Single-cell atlas of glioblastoma across the tumour-to-brain axis reveals the distinct cellular landscape of infiltrative disease

Single-cell atlas of glioblastoma across the tumour-to-brain axis reveals the distinct cellular landscape of infiltrative disease | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Biological Sciences - Article Single-cell atlas of glioblastoma across the tumour-to-brain axis reveals the distinct cellular landscape of infiltrative disease Teresa Purzner, Kaytlin Andrews, Jacob Howran, Douglas Quilty, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9590848/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract In most cancers, the tissue removed at surgery forms the basis of our molecular understanding of the disease, yet it is the residual disease left behind that drives recurrence and determines patient outcomes. This disjunction is particularly consequential in glioblastoma, a diffusely infiltrative brain tumour in which recurrence is inevitable even following maximal resection. While single-cell profiling has revealed remarkable heterogeneity within resected glioblastoma, it remains unknown which malignant states, immune populations, and microenvironmental features characterize the occult infiltrative disease beyond the surgical margin that ultimately drives recurrence and treatment failure. Here, leveraging rare en bloc lobectomy specimens that preserve the complete continuum from tumour core to macroscopically normal brain, we profiled 472,089 single cells across 32 spatially matched biopsies from five glioblastoma patients, revealing that this residual disease constitutes a biologically distinct entity from both tumour and normal brain. Malignant cell composition shifts dramatically across the tissue axis: mesenchymal-like cells dominate the tumour bulk, while neural progenitor-like and oligodendrocyte precursor-like cells predominate in distal grey and white matter respectively, each mirroring the transcriptional identity of the dominant normal cell type in their zone and rendering them resistant to different and largely non-overlapping therapeutic approaches. The immune landscape undergoes a spatial inversion from blood-derived macrophage dominance at the tumour core to microglial dominance distally, with progressive regulatory T cell enrichment and near-complete loss of checkpoint target expression, providing a spatial context for the challenges that have been faced with immunotherapy in glioblastoma. Non-malignant cells throughout radiographically normal parenchyma exhibit transcriptional stress before any treatment, suggesting a therapeutic window for neuroprotection. These findings provide a framework for therapeutic strategies matched to the biology of residual disease, and illustrate a principle likely applicable across solid tumours: that meaningful progress against recurrence requires studying not what is removed, but what remains. Biological sciences/Cancer/CNS cancer Biological sciences/Cancer/Tumour heterogeneity Biological sciences/Cancer/Cancer microenvironment Biological sciences/Neuroscience/Diseases of the nervous system/Cancer in the nervous system Health sciences/Diseases/Cancer/CNS cancer Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Glioblastoma (GB) is the most common malignant brain tumour in adults, with a median survival of 14-17 months and 5-year survival of 5% despite maximal surgical resection, 1-4 radiotherapy, and temozolomide chemotherapy. Recurrence typically occurs within 7 months and is inevitable irrespective of the extent of resection, historically documented even following hemispherectomy, underscoring that GB is diffusely infiltrative at diagnosis. 5-7 Despite this, previous human GB studies have been limited to primary tumour tissue removed at the time of surgery. As a result, current models of GB are built on what we remove, rather than what remains after surgery. Yet a patient's clinical course, including progression patterns, treatment resistance, and survival, is dictated by this residual disease, which until now has remained beyond the reach of systematic study, obtainable only through rare surgical procedures that preserve the anatomical continuity between tumour and the surrounding brain. This occult infiltrative disease (OID), composed of sparse tumour cells infiltrating normal-appearing brain beyond the resection margin, exists in a microenvironment fundamentally distinct from the densely cellular tumour bulk. Consequently, therapeutic targets identified from resected tissue may not fully apply to the disease driving recurrence. Single-cell and spatial transcriptomic profiling of resected GB tissue has revealed remarkable heterogeneity, identifying four principal malignant cell states (astrocyte-like, mesenchymal-like, neural progenitor-like, and oligodendrocyte-like), 8-10 that differ in their proliferative capacity, therapy resistance, and immune interactions. These studies have further characterized the immune and stromal microenvironment, and mapped regional variation in cellular composition within the tumour mass. Yet it remains unknown which of these cellular states and microenvironmental features are present beyond the resectable margin, in the tissue where recurrence arises. 9,10 Here we address this gap by leveraging rare surgical specimens that provide access to the complete GB tissue continuum. En bloc lobectomies, in which the tumour is resected together with the surrounding lobe, provide anatomically intact specimens that preserve spatial relationships spanning tumour core, invasive margin, and distal infiltrated brain. Such specimens are exceptionally rare; to date, only four centres worldwide have published case series of GB lobectomy, collectively reporting fewer than 200 patients at accrual rates of 3–4 cases annually. Our cohort of five IDH-wildtype GB patients, together with non-oncological lobectomy controls, comprises 32 spatially resolved specimens across the tissue continuum. Using multiplexed single-cell RNA sequencing of these spatially matched fresh-frozen biopsies, we profiled 472,089 cells across five anatomically defined spatial zones: two tumour compartments (tumour core and tumour edge) and three OID compartments (adjacent white matter, AWM; distal white matter, dWM; and distal grey matter, dGM), together with corresponding zones from non-oncological controls. We reveal that the OID is distinct from both tumour and normal brain: malignant cells adopt transcriptional programs mirroring the dominant cell type of the region they occupy, the immune landscape transitions from blood-derived myeloid dominance toward brain-resident microglial composition, and non-malignant cells exhibit transcriptional stress with evidence of incomplete compensation. Together, these findings suggest that the biological extent of GB substantially exceeds the radiographically defined tumour margin, and that the tissue in which recurrence arises is already transcriptionally and immunologically distinct from normal brain before any treatment is administered. This spatially resolved characterization provides a framework for therapeutic strategies designed around the biology of residual disease. Results OID compartments are distinct from both tumor and normal brain To characterize the cellular landscape of GB across the complete tissue continuum, we performed single-cell RNA sequencing on spatially matched biopsies from five IDH-wildtype, treatment-naïve GB patients undergoing en bloc temporal lobectomy and one non-oncological lobectomy control (Fig. 1A-C; Supplementary Table 1). For each tract, biopsies were collected along a defined anatomical axis: tumour core, tumour edge, adjacent white matter (AWM), distal white matter (dWM), and distal grey matter (dGM), with the latter two zones located in macroscopically normal-appearing brain parenchyma outside the contrast-enhancing margin. After quality control and doublet removal, 472,089 cells were retained (median 1,532 genes per cell; median 2,375 UMIs per cell). Cells from equivalent anatomical zones co-localized across patients in UMAP space (Fig. 1D), with integration quality assessed by Local Inverse Simpson Index (LISI) scoring: integration LISI improved from 1.48 (pre-Harmony) to 1.64 after batch correction (maximum 2.0), while cluster LISI remained near-optimal at 1.10 (ideal 1.0), confirming that patient-of-origin effects were reduced without obscuring biologically meaningful cell-type boundaries (Supplementary Table 2). Malignant cells were identified by characteristic GB copy-number alterations and reference-based classification (Methods; Supplementary Fig. 1; Supplementary Table 3). Multi-method cell type annotation combining GBmap SCANVI reference mapping, CellTypist classification, CNV inference, and marker panel scoring (Methods) resolved 18 transcriptionally distinct populations spanning malignant GB states, neural populations, and immune infiltrates (Fig. 1E; Supplementary Fig. 2). Cell type composition shifted progressively across the spatial axis (Fig. 1F; Supplementary Table 4). Malignant cells comprised 75.2% of cells at the tumour core, declining to 22.1% in AWM, approximately 20% in dWM, and 10% in dGM. Oligodendrocytes showed the inverse pattern, rising from less than 5% at the core to 41.0% in dWM, declining to 22.5% in dGM. Neurons and astrocytes were nearly absent from the tumour core (neurons 0.1%, astrocytes 0.2%), with partial recovery in dGM (neurons 25.1%, astrocytes 12.0%), and reached their highest proportions in GM(Ctrl) (neurons 32.2%, astrocytes 12.6%). Notably, while total immune cell density relative to non-malignant cells returned to near-control levels in the OID (AWM 1.06x, dWM 0.89x, dGM 0.85x of control), the immune compartment was qualitatively remodeled at every zone: monocyte-derived macrophages remained 25- to 33-fold elevated above control in AWM and dWM, while microglia were reciprocally depleted, indicating that tumour-associated immune remodeling extends into the OID compartments even where overall immune density appears normal. Beyond compositional shifts, cells in the OID compartments occupy a transcriptionally intermediate position between tumour and normal brain. Wasserstein distance analysis of zone-level transcriptional profiles confirmed a continuous spatial gradient (Spearman ρ = 0.94, p = 0.005 across all zones) rather than a binary tumour–normal boundary (Supplementary Fig. 3). Tumour-patient dWM retained 19.9% of the transcriptional distance from tumour core to WM(Ctrl); dGM retained 7.2% of the distance to GM(Ctrl). Infiltrating malignant cells mirror the transcriptional identity of their host tissue compartment Malignant cell subtype composition varied markedly and non-uniformly across the spatial axis (Fig. 2A). Mesenchymal subtypes were concentrated in the tumour mass: 93.7% of MES1-like cells and 97.2% of MES2-like cells were located at the core or edge, with mesenchymal subtypes together comprising only 7.2% of malignant cells in dGM. By contrast, NPC2-like and OPC-like cells together accounted for 60.1% of all malignant cells at dGM and 21.4% at dWM, comprising the dominant malignant population within the OID compartments. Within the NPC2-like population itself, 61.0% of cells resided in dGM (compared with only 3.5% at the tumour core; Spearman ρ=0.90, p=0.037); OPC-like cells showed the complementary pattern with 26.3% at AWM and 44.2% at dWM (Kruskal-Wallis H (KW H) = 13.3, p = 0.010). Each subtype was thus enriched in the zone whose dominant normal cell type shares its transcriptional lineage. AC-like cells were present across all zones, consistent with their numerical dominance as the largest subtype rather than specific enrichment in any compartment. This spatial distribution was consistent across