BIOM-101. STRATIFICATION OF GLIOBLASTOMA PROGRESSORS USING MULTI-OMIC PROFILING AND GRAPH NEURAL NETWORKS

Abstract Patients with glioblastoma exhibit highly variable responses to standard chemoradiation, with approximately 30% progressing during treatment and a small subset surviving beyond 5 years. Despite this heterogeneity, non-invasive biomarkers to predict early treatment response remain limited. Chemical exchange saturation transfer (CEST) MRI, sensitive to metabolic and treatment-induced changes, has shown promise in identifying early, standard, and late progressors prior to treatment initiation. In this study, we aimed to characterize the molecular profiles of CEST-defined subgroups and develop predictive models for treatment stratification. A cohort of 180 patients with primary, IDH wild-type glioblastoma underwent CEST imaging at multiple timepoints. Tumor and matched normal tissue were profiled using whole genome sequencing, enzymatic methyl-seq, and RNA-seq. Clinical variables (age, sex, MGMT promoter status, extent of resection, ECOG score) were collected. Kaplan-Meier and Cox proportional hazards analyses were used to evaluate clinical and molecular correlates of progression-free survival (PFS).To integrate imaging, clinical, and molecular data, we developed a graph neural network (GNN)-based model for survival prediction and risk stratification. Our architecture incorporates pathway-level pooling and cross-modal attention, capturing synergistic patterns between RNA expression and DNA methylation. The model demonstrated superior concordance index and risk group separation compared to published ML models. SHAP-based interpretability analyses revealed that genes involved in DNA repair, glucose transport, and arginine metabolism were among the top contributors to risk predictions. Early progressors had a median PFS of 142 days versus 832 days for late progressors (p<0.0001), with distinct gene expression and epigenetic signatures. A multigene signature derived from the GNN was prognostic of both PFS and overall survival. Collectively, our findings suggest that integrating multi-omic profiles with deep learning on radiogenomic features may enable early, individualized treatment adaptation in glioblastoma.