© The Authors, published by EDP Sciences, 2026

Single-cell genomics has dismantled the idea that every glioblastoma (GBM) belongs neatly to one of four bulk‐expression “TCGA subtypes” and instead reveals a fluid landscape in which multiple developmental programs, lineage origins, and injury-response signatures intermingle within the same tumour. Pioneering work by Patel and colleagues first uncovered profound intratumoural heterogeneity of GBM in gene expression, signalling pathway activity, and stemness states [1]. Building on extensive single-cell studies, today’s GBM state model recognises four flexible malignant programs: Neural-progenitor-like (NPC; or radial glia-like, Rgl-like), oligodendrocyte-progenitor-like (OPC), astrocyte-like (AC), and mesenchymal-like (MES). We refer to these malignant transcriptional states as developmental malignant programs, to emphasise that GBM programmes reflect distorted neurodevelopmental lineages and remain plastic under genetic and microenvironmental influence. The balance among these programmes and their interconversion is shaped by genetic drivers and local microenvironmental cues rather than by fixed “box” subtypes [2]. Subsequent single-cell studies and conceptual reviews have refined this picture into complementary classification frameworks: Plastic cell-state continua, lineage-of-origin hierarchies, and dynamic transition maps, all of which anticipate spatial patterning and diverge from the static tumour-level taxonomy that arose from bulk TCGA data [3, 4]. Consistent with that expectation, integrative spatial transcriptomics (10x Visium), spatial proteomics (CODEX), and computation now show that these malignant states are locally enriched and assemble into a reproducible, hypoxia-graded five-layer architecture that defines the spatial organization of GBMs [5]. These advances pave the way toward constructing a spatially resolved, multi-omics atlas that offers a holistic view of GBMs.

The bulk-expression legacy: TCGA’s four transcriptional subtypes

The TCGA landmark analysis partitioned GBM into Proneural (PN), Neural (NE), Classical (CL), and Mesenchymal (MES) subgroups based on bulk transcriptome profiles, and each subgroup is associated with characteristic driver lesions [6]. The alterations of PDGFRA and point mutations in IDH1 were observed in the Proneural class, which also exhibited high expression of oligodendrocyte development markers such as OLIG2 and SOX genes. Classical tumours typically harbour high-level EGFR amplification and upregulate cell cycle regulators such as CCND1, often with suppression of the TP53 pathway. The feature of the Mesenchymal class was lower expression levels of NF1, and enrichment of mesenchymal and inflammatory signatures, including CHI3L1 (YKL-40), CD44, and STAT3 activation. The Neural class was initially defined by the expression of neuronal genes such as NEFL, SYT1, and GABRA1 [6]. However, follow-up work showed that this signal is not tumour-intrinsic, and this Neural class largely reflects admixture of normal brain tissue rather than a malignant programme [7]. Nevertheless, the TCGA classification remained limited because of its high reliance on bulk RNA profiling, which represents the average gene expression of all malignant and non-malignant cells, thereby concealing the intratumoural heterogeneity and immune interactions [8]. Key elements of developmental hierarchy and transcriptional diversity that are evident in single-cell datasets remain obscured in bulk analyses [9]. Moreover, such bulk profiling depicted only a static snapshot, overlooking the dynamic plasticity and co-existence of multiple transcriptional programs within individual tumours [10, 11]. Even so, TCGA subtypes remain clinically useful because they summarize the dominant signal in a surgical specimen and correlate with patient survival and therapy response.

Studies have shown that the TCGA classification system aids in understanding tumour biology and potential therapeutic responses; however, its independent value in predicting long-term prognosis is limited. The relatively “favourable prognosis” associated with the Proneural subtype is primarily attributable to cases with IDH mutations, whereas the Mesenchymal subtype often indicates poor outcomes, and the Neural subtype may reflect normal tissue components. In contrast, IDH status, 1p/19q co-deletion, G-CIMP methylation subtype, TERT promoter mutations, and WHO grade/molecular diagnosis more robustly stratify patient outcomes: IDH-mutated cases, particularly those with 1p/19q co-deletion, exhibit the best survival; G-CIMP-high is associated with better prognosis than G-CIMP-low; while IDH-wild-type tumours with TERT mutations or classic glioblastoma molecular features show the worst prognosis. Therefore, clinical evaluation should prioritize IDH status, 1p/19q co-deletion, G-CIMP, TERT, and WHO grading, with transcriptomic subtypes serving as supplementary information to infer driver pathways and potential differences in immune or treatment responses [6, 12, 13].