all five GB patients and robust to leave-one-out patient exclusion (Supplementary Table 5). Beyond the compositional differences described above, cells of the same malignant subtype show substantial regional transcriptional differences, with the genes most strongly upregulated in each subtype varying systematically across tumour and OID compartments. Within OPC-like cells, oligodendrocyte lineage genes were most highly expressed in white matter zones relative to the tumour core, with MBP and MOBP showing 6-fold and 10-fold increases respectively from the tumour core to dWM. Within NPC2-like cells, neuronal identity genes showed the strongest expression in grey matter, with GRIN1 increasing 74-fold from the tumour core to dGM and RBFOX3 expression absent at the tumour core (mean=0.00) rising to 0.66 in dGM. The small number of MES1-like cells present in distal zones showed progressively lower expression of mesenchymal markers, including CHI3L1 (3.74 at Core, declining to 0.94 in dGM) and VEGFA (3.87 at Core, declining to 0.95 in dGM). AC-like cells showed comparatively modest spatial variation in AC marker expression, with core identity genes CST3 (4.67 to 2.54) and BCAN (2.82 to 1.58) showing gradual decline across zones (Fig. 2B). Notably, concordant patterns were observed when the analysis was repeated using the Neftel et al. (2019) 10 gene sets that originally defined these subtypes, confirming the robustness of these findings to gene set choice (Supplementary Fig. 4). Similarly, malignant subtype cycling fractions within the tumour compartment were closely concordant with those observed when reanalyzing data from Darmanis et al. (2017) 8 using an identical cell cycle scoring pipeline across an independent cohort profiled on a distinct platform (Smart-seq2 vs 10x Flex; AC-like: 35.1% vs 36.0%, n=467 vs n=47,710; MES1-like: 33.5% vs 37.6%, n=403 vs n=42,552; Supplementary Fig. 5, Supplementary Tables 6 and 7), supporting the reproducibility of subtype-specific cell cycle scoring. 8,10 Of eight published GB transcriptional programs examined here, spatial variation was most prominent in hypoxia (Gavish et al. 2023; MP7) 11 declining distally in MES2-like cells (0.814 at Core to 0.165 in dGM), interferon response peaking at the tumour edge in MES1-like cells (0.263 at Edge, declining to near-zero in OID zones), and both developmental stemness and invasion/tumour microtube score peaking at AWM in AC-like cells (0.423 and 0.351 respectively, declining to 0.137 and 0.174 in dGM; Fig. 2D; Supplementary Fig. 6). HIF1α-driven paracrine signaling provides a candidate mechanism by which hypoxia-induced stress propagates into non-hypoxic distal compartments, consistent with reports of HIF1α target genes differentially expressed across the GB core-margin axis. 12,13 NPC2-like cells are the dominant malignant cell type in dGM, where they exhibit a markedly reduced proliferative fraction compared with the tumour bulk. Cell cycle phase scoring across all 144,089 malignant cells (Tirosh et al., 2016; Supplementary Fig. 7) 14 revealed that NPC2-like cycling fraction (S+G2M) fell from 55.2% at the tumour core to 19.7% in dGM (Spearman ρ = -0.90, p = 0.037 across tumour zones; K-W H = 355, p = 1.2 x 10⁻⁷⁵), substantially below the all-malignant mean at every distal zone (Core 40.5%, Edge 35.2%, AWM 29.2%, dWM 31.7%, dGM 27.5%) (Fig. 2C). This quiescent profile was accompanied by progressive enrichment of the neural mimicry transcriptional program. Analysis of all spatially variable genes within NPC2-like cells enriched for chemical synaptic transmission distally (p=3.7x10⁻¹³) (Supplementary Fig. 8; Supplementary Tables 8 and 9). Consistent with this, genes gained distally in NPC2-like cells showed significant overlap with the top 100 excitatory neuron marker genes in our dataset (87/100 excitatory neuron markers; Fisher's exact p=1.0x10⁻¹⁴⁶) (Supplementary Fig. 9; Supplementary Table 10). Similarly, the neural mimicry program, defined by neuronal identity genes including NRGN, RBFOX3, NEFL, NEFM, STMN2, and MAPT, showed the highest expression in NPC2-like cells at every distal zone, peaking at dGM (mean score +0.757; Core mean +0.003), while remaining near-absent in mesenchymal subtypes throughout (Fig. 2E). 14 In contrast to the quiescent NPC2-like population in grey matter, OPC-like cells in white matter zones maintained high proliferative activity. Cycling fraction for OPC-like cells remained at 58.4% in dWM, 29.7 percentage points above the all-malignant mean of 28.7% in the same zone (OPC-like > non-OPC malignant in all five tumour patients; one-sided paired Wilcoxon p=0.031), and comparable to levels observed at the tumour core (65.7%) and edge (54.0%). OPC lineage score showed a systematic increase in white matter zones, rising from 0.086 at the tumour core to 0.884 at dWM, the highest score of any malignant subtype at any zone (Fig. 2F), while invasion/tumour microtube program score remained near-absent (mean 0.018 at AWM). This contrasts with AC-like cells, which carried the highest invasion/tumour microtube program score at AWM (mean 0.351). 15 Consistent with this, the top spatially variable genes within OPC-like cells in white matter zones included canonical myelinating oligodendrocyte genes MBP (6-fold increase from Core to dWM), MOBP (10-fold) and MYRF (7-fold). In keeping, genes gained distally in OPC-like cells showed significant overlap with the top 100 oligodendrocyte marker genes in our dataset (92/100; Fisher's exact p = 2.0x10⁻¹⁹³) (Supplementary Fig. 9; Supplementary Table 10). 15 OPC-like cells in white matter and NPC2-like cells in grey matter thus present distinct transcriptional profiles: proliferative and oligodendrocyte-like in one compartment, quiescent and neuron-like in the other. The immune microenvironment of GB demonstrates spatial variation Immune populations were identified across all spatial zones, including monocyte-derived macrophages (MDMs), microglia, T cells, neutrophils and granulocytes (Fig. 3A). While all major immune populations were detectable throughout the tissue axis, their composition and density underwent striking spatial remodeling (Fig. 3A; Supplementary Fig. 10). MDMs were the dominant immune population at the tumour core and edge, reaching 109.9 and 176.5 cells per 1,000 total cells respectively. MDM density declined progressively across the OID but remained substantially elevated relative to control at every zone: 71.8 cells per 1,000 in AWM and 56.1 in dWM (33- and 26-fold above WM(Ctrl)) , and 10.2 in dGM (3.7-fold above GM(Ctrl)). Microglia showed the inverse pattern: virtually absent from the tumour core and edge (<1 cell per 1,000 total cells, compared with 255.6 in white matter control), with partial recovery across the OID, reaching 57% of WM(Ctrl) levels in AWM, 49% in dWM, and 57% in dGM where it is the dominant immune cell type. The microglia:MDM ratio reflected this spatial inversion, with a ratio of 9:1 in dGM (compared with 58:1 in control grey matter), 2.2:1 in dWM, 2.0:1 in AWM, <0.1:1 at the tumour edge and <0.1:1 at the tumour core. These data reveal that MDM-driven myeloid remodeling extends substantially beyond the tumour mass, with the OID compartments representing a transitional zone between tumour-associated MDM dominance and the microglial-dominant composition of normal brain. T cell abundance demonstrated spatial variance. T cells at the tumour core comprised 23.7% of immune cells (1.6-fold above white matter control), then decreased to levels comparable to white matter control at the tumour edge (16.3% vs 14.9%) and 14.2%, 18.4%, and 32.3% in AWM, dWM, and dGM respectively (Supplementary Fig. 10; Supplementary Table 4). T cell composition exhibited a marked shift across spatial zones (Fig. 3B). CD8+ cytotoxic T cells comprised 14.8% of T cells in the tumour core, rising to 17.5% at the tumour edge before declining to 11.6% and 9.9% in AWM and dWM respectively, and only 4.4% in dGM. Regulatory T cells showed the inverse pattern, rising from 9.7% at the tumour edge to 18.2% and 9.3% in AWM and dWM respectively, and up to 39.8% in dGM, approaching the fraction seen in grey matter control (40.7%). The CD8:Treg ratio thus inverted across the tissue axis, from 1.9:1 at the tumour edge to 0.1:1 in dGM (0.2:1 in GM(Ctrl)). These data reveal that the OID compartments harbor a T cell population that is not merely depleted but compositionally shifted toward regulatory dominance. Expressed as a proportion of CD45+ immune cells (0.9–1.1% across all zones, KW p=0.71) or as absolute density per 1,000 total cells (1.3–2.9, KW p=0.47), Treg abundance was not significantly different between zones or between tumour-patient dGM and GM(Ctrl). The apparent Treg enrichment in dGM as a fraction of the T cell pool thus reflects local depletion of conventional CD4+ and CD8+ T cells rather than expansion of the Treg compartment itself. Among non-conventional lymphocyte populations, NK cells and double-negative T cells (DNTs) were detected at low but consistent frequencies across all spatial zones, comprising 21.8% and 22.7% of T-lineage cells in dWM and dGM respectively (Fig. 3B). NK cells and DNTs, when cytotoxic, engage targets through MHC-unrestricted mechanisms, though functional activity cannot be inferred from transcriptional profiling alone. T cell checkpoint receptor expression followed a consistent spatial gradient (Fig. 3C). The three principal checkpoint receptor targets, PDCD1 (PD-1), HAVCR2 (TIM-3), and TIGIT, all showed significant spatial variation (K-W H = 701, 392, and 865 respectively; all p < 10⁻⁸¹), peaking at the tumour edge (7.4%, 7.9%, and 9.2% of T cells respectively) and declining progressively across the OID: 2.9%, 4.4%, and 4.7% in AWM; 2.9%, 3.6%, and 5.7% in dWM; and 0.4%, 1.8%, and 0.5% in dGM, approaching levels observed in GM(Ctrl) (0.3%, 1.4%, 0.3%). Co-expression of two or more checkpoint receptors showed a similar spatial pattern: 4.1% of T cells at the tumour edge expressed multiple checkpoint markers, declining to 1.5% in AWM, 1.8% in dWM, and 0.1% in dGM (Supplementary Table 11). PD-L1 (CD274), the principal ligand for PD-1, was correspondingly elevated on malignant cells at the tumour edge (14.3% positive) and core (7.5%), declining sharply to 1.9% in dWM and 1.2% in dGM (KW H = 3,524, p ≈ 0), establishing spatial co-localization of checkpoint ligand and receptor expression at the tumour margin. These data reveal that the molecular signatures of T cell exhaustion and immune evasion are concentrated at the tumour margin and progressively diminish across the OID, a spatial gradient observed consistently across all five GB patients (Supplementary Table 11). Non-malignant cells exhibit spatially graded transcriptional stress Neurons were present at very low frequency in the tumour core and edge (n=28 and n=77 total respectively) precluding reliable zone-level analysis, but were abundant in AWM (n=1,949), dWM (n=3,077), dGM (n=19,455) and GM(Ctrl) (n=18,844), providing adequate statistical power for spatial comparisons across AWM through GM(Ctrl) (Fig. 4A). In macroscopically normal-appearing dGM, 39.6% of neurons scored above zero for the stress program, 11 compared with 6.7% of GM(Ctrl). The same pattern held in AWM (81.8%) and dWM (50.7%), and in both excitatory (40.2% in dGM) and inhibitory subtypes (38.3%; Supplementary Table 12), consistent across all leave-one-out patient-exclusion scenarios. The stress program was significantly elevated in OID compartments relative to zone-matched non-oncological control (KW H=12,817; AWM FDR ≈ 0, dWM FDR = 2.6x10⁻¹⁶³, dGM FDR ≈ 0; Supplementary Fig. 11; Supplementary Table 12), with mean scores positive across all OID compartments (dGM mean +0.023) and negative in GM(Ctrl) (mean -0.107) (Supplementary Table 5; Supplementary Table 12). Stress scores peaked at AWM (mean +0.286) and declined through dWM (mean +0.103) and dGM (mean +0.107) (Fig. 4B), yet remained above zero in every GB patient zone. In contrast, synaptic program expression in dGM neurons was comparable to GM(Ctrl) (99.6% vs 96.5% respectively; Supplementary Fig. 11). The GB-associated field effect thus selectively activates a stress and injury transcriptional response while leaving core synaptic identity intact. Beyond the six scored programs, agnostic spatial variable gene analysis identified 3,776 genes with significant expression variation across spatial zones in neurons (Kruskal-Wallis H, FDR 0.3), of which 1,306 were elevated in tumour-patient zones and 881 depleted relative to non-oncological control, confirming that the neuronal field effect encompasses broad transcriptional reprogramming beyond the stress response alone (Supplementary Fig. 12). Additionally, established diffusible inducers of cell stress including VEGF, HIF1α, IL-6, IL-8 (CXCL8), CCL-2, and TGF-β1 are upregulated in the tumour mass relative to non-oncological control, with residual elevation persisting through dWM and dGM, consistent with previous reports of these factors as components of the established GB secretome. 