Decision support for treatment of patients with glioblastoma: A systems biology approach enabled by explainable AI.

2081 Background: Glioblastoma (GBM) patients (pts) have median survival of 15 months and 5-year survival <7% with standard therapy. Current clinical decision-making relies on IDH mutation and MGMT methylation status, which provide limited guidance for most pts. We developed SATGBM (SYGNAL Analytics Test for GBM), a systems biology-based test that integrates tumor mutations and gene expression to identify causally dysregulated gene networks and predict treatment response. Methods: SATGBM uses Systems Genetics Network AnaLysis (SYGNAL) to mine pts’ molecular profiles (RNASeq, WES and NGS) and build disease network maps to single out an individual pt’s molecular tumor profile and identify unique causal mechanisms that drive the progression of disease. Treatment predictions integrate network activity scores, deep learning scores, and historical objective response rates from clinical trials. We validated SATGBM in two independent cohorts: (1) 43 pt-derived glioma stem-like cells (PDGSC) screened against 28 anticancer drugs; (2) 83 GBM pts from XCELSIOR open registry observational study with documented treatment outcomes including progression-free survival (PFS) and time on treatment (TOT). Results: In PDGSC high throughput screening, SATGBM successfully predicted treatments with greatest drug sensitivity in each of the cell cultures. SATGBM predictions separated responders vs non-responders: median IC50 2.70×10⁻⁷ vs 1.00×10⁻³, Mann–Whitney p=2.1×10⁻⁴¹, Pearson r = -0.38, AUC = 0.79. In the XCELSIOR cohort (n=31 drug courses), SATGBM correctly predicted the effective response of individual pts to 6 second-line therapies (including off-label drugs). Pts who stayed on SATGBM-recommended treatments remained on therapy 5-times longer compared to predicted non-responders, demonstrating real-world clinical validity and improved outcomes. SATGBM predictions yielded median TOT of 260 vs 50.5 days, p=0.0021; sensitivity/specificity were 60.0% / 95.2% and AUC = 0.79 (benefit = ToT ≥120 days). Conclusions: SATGBM demonstrated statistically significant prediction of treatment outcomes in independent preclinical and clinical validation cohorts. Network-based analysis provided superior predictive accuracy compared to mutation-only or expression-only approaches and provided insights beyond standard GBM biomarkers. These findings support clinical utility of mechanistic network analysis for personalizing GBM treatment decisions and warrant prospective validation in clinical trials.