Tumor–neuronal hybrid programs reveal a spatially adaptive malignant state in glioblastoma

Abstract

Objective Glioblastoma (GBM) exhibits marked cellular plasticity, including partial acquisition of neuron-associated transcriptional features. We tested whether public GBM spatial and single-nucleus datasets reveal a tumor-neuronal program with spatial organization and distinction from canonical proneural-like identity. Methods We re-analyzed public multi-region GBM spatial transcriptomic data from Greenwald et al. and used healthy cortex data from Ravi et al. as a non-malignant reference. Tumor-neuronal (TN) classifications were defined by concurrent enrichment of malignancy-associated and neuronal-associated gene programs using permutation-derived empirical thresholds. Classification stability was assessed by bootstrap resampling, repeated threshold estimation, and Jaccard overlap. Greenwald-derived CNA/malignancy scores were used to evaluate tumor-associated signal in TN-classified spatial units. External validation was performed in an independent Wang malignant GBM-cell/single-nucleus dataset, including held-out TN-specific scoring after excluding genes used in the original scoring programs. Exploratory bulk-cohort analysis was performed in TCGA GBM. Results TN-classified spatial units were detected across tumor regions at higher frequency than healthy cortex and showed regional organization within the spatial discovery specimen. They expressed synaptic and vesicle-associated genes, including SYT1 and STMN2, while retaining malignant/glial-associated features such as GFAP, VIM, and CLU. Mature neuronal subtype markers were less enriched, supporting partial neuronal mimicry rather than lineage conversion. TN-associated genes formed a connected protein-interaction network centered on synaptic, neurite-associated, and cytoskeletal proteins. CNA/malignancy-score analysis showed tumor-associated signal in TN-classified spatial units. In the Wang malignant GBM-cell validation dataset, TN-classified cells/nuclei retained held-out TN-specific enrichment relative to non-TN and proneural-high non-TN malignant cells/nuclei. Conclusion Public spatial and single-nucleus GBM datasets identify a tumor-associated neuronal mimicry program that is spatially organized in a GBM discovery specimen and detectable in an independent malignant-cell validation dataset. This program is proneural-adjacent but not reducible to canonical proneural identity.