A multimodal AI framework integrating spatial omics and radiomics for recurrence prediction in glioblastoma

Glioblastoma (GBM) has a dismal prognosis, yet times to recurrence vary substantially. We integrated multimodal data to identify tumor- and microenvironment-level features associated with timing of recurrence.

Methods:

Primary GBMs tumors with documented time to recurrence were analyzed by Xenium spatial transcriptomics (ST). ST data were processed with Seurat, manually annotated, and organized into five reproducible spatial niches. Differential expression was performed within tumor-dense niches comparing that shortest recurrence (SR) and those with the longer time to recurrence (LR). ST embeddings were extracted using the graph attention network based Novae foundation model (FM), while embeddings from the pathology whole-slide Hematoxylin and Eosin (H&E) images were obtained using the vision transformer-based UNI FM and reviewed by a neuropathologist. CODEX spatial proteomics (SP) embeddings will be generated using the vision transformer-based KRONOS FM, and pre-operative MRI features will be incorporated via a PyRadiomics pipeline. Multimodal embeddings from all platforms are being used to train an interpretable tree-based survival model predicting time to recurrence.

Results:

ST analyses show that there are marked microenvironmental differences between SR and LR tumors. SR tumors showed malignant-cell infiltration across nearly all spatial niches, indicating disruption of immune and stromal architecture. LR tumors retained distinct vascular, immune, and endothelial-rich regions, reflecting maintained compartmentalization; H&E review confirmed these patterns. Malignant-cell composition also diverged: SR tumors were dominated by AC-like cells (83%) with minimal MES-like (1%), OPC-like (1%), and NPC-like (15%) states. LR tumors were more heterogeneous, with balanced AC-like (38%) and MES-like (36%) fractions and increased OPC-like (16%) and NPC-like (9%) representation. Pseudo-bulk differential expression identified 367 genes (313 up in SR; 54 in LR). SR tumors overexpressed CD38, HLA-DQA1, CX3CR1, TMEM119, and P2RY12, indicating microglial activation and inflammatory signaling. LR tumors upregulated POSTN, LOX, ACKR1, VEGFA, and CAV1, consistent with extracellular matrix organization, angiogenesis, and vascular remodeling. SR GBM exhibits widespread tumor infiltration and immune activation, whereas LR tumors preserve structured vascular-immune niches and more diverse malignant-cell states.

Conclusion:

ST, SP, pathology slide H&E imaging, and MRI radiomics provides complementary insights into the cellular and microenvironmental determinants of GBM recurrence. While multimodal model training involving embeddings from all the modalities is ongoing, the preliminary findings highlight the importance of spatial organization of the tumor-immune-vascular interface as a key determinant of recurrence timing.