A prognostic neural epigenetic signature in high-grade glioma

The study investigates whether an epigenetically defined neural signature within IDH-wild-type high-grade gliomas (especially glioblastoma) reflects malignant neural-lineage features that integrate into neuron-to-glioma networks and whether this signature predicts clinical outcomes. Prior work has shown neuron-to-glioma synapse formation and tumor-driven hyperexcitability, with higher functional connectivity linked to poorer survival. High-grade gliomas contain malignant and nonmalignant cells; thus, bulk DNA methylation deconvolution can infer cellular composition. The authors hypothesized that an epigenetic neural signature would stratify glioblastomas into clinically relevant subgroups (low- vs high-neural), differing in molecular programs, microenvironmental architecture, connectivity, and prognosis, and that the signature would generalize to diffuse midline glioma.

Previous studies established that glioma cells integrate into neural circuits via glutamatergic synapses and remodel neuronal activity, promoting tumor progression and epilepsy. High functional connectivity in glioblastoma correlates with poor survival. Distinct malignant cell states (connected vs unconnected) and neural circuit engagement have been described, including callosal projection neuron-driven infiltration. Epigenome profiling has successfully classified CNS tumors, and reference methylation atlases enable deconvolution of cell-type contributions. These foundations support assessing an epigenetic neural signature as a clinically meaningful classifier and biomarker.

• Cohorts: Combined epigenetically profiled CNS tumor datasets from Capper et al. and institutional cohorts, plus healthy tissue references. Primary glioblastoma clinical cohort: 363 IDH-wild-type patients treated with surgery and standard chemoradiotherapy; external validation in TCGA-GBM (n=187). Pediatric/adolescent H3K27-altered diffuse midline glioma cohort combined from institutional and published datasets (n=72 total).
• Epigenetic neural signature and deconvolution: Applied Moss et al. cell-type methylation reference to Illumina 850k/450k arrays using non-negative least squares to estimate neural proportion. Dichotomized glioblastomas using the median neural score cutoff 0.41 (validated across cohorts) into low- and high-neural groups.
• Statistics: Survival analyses (Kaplan–Meier, log-rank) for OS and PFS; multivariate Cox regression for independent prognostic value. Differential methylation analysis identified CpGs separating neural groups. Tumor purity and immune/microglia signatures inferred from methylation data; copy-number and 201-gene NGS profiling assessed genomic differences.
• Multi-omics integration: Paired methylation and RNA-seq from GBM (n=86) analyzed via WGCNA to identify expression modules correlated with neural score; GO enrichment assessed synaptic/neural processes. Mapped module eigengenes to single-cell references (GBMap malignant atlas and Allen human motor cortex) to assign lineage enrichment (neurons, NPC-like, OPC-like, oligodendrocytes).
• Spatial transcriptomics and GNN: Visium FFPE SRT with paired methylation (n=24 tumors) used to construct spatial subgraphs (three-hop neighborhoods). Trained a graph neural network (GIN-based) on 1,000 microenvironments to predict neural score from local transcriptional landscapes (R^2=0.99, F1=0.98). Classified microenvironments as neural high/low and deconvolved cell-type composition (Cell2Location/CytoSpace).
• Proteomics and histology: LC-MS/MS proteomics (n=28) with WGCNA modules; immunohistochemistry for OLIG2, NeuN, Sox2, GFAP.
• Functional connectivity imaging: Magnetoencephalography (MEG) and resting-state fMRI in GBM patients to measure peritumoral connectivity and relate to neural subgroup.
• In vitro/in vivo functional assays: Patient-derived GBM neurosphere cultures assessed for neural signature conservation. Orthotopic xenografts in immunodeficient mice compared low- vs high-neural survival, proliferation (Ki67/HNA), migration/infiltration (corpus callosum spread), synapse puncta colocalization (synapsin/PSD95), and electron microscopy verification of neuron-to-glioma synapses. Neuron–glioma co-cultures assayed for synaptic puncta and EdU proliferation.
• Drug sensitivity: Screened 27 agents in GBM cell lines; measured cleaved caspase-3 and cell/spheroid size.
• Biomarkers: Measured serum/plasma BDNF levels and correlated with neural signature and seizures; isolated extracellular vesicles (EVs) from plasma/CSF; profiled EV-DNA and cfDNA methylation for neural signature detection.
• Spatiotemporal stability: Multi-site stereotactic biopsies per tumor and matched primary–recurrence pairs assessed for neural subgroup consistency.

