Machine learning-based identification of extracellular matrix-related prognostic subtypes in SHH-activated medulloblastoma

Medulloblastoma (MB) is the most common malignant pediatric brain tumor, with the Sonic Hedgehog (SHH) subgroup exhibiting heterogeneous clinical outcomes. The extracellular matrix (ECM) critically modulates tumor behavior, yet its contribution to SHH-activated MB heterogeneity remains poorly defined. We integrated bulk microarray expression profiles from GSE85217 and single-cell RNA-sequencing data (GSE155446) from SHH-activated MB. ECM-related gene sets were retrieved from MsigDB. Non-negative matrix factorization (NMF) was applied to microarray data to delineate ECM subtypes, followed by weighted gene co-expression network analysis (WGCNA) and differential expression analysis to identify core genes. A prognostic model was constructed using 101 machine-learning algorithm combinations. Single-cell data were processed with Seurat and Harmony, cell-type annotations were transferred via the Scissor algorithm, and cell–cell communication was inferred with CellChat. Two distinct ECM-associated subtypes (C1 and C2) were identified within SHH-activated MB. C2 displayed a poor prognosis (log-rank P = 0.00015. The top 50 ranked by the MCC algorithm were used to build a Ridge-based prognostic signature that independently predicted overall survival (P < 0.0001). Single-cell mapping showed that the adverse C2 signature was primarily harbored by complement-activated myeloid cells. These cells engaged in strong MHC-II-mediated interactions with dendritic-like and M2-like myeloid cells, implicating myeloid-driven ECM remodeling in SHH-activated MB progression. ECM heterogeneity defines clinically relevant SHH-activated MB subtypes. A myeloid-centric, complement-activated microenvironment underpins the aggressive C2 subtype, offering tractable therapeutic targets to overcome current treatment limitations.