13,16-19 TNF-α showed the inverse spatial pattern, preferentially expressed by macrophages and microglia in distal compartments rather than at the tumour mass. 20 Each of these inducers engages stress signaling via the kinase arms profiled above. 12,21,22 The neuronal stress response was not driven by a single upstream pathway. Scoring neurons across five integrated stress response (ISR) module programs revealed co-activation of the PERK/UPR arm (BiP/HSPA5 elevated in 18.9% of stressed neurons), the HRI/mitochondrial UPR arm (ATF5, HMOX1), and an excitotoxicity/immediate-early gene signature (NPAS4, NR4A1) that was the dominant signature across all cell types (Supplementary Fig. 13). The PKR/interferon arm was not differentially activated in MP6-classified stressed neurons (Δ ≈ 0; Mann-Whitney p = n.s.), arguing against type-I interferon signalling as a major driver. Although the strongest ISR-arm activation was observed in neurons, scoring the same kinase-arm modules across all cell types showed detectable activation across all four profiled cell types (neurons, OPCs, oligodendrocytes, astrocytes; Supplementary Figure 13C), supporting a shared upstream paracrine driver rather than a neuron-restricted mechanism. The CHOP:GADD34 ratio in stressed neurons indicated an active resolution phase rather than chronic terminal stress. The skewed ratio toward GADD34 over CHOP, together with the predominance of immediate-early and excitotoxicity transcripts over committed pro-apoptotic effectors (e.g., BAX, BAK1, BIM all near baseline), is consistent with a pro-survival ISR signature rather than a commitment to apoptosis. Ligand-receptor analysis identified tumour-associated macrophages (MDM) and AC-like tumour cells as the principal candidate sources of upstream ISR-activating signals, expressing TGFB1, CXCL12, IL1B, and CSF1 at highest levels in Core and Edge zones with expression declining across the spatial axis (Supplementary Fig. 14). To identify a discrete stressed neuron population, neurons were classified as stressed if their MP6 stress score exceeded a zone-matched threshold (non-oncological control neuron mean + 2 SD, computed separately for white matter and grey matter zones to account for baseline transcriptional differences; Methods). This identified 8,401 stressed neurons (17.8% of all 47,262 neurons). Comparing stressed neurons to unstressed neurons by pseudobulk differential expression (DESeq2; n=5 tumour patients) identified 169 genes significantly upregulated and 24 downregulated in the stressed population (adjusted p 0.5), with pathway enrichment confirming activation of TNFα Signaling via NF-κB (FDR = 1.6x10⁻⁷¹), Hypoxia (FDR = 6.8x10⁻¹⁶), p53 Pathway (FDR = 2.4x10⁻⁸), and Apoptosis (FDR = 1.4x10⁻⁷) programs (Supplementary Table 13; Supplementary Fig. 15). The proportion of stressed neurons was elevated in AWM relative to WM(Ctrl) (39.6% vs 4.5%), in dGM relative to GM(Ctrl) (32.3% vs 3.9%), and in dWM relative to WM(Ctrl) (14.1% vs 4.5%) (Supplementary Fig. 16), suggesting that tumour-associated neuronal stress extends throughout macroscopically normal brain parenchyma in GB patients. Astrocytes demonstrated a progressive reactive gradient that tracked distance from the tumour mass (Fig. 4C). At the tumour edge, 96.7% of astrocytes occupied a reactive state (Reactive A1, A2, or DAA), declining to 88.5% in AWM, 70.5% in dWM, and 61.5% in dGM (Spearman ρ = -0.900, p = 0.037 across tumour zones), approaching but remaining above the 56.9% observed in non-oncological GM(Ctrl) (Core 96.2%; n=80, interpreted with caution given limited recovery). This gradient represents the steepest and most spatially coherent cellular transition in the dataset, robust to leave-one-out exclusion with a minimum AWM-to-dGM difference of 24.0 percentage points (Supplementary Table 5). The A1 neurotoxic state, associated with complement pathway activation and neuronal injury, 23 was the dominant reactive state in AWM (42.8% of astrocytes) and remained elevated above control in both dWM (27.9%) and dGM (13.8% vs 12.0% in GM(Ctrl)). The disease-associated astrocyte (DAA) state, characterized by upregulation of SERPINA3 (logFC 7.63) and SPHK1 (logFC 6.38) relative to homeostatic astrocytes within the dataset, was detectable across all GB zones (26.2% in dGM vs 23.2% in GM(Ctrl)). Astrocytic reactivity, including neurotoxic and disease-associated states, is therefore a region-wide property of the GB-affected brain that persists into macroscopically normal tissue. Consistent with this, agnostic spatial variable gene analysis identified 2,615 genes with significant spatial variation in astrocytes, the second-highest of any non-malignant cell type, of which 1,659 were tumour-elevated, enriching for regulation of apoptotic processes (FDR = 4.4 x 10^-6) and transcriptional repression (FDR = 6.7 x 10^-7), indicating that reactive astrocytes in the OID are broadly transcriptionally reprogrammed beyond the scored reactive state markers (Supplementary Fig. 12). Oligolineage cells demonstrated widespread stress (Fig. 4D; Supplementary Fig. 17). Stress program scores were elevated in oligodendrocytes at the Core (+0.262) and remained significantly elevated at AWM (mean +0.083; MWU FDR = 3.2x10⁻²⁴⁷) and dWM (mean +0.031; MWU FDR = 4.6x10⁻⁴⁵), but did not extend robustly to dGM (mean -0.004; KW H = 10,741 across all zones, p ≈ 0). OPCs showed a spatially broader stress response, elevated across tumour zones (Core +0.235, AWM +0.117, dWM +0.165, dGM +0.051; KW H = 6,170, p ≈ 0) (Fig. 4E), well above WM(Ctrl) in white matter zones and above GM(Ctrl) in dGM; notably, OPC stress scores in dWM (+0.165) exceeded those of mature oligodendrocytes in the same zone (+0.031) (MWU p approximately 0), suggesting the progenitor compartment is more affected than the mature population. Beyond the stress program, three additional findings were consistent across all leave-one-out patient-exclusion scenarios (Supplementary Table 5): myelin and maturity marker scores (MBP, PLP1, MOG) were elevated rather than suppressed in GB patient zones relative to control (AWM difference +0.28; dGM difference +0.16; LOO-robust at all zones); remyelination signaling was elevated vs matched control in white matter zones (AWM +0.08, dWM +0.05; LOO-robust) but reduced specifically at dGM (-0.03; LOO-robust); and GB patient oligodendrocytes retained higher progenitor marker scores than non-oncological control oligodendrocytes in all five zones (AWM difference +0.10, dGM difference +0.11; LOO-robust at all zones). An inflammatory response programme was modestly elevated in tumour-proximal zones in both oligodendrocytes (Core +0.09, Edge +0.14) and OPCs (Core +0.08, Edge +0.08), recovering toward non-oncological control levels in distal compartments and consistent with the spatially restricted myeloid signal at the tumour mass (Fig. 4D, 4E). Agnostic spatial variable gene analysis corroborated these findings, identifying 1,303 spatially variable genes in oligodendrocytes and 2,133 in OPCs. Among OPC tumour-elevated genes, pathway enrichment identified transcriptional regulation (FDR = 5.3 x 10^-7), while tumour-depleted OPC genes enriched for synapse organization and nervous system development (FDR = 1.5 x 10^-5), suggesting that OPCs in tumour-proximal zones lose normal supportive functions alongside their stress activation. Oligodendrocyte tumour-elevated genes enriched for sterol biosynthesis (FDR = 2.3 x 10^-4), consistent with the paradoxical upregulation of myelin maturity markers (Supplementary Fig. 12). Discussion Single-cell atlases of resected glioblastoma tissue have defined the malignant cellular states, immune composition, and transcriptional heterogeneity of the tumour mass in considerable detail. 8-10,24,25 Yet recurrence is driven not by the tissue removed at surgery but by the occult infiltrative disease (OID) that persists beyond the resection margin. Here we reveal that this OID is biologically distinct from both tumour bulk and normal brain. Infiltrating GB cells adopt transcriptional programs that mirror the dominant cell type of the region they occupy: OPC-like cells in white matter, NPC2-like cells in grey matter. This transcriptional resemblance is supported at the gene level by significant overlap between spatially gained genes and independently-derived normal cell type markers (OPC-like: 92/100 oligodendrocyte markers, p=2.0x10⁻¹⁹³; NPC2-like: 87/100 excitatory neuron markers, p=1.0x10⁻¹⁴⁶). Concurrently, the surrounding parenchyma acquires attributes of the tumour microenvironment, with myeloid remodeling, neuronal stress, and astrocytic reactivity extending well beyond the contrast-enhancing margin, spatial patterns observed consistently across all five profiled patients and robust to leave-one-out exclusion. The chimeric tissue identity of the OID, with malignant cells assuming host phenotypes while host tissue assumes tumour-associated features, has direct therapeutic implications which we address in turn. NPC2-like cells are enriched in distal grey matter, where they adopt a progressively quiescent state, with cycling fraction declining monotonically from tumour core to dGM (Fig. 2C). This spatial gradient of quiescence has important therapeutic implications. Slow-cycling glioma cells survive temozolomide and radiation, persist through treatment, and can regenerate the tumour mass 26,27 ; a non-cycling population residing outside the standard radiation field therefore represents a reservoir from which recurrence would be expected to arise. The therapeutic challenge is compounded by the neural mimicry program carried by NPC2-like cells, a transcriptional resemblance to neurons so deep that 87% of excitatory neuron markers are gained distally (Fisher's exact p=1.0x10⁻¹⁴⁶). This expression pattern increases with distance from the tumour, peaking in dGM where NPC2-like cells and neurons co-occupy the same spatial compartment and where Treg frequency is highest. Central self-tolerance is established in part through AIRE-mediated thymic expression of tissue-restricted antigens, including neuronal proteins, leading to deletion of T cells reactive against these self-antigens. 