ID #465 Immune gene signatures predict B7-H3 expression in diffuse midline glioma and medulloblastoma

Abstract

Diffuse midline glioma, H3K27-altered (DMG) and medulloblastoma (MB) are aggressive central nervous system (CNS) tumors needing novel, targeted therapies. B7-H3 is an attractive immunotherapy target due to its surface expression on these tumors; therefore, it has been evaluated in chimeric antigen receptor (CAR) T trials. However, whether specific immune gene signatures correlate with B7-H3 status (and reveal tumor-specific microenvironmental differences or resistance mechanisms) have not been systematically defined and would be critical to know prior to B7-H3 CAR T cell trial enrollment. Here, we developed a machine learning (ML) framework to identify B7-H3-related immune gene signatures and to characterize immune evasion mechanisms.

We analyzed bulk RNA-seq data from 895 patients (433 DMG, 462 MB) available from the OpenPedCan database (PedcBioPortal). A novel five-stage feature selection funnel framework, integrating logistic regression, random forest, SHAP analysis, spearman correlation and data integrity check identified six key immune gene predictors of high/low B7-H3 transcription. Models were evaluated using ROC-AUC and were validated against published literature.

Despite shared B7-H3 overexpression, DMG and MB exhibited fundamentally different immune signatures. In DMG, high B7-H3 tumors demonstrated coordinated upregulation of a “triple shield”: CD24, CD47, and PVR, contributing 48.5% of model’s predictive power and suggesting multi-layered immune blockade of phagocytic and NK cell clearance. In contrast, high B7-H3 MB tumors exhibited an “immune desert” phenotype with high PVR, CD47, and LAG3, with low PTPRC, indicating immune exclusion. Low B7-H3 tumors in both types showed elevated exhaustion markers HAVCR2, CD274 with high immune infiltration. RF models achieved high predictive accuracy (DMG: AUC = 0.993; MB: AUC = 0.952).

These results demonstrate that B7-H3 expression may be associated with tumor-specific immune evasion architectures, providing a foundation for tumor-specific immunotherapy stratification and combinatorial treatment strategies targeting B7-H3 alongside tumor-specific immune checkpoints in pediatric CNS tumors.