A cell-state framework: Four developmental programmes with high plasticity

Single-cell RNA-seq overturned the subtype dogma by demonstrating that every GBM contains mixtures of NPC-, OPC-, AC-, and MES-like states regardless of its bulk identity [14]. In IDH wildtype glioblastoma, the NPC-like state is defined by high levels of neuro-developmental genes such as SOX2 and ASCL1, reflecting a proliferative, stem-like phenotype [2]. This state shows high expression of cell cycle regulators (e.g., MKI67, TOP2A) and is typically enriched in the tumour core, often co-localizing with high Ki67 immunoreactivity. The OPC-like state cells are highly proliferative and tumour-propagating, especially in H3K27M-gliomas, which often linked to PDGFRA overexpression [9]. These cells also express markers such as OLIG1, OLIG2, and NG2 (CSPG4), and transcriptionally resemble oligodendrocyte precursor cells found in normal neurodevelopment. They are thought to represent a “stem-like” reservoir that seeds intratumoural heterogeneity. The AC-like programme is associated with a more differentiated and less proliferative phenotype and is characterised by EGFR expression. These cells often occupy the tumour margin, express astrocytic markers such as GFAP and AQP4, and are thought to arise from lineage progression of NPC/OPC-like states under normoxic and less inflammatory conditions. The MES-like state expresses CHI3L1 (YKL-40), FN1, and CD44, which present the environments of hypoxia, stress, and increased glycolysis, respectively [2]. This state is also enriched for inflammatory signalling and exhibits upregulation of transcription factors such as STAT3 and CEBPB. MES-like cells are frequently found in peri-necrotic, hypoxic niches and have been implicated in treatment resistance and immune evasion.

Additionally, the distributions of cell states are much closer to the “functional unit” than bulk subtypes. The majority of GBMs were predominantly composed of NPC-like plus OPC-like cells, or of AC-like plus MES-like cells, reflecting lineage coherence and developmental coordination across tumour regions. Among these, the combination of OPC-like and NPC-like corresponds to the PN subtype in TCGA, which is characterized by elevated expression of neural stem and progenitor cell genes, lower immune infiltration, and generally better prognosis compared to the MES-like enriched GBMs [2].

Moreover, genetic lesions bias these proportions. For example, EGFR amplification favours AC-like cells, whereas NF1 loss drives MES-like programmes, yet environmental factors such as hypoxia or cytokines can push cells across state boundaries, emphasizing plasticity over lineage lock-in [15]. Suvà and Tirosh consequently recast the “glioma stem-cell model” as a dynamic state machine rather than “quiescent stem cells”, in which stemness is a transient property rather than a fixed identity [16]. Using large time-series barcoding, Schiffman et al. quantified the probability that one malignant state converts into another and found a directional bias from NPC/OPC towards AC/MES under stress, implying quasi-developmental flows rather than random walks [15]. Their mathematical model segregates cell phenotypes into three axes: Heritability, plasticity, and transition rate, offering a quantitative add-on to Neftel’s qualitative map. In glioblastoma (GBM), different cell types exhibit distinct response characteristics to the tumor microenvironment. Neural precursor-like (NPC-like) and oligodendrocyte precursor-like (OPC-like) cells display high plasticity and are susceptible to regulation by hypoxia, inflammatory factors, and growth factors (such as PDGF and EGFR signaling), thereby transitioning into astrocyte-like (AC-like) or mesenchymal-like (MES-like) states [2]. Under hypoxia and the influence of angiogenic factors (e.g., VEGFA), AC-like cells demonstrate enhanced angiogenic and migratory capacities, while molecules such as JUN and CXCR4 help maintain their stem cell-like properties [17]. MES-like cells are sensitive to hypoxia, inflammation, and immune-related factors; loss of NF1 and signaling molecules such as TNF and IL1A/IL1B can promote their invasiveness and contribute to the formation of an immunosuppressive microenvironment [15]. These observations indicate that GBM cell states are regulated by environmental factors and exhibit high plasticity and heterogeneity.

Single-cell genomics has reframed the glioma-stem-cell paradigm: Instead of a rigid hierarchy, malignant cells behave as a dynamic state-machine whose NPC-like, OPC-like, AC-like, and MES-like programmes interconvert along predictable trajectories, and quantitative bar-coding now measures the speed and direction of those shifts—clarifying why therapies that hit one programme are quickly outflanked as cells “flow” into alternate identities [2, 18]. Far from discarding the familiar TCGA bulk taxonomy (Proneural, Classical, Mesenchymal, Neural), this insight embeds it in a richer, four-dimensional framework that traces how genetic lesions, epigenetic memory and treatment pressure remodel cell states through space and over time, making the bulk subtypes merely a snapshot of whichever programme dominates at biopsy [19, 20].