28,29 NPC2-like cells expressing these neuronal identity genes in distal grey matter may therefore escape T cell-mediated recognition through the same mechanism that protects normal neurons from autoimmune attack, rendering them resistant to both neoantigen-based and T cell-redirecting immunotherapies while their quiescent state also renders them resistant to standard cytotoxic therapy. The white matter OID compartments are characterized by enrichment of OPC-like cells which, unlike the quiescent NPC2-like population in grey matter, maintain a high proliferative index rendering them, in principle, both radiosensitive and chemosensitive. In the white matter, then, the therapeutic challenge is not tumour cell biology but drug delivery and radiation targeting: these zones lie outside the standard radiation field and while temozolomide crosses the blood-brain barrier it achieves only approximately 20% of plasma concentrations in normal brain interstitium, 30,31 substantially below the levels reached in the tumour mass where BBB integrity is compromised. 32 Together, the two-phenotype model predicts that effective treatment of the OID will require combination approaches: anti-proliferative strategies directed at the cycling OPC-like population, and non-cell-cycle-dependent approaches directed at quiescent NPC2-like cells. For the latter, differentiation therapy represents a rational strategy, with BMP4 and retinoic acid both shown to induce terminal differentiation of glioma stem cells, reduce self-renewal capacity, and sensitize cells to cytotoxic therapy. 33-35 Alternatively, strategies that force quiescent cells into cycle, such as checkpoint kinase inhibition or metabolic modulation, could render them susceptible to radiation and temozolomide. 36-38 CHK1 inhibition represents a well-characterized approach in this category: glioma stem cells survive radiation through preferential activation of the CHK1-mediated DNA damage checkpoint, and pharmacological CHK1 inhibition reverses this radioresistance by abrogating the protective G2/M arrest; CHK1 inhibition similarly potentiates temozolomide cytotoxicity 5-fold by preventing the G2 arrest that otherwise allows DNA damage repair. 27,39 Metabolic modulation offers a complementary strategy: slow-cycling glioma cells depend on oxidative phosphorylation, and metformin-induced AMPK activation has been shown to promote differentiation of stem-like cells into non-tumorigenic populations. 40,41 Tumour-associated macrophages and microglia are the dominant immune population in GB, and recent single-cell studies have begun to distinguish their distinct functional states and spatial distributions within the tumour.42-45 Our data extend these observations across the complete tissue axis, revealing a striking spatial inversion of the myeloid compartment: at the tumour core MDMs outnumber microglia by more than 100-fold, while in dGM the composition inverts to 90% microglia and 10% MDM (microglia:MDM ratio 9:1). T cell composition shifts in parallel, with CD8+ cytotoxic T cells declining from 17.5% at the tumour edge to 4.4% in dGM while regulatory T cells rise from 9.0% to 35.4%. Tumour-infiltrating Tregs maintained CTLA-4, TIGIT, and LAG-3 expression across core, edge, and AWM zones, consistent with the activated Treg phenotype described in human cancer, while PD-1 remained minimal on this population and was instead carried predominantly by CD8+ T cells. The principal checkpoint receptor targets (PD-1, TIM-3, and TIGIT) peak at the tumour edge and decline to levels approaching normal brain in dGM (0.4-1.8%), a spatial gradient observed consistently across all five patients. These findings have direct implications for immunotherapy in GB, which we address in turn. 42-45 The three checkpoint receptors described above are the principal targets of checkpoint inhibitor trials in glioblastoma, selected on the basis of expression profiling in resected tumour. To date, however, every phase III checkpoint inhibitor trial in glioblastoma has failed to demonstrate a survival benefit, a finding confirmed by a recent meta-analysis of 12 randomized controlled trials. 46-49 While the failure of checkpoint blockade in glioblastoma is undoubtedly multifactorial, our data provide some spatial context. Checkpoint therapies are designed around the biology of the tumour mass and its immediate margin, where our data confirm high checkpoint expression, exhausted T cells, and MDM-dominated immunosuppression, providing a rational biological substrate for blockade. Consistent with this, Nivolumab (anti-PD-1) reaches saturating levels on intratumoural T cells in glioblastoma and induces local T cell activation and proliferation, and neoadjuvant PD-1 blockade promotes intratumoural T cell clonal expansion and cDC1 activation. 50-53 In the OID, however, PD-1, TIM-3, and TIGIT are near-absent (0.4-1.8% in dGM, approaching normal brain levels), and PD-L1 on malignant cells declines from 14.3% at the tumour edge to 1.2% in dGM, indicating that the molecular substrate for local checkpoint reinvigoration is largely confined to the tumour mass and its immediate margin. 46-53 Importantly, the principal mechanism of PD-1 blockade is thought to operate not locally but through progenitor exhausted T cells residing in tumour-draining lymph nodes, which proliferate, differentiate, and traffic to the tumour after systemic blockade. 54-59 This pathway remains intact irrespective of local checkpoint expression in the OID. Yet even if checkpoint blockade successfully generates tumour-reactive effector T cells, those cells must traffic to and function within the OID to eliminate residual disease, and our data suggest that this effector phase is substantially limited. Antigen-reactive T cells are concentrated at the tumour edge rather than in the OID compartments, consistent with 5-ALA-guided spatial profiling of glioblastoma showing preferential enrichment of PD-1⁺CD103⁺ tissue-resident T cells in the intermediate and marginal tumour layers rather than in distal tissue. 60,61 The OID thus presents compounding effector-phase barriers: Treg dominance (CD8:Treg ratio of 0.1:1 in dGM), persistent MDM-mediated immunosuppression (25- to 33-fold above control in AWM and dWM), and the neural mimicry program of NPC2-like cells, which may exploit central tolerance mechanisms to evade MHC-restricted recognition. 15,62-64 As such, patients with subtotal resections, in whom checkpoint-high, MDM-dominated tumour bulk remains in situ, may represent a more biologically appropriate population for checkpoint inhibitor trials. These effector-phase barriers are not unique to checkpoint blockade. Neoantigen vaccines, bispecific T cell engagers, and other T cell-redirecting strategies face the same compounding limitations in the OID, as each ultimately depends on endogenous T cells recognizing and eliminating tumour cells. MHC-unrestricted approaches such as CAR-T platforms targeting glioblastoma-specific surface antigens circumvent the antigen visibility problem, although antigen loss and intratumoural heterogeneity remain well-documented challenges. 65-69 Regardless of modality, the pronounced Treg dominance in the OID suggests that any immunotherapy strategy directed at residual disease may benefit from concurrent efforts to shift the CD8:Treg ratio toward cytotoxic predominance. 63 Beyond the malignant and immune compartments, non-malignant cells are transcriptionally altered in spatially structured patterns extending from the tumour margin throughout the OID. Neurons, astrocytes, and oligodendrocytes each exhibit transcriptional changes in tissue that is macroscopically and radiographically normal-appearing, constituting a region-wide field effect that was consistent across all five patients and robust to leave-one-out exclusion in each cell type examined. Neurons exhibit a stress/injury transcriptional response that is elevated across all OID compartments relative to non-oncological controls, peaking at AWM and persisting into macroscopically normal dGM. This extends prior immunohistochemical observations of stress-response protein upregulation in tumour-infiltrated cortex to well outside regions of direct tumour contact.70 Critically, synaptic program expression in dGM neurons is preserved relative to controls, suggesting that core synaptic identity remains intact despite widespread stress activation. This dissociation between stress activation and preserved synaptic identity is consistent with a sublethal injury state in which neurons are functionally compromised but not yet irreversibly damaged, and defines a potential therapeutic window. The character of this response implicates diffusible signals from the tumour and the microenvironment rather than direct hypoxic injury, with immediate-early transcription factors and molecular chaperones dominating over canonical HIF1A-driven hypoxia genes. The pattern of ISR kinase arm co-activation is consistent with specific tumour-derived diffusible signals known to be elevated in the GB microenvironment: VEGF and HIF1α engaging the PERK/UPR arm, IL-6 and IL-8 activating NF-κB signaling, and TGF-β1 contributing to GCN2-arm metabolic stress. 70,71 Astrocytes exhibit a progressive reactive gradient extending throughout the OID, from 88.5% in the AWM to 61.5% in the dGM, where the total reactive fraction remains above non-oncological controls. The disease-associated astrocyte (DAA) and A1 neurotoxic states are the dominant reactive subtypes across OID compartments, characterized respectively by upregulation of SERPINA3 and complement pathway activation. These findings are consistent with emerging evidence that tumour-associated reactive astrocytes are not passive bystanders but active participants in GB pathogenesis: astrocyte-derived CCL2 and CSF1 govern tumour-associated macrophage recruitment, astrocyte-derived cholesterol supports glioma cell survival, and a distinct astrocyte subset limits tumour immunity by inducing T cell apoptosis through TRAIL driven by tumour-derived IL-11 signaling through STAT3. 72,73 DAA and A1 astrocytes in the OID therefore represent co-existing mechanisms of neuronal injury and immunosuppression that persist independently of direct malignant cell contact. These reactive glia, along with microglia and macrophages, are also primary sources of reactive oxygen species: the ROS response program is upregulated in a gradient-dependent manner, peaking at AWM in microglia and persisting through the OID in astrocytes, providing a spatially graded cell-extrinsic inducer of stress in surrounding tissue. 74,75 Together, these glial changes contribute to the region-wide transcriptional stress response compounding the neuronal injury described above. 72-75 Oligodendrocyte lineage cells exhibit a spatially restricted stress response concentrated in white matter zones, but the more striking finding is a pattern of active but failing compensation. More revealing is the combination of paradoxically upregulated myelin maturity markers alongside reduced remyelination signaling and elevated progenitor marker scores, a pattern consistent with active but failing compensation: mature oligodendrocytes are attempting to maintain existing myelin at the transcriptional level while OPCs fail to differentiate, impairing the capacity for white matter repair. Together, these findings describe a brain under sustained transcriptional stress across multiple cell types and spatial zones, in tissue that appears normal by conventional macroscopic and radiographic assessment. Current neuroprotective strategies in GB have focused predominantly on mitigating treatment-induced toxicity, including radioprotectant agents, hippocampal-sparing radiation techniques, and anti-glutamatergic combinations, and were not designed around the pre-treatment molecular landscape of the affected parenchyma. 