• Classification and survival: Using a cutoff neural proportion of 0.41, 363 clinical GBM patients stratified into low vs high neural showed significantly shorter survival for high-neural: internal cohort OS median 14.2 vs 21.2 months (P<0.0001) and PFS median 6.2 vs 10.0 months (P=0.02). External TCGA cohort OS median 12.0 vs 17.1 months (P<0.01). Neural subgroup was an independent prognostic factor: OS OR 1.96 (95% CI 1.45–2.64, P<0.01); PFS OR 1.51 (95% CI 1.13–2.02, P<0.01). Other lymphoid/myeloid signatures did not associate with survival.
• Epigenetic/transcriptomic features: High-neural tumors were hypomethylated at CpGs in invasivity, neuronal synapse formation, and trans-synaptic signaling gene sets; synapse-related genes were upregulated. WGCNA identified modules (green, cyan, midnight blue) strongly correlated with neural score (R^2 up to 0.67), enriched for synaptic functions (e.g., GRIN3A, SYT4, SNAP25) and neuronal differentiation (NEUROD2), and calcium-dependent adhesion.
• Cell states and lineage: Module projections enriched in neural lineage cells (healthy neurons, malignant NPC-like, OPC-like) and oligodendrocytes. Neural signature correlated positively with DNA tumor purity and negatively with microglia and immune signatures. Non-reference deconvolution showed higher stem/progenitor state in high-neural GBM (28.05%) vs all new GBM (17.31%) and low-neural (14.14%). No major copy-number differences between subgroups; mutations more frequent in high-neural for PIK3CA (0% vs 15.0%) and TP53 (9.23% vs 31.67%).
• Spatial architecture and GNN: GNN predicted neural scores from spatial transcriptomics with R^2=0.99; 41.2% of samples contained mixed neural high/low microenvironments, yet most samples had a predominant type (95.8%). High-neural microenvironments were enriched for NPC-like, astrocyte-like, OPC-like cells and oligodendrocytes.
• Proteomics/histology: High-neural GBM proteomes enriched for synaptic transmission proteins and OPC-/NPC-/astrocyte-like characteristics; higher fraction of OLIG2+ tumor cells in high-neural; sparse NeuN+ infiltration comparable between groups.
• Network integration and connectivity: High-neural GBM xenografts showed increased colocalization of pre/post-synaptic puncta (P=0.0008) and higher EM-verified glioma–neuron synapses (P=0.008). In neuron co-cultures, high-neural cells had increased PSD95–synapsin colocalization (P=0.0007). Clinically, high-neural subgroup had higher peritumoral connectivity on MEG and rs-fMRI (MEG-based high vs low functional connectivity groups differed in neural signature, P=0.0327; rs-fMRI peritumoral connectivity correlation with neural signature P=0.05; group difference P=0.0416). No increased tumor-to-contralateral connectivity.
• Model translatability and growth: Neural signature conserved in 82.3% of cultures relative to original tumors. High-neural tumors led to shorter mouse survival in internal PDX (median 57.5 vs 114 days, P=0.0009) and external datasets (median 58 vs 74 days, P=0.001). Higher proliferation in vivo (Ki67/HNA, P=0.00819) and in neuron co-cultures (EdU increased for high-neural with neurons, P=1.72×10^-5). Greater migration in vitro (P=0.0115) and increased in vivo spread into corpus callosum (P<0.0004).
• Stability: Among 34 patients (3–7 spatial biopsies each), 67.6% were purely low or high neural; 29.4% predominantly one group. In 39 paired primary–recurrence cases, 79.5% retained the same neural subgroup.
• Therapeutic implications: Drug screen trends suggested potential sensitivity differences (e.g., lomustine, JNJ10198400, cyclosporine in high-neural; talazoparib in low-neural) but no statistically significant differences overall. Extent of resection (EOR) benefit differed by subgroup: in low-neural GBM, near-GTR/GTR conferred significant survival benefit vs partial resection (P<0.001), whereas near-GTR benefit was not observed in high-neural (multivariate analyses confirmed). MGMT promoter methylation improved survival in both, with a larger effect in low-neural (median OS difference 12 months; P<0.0001).
• Biomarkers: Serum/plasma BDNF levels were higher in high-neural GBM than low-neural, meningioma, and healthy donors; BDNF correlated with neural signature (P<0.01, R^2=0.28) and inversely with immune signature, and associated with seizures at diagnosis (P=0.02) and follow-up (P<0.001). Neural signature detectable in circulating EV-DNA (but not cfDNA), with higher incidence in high-neural tumors; EV counts correlated with neural signature (P=0.01). CSF EV-DNA neural scores more comparable to tissue than plasma.
• Generalization to DMG: In H3K27-altered DMG (n=72), high-neural tumors exhibited hypomethylation of trans-synaptic genes, stem/glial cell state association, increased synaptic gene expression (P=0.01), and significantly worse survival (median OS 8.5 vs 16.5 months, P<0.01).