Lineage-of-origin and injury-response classifications

Beyond the studies of TCGA bulk subtypes and cellular states, numerous studies have further explored the heterogeneity of GBMs by highlighting two additional dimensions: Lineage of origin and injury response programmes.

Malignant GBM cells that induce different programs express distorted versions of them, which implicates their developmental lineage-of-origin, a principle supported by spontaneous and genetically engineered mouse models [3, 21]. For instance, as the H3K27M mutation suppressed the function of PRC2, impairing normal cellular differentiation and leading to the accumulation of immature OPC-like cells, H3K27M-mutant gliomas were largely dominated by OPC-like cells [9]. Subsets of outer radial glia-like cells had also been identified in high-grade glioblastomas, contributing to intratumoural heterogeneity [22]. Similarly, in IDH-mutant diffuse gliomas, single-cell analyses trace differentiation from radial-glia–like precursors toward astrocytic and oligodendroglial lineages, and this lineage progression spans genetically distinct subclones, implying that lineage determination can be partly decoupled from genotype [23, 24]. Beyond the classical CNS lineages, neural-network projections of single-cell atlases identify a neural crest–perivascular programme (NCperiV) marked by extracellular-matrix and perivascular markers such as COL3A1, COL5A2, and NES [25], and this programme is further supported by single-cell data and in vivo models, including spontaneous mouse systems [26, 27]. Experimental lineage tracing further shows that gliomagenesis can mimic a neural-crest-like injury response, with SOX10-positive neural crest-derived cells seeding invasive fronts and mesenchymal transition [26]. Meanwhile, injury and stress-response programmes, including hypoxia signalling, heat-shock responses, and the unfolded protein response, recur across GBM and other cancers, reflecting both intrinsic stress and microenvironmental cues [28]. Driven by factors like AP-1 (FOS, JUN), these programs are broadly conserved across tumour types and reflect both intrinsic and microenvironmental stress signals [3, 29]. In this context, the MES-like programme often reads as an injury-response state, with up-regulation of CD44, CHI3L1 (YKL-40), and ANXA2; it can be initiated by NF1 loss and reinforced by inflammatory signals, including TNF-α and cues from macrophages/microglia [2, 7]. Spatial and functional studies show that MES-like reprogramming is enriched in hypoxic, immune-infiltrated territories and tends to become more prominent after treatment, which helps explain the frequent shift toward mesenchymal programmes at recurrence [30, 31].

Thus, in GBM, the term “mesenchymal” spans two distinct entities: A stress environmental programme induced by hypoxia, NF-κB, and myeloid signals (here called MES-transition), and a developmental lineage programme with neural crest and perivascular features (here called MES-NCperiV). These programmes can co-exist within the same tumour but reflect different biology.

Micro-environment-informed extensions

Although most classification schemes above focus on malignant cells, recent single-cell atlases of the blood–brain and blood-tumour barriers (BTB) resolve endothelial and mural sub-lineages that pair preferentially with NPC- or MES-rich tumour territories, adding a vascular dimension to GBM classification and druggability [27, 32]. These datasets map tumour-type-specific endothelium and pericytes across IDH-mutant and IDH-wild-type gliomas and brain metastases, and define BTB states that likely modulate permeability and immune entry, thereby refining predictions of therapy response. In parallel, stromal gene programs carry prognostic and immunologic signals: A five-gene(ITGA5, MMP14, FN1, COL5A1, and COL6A1) CAF/ECM signature stratifies survival and tracks reduced predicted responsiveness to immunotherapy, consistent with the role of collagen-rich extracellular matrix (ECM) in fostering invasion and immunosuppression. Bringing these vascular and stromal taxonomies together with malignant state maps offers a more complete framework for patient stratification. This framework links where a tumour is permeable and immune permissive to which cell programmes dominate, and supports the design of microenvironment-targeted combinations alongside state-directed therapies [33].

When single-cell profiles are deconvoluted back to bulk resolution, the relative abundance of NPC/OPC states aligns with Proneural signatures, high MES fractions reproduce Mesenchymal profiles, and Classical tumours are enriched for AC-like cells driven by EGFR signalling. Hence, TCGA categories capture the dominant programme but mask intra-tumour diversity and dynamic transitions. Recognising that plasticity, lineage origin, and niche interaction all contribute to phenotype, recent work argues for a multi-layered nomenclature that records (i) developmental state, (ii) genetic driver, and (iii) spatial context for each tumour cell [3, 25].