76 Our data reveal that neuronal stress is already established before any treatment is administered, is driven by multiple parallel mechanisms including co-activation of distinct ISR kinase arms (PERK/UPR, HRI/mitochondrial), A1 astrocytic neurotoxicity, and oligolineage compensation failure, and extends into tissue outside any treatment field. Glutamate-mediated excitotoxicity through system is well-established as one contributor to peritumoral neuronal injury, 77-79 but the co-activation of multiple ISR kinase arms identified here suggests this is one of several parallel mechanisms, and that single-agent neuroprotective strategies targeting individual pathways may be insufficient. That affected neurons retain synaptic identity despite widespread stress activation suggests a therapeutic window may exist. Neuroprotective intervention informed by the specific injury mechanisms identified here could preserve cognitive and functional outcomes in a patient population for whom quality of life during the highest-functioning period after diagnosis remains an unmet clinical priority. These findings provide a spatially informed blueprint for combination therapies designed around the biology of each compartment: anti-proliferative strategies for cycling OPC-like cells in white matter, non-cell-cycle-dependent approaches including MHC-unrestricted immunotherapies for quiescent NPC2-like cells in grey matter and neuroprotective intervention directed at the specific injury mechanisms identified in the surrounding parenchyma. Additionally, therapies should reflect the disease present at the time of treatment. Gross total resection leaves predominantly OID with quiescent invaders and low checkpoint target expression, while subtotal resection retains mesenchymal-dominant, checkpoint-high tumour bulk. Current protocols often apply uniform adjuvant treatment approaches regardless of what remains, and clinical trials that select for patients with gross total resections may paradoxically select for the population whose residual disease is least amenable to the therapies under investigation. Stratification by extent of resection may be necessary to match therapeutic strategy to the biology actually present. Together these findings recontextualize the cellular states, immune biology, and therapeutic vulnerabilities previously characterized in resected tumour, providing a framework for GB therapeutics designed not in response to the tumour removed, but in anticipation of recurrence driven by the disease that remains. Methods Study design and patient cohort Tissue samples were obtained from patients who underwent en bloc temporal lobectomy at [Institution] and consented to tissue donation (REB [withheld for blinding]). Five patients with histologically confirmed glioblastoma (IDH-wildtype, WHO Grade 4; Supplementary Table 1) and one non-oncological control patient (BT077), who underwent temporal lobectomy for medically refractory epilepsy, were enrolled under the same protocol. Written informed consent was obtained from all participants. For each GB patient, five spatially matched biopsies were collected along a defined anatomical axis: tumour core, tumour edge, adjacent white matter (AWM; peri-tumoural white matter immediately bordering the tumour mass, resected en bloc with the specimen), distal white matter (dWM), and distal grey matter (dGM). dWM and dGM were sampled from macroscopically normal-appearing brain parenchyma outside the contrast-enhancing region on pre-operative MRI. One patient (BT074) contributed two independent biopsy series collected from spatially distinct regions of the same tumour (BT074-A and BT074-C), providing a deliberate within-patient assessment of sampling reproducibility across the spatial axis. The non-oncological control patient provided biopsies from white matter and grey matter locations. All tissue was snap-frozen immediately following resection and stored at -80°C. Single-cell RNA sequencing Single-cell RNA sequencing was performed using the 10x Genomics Chromium Flex probe-based library preparation kit, which enables high-quality capture from fresh-frozen tissue and multiplexing of up to 16 samples per sequencing run. All 32 samples were distributed across two 16-plex pools (Pool_152 and Pool_153), reducing batch effects introduced by sequential sample processing. Libraries were sequenced on an Illumina NovaSeq X Plus to a target depth of 30,000 reads per cell. Raw sequencing data were processed using Cell Ranger 10.0.0 with the human GRCh38 reference genome and the 10x Flex probe set (v1.0, 18,108 probes). Sample demultiplexing was performed using the Flex multiplexing algorithm within Cell Ranger. Quality control and doublet removal Per-sample quality control was performed using Scanpy 1.12.80 Cells were filtered to retain those with 205–13,700 genes/cell detected and fewer than 25% mitochondrial probe counts. Mitochondrial content was low throughout (cohort median 0.17%), consistent with the probe-based capture chemistry of the Flex kit, which includes only 12 mitochondrial probe targets. Doublets were identified using Scrublet 0.2.3 with per-sample threshold optimization, followed by manual review of doublet score distributions.81 Cells flagged as doublets by both automatic classification and manual review were excluded. 4 of the 32 samples with low cell yield were retained after confirming that their inclusion did not alter the direction or magnitude of central findings in leave-one-out sensitivity analysis (Supplementary Table 5). After quality control, 472,089 cells were retained. 80,81 Normalization, dimensionality reduction, and batch correction Count matrices were normalized to a target sum of 10,000 counts per cell and log-transformed (log1p) using Scanpy. Highly variable genes were identified using the Seurat flavour dispersion method (n=3,000 HVGs), restricted to genes that were highly variable in at least one sample (batch_key=sample_id; 32 samples). Principal component analysis was performed using randomized SVD (50 components). Batch correction was performed using Harmony 0.2.0, 82 with pool identity (Pool_152, Pool_153; 2 pools) as the batch covariate. Pool identity rather than patient identity was used as the Harmony batch covariate to preserve inter-patient biological variation, which was subsequently assessed through leave-one-out sensitivity analysis. Integration quality was assessed using the Local Inverse Simpson's Index (LISI): integration LISI improved from 1.48 to 1.64 post-correction, while cluster LISI remained at a median of 1.0, indicating effective batch mixing without loss of biological signal. UMAP embedding was computed on the Harmony-corrected neighbour graph (n_neighbors=15, min_dist=0.3) using UMAP 0.5.11. 83 Cell clustering Leiden community detection was applied to the Harmony-corrected neighbour graph at four resolutions (r=0.2, 0.5, 1.0, 2.0), yielding 14, 22, 41, and 73 clusters respectively. The r=1.0 resolution (41 clusters) was used as the primary clustering for cell type annotation. The r=0.5 resolution was used for CellTypist label propagation and the r=2.0 resolution was used for differential expression analyses requiring finer cluster granularity. Cell type annotation A multi-method annotation strategy was used, leveraging each tool for the cell populations where it performs most reliably. Malignant cell states. Malignant GB cells were annotated using the GBmap SCANVI reference model (Ruiz-Moreno et al., 2025), trained on a curated glioblastoma single-cell reference atlas. 24 GBmap assigns cells to six malignant states based on the Neftel et al. (2019) classification: AC-like, MES1-like, MES2-like, NPC1-like, NPC2-like, and OPC-like. 10 Cells with GBmap posterior probability below 0.5 were flagged as low confidence. Malignant cell classification was further refined using copy-number variation analysis and post-hoc revision, as detailed below. Normal neural populations. Neurons, oligodendrocytes, OPCs, astrocytes, endothelial cells, and vascular leptomeningeal cells were annotated using CellTypist 1.7.1 with the Adult Human MTG model, 84 applied with cluster majority-vote label propagation at Leiden r=0.5 resolution. CellTypist was used in preference to GBmap for these populations because the GBmap reference is tumour-biased and underperforms on normal brain cell types, for which distal white matter and grey matter samples are enriched. Immune populations. Major immune populations (microglia, macrophages, T cells, dendritic cells) were initially assigned by GBmap. Fine immune subpopulation annotation (18 populations) was performed using custom marker panel scoring with sc.tl.score_genes, followed by Leiden cluster modal assignment to resolve individual cells to the most frequent population within their cluster. Marker panels were derived from Villani et al. (2017) 85 for dendritic cell subtypes, Friebel et al. (2020) 86 for myeloid populations, Müller et al. (2017) 87 for monocyte subtypes, and Zheng et al. (2017) 88 for lymphocyte subtypes. Scoring-based annotation assigns a subset of activated T cells to dendritic cell populations due to overlapping marker expression; for T cell composition analyses, cluster-level modal assignment was therefore used, which correctly separates CD4 T, CD8 T, regulatory T, double-negative T, and NK cell populations. Regulatory T cells were defined by co-expression of FOXP3, IL2RA (CD25), CTLA4, and IKZF2 (Helios) within CD4-expressing clusters; double-negative T cells by CD3D/CD3E/TRAC positivity with neither CD4 nor CD8A/CD8B; NK cells by NKG7, KLRD1, and GNLY in the absence of CD3 transcripts. T cell composition statistics and per-subtype marker scoring. Cell composition per zone was computed at three denominators where specified: (i) absolute density per 1,000 total cells, (ii) percentage of CD45+ immune cells (sum of all annotated immune populations), and (iii) percentage of the T cell pool (CD4_T + CD8_T + Treg + DNT + NK_cell). For T cell composition statistical comparisons across zones, biopsy-level (sample_id-aggregated) fractions were used (n = 5 GB patients x 5 tumour zones; 4 BT077 biopsies for WM(Ctrl); 3 BT077 biopsies for GM(Ctrl)). Kruskal-Wallis H tests compared the five tumour zones for each subtype and for total T cell %CD45+; pairwise Mann-Whitney U tests compared each tumour zone to its matched non-oncological control (WM(Ctrl) for AWM and dWM; GM(Ctrl) for dGM; pooled WM+GM control for Core and Edge), with Benjamini-Hochberg correction within each subtype's pairwise test family. Treg abundance was additionally tested under six denominators (% of T cell pool, % of CD4 lineage [Treg / (CD4_T + Treg)], % of CD45+, per 1,000 total cells, % of total cells, % of non-malignant cells) to distinguish absolute population changes from denominator effects driven by changes in the surrounding T cell compartment. Per-subtype checkpoint receptor expression (CD4 T, CD8 T, Treg) was computed as the fraction of cells with normalised expression > 0 for each receptor (PDCD1/PD-1, HAVCR2/TIM-3, TIGIT, CTLA4, LAG3) per zone; cells per subtype x zone with n < 10 were flagged as low-confidence and shown as open markers in figures. PD-L1 (CD274) was analysed on its source populations (malignant cells, microglia, macrophages) rather than on T cells, since PD-L1 is the ligand and is expressed by antigen-presenting / target cells. CD4+ T cells were further characterised by helper versus cytotoxic gene-set scoring: a helper score (IL7R, CCR7, LEF1, TCF7, SELL, CD40LG, CXCR5) and a cytotoxic score (GZMB, GZMA, GZMK, PRF1, GNLY, IFNG, NKG7, CCL5), each computed via sc.tl.score_genes against the same expression-stable background as other program scores. Copy-number variation analysis Patient-level malignancy was first validated by pseudo-bulk CNV analysis: per-patient counts were aggregated across malignant and matched-normal cells, normalised to CPM, and the per-gene log₂(tumour/reference) ratio smoothed by rolling median (window=100 genes). Six lineage-matched reference panels were used (AC- and MES1/2-like