The findings validate an epigenetically defined neural signature as a stable, clinically relevant marker in high-grade gliomas. High-neural glioblastomas exhibit synaptogenic programs, malignant neural lineage (NPC/OPC/astrocyte-like) characteristics, lower immune infiltration, stronger integration into neuron-to-glioma networks, and worse outcomes. The signature remains conserved across space, time, and preclinical models, supporting its robustness over more plastic transcriptional states. Increased functional connectivity and elevated BDNF levels provide convergent clinical correlates of neural integration. The subgroup-specific impact of surgical resection suggests that high-neural tumors may require maximal resection to derive survival benefit, while low-neural tumors benefit from near-GTR. Detection of the neural signature in EV-DNA and correlation with serum BDNF suggest minimally invasive biomarkers for stratification. Extension to H3K27-altered DMG underscores broader relevance across high-grade gliomas. These results support using neural signature stratification to guide prognosis and potentially tailor neuroscience-focused therapies (e.g., targeting synaptic signaling, BDNF–TrkB axis) and surgical planning.

Defining a high-neural epigenetic signature in IDH-wild-type gliomas identifies a malignant NPC/OPC/astrocyte-like state linked to synaptic integration, high functional connectivity, and poor survival. The signature is stable across space and time, conserved in models, and detectable via circulating biomarkers (EV-DNA, serum BDNF). Clinically, neural classification refines prognosis and suggests subgroup-specific benefits from extent of resection, with high-neural tumors likely requiring maximal resection. Validation in H3K27-altered DMG indicates generalizability. Future work should develop imaging-based predictors of the neural signature (MEG/rs-fMRI), evaluate targeted disruption of neuron-to-glioma synapses and BDNF–TrkB signaling, and advance liquid biopsy approaches for monitoring and stratification.

• The neural signature reference was derived from a single cortical neuron reference built from three IDAT files; broader references could improve discrimination and generalizability.
• Although analyses addressed neuronal contamination (purity correlations, NeuN+ scarcity, cell culture preservation, reference-free deconvolution), residual confounding by nonmalignant neurons cannot be entirely excluded.
• Drug sensitivity trends did not reach statistical significance; larger panels and matched in vivo validation are needed.
• EV-DNA neural scores in plasma were markedly lower than tissue and CSF; sensitivity and standardization for liquid biopsy applications require refinement.
• Heterogeneity remains, with some tumors exhibiting mixed microenvironments; while subgroup stability was high, a minority switched at recurrence.
• Affiliation of mutations (e.g., PIK3CA, TP53) to specific functional effects within the neural subgroup needs mechanistic validation.