What remained unclear until recently was where these states reside, how they interact, and how clonal evolution unfolds in tissue. Spatial multi-omics answers this by pinning flexible programmes to their anatomical niches, showing how local ecology constrains or channels state switching and turning descriptive single-cell maps into geography-aware biology [5, 34]. In glioblastoma, integrated spatial transcriptomics and proteomics reveal that cell states assemble into a reproducible, oxygen-graded, five-layer architecture from necrotic core to infiltrative rim, with hypoxia acting as a long-range organiser; complementary 3-D sampling maps these programmes and clones across the whole tumour [5]. This framework suggests testable links between developmental plasticity and spatial ecology: NPC/OPC programmes should concentrate in oxygenated belts near functional vasculature, this region also harbours a subset of AC-like cells derived from NPCs; MES injury programmes should peak in hypoxic and macrophage-rich zones, the hypoxic environment in this region coexists with the presence of AC-like features [17], and blood–tumour-barrier sub-lineages of endothelial and mural cells should help predict regional drug penetration and immune entry.

In glioblastoma (GBM), hypoxia serves as a critical driver shaping the tumour microenvironment. Insufficient oxygen supply in necrotic areas and densely vascularized yet dysfunctional regions promotes the polarization of tumour-associated macrophages (TAMs) into a hypoxia-responsive phenotype (Hypoxia-TAM). These cells upregulate hypoxia-responsive genes and secrete adrenomedullin (ADM), which disrupts adherent junctions between endothelial cells (ECs), increases vascular permeability, and consequently leads to vascular dysfunction and impaired drug distribution. Meanwhile, hypoxia induces the activation of the HIF pathway in tumour cells, pericytes, and ECs, further promoting angiogenesis, metabolic reprogramming, and immunosuppression [35]. In addition to neovascularization, GBM cells invade normal blood vessels through vascular co-option and interact directly with pericytes via flectopodia, altering pericyte contractility and resulting in abnormal vascular structure and blood flow. This enables tumour dissemination via co-opted vessels even under VEGF inhibition [36]. During this process, ECs not only express molecules such as VEGF and Ang-2 in response to hypoxia or inflammatory signals, thereby promoting angiogenesis, but also induce reduced pericyte coverage and enhanced macrophage infiltration via Ang-2, exacerbating vascular instability and immunosuppressio [37]. These vascular abnormalities not only compromise the homogeneity of drug distribution but also enhance treatment resistance. Blocking ADM has been shown to partially restore vascular stability and improve drug delivery. Clinically, the abundance of Hypoxia-TAMs or the expression level of ADM is closely associated with poor prognosis [38]. Prospective, matched spatial and longitudinal studies can now trace state switching along these gradients in patients and models, closing the loop between lineage dynamics and anatomical niche [27, 32].

Thus, spatial multi-omics now allows us to stand back and view GBM as an integrated, three-dimensional ecosystem rather than a loose collection of plastic single cells. By stitching transcriptomic, proteomic, metabolomic, and genomic layers directly onto surgical coordinates, recent studies instead portray GBM as a hierarchically ordered “malignant organism” whose architecture, metabolism, and immune defences differ predictably from its necrotic core to its infiltrative rim [39, 40].

From single-cell snapshots to whole-tumour cartography

While single-cell transcriptomics transformed our view of cellular heterogeneity in GBM, it provided a fragmented picture—molecularly rich but spatially blind. States such as NPC-like, OPC-like, AC-like, and MES-like were first defined without anatomical context, limiting insight into how these populations are arranged within tissue. Spatial transcriptomics and multiplexed proteomics address this gap by mapping transcriptional and phenotypic states in situ across intact tumour sections [2, 41, 42].

The conceptual pivot began with in situ transcriptomics platforms such as 10x Visium and Slide-seq, which anchor the single-cell lineage programmes to precise tissue coordinates. Ravi and colleagues combined Visium with MALDI mass spectrometry imaging and imaging mass cytometry to show that lineage programmes cluster into metabolically and immunologically distinct neighbourhoods that adapt to local hypoxia or inflammation [43]. This shift was concluded as a “holistic view of the malignant organism”, arguing that therapeutic design must account for macroscopic tumour order [40]. Mathur and co-workers extended this approach into three dimensions using navigation guided sampling across ten GBMs, showing that early drivers such as MDM4 changes can span the whole tumour while later events like EGFR gains may be regionally restricted [39].

Multi-omics integration is central to this paradigm. CODEX and imaging mass cytometry overlay dozens of proteins on each transcriptomic map and, in the largest survey to date, revealed a conserved five-layer scaffold organised by oxygen gradients, with hypoxia acting as a long-range tissue organiser [5]. Long-read spatial RNA-seq uncovers isoform-specific vulnerabilities. For example, FAM20C variants mark radial-glia-like cells at neuron-rich invasive edges. However, partial metabolite imaging and co-registered proteomics paint glycolytic and lipid ridges that align with hypoxic and immune niches [43, 44]. Together, these advances turn single-cell state catalogues into geography-aware biology.