vs. oligodendrocyte; NPC-like vs. neuron; OPC-like vs. OPC). Significant events (|Z|>2 vs the control patient) confirmed hallmark GB alterations (chr7 gain, chr10 loss, focal EGFR amplification) in all five patients (Supplementary Fig. 18; Supplementary Table 3). per-cell CNV calling and tumour classification Higher-resolution per-cell CNV analysis used cnv_group_flex. Each GB subtype was compared to an empirically-matched BT077 reference cell type (AC and MES vs astrocytes; OPC vs oligodendrocytes; NPC1 vs OPCs; NPC2 vs excitatory neurons), with mappings selected by minimising the residual SD of the genome-wide log₂ ratio after drift correction. Three bias-correction layers were applied: sex chromosomes and chrM were excluded; a 138-gene empirical blacklist removed cell-type markers that mimic CNV signal (e.g. GAD2, HTR2A, CNTNAP3B); and a genome-wide drift correction subtracted the median log₂ ratio over expression-stable genes to mitigate the −0.10 to −0.25 baseline shifts seen in tumour-adjacent neurons. Stage 1. Cells were partitioned into pseudobulk groups of k = 50 within each (patient, cell type, zone) stratum. Per-group regional scores were tested against bootstrap nulls (2,000 iterations) of matched BT077 cells, with BH-corrected two-sided p-values, at three resolutions: chromosome, arm (39 autosomal arms), and focal (15 GB-relevant GISTIC regions ±1 Mb). Marker libraries. For each patient, markers were derived from group-level penetrance in the GBM_AC core/edge cohort, combining drift-corrected arm markers with pre-drift focal markers (drift correction attenuates focal-amp signal). Library sizes ranged from 8 to 17 markers per patient. Stage 2. To recover per-cell resolution at k = 50, an iterative sort-and-group algorithm (50 seed rounds, 30 ranked rounds with 10% jitter) accumulated per-marker firing as an exponentially-weighted moving average (α = 0.3). Per-marker scores were averaged within four directional tiers (arm gain/loss, focal gain/loss) and combined as mean_score = 0.7·max(tier means) + 0.3·mean(remaining), bounded [0, 1]. The directional split prevents asymmetric profiles (e.g. NPC2-like loss-only) from being diluted by silent opposite-sign markers. Classification. Cell-type-specific p50 and p99 thresholds were derived per patient by applying the marker library to BT077 controls. Cells were classified normal (< p50), tumour (≥ p99), or indeterminate; cell types with p99 ≥ 0.90 were flagged unreliable. For subtypes lacking a matched BT077 reference (e.g. GBM_NPC2), fallback thresholds (p50 = 0.10, p99 = 0.50) were used (Supplementary Fig. 18). Post-hoc cell type revision. GBmap classified 92,657 cells as AC-like; however, a subset of these were non-malignant astrocytes misclassified due to shared transcriptional features between reactive astrocytes and the AC-like malignant state. A two-tier revision was applied exclusively to AC-like cells to identify and reclassify non-malignant astrocytes. In Tier 1, AC-like cells were reclassified as astrocytes if k-nearest neighbor label transfer from non-malignant cells assigned an astrocyte identity with transfer confidence ≥0.50 (n=16,367 cells; 3.5% of total; mean confidence 0.89). In Tier 2, an additional 33 cells were reclassified based on homeostatic astrocyte marker scoring above the non-oncological control median combined with low malignancy marker expression. All remaining AC-like cells (n=76,290) were retained as malignant. The revised annotation was used throughout all downstream analyses (Supplementary Table 2). Sub-compartment annotation Neuron subtypes. Of 47,262 neurons identified by CellTypist, 8,931 with pan-neuronal marker expression (mean of SNAP25, SYN1, SYN2, SYT1, STMN2, TUBB3, MAP2) below the non-oncological control 25th percentile were reclassified as non-neuronal, leaving 38,331 neurons for subtype analysis. These were subtyped into six classes, excitatory layer subtypes (Ex_L2/3, Ex_L4, Ex_L5/6) and inhibitory subtypes (Inh_PV, Inh_SST, Inh_VIP/LAMP5), using marker panels from Hodge et al. (2019) 89 and the Allen Brain Cell Atlas. Assignment was performed by maximum marker panel score with Leiden cluster modal propagation. Astrocyte states. Astrocytes (n = 25,799) were further classified into reactive states using published marker gene panels scored with sc.tl.score_genes: A1 neurotoxic (C3, SERPING1, GBP2, PSMB8, SRGN, AMIGO2) 23 , A2 neuroprotective (EMP1, S100A10, SPHK1, CD109, PTGS2, TM4SF1) 23 , Disease-Associated Astrocyte (DAA; SERPINA3, SPHK1, CD44, VIM, GFAP) 90 , and Homeostatic (AQP4, GJA1, SLC1A3) 91 . Initial assignment (v2) classified each astrocyte to the state with the highest positive score, with cells co-expressing multiple reactive markers labeled "Mixed" (n = 2,430; 9.4%) and cells with no score above threshold labeled "Uncertain" (n = 7,362; 28.5%). A refinement step (v3) resolved these categories: Mixed cells were assigned to their single highest-scoring reactive state (redistributing to A1 n = +876, A2 n = +1,407, DAA n = +147); Uncertain cells with any reactive score > 0 were assigned to the highest reactive state (n = 2,195; mostly DAA), those with homeostatic score > 0.5 were assigned Homeostatic (n = 2,109), and the remainder were assigned "Astrocyte_Low" (n = 3,058), representing cells with no clear reactive or homeostatic signature. All astrocyte reactive percentages reported in the main text use v3 annotation. The v3 reclassification does not alter any biological conclusion; the reactive gradient is preserved and the relative ordering of states is unchanged (Supplementary Table 5). Oligodendrocyte and OPC program scoring. Transcriptional programs were scored across oligodendrocytes (n=104,516) and OPCs (n=26,769) using sc.tl.score_genes. Five programs were evaluated: stress (Gavish et al. 2023, MP6; 49 genes) 11 , myelin maturity (MBP, PLP1, MOG), progenitor markers (PDGFRA, CSPG4, OLIG2), remyelination signaling, and inflammatory response. Unlike astrocytes, which exhibit discrete reactive states, oligodendrocytes and OPCs express stress and maturation markers along a continuous gradient; program scores were therefore used to capture this continuum rather than imposing categorical state boundaries. Between-zone differences were assessed using Kruskal-Wallis H test with Benjamini-Hochberg FDR correction. Immune fine populations. Fine immune annotation (18 populations; n=119,164 cells) was performed as described above. Rescue rate was 100%; all cells were assigned to a population via Leiden cluster modal annotation. Cell cycle scoring Cell cycle phase was scored for all 472,089 cells using the Tirosh et al. (2016) S-phase and G2M-phase gene lists via sc.tl.score_genes_cell_cycle in Scanpy. 14 Cells were classified as S, G2M, or G1 phase based on the maximum of S-score and G2M-score relative to a threshold of 0. Cycling fraction per subtype per zone was defined as the percentage of cells in S or G2M phase. GB malignant state program scoring Eight transcriptional programs were scored across all 144,089 malignant cells using sc.tl.score_genes (Scanpy 1.12). Gene lists were obtained from published sources: Hypoxia (MP7; 27 genes) 11 , EMT-I Mesenchymal (MP15; 30 genes) 11 , EMT-III S100/Annexin (MP17; 30 genes) 11 , Heat-Shock Stress (MP9; 29 genes) 11 , Interferon Response (14 genes) 11 , Invasion/Tumor Microtube (17 genes) 15 , Developmental Stemness (14 genes) 92 , and Neural Crest-Like identity (12 confirmed NCC markers) 64 . Gavish meta-program gene lists were obtained from the 3CA Supplementary Table (top 30 genes per meta-program, restricted to genes present in the 10x Flex probe panel). The Hamed NCL gene list comprises the 12 NCC identity markers confirmed in the original publication; injury-associated genes were excluded as they derive from a mouse injury model and include immediate-early genes susceptible to dissociation artifacts. 93 Full gene lists are provided in Supplementary Table 14. Between-zone differences were assessed using Kruskal-Wallis H test; spatial gradients were confirmed at the patient level using Friedman's test on pseudobulk means (n=5 patients x 5 tumour zones; Methods, Sensitivity analyses). Two additional cell-state-specific programmes were scored on the same 144,089 malignant cells to capture the lineage-resemblance signatures referenced in the spatial mimicry analysis: a neural mimicry panel comprising 12 neuronal identity markers (NRGN, RBFOX3, NEFL, NEFM, STMN2, MAPT, CNTNAP2, NEFH, SYT1, SNAP25, JPH3, MAP2) capturing the NPC2-like neuronal transcriptional resemblance, and an OPC lineage panel comprising 9 oligodendrocyte and OPC identity markers (MBP, PLP1, MOG, PDGFRA, CSPG4, OLIG2, SOX10, SOX8, NKX2-2) capturing the OPC-like oligodendroglial resemblance. Both panels were scored using sc.tl.score_genes with the same expression-stable background as the eight published programmes above. GB malignant subtype marker gene identification. Marker genes for each GB malignant subtype were identified by Wilcoxon rank-sum test (sc.tl.rank_genes_groups, method="wilcoxon"), comparing each subtype against all other malignant cells (n=144,089 total). Tests were performed separately for upregulated and downregulated genes. The top 5 upregulated and top 5 downregulated genes per subtype (ranked by log2 fold-change, filtered to adjusted p<0.05 and minimum detection rate 10%) were selected for visualization. Mean log-normalized expression per subtype per spatial zone was z-scored across zones and clipped at ±3 for display. Spatially variable gene analysis. To identify genes whose expression varies systematically across spatial zones, Kruskal-Wallis H tests were applied independently to each malignant subtype and non-malignant cell population. For each cell type, genes expressed in at least 3% of cells were tested for expression variation across spatial zones (5 tumour zones for malignant subtypes; 7 zones including controls for non-malignant populations). Multi-test correction was performed using Benjamini-Hochberg FDR (FDR0.3). Genes were classified as upregulated or downregulated at the infiltration front based on Spearman rank correlation between expression and spatial zone order. Malignant and non-malignant cells were analyzed separately because they have fundamentally different spatial distributions: malignant cell density decreases from Core to distal grey matter, while non-malignant cell density increases; a gene positively correlated with zone in one compartment may show the opposite pattern in the other. Cell type marker overlap test. To assess which normal cell type each malignant subtype most closely resembles transcriptionally, the spatially variable genes of each malignant subtype were compared to normal cell type marker genes. Normal cell type markers were independently derived from the 328,000 non-malignant cells using Wilcoxon rank-sum tests (sc.tl.rank_genes_groups) for each of 8 normal cell types. The overlap between each malignant subtype's spatially variable genes and each normal cell type's markers was assessed using Fisher's exact test, with odds ratios and FDR-corrected p-values reported (Supplementary Fig. 9; Supplementary Table 10). Neuron transcriptional program scoring. Cellular stress was scored across all non-malignant cell types using the Gavish et al. (2023) Stress 1 meta-program (MP6; 49 genes from the 3CA Supplementary Table, of which 49 were present in the 10x Flex probe panel). 