Integration of Single-Cell Multi-omics and Machine Learning Identifies Predictive Biomarkers for Immunotherapy Response in Glioblastoma

Glioblastoma (GBM) remains one of the most lethal primary central nervous system malignancies, characterized by profound intratumoral heterogeneity, profound immunosuppression, and resistance to immune checkpoint blockade (ICB). Despite aggressive multimodal interventions, therapeutic response rates to anti-PD-1/PD-L1 regimens remain dismal. In this study, we developed an integrative computational framework combining single-cell RNA sequencing (scRNA-seq) and single-cell assay for transposase-accessible chromatin sequencing (scATAC-seq) data from 38 primary and recurrent GBM cohorts to identify cellular and epigenetic determinants of immunotherapy efficacy. By applying an ensemble machine learning architecture incorporating extreme gradient boosting (XGBoost) and graph neural networks, we resolved a distinct subpopulation of immunosuppressive SPP1⁺/CD163⁺ tumor-associated macrophages tightly coupled with exhausted TCF7^lowHAVCR2^high CD8⁺ T cells through aberrant ligand-receptor signaling axes. Furthermore, our machine learning pipeline prioritized an 8-gene transcriptional and chromatin accessibility signature—designated the Glioblastoma Immunotherapy Predictive Index (GIPI)—that robustly predicted ICB clinical response with an area under the receiver operating characteristic curve (AUC) of 0.912 in an independent validation cohort. Epigenetic motif enrichment revealed that the AP-1 and RUNX1 regulons drive the chromatin remodeling underlying therapy-refractory cellular states. Our findings demonstrate that integrating single-cell multi-omics with explainable machine learning uncovers actionable regulatory mechanisms and provides a clinically translatable biomarker panel for stratifying GBM patients undergoing immunotherapy.