A five-ring malignant organism

Building on spatial transcriptomic maps, Greenwald et al. proposed a unifying framework in which GBM forms a five-layer concentric architecture, akin to a malignant pseudo-organ [5]. This “five-ring” architecture reflects both the spatial segregation of transcriptional states and their associated microenvironments, revealing a functional geography of the tumour. Notably, tumours (and regions) with minimal hypoxia show weaker large-scale organisation, underscoring context dependence.

Across platforms and cohorts, GBM consistently unfolds as concentric territories (Figure 1). At the centre lies the necrotic core, a hypoxic, acidotic region with extensive tissue necrosis, enriched for MES-like tumour cells, including MES-transition and MES-NCperiV states, alongside inflammatory myeloid populations and stress response gene signatures [5]. This innermost zone is a necrotic, acidotic core dominated by mesenchymal hypoxic cells and over-expressing checkpoint molecules such as CD276/B7-H3, which drive tumour stemness and resistance to therapy [45]. Surrounding this core is a hypoxia-associated rim where we propose MES-transition programmes are more frequent, whereas the precise prevalence and positioning of MES-NCperiV remain to be established and will require matched spatial validation in larger cohorts. This is consistent with evidence that hypoxia and inflammatory cues promote mesenchymal reprogramming and with spatial maps that place injury response signals near perinecrotic zones [5]. It is enriched for MES- and AC-like cells that remodel extracellular matrix through TGF-β signalling and seed invasion into normal brain [43]. The adjacent viable tumour ring harbors proliferative NPC-like and OPC-like cells, and may also include emerging or transdifferentiating MES-NCperiV cells near perivascular regions, reflecting zones of self-renewal and expansion. This belt harbours aberrant vessels, pericytes, and proliferating tumour cells, acting both as a perfusion hub and an immune checkpoint corridor where checkpoint expression is concentrated in perivascular and angiogenic belts [5]. Perivascular stromal features compatible with MES-NCperiV may occur near vessels, but their consistent co-localisation across patients is not yet proven at high resolution.

Figure 1 Concentric spatial architecture of GBM and enrichment of developmental malignant programs and niches. Schematic of a glioblastoma section rendered as five concentric territories (L1–L5), integrating developmental malignant programs (NPC-like, OPC-like, AC-like, and mesenchymal variants), immune–vascular niches, and example markers. This five-layer scaffold is recapitulated across cohorts, but exact state boundaries vary between patients and are constrained by current spatial resolution. L1 Necrotic/Hypoxic core: Acidotic tissue with extensive necrosis, enriched for MES-like tumour cells, including injury-response MES-transition and neural-crest/perivascular MES-NCperiV, together with inflammatory myeloid cells and stress-response signatures; checkpoint proteins such as B7-H3/CD276 and metabolic features such as MCT1/lactate are highlighted. L2 Hypoxic rim: Perinecrotic territory with stronger hypoxia signals (HIF-1) and ECM remodelling (TGF-β). We propose that MES-transition programs are more frequent here, consistent with links between hypoxia, inflammation, and mesenchymal reprogramming. L3 Angiogenic/immune belt: A viable ring with proliferative NPC-like and OPC-like programs adjacent to aberrant vessels and pericytes; this zone functions as a perfusion hub and an immune-checkpoint corridor where PD-L1 and CTLA-4 are often elevated. Perivascular stromal features compatible with MES-NCperiV may appear near vessels. L4 Neuro-developmental zone: Oxygenated territory enriched for NPC-like and OPC-like developmental programs that mirror embryonic lineages (illustrative oncogenic drivers MYC/KRAS shown). L5 Infiltrative margin: Tumour–brain interface containing reactive glia, neurons, and scattered NPC-like cells; mature neurons at the edge explain the historical “Neural” bulk signal. Marker key: MES-transition (e.g., CHI3L1/YKL-40, CD44), MES-NCperiV (e.g., COL5A2, NES), AC like (GFAP, AQP4), NPC like (ASCL1, SOX2), OPC like (OLIG2, NG2/CSPG4). Together, the layers depict a reproducible, hypoxia-graded organisation that ties developmental malignant programs to microenvironments and helps explain region-specific therapy response. (The details are shown in Video 1 at the bottom of the article as this speculation.)

