AI Maps Glioma Ecosystems and Exposes a Macrophage-Driven Vulnerability in Glioblastoma

A team at Chinese PLA General Hospital and Hong Kong Metropolitan University has combined artificial intelligence with single-cell and spatial multi-omics to define three clinically relevant ecosystem states in glioma. The work, accepted for publication in Frontiers in Cell and Developmental Biology, integrated public multi-omics datasets with the researchers’ own single-nucleus RNA sequencing from 12 tumors, spatial transcriptomics from 6 IDH-wildtype glioblastomas, multiplex immunofluorescence staining across 138 formalin-fixed paraffin-embedded cases, and laboratory co-culture assays pairing glioma cells with macrophages.

Using interpretable machine learning to fuse the single-nucleus, spatial and tissue-level data, the team identified three distinct glioma ecosystem states. One state stood out for its hostile composition. Termed the hypoxic myeloid-mesenchymal state, it showed the highest scores for mesenchymal-like tumor programs, hypoxia signaling, myeloid cells expressing the secreted proteins SPP1 and APOE, and temozolomide-resistance signatures. It also displayed the lowest density of CD8-positive T cells and the shortest measured distance between tumor cells and myeloid cells, indicating unusually close physical contact between the two compartments.

The prognostic weight of this state was tested in a cohort of 96 IDH-wildtype glioblastomas. After statistical adjustment for major clinical and molecular covariates, the hypoxic myeloid-mesenchymal state independently predicted shorter progression-free survival, with a hazard ratio of 2.19 (95% confidence interval 1.22 to 3.92), and shorter overall survival, with a hazard ratio of 2.07 (95% confidence interval 1.12 to 3.81). Patients whose tumors were classified in this state therefore faced roughly double the risk of progression or death compared with otherwise similar patients in the cohort.

The spatial architecture captured by the analysis links the aggressive tumor program to a specific immune niche. In the hypoxic myeloid-mesenchymal state, myeloid cells rich in SPP1 and APOE cluster tightly around tumor cells, while CD8-positive T cells are sparse in the vicinity. According to the authors, SPP1 signals through CD44 on tumor cells and activates downstream FAK and SRC pathways, a cascade they describe as an actionable therapeutic vulnerability embedded in the tumor’s immune microenvironment.

To probe that vulnerability experimentally, the researchers cultured glioma cells under macrophage-conditioned conditions and treated them with defactinib, a FAK inhibitor, combined with temozolomide. The combination reduced cell viability by 56.0%, colony formation by 61.3% and cell migration by 57.8% relative to controls. These co-culture results suggest that blocking the FAK-SRC signaling axis may reduce the malignant behavior of glioma cells in a myeloid-rich environment, although the findings remain at the preclinical cell-culture stage and do not establish clinical benefit.

The methodological design combines data types that capture different scales of tumor biology. Single-nucleus RNA sequencing resolves gene expression in individual tumor and immune cells, spatial transcriptomics preserves the physical coordinates of that expression, and multiplex immunofluorescence across 138 archived cases extends the analysis to routine pathology material. This combination is what allowed the algorithm to quantify relationships such as tumor-to-myeloid distance rather than merely counting cell types, and to define ecosystem states that can, in principle, be assessed in standard clinical specimens.

The authors frame the work as a step toward individualized therapeutic stratification in glioblastoma. By showing that the immune niche is a spatially organized feature associated with temozolomide resistance and survival, the study argues that patients could be characterized by ecosystem state in addition to genomic alterations. The identified SPP1-CD44/FAK-SRC signaling axis provides a concrete molecular link between the myeloid niche and the treatment-resistant phenotype observed in the hypoxic myeloid-mesenchymal state.

Several caveats temper the findings. The single-nucleus sequencing cohort comprised 12 tumors and the spatial transcriptomics cohort 6 glioblastomas, small numbers that the larger 138-case immunofluorescence analysis partially offsets but cannot fully replace. The defactinib results come from cell-culture assays, not animal models or patients, and the hazard ratios derive from a single institutional cohort of 96 patients. The article’s final formatted version is still pending publication, so conclusions may be refined. Even so, the study offers a measurable, prognostic and potentially druggable map of how tumor cells and myeloid cells are organized within glioblastoma tissue.

Subject of Research: AI-driven single-cell and spatial multi-omics classification of glioma ecosystem states and immune-niche-dependent therapeutic vulnerabilities in glioblastoma

Article Title: Defining glioma ecosystem states with AI-driven single-cell and spatial multi-omics: immune-niche reprogramming and therapeutic vulnerabilities

Article References: Qin, Z., Zhang, Y., Wen, B., Li, J., Chen, Z., Li, X., & Kong, D. (2026). Defining glioma ecosystem states with AI-driven single-cell and spatial multi-omics: immune-niche reprogramming and therapeutic vulnerabilities. Frontiers in Cell and Developmental Biology, 14, Article 1943227. https://doi.org/10.3389/fcell.2026.1943227

Image Credits: AI Generated

DOI: 10.3389/fcell.2026.1943227

Keywords: glioblastoma, glioma, spatial transcriptomics, single-nucleus RNA sequencing, machine learning, tumor microenvironment, macrophages, SPP1, FAK inhibitor, temozolomide, prognosis, defactinib