11 The same program was applied to neurons, oligodendrocytes, OPCs, and astrocytes for consistency. Five additional neuron programs were scored using published gene sets: synaptic function (SYP, SYT1, SNAP25, VAMP2, STX1A, DLG4, NRXN1, BSN, PCLO, SHANK2), neuroplasticity/LTP (BDNF, NTRK2, ARC, HOMER1, CAMK2A, CAMK2B, CREB1, MAPK3, FMR1, SYNGAP1), oxidative stress (SOD1, SOD2, HMOX1, NQO1, GPX4, PRDX1, PRDX2, CAT, TXNRD1), neurodegeneration (APP, PSEN1, MAPT, SNCA, SQSTM1, TARDBP, CLU, VCP), and glutamate/excitotoxicity (GRIA1, GRIA2, GRIN1, GRIN2A, GRIN2B, SLC1A2, SLC1A3, GLS, GLUL). Neurons were classified as stressed if their MP6 score exceeded a zone-matched threshold defined as the non-oncological control neuron mean + 2 standard deviations, computed separately for white matter zones (threshold = 0.241) and grey matter zones (threshold = 0.102) to account for baseline differences in neuronal transcriptional state between tissue compartments. Integrated stress response (ISR) module scoring. To dissect the upstream pathways of MP6 stress activation in neurons, five ISR kinase-arm and effector modules were independently scored on the 47,262 CellTypist-identified neurons (reduced to 38,331 after pan-neuronal marker filtering): the PERK/UPR arm (HSPA5/BiP, DNAJB9, HERPUD1, ERO1A, EDEM1, HYOU1, SEC61B, DERL1, DERL2, P4HB, PDIA3, PDIA4, PDIA6, CALR, XBP1); the GCN2/kynurenine arm (ASNS, PHGDH, PSAT1, PSPH, SLC7A5, SLC1A5, CBS, CTH, MTHFD2, CYP1A1, CYP1B1, AHRR); the PKR/interferon arm (IFIT1, IFIT2, IFIT3, OAS1, OAS2, OAS3, MX1, MX2, ISG15, RSAD2, DDX58, HERC5, ISG20, GBP1, GBP2); the HRI/mitochondrial-UPR arm (ATF5, HSPD1, HSPE1, LONP1, CLPP, BNIP3L, PINK1, HMOX1, NQO1, GCLC, GCLM, TXNRD1, SRXN1); and an excitotoxicity / immediate-early gene module (NPAS4, ARC, FOSB, FOSL1, EGR3, BDNF, NR4A1, NR4A2, NR4A3). Module scores were computed using sc.tl.score_genes (use_raw=False); between-group comparisons (MP6-classified stressed vs non-stressed neurons) used the two-sided Mann-Whitney U test. Per-marker positivity rates within stressed neurons (e.g. BiP/HSPA5 firing in 18.9% of stressed neurons) were computed as the fraction with normalised expression > 0. Non-malignant cell spatial transcriptional analysis. Agnostic spatial variable gene analysis was applied to four non-malignant cell populations: neurons (n = 47,262), oligodendrocytes (n = 104,516), OPCs (n = 26,769), and astrocytes (n = 25,799). For each cell type, genes expressed in at least 3% of cells were tested for expression variation across all 7 spatial zones (including non-oncological control) using Kruskal-Wallis H test with BH FDR correction (FDR 0.3). Genes were classified as tumour-elevated (higher expression in tumour-proximal zones; Spearman rho 0.3). Pathway enrichment on each direction-specific gene list was performed using gseapy enrichr against GO Biological Process 2023 and MSigDB Hallmark 2020 gene sets. Pseudobulk differential expression Pseudobulk differential expression analysis was performed using PyDESeq2 0.5.4. For each comparison, raw counts were aggregated per patient per zone per cell type to generate pseudobulk profiles. 94 Comparisons included: (1) distal non-malignant cells vs. the non-oncological control, (2) adjacent white matter non-malignant vs. distal non-malignant, (3) adjacent white matter malignant vs. tumour edge malignant, (4) glioblastoma microglia vs. non-oncological control microglia, and (5) stressed vs. non-stressed neurons. Genes with adjusted p0.5 were considered differentially expressed. Wasserstein distance analysis Transcriptional intermediacy of spatial zones was quantified using the Wasserstein distance (earth mover's distance) between zone-level UMAP distributions. For each spatial zone, 5,000 cells were subsampled from the 2D UMAP embedding. Pairwise Wasserstein distances were computed using scipy.stats.wasserstein_distance, averaged across both UMAP dimensions (Supplementary Fig. 3). Cell-cell communication analysis Ligand-receptor interaction analysis was performed using LIANA 1.7.1 on the tumour microenvironment subset (Core + Edge cells) and the infiltration front subset (Edge + AWM cells). 95 The full 472,089-cell dataset was not processed due to memory constraints. LIANA was run with the consensus resource and five scoring methods (CellChat, NATMI, SingleCellSignalR, Connectome, log2FC; n_perms=50); interactions reported are those significant across at least three methods (FDR<0.05). Sensitivity and robustness analyses Leave-one-out (LOO) sensitivity analysis was performed to assess whether key quantitative findings were robust to the exclusion of individual patients. Each of the five glioblastoma patients was excluded in turn, and the relevant statistic was recomputed on the remaining four patients. BT074-A and BT074-C (biological replicates) were additionally excluded together as a sixth scenario. A finding was considered robust if the direction of effect was preserved across all LOO iterations. Pseudobulk sensitivity analysis was performed by aggregating program scores to patient-level means and testing with Friedman's test (non-parametric repeated-measures ANOVA; n=5 patients x 5 tumour zones). Pairwise comparisons used Wilcoxon signed-rank tests paired by patient (Supplementary Table 5). Software Used All statistical analyses were performed in Python 3.10 using scipy 1.16.3, numpy 2.3.5, and statsmodels 0.14.6. Unless otherwise stated: between-group comparisons used the Mann-Whitney U test (two-sided); multiple comparison correction used the Benjamini-Hochberg false discovery rate procedure; significance threshold was FDR<0.05. Kruskal-Wallis H test was used to assess overall between-zone differences for program score analyses. Spearman rank correlation was used for continuous variable associations. Cohen's d effect sizes for the power calculation were computed as mean delta / SD of within-patient deltas (paired design). All figures were generated using matplotlib 3.10.8 and seaborn 0.13.2. Generative AI assistance: Claude Code (Anthropic; model: Claude Opus 4.6) was used for copy-editing of manuscript prose, assistance with analysis-code refactoring, and debugging. All AI-generated text and code were reviewed, edited, and verified by the authors, who take full responsibility for the accuracy and integrity of the work. Declarations Ethics statement All tissue samples were obtained under written informed consent following protocols approved by the [Institution] Research Ethics Board (REB IDs withheld for double-blind review; provided to editorial office). Procedures conformed to the Declaration of Helsinki and applicable national guidelines. Data availability Raw and processed single-cell RNA-sequencing data have been deposited in the Gene Expression Omnibus (GEO) under accession number [withheld for double-blind review; reviewer access provided to editorial office]. The annotated AnnData object (06_publication.h5ad; 472,089 cells x 18,108 genes) and per-figure summary tables are available at [withheld for review]. All publicly available reference datasets (Darmanis et al. 2017; Siletti et al. 2023) are cited in the references and accessed via their original deposition. Code availability All analysis code, including notebooks for figure generation and the per-cell CNV pipeline (cnv_group_flex), is available at [link provided to editorial office]. The repository includes a Conda environment specification (scrna.yml) and per-script README files. Software dependencies and exact versions are documented in the Methods. Author contributions [Author contributions removed for double-blind review; provided in cover letter to editorial office.] Competing interests The authors declare no competing interests. References Stupp, R. et al. Radiotherapy plus concomitant and adjuvant temozolomide for glioblastoma. New England Journal of Medicine 352 , 987–996 (2005). https://doi.org/10.1056/NEJMoa043330 Price, M. et al. CBTRUS Statistical Report: Primary Brain and Other Central Nervous System Tumors Diagnosed in the United States in 2018–2022. Neuro-Oncology 27 , iv1–iv66 (2025). https://doi.org/10.1093/neuonc/noaf194 Wen, P. Y. et al. Glioblastoma in adults: A Society for Neuro-Oncology (SNO) and European Society of Neuro-Oncology (EANO) consensus review on current management and future directions. Neuro-Oncology (2025). https://doi.org/10.1093/neuonc/noaf177 Louis, D. N. et al. The 2021 WHO Classification of Tumors of the Central Nervous System: a summary. Neuro-Oncology 23 , 1231–1251 (2021). https://doi.org/10.1093/neuonc/noab106 Jr., E. B. & Karnosh, L. J. Cerebral Hemispherectomy. Journal of Neurosurgery 6 (1949–07–01). https://doi.org/10.3171/jns.1949.6.4.0285 Gardner, W. J. Removal of the Right Cerebral Hemisphere for Infiltrating Glioma. Journal of the American Medical Association 101 (1933/09/09). https://doi.org/10.1001/jama.1933.02740360003002 Delgado-López, P. D. & Corrales-García, E. M. Survival in glioblastoma: a review on the impact of treatment modalities. Clinical and Translational Oncology 2016 18:11 18 (2016–03–10). https://doi.org/10.1007/s12094-016-1497-x Darmanis, S. et al. Single-Cell RNA-Seq Analysis of Infiltrating Neoplastic Cells at the Migrating Front of Human Glioblastoma. Cell Reports 21 (2017/10/31). https://doi.org/10.1016/j.celrep.2017.10.030 Couturier, C. P. et al. Single-cell RNA-seq reveals that glioblastoma recapitulates a normal neurodevelopmental hierarchy. Nature Communications 11 (2020). https://doi.org/10.1038/s41467-020-17186-5 Neftel, C. et al. An Integrative Model of Cellular States, Plasticity, and Genetics for Glioblastoma. Cell 178 (2019/08/08). https://doi.org/10.1016/j.cell.2019.06.024 Gavish, A. et al. Hallmarks of transcriptional intratumour heterogeneity across a thousand tumours. Nature 2023 618:7965 618 (2023–05–31). https://doi.org/10.1038/s41586-023-06130-4 Chédeville, A. L. et al. Investigating Glioblastoma Response to Hypoxia. Biomedicines 2020, Vol. 8, Page 310 8 (2020–08–27). https://doi.org/10.3390/biomedicines8090310 Feldman, L. Hypoxia within the glioblastoma tumor microenvironment: a master saboteur of novel treatments. Frontiers in Immunology 15 (2024/06/26). https://doi.org/10.3389/fimmu.2024.1384249 Tirosh, I. et al. Single-cell RNA-seq supports a developmental hierarchy in human oligodendroglioma. Nature 2016 539:7628 539 (2016–11–02). https://doi.org/10.1038/nature20123 Venkataramani, V. et al. Glioblastoma hijacks neuronal mechanisms for brain invasion. Cell 185 , 2899–2917.e2831 (2022). https://doi.org/10.1016/j.cell.2022.06.054 D'Alessio, A. et al. Analysis of angiogenesis related factors in glioblastoma peritumoral tissue and their derived cancer stem cells. Oncotarget 7 (2016–10–01). https://doi.org/10.18632/oncotarget.12398 Jarmuzek, P. et al. Cytokine Profile in Development of Glioblastoma in Relation to Healthy Individuals. International Journal of Molecular Sciences 2023, Vol. 24, Page 16206 24 (2023–11–11). https://doi.org/10.3390/ijms242216206 Hassel, B., Niehusmann, P., Halvorsen, B. & Dahlberg, D. Pro-inflammatory cytokines in cystic glioblastoma: A quantitative study with a comparison with bacterial brain abscesses. With an MRI investigation of displacement and destruction of the brain tissue surrounding a glioblastoma. Frontiers in Oncology 12 (2022/07/29). https://doi.org/10.3389/fonc.2022.846674 Sasaki, A. et al. Secretion of Transforming Growth Factor-β1 and -β2 by Malignant Glioma Cells. Neurologia medico-chirurgica 35 (1995). https://doi.org/10.2176/nmc.35.423 Wei, Q. et al. TNFα secreted by glioma associated macrophages promotes endothelial activation and resistance against anti-angiogenic therapy. Acta Neuropathologica Communications 2021 9:1 9 (2021–04–14). https://doi.org/10.1186/s40478-021-01163-0 Soni, H. et al. PERK-mediated expression of peptidylglycine α-amidating monooxygenase supports angiogenesis in glioblastoma. Oncogenesis 2020 9:2 9 (2020–02–13). https://doi.org/10.1038/s41389-020-0201-8 Colwell, N. et al. Hypoxia in the glioblastoma microenvironment: shaping the phenotype of cancer stem-like cells. Neuro-Oncology 19 (2017/07/01). https://doi.org/10.1093/neuonc/now258 Liddelow, S. A. et