Further out lies the invasive margin, an oxygen-rich neuro-developmental zone, dominated by OPC-like phenotypes, packed with neural and oligodendrocyte progenitor-like cells that mirror foetal brain lineages and fuel expansion via MYC- and KRAS-linked metabolism [5]. Finally, the outermost reactive brain zone comprises non-malignant parenchymal cells, such as microglia, astrocytes, neurons, and endothelial cells. These malignant parenchymal cells exhibit inflammation- and stress-associated transcriptional changes in response to tumour proximity, and intermix with radial glia-like cancer stem cells that exploit neuron-derived cues to support infiltration and territorial expansion [5, 44]. Together, these spatially ordered compartments define GBM as a highly structured and ecologically diverse malignancy, with important implications for therapeutic targeting. The spatial architecture of glioblastoma integrates malignant cell states with distinct microenvironmental features, which underlie differential therapy resistance across regions and highlight the need for spatially informed treatment strategies.

Within these layers, the perivascular niche emerges as an immune-privileged sanctuary: Endothelial cells adopt immunosuppressive states (including checkpoint expression), pericytes provide trophic cues such as TGF-β, and CXCL12 gradients attract and maintain stem-like glioma cells, collectively sustaining stemness, invasion, and immune evasion [5]. Hypoxic zones recruit macrophages through the CCL8/IL-1β axis; once inside, excess lactate exported by tumour cells stabilises HIF-1α and installs an immunosuppressive gene programme through histone lactylation, locking macrophages into an M2 fate [46–48]. Spatial transcriptomics also resolves a gradient of T-cell exhaustion: Stem-like, TCF1-positive CD8⁺ cells linger near vessels, whereas TOX- and TIM-3-high terminally exhausted cells concentrate in the hypoxic core as antigen presentation persists, lowering the progenitor-to-terminal ratio that predicts checkpoint response [5].

Therapeutic lessons from spatial logic

The spatial compartmentalization of glioblastoma (GBM) offers not only biological insights but also therapeutic opportunities. Distinct tumour zones exhibit unique combinations of cellular states, microenvironmental cues, and vulnerabilities, suggesting that effective treatment may require spatially informed strategies rather than uniform systemic approaches.

The fundamental differences between glioblastoma (GBM) and other intracranial malignancies primarily lie in their cellular origins and driver molecular events. GBM, along with low-grade astrocytomas and oligodendrogliomas, belongs to the category of primary central nervous system tumours; however, it exhibits the highest degree of malignancy and is classified as WHO grade 4 [49]. In recent years, the core of diagnosis and differentiation has progressively shifted from traditional histology toward molecular subtyping. Among these, the mutation status of the IDH1/2 genes serves as the most critical distinguishing marker: The vast majority of GBM cases (approximately 90%) are IDH wild-type [50, 51], which generally indicates high aggressiveness and poor prognosis, whereas low-grade gliomas are mostly IDH-mutant, correlating with longer overall survival. On the other hand, combined 1p/19q chromosomal codeletion is a diagnostic hallmark of oligodendroglioma and is never observed in GBM. Furthermore, TERT promoter mutations and EGFR amplification are frequently found in IDH wild-type GBM and represent profound drivers of unlimited proliferation and malignant progression [52].

In contrast, the distinctions between GBM and brain metastases are more straightforward. Brain metastases originate from tumour cells that disseminate to the brain from primary cancers in other organs (such as the lung, breast, or melanoma). Consequently, their molecular genealogies largely reflect the driver genetic alterations of the primary tumour, rather than glioma-specific variants [53]. For instance, EGFR mutations commonly seen in lung cancer, HER2 amplification in breast cancer, or the BRAF V600E mutation in melanoma can be detected in their corresponding brain metastases, whereas typical glioma-associated alterations such as IDH mutations and TERT promoter mutations are absent [54]. This difference in origin directly dictates divergent treatment strategies: The management of brain metastases should be guided by the molecular drivers of the primary tumour, often involving targeted or immunotherapy, whereas the standard treatment for GBM remains maximal safe resection combined with radiotherapy and temozolomide chemotherapy, with targeted therapies currently showing limited efficacy.

Additionally, significant differences exist in the tumour microenvironment between the two. GBM is characterized by a highly immunosuppressive microenvironment, featuring impaired immune cell function and extensive involvement of tumor-associated macrophages and glial cells. This distinct immune ecology partly explains why immune checkpoint blockers demonstrate efficacy in some patients with brain metastases but yield minimal benefit in GBM. Molecular pathological testing holds profound clinical value for accurately distinguishing between different types of intracranial tumors, formulating individualized treatment plans, and predicting prognosis.

