biomarker that was not detected by biopsy-based molecular classification alone. Together, these findings establish a graph-based imaging framework for spatially resolved molecular classification of invasive IDHwt HGG and demonstrate that whole-tumor molecular state architecture carries prognostic information beyond conventional tissue sampling.