Unveiling Transcriptional Heterogeneity and Discovering Footprints in Pediatric Medulloblastoma via Spatial RNA-Seq and Machine Learning

Medulloblastoma is the most common malignant pediatric brain tumor and is classified into four molecular subgroups: WNT, SHH, Group 3, and Group 4. Accurate characterization and classification of these subtypes are critical for prognosis and personalized treatment. In this study, we integrate spatial transcriptomics, machine learning, and explainable artificial intelligence to comprehensively analyze the spatial and transcriptional heterogeneity of the most prevalent and aggressive medulloblastoma subtypes (SHH, Group 3, and Group 4) in Mexican pediatric patients. By implementing a multiscale framework, we uncover extensive intra- and intertumoral variability that challenges traditional molecular classification. Our approach reveals subtype-specific gene expression signatures, identifies a core set of robust biomarkers and highlights distinct spatial patterns of tumor organization, including invasive and metabolically active niches. Notably, we demonstrate that Group 3 tumors, despite being the most aggressive, exhibit the lowest transcriptional heterogeneity, suggesting a streamlined malignant program centered on invasion and plasticity. These findings underscore the scope of bulk molecular profiling and support the implementation of spatially informed diagnostic tools and personalized therapeutic strategies that account for the complex architecture and functional diversity within MB tumors.