Different regions within the tumour display differential sensitivities to therapy. The hypoxic, MES-like core is enriched for stress-response programs and often exhibits resistance to radiotherapy and immune clearance [2, 5]. In contrast, the outer viable rim, composed largely of proliferative NPC- and OPC-like cells, may be more susceptible to targeted therapies aimed at cell cycle or developmental pathways. The infiltrative edge, dominated by OPC-like or radial glia-like tumour cells intermingled with reactive glia and neurons, presents a particular therapeutic challenge due to low proliferative index and diffuse infiltration beyond the resection margin [3].

The MES-like state, which often colocalizes with tumour-associated macrophages (TAMs), is notably resistant to immune-based therapies. These regions show high expression of immunosuppressive and injury response genes, classically CD44 and CHI3L1 (YKL-40), and, in several datasets, ANXA2, and are frequently associated with NF1 loss or cytokine-driven reprogramming via NF-κB and STAT3 signalling [7, 55]. In line with this biology, TAMs can push GBM cells toward mesenchymal programmes, and immunotherapy itself can select for a MES shift, both of which undermine checkpoint efficacy. Accordingly, combinations that target myeloid function (for example, CSF1R blockade with PD-1 inhibition) or interrupt mesenchymal transition pathways (NF-κB, STAT3, IL-6) are being pursued to restore sensitivity [15, 16]. Within this context, locoregional, dual-target CAR-T approaches are emerging. Within this context, locoregional, dual-target CAR-T approaches are emerging. Early phase results with bivalent EGFR/IL13Rα2 CAR-T cells delivered intrathecally/intraventricularly in recurrent GBM show feasibility, on-target activity in the CNS compartment, and the practical advantage of multi-antigen coverage to mitigate antigen escape. Such delivery achieves high CNS exposure compared with peripheral infusion and is compatible with modular scFv designs for retargeting other compartments; combinations with myeloid-modulating or anti-inflammatory agents are a rational next step to address microenvironment-mediated resistance [56]. In addition to identifying vulnerabilities, spatial mapping also reveals that cellular plasticity itself is a key resistance mechanism. Transitions between proliferative and stress-adapted states allow tumour cells to escape targeted therapies. Recent work has proposed targeting the regulators of state transition as a strategy to restrict adaptive reprogramming [15], such as TGF-β signalling or chromatin modulators. In parallel, single-cell lncRNA profiling has revealed intratumoural subtype diversity even in spatially proximate regions, reinforcing the need for combinatorial therapies that address molecular heterogeneity beyond physical compartments [57].

These insights underscore the limits of uniform drug delivery across a spatially heterogeneous tumour. Local delivery systems, including convection enhanced infusion, implantable hydrogels, and nanoparticle carriers, can help overcome regional barriers, especially in hypoxic or immune-privileged niches [58]. Among nanoparticle approaches, lipid-based drug delivery systems are particularly promising [59]. By leveraging diverse lipid materials, they can reduce toxicity to normal tissue, prolong drug half-life, enhance solubility, improve pharmacokinetics, and increase bioavailability, with growing evidence in GBM models. However, the enhanced permeability and retention (EPR) effect is variable and often weak in GBM because BTB permeability is heterogeneous, and delivery improves when active transport strategies such as receptor-mediated transcytosis (for example, transferrin receptor or LDLR/LRP1 ligands) are used. Targeting specific tumour regions can be achieved passively, leveraging the EPR effect inherent to the disorganized vasculature and impaired lymphatic drainage of cancer tissue, or actively by incorporating specific ligands (e.g., antibodies, peptides, small molecules) onto the lipid surface during liposome fabrication to bind receptors overexpressed on distinct GBM cell states or niche components [60]. At present, the exact spatial placement and boundaries of stress-induced mesenchymal transition (MES-transition) and neural crest perivascular (MES-NCperiV) programmes are not clearly defined. Therefore, systematically linking spatial layers to developmental malignant programmes will be a key focus of future work. Moreover, advances in image-guided therapy and digital pathology may allow dynamic treatment planning based on evolving spatial maps. Ultimately, integrating spatial logic into clinical workflows holds promise for precision oncology: Combining molecular state, microenvironment, and anatomical context to guide compartment-specific interventions. This paradigm shift, from treating GBM as a bulk entity to dissecting and targeting its internal architecture, may inform more effective and durable therapeutic strategies [3].

Mapping molecular targets onto this cartograph clarifies why many monotherapies fail. Hypoxia-activated prodrugs and glycolytic blockade should be directed toward the inner layers, where oxygen tension is lowest and mesenchymal clones dominate [61], whereas dual VEGF/TGF-β or CXCL12 blockade promises to collapse the perivascular stem-cell niche and sensitise tumours to radiation [43]. Metabolic-immune coupling points to MCT1 inhibition or lactate-scavenging strategies that re-programme macrophages and lift immune suppression, particularly when combined with PD-1 blockade [48]. Finally, spatial checkpoint maps argue for multiplexed antibodies, such as concurrent B7-H3 and PD-1 blockade, to strike both core and rim hotspots simultaneously [45].

Outlook

Longitudinal, four-dimensional datasets already suggest that chemoradiotherapy can flatten hypoxic layering while expanding fibrovascular and immune niches, which underscores the need for real-time spatial monitoring during treatment [39]. Coupling these atlases with MRI radiogenomics may soon allow clinicians to predict a tumour’s spatial postcode non-invasively, while organoid models replicating the five-ring blueprint are beginning to serve as functional test-beds for niche-targeted drug cocktails [44]. In parallel, organoid and explant systems that recapitulate GBM spatial niches are emerging as functional test beds; newer engineered GBM organoids and 3D spheroids capture hypoxia, perivascular cues, and therapy responses and can be used to evaluate niche-targeted drug combinations [22]. Advances in spatial proteomics, metabolomics, and AI-driven tissue segmentation are also increasing atlas resolution and enabling integrative, multi-scale modelling of tumour ecosystems [62]. In sum, spatial multi-omics reframes GBM as an organised, evolving entity whose vulnerabilities are written not only in its genes but also in its geography, and exploiting that geography is becoming a central principle of precision neuro-oncology.

Funding

This research was funded by Åke Wiberg research grant, and Karolinska Institute research grant to Y.H.; Heilongjiang Postdoctoral Launch Fund, "Support Program for Basic Research of Excellent Young Teachers" of provincial undergraduate universities in Heilongjiang Province, and Clinical Research Special Funding of Wu Jieping Medical Foundation to L.L.

Conflicts of interest

The authors declare that there are no conflicts of interest in the article.

Data availability statement

The article was written without ethical implications or exclusive data.

Author contribution statement

Rui Yan, Xiaowei Song, Yishan Hu: Writing – original draft; Yi Wang, Yanju Lv, Kaiteng Jiang: Writing – review, editing; Li Li, Yizhou Hu: editing – conceptualization.

Ethics approval

The article was written without ethical implications.

Supplementary material

Video 1: A brief overview of the five-layer structure of glioma. Access here

References

All Figures

Figure 1 Concentric spatial architecture of GBM and enrichment of developmental malignant programs and niches. Schematic of a glioblastoma section rendered as five concentric territories (L1–L5), integrating developmental malignant programs (NPC-like, OPC-like, AC-like, and mesenchymal variants), immune–vascular niches, and example markers. This five-layer scaffold is recapitulated across cohorts, but exact state boundaries vary between patients and are constrained by current spatial resolution. L1 Necrotic/Hypoxic core: Acidotic tissue with extensive necrosis, enriched for MES-like tumour cells, including injury-response MES-transition and neural-crest/perivascular MES-NCperiV, together with inflammatory myeloid cells and stress-response signatures; checkpoint proteins such as B7-H3/CD276 and metabolic features such as MCT1/lactate are highlighted. L2 Hypoxic rim: Perinecrotic territory with stronger hypoxia signals (HIF-1) and ECM remodelling (TGF-β). We propose that MES-transition programs are more frequent here, consistent with links between hypoxia, inflammation, and mesenchymal reprogramming. L3 Angiogenic/immune belt: A viable ring with proliferative NPC-like and OPC-like programs adjacent to aberrant vessels and pericytes; this zone functions as a perfusion hub and an immune-checkpoint corridor where PD-L1 and CTLA-4 are often elevated. Perivascular stromal features compatible with MES-NCperiV may appear near vessels. L4 Neuro-developmental zone: Oxygenated territory enriched for NPC-like and OPC-like developmental programs that mirror embryonic lineages (illustrative oncogenic drivers MYC/KRAS shown). L5 Infiltrative margin: Tumour–brain interface containing reactive glia, neurons, and scattered NPC-like cells; mature neurons at the edge explain the historical “Neural” bulk signal. Marker key: MES-transition (e.g., CHI3L1/YKL-40, CD44), MES-NCperiV (e.g., COL5A2, NES), AC like (GFAP, AQP4), NPC like (ASCL1, SOX2), OPC like (OLIG2, NG2/CSPG4). Together, the layers depict a reproducible, hypoxia-graded organisation that ties developmental malignant programs to microenvironments and helps explain region-specific therapy response. (The details are shown in Video 1 at the bottom of the article as this speculation.)

In the text