al. Neurotoxic reactive astrocytes are induced by activated microglia. Nature 2017 541:7638 541 (2017–01–18). https://doi.org/10.1038/nature21029 Ruiz-Moreno, C. et al. Charting the single-cell and spatial landscape of IDH-wild-type glioblastoma with GBmap. Neuro-Oncology 27 (2025/10/14). https://doi.org/10.1093/neuonc/noaf113 Patel, A. P. et al. Single-cell RNA-seq highlights intratumoral heterogeneity in primary glioblastoma. Science 344 (2014–06–20). https://doi.org/10.1126/science.1254257 Chen, J. et al. A restricted cell population propagates glioblastoma growth after chemotherapy. Nature 2012 488:7412 488 (2012–08–01). https://doi.org/10.1038/nature11287 Bao, S. et al. Glioma stem cells promote radioresistance by preferential activation of the DNA damage response. Nature 2006 444:7120 444 (2006–10–18). https://doi.org/10.1038/nature05236 Bluestone, J. A. & Anderson, M. Tolerance in the Age of Immunotherapy. New England Journal of Medicine 383 (2020–09–17). https://doi.org/10.1056/NEJMra1911109 Handel, A. E. et al. The role of thymic tolerance in CNS autoimmune disease. Nature Reviews Neurology 2018 14:12 14 (2018–11–19). https://doi.org/10.1038/s41582-018-0095-7 Ostermann, S. et al. Plasma and Cerebrospinal Fluid Population Pharmacokinetics of Temozolomide in Malignant Glioma Patients. Clinical Cancer Research 10 (2004/06/01). https://doi.org/10.1158/1078-0432.CCR-03-0807 Portnow, J. et al. The Neuropharmacokinetics of Temozolomide in Patients with Resectable Brain Tumors: Potential Implications for the Current Approach to Chemoradiation. Clinical Cancer Research 15 (2009/11/15). https://doi.org/10.1158/1078-0432.CCR-09-1349 Rosso, L. et al. A New Model for Prediction of Drug Distribution in Tumor and Normal Tissues: Pharmacokinetics of Temozolomide in Glioma Patients. Cancer Research 69 (2009/01/01). https://doi.org/10.1158/0008-5472.CAN-08-2356 Piccirillo, S. G. M. et al. Bone morphogenetic proteins inhibit the tumorigenic potential of human brain tumour-initiating cells. Nature 2006 444:7120 444 (2006/12). https://doi.org/10.1038/nature05349 Ying, M. et al. Regulation of glioblastoma stem cells by retinoic acid: role for Notch pathway inhibition. Oncogene 2011 30:31 30 (2011–03–07). https://doi.org/10.1038/onc.2011.58 Campos, B. et al. Differentiation Therapy Exerts Antitumor Effects on Stem-like Glioma Cells. Clinical Cancer Research 16 (2010/05/15). https://doi.org/10.1158/1078-0432.CCR-09-1800 Tallman, M. M. et al. Improving Localized Radiotherapy for Glioblastoma via Small Molecule Inhibition of KIF11. Cancers 2023, Vol. 15, Page 3173 15 (2023–06–13). https://doi.org/10.3390/cancers15123173 Shen, H., Hau, E., Joshi, S., Dilda, P. J. & McDonald, K. L. Sensitization of Glioblastoma Cells to Irradiation by Modulating the Glucose Metabolism. Molecular Cancer Therapeutics 14 (2015/08/01). https://doi.org/10.1158/1535-7163.MCT-15-0247 Ahmed, S. U. et al. Selective Inhibition of Parallel DNA Damage Response Pathways Optimizes Radiosensitization of Glioblastoma Stem-like Cells. Cancer Research 75 (2015/10/15). https://doi.org/10.1158/0008-5472.CAN-14-3790 Hirose, Y., Berger, M. S. & Pieper, R. O. Abrogation of the Chk1-mediated G2 Checkpoint Pathway Potentiates Temozolomide-induced Toxicity in a p53-independent Manner in Human Glioblastoma Cells1. Cancer Research 61 , 5843–5849 (2001). Hoang‐Minh, L. B. et al. Infiltrative and drug‐resistant slow‐cycling cells support metabolic heterogeneity in glioblastoma. The EMBO Journal 2018 37:23 37 (2018–10–15). https://doi.org/10.15252/embj.201798772 Sato, A. et al. Glioma-Initiating Cell Elimination by Metformin Activation of FOXO3 via AMPK. Stem Cells Translational Medicine 1 (2012/11/01). https://doi.org/10.5966/sctm.2012-0058 Zhang, P. et al. Therapeutic targeting of tumor-associated myeloid cells synergizes with radiation therapy for glioblastoma. Proceedings of the National Academy of Sciences 116 (2019–11–19). https://doi.org/10.1073/pnas.1906346116 Khan, F. et al. Macrophages and microglia in glioblastoma: heterogeneity, plasticity, and therapy. The Journal of Clinical Investigation 133 (2023/01/03). https://doi.org/10.1172/JCI163446 Kumar, S. et al. Molecular heterogeneity of Glioblastoma-associated microglia and macrophages and myeloid-derived suppressor cells: Insights from single-cell omics and therapeutic implications. Critical Reviews in Oncology/Hematology 220 (2026/04/01). https://doi.org/10.1016/j.critrevonc.2026.105190 Akay, F. & Saleh, M. Rational therapeutic targeting of myeloid cells in glioblastoma: challenges and perspectives. Frontiers in Immunology 16 (2025/06/26). https://doi.org/10.3389/fimmu.2025.1472710 Gómez, E. G. et al. Efficacy and safety of immune checkpoint inhibitors and mTOR inhibitors as targeted therapy for glioblastoma: A systematic review and meta-analysis of randomized clinical trials. Neurosurgical Review 2026 49:1 49 (2026–01–03). https://doi.org/10.1007/s10143-025-04043-6 Melo, S. M. d. et al. Anti‐PD‐1 and anti‐PD‐L1 antibodies for glioma - de Melo, SM - 2025 | Cochrane Library. Cochrane Database of Systematic Reviews 2025 (2025). https://doi.org/10.1002/14651858.CD012532.pub2 Zeng, Y.-F. et al. The efficacy and safety of anti-PD-1/PD-L1 in treatment of glioma: a single-arm meta-analysis. Frontiers in Immunology 14 (2023/04/14). https://doi.org/10.3389/fimmu.2023.1168244 Tykocki, T. Immunotherapy failure in glioblastoma: A systematic review and meta-analysis of randomized controlled trials. Critical Reviews in Oncology/Hematology 222 (2026/06/01). https://doi.org/10.1016/j.critrevonc.2026.105279 Skadborg, S. K. et al. Nivolumab Reaches Brain Lesions in Patients with Recurrent Glioblastoma and Induces T-cell Activity and Upregulation of Checkpoint Pathways. Cancer Immunology Research 12 (2024/09/01). https://doi.org/10.1158/2326-6066.CIR-23-0959 Lee, A. H. et al. Neoadjuvant PD-1 blockade induces T cell and cDC1 activation but fails to overcome the immunosuppressive tumor associated macrophages in recurrent glioblastoma. Nature Communications 2021 12:1 12 (2021–11–26). https://doi.org/10.1038/s41467-021-26940-2 Cloughesy, T. F. et al. Neoadjuvant anti-PD-1 immunotherapy promotes a survival benefit with intratumoral and systemic immune responses in recurrent glioblastoma. Nature Medicine 2019 25:3 25 (2019–02–11). https://doi.org/10.1038/s41591-018-0337-7 Schalper, K. A. et al. Neoadjuvant nivolumab modifies the tumor immune microenvironment in resectable glioblastoma. Nature Medicine 2019 25:3 25 (2019–02–11). https://doi.org/10.1038/s41591-018-0339-5 Nose, Y. et al. PD-1 antibody-bound progenitor-exhausted CD8+ T cells in lymph nodes boost PD-1-blockade anti-tumor immunity in gastrointestinal cancer. Nature Communications 2026 (2026–04–08). https://doi.org/10.1038/s41467-026-70751-2 Tsui, C. et al. Lymph nodes fuel KLF2-dependent effector CD8+ T cell differentiation during chronic infection and checkpoint blockade. Nature Immunology 2025 26:10 26 (2025–09–15). https://doi.org/10.1038/s41590-025-02276-7 Wijesinghe, S. K. M. et al. Lymph-node-derived stem-like but not tumor-tissue-resident CD8+ T cells fuel anticancer immunity. Nature Immunology 2025 26:8 26 (2025–07–29). https://doi.org/10.1038/s41590-025-02219-2 van Pul, K. M., Fransen, M. F., van de Ven, R. & de Gruijl, T. D. Immunotherapy Goes Local: The Central Role of Lymph Nodes in Driving Tumor Infiltration and Efficacy. Frontiers in Immunology 12 (2021/03/01). https://doi.org/10.3389/fimmu.2021.643291 Dammeijer, F. et al. The PD-1/PD-L1-Checkpoint Restrains T cell Immunity in Tumor-Draining Lymph Nodes. Cancer Cell 38 (2020/11/09). https://doi.org/10.1016/j.ccell.2020.09.001 Miller, B. C. et al. Subsets of exhausted CD8+ T cells differentially mediate tumor control and respond to checkpoint blockade. Nature Immunology 2019 20:3 20 (2019–02–18). https://doi.org/10.1038/s41590-019-0312-6 Vanni, A. et al. 5‐ALA Assisted Surgery of Human Glioblastoma Samples Reveals an Enrichment of T Cells Expressing PD‐1 and CD103 in the Intermediate and Marginal Layers. European Journal of Immunology 55 (2025/06/01). https://doi.org/10.1002/eji.202451681 Romagnoli, G. et al. CD8+CD103+PD1+TIM3+ T cells in glioblastoma microenvironment correlate with prognosis. Immunology 171 (2024/02/01). https://doi.org/10.1111/imm.13710 Pedersen, V. et al. Tumor and immune cell distribution in the tumor core and outer part of glioblastoma, IDH wildtype. Journal of Neuro-Oncology 2025 175:3 175 (2025–09–25). https://doi.org/10.1007/s11060-025-05232-5 Galvez-Cancino, F. et al. Regulatory T cell depletion promotes myeloid cell activation and glioblastoma response to anti-PD1 and tumor-targeting antibodies. Immunity 58 (2025/05/13). https://doi.org/10.1016/j.immuni.2025.03.021 Hamed, A. A. et al. Gliomagenesis mimics an injury response orchestrated by neural crest-like cells. Nature 2025 638:8050 638 (2025–01–01). https://doi.org/10.1038/s41586-024-08356-2 Karschnia, P. et al. Chimeric Antigen Receptor T Cells for Glioblastoma. Neurology 97 (2021–August–3). https://doi.org/10.1212/WNL.0000000000012193 Hayat, M. A., Yu, S., Tao, G., Ghani, M. U. & Hu, J. Chimeric antigen receptor T cell therapy for glioblastoma: overcoming current barriers and strategies to enhance efficacy for therapeutic implications. Biochemical Pharmacology 248 (2026/06/01). https://doi.org/10.1016/j.bcp.2026.117851 Pant, A., Lim, M., Pant, A. & Lim, M. CAR-T Therapy in GBM: Current Challenges and Avenues for Improvement. Cancers 2023, Vol. 15, Page 1249 15 (2023–02–16). https://doi.org/10.3390/cancers15041249 Ismail, F. S. et al. Current and Future Roles of Chimeric Antigen Receptor T-Cell Therapy in Neurology. JAMA Neurology 82 (2025/01/01). https://doi.org/10.1001/jamaneurol.2024.3818 Luksik, A. S. et al. CAR T Cell Therapy in Glioblastoma: Overcoming Challenges Related to Antigen Expression. Cancers 2023, Vol. 15, Page 1414 15 (2023–02–23). https://doi.org/10.3390/cancers15051414 Goel, S., Wharton, S. B., Brett, L. P. & Whittle, I. R. Morphological changes and stress responses in neurons in cerebral cortex infiltrated by diffuse astrocytoma. Neuropathology 23 (2003/12/01). https://doi.org/10.1046/j.1440-1789.2003.00510.x Pakos‐Zebrucka, K. et al. The integrated stress response. The EMBO Reports 2016 17:10 17 (2016–09–14). https://doi.org/10.15252/embr.201642195 Perelroizen, R. et al. Astrocyte immunometabolic regulation of the tumour microenvironment drives glioblastoma pathogenicity. Brain 145 (2022/09/14). https://doi.org/10.1093/brain/awac222 Faust Akl, C. et al. Glioblastoma-instructed astrocytes suppress tumour-specific T cell immunity. Nature 2025 643:8070 643 (2025–05–21). https://doi.org/10.1038/s41586-025-08997-x Conti, A. et al. Role of Inflammation and Oxidative Stress Mediators in Gliomas. Cancers 2010, Vol. 2, Pages 693-712 2 (2010–04–26). https://doi.org/10.3390/cancers2020693 Olivier, C., Oliver, L., Lalier, L. & Vallette, F. M. Drug Resistance in Glioblastoma: The Two Faces of Oxidative Stress. Frontiers in Molecular Biosciences 7 (2021/01/27). https://doi.org/10.3389/fmolb.2020.620677 Kumaria, A. & Ashkan, K. Novel therapeutic strategies in glioma targeting glutamatergic neurotransmission. Brain Research 1818 (2023/11/01). https://doi.org/10.1016/j.brainres.2023.148515 Sokolov, E., Dietrich, J. & Cole, A. J. The complexities underlying epilepsy in people with glioblastoma. The Lancet Neurology 22 (2023/06/01). https://doi.org/10.1016/S1474-4422(23)00031-5 Oh, M. C. et al. Overexpression of Calcium-Permeable Glutamate Receptors in Glioblastoma Derived Brain Tumor Initiating Cells. PLOS ONE 7 (Oct 23, 2012). https://doi.org/10.1371/journal.pone.0047846 Buckingham, S. C. & Robel, S. Glutamate and tumor-associated epilepsy: Glial cell dysfunction in the peritumoral environment. Neurochemistry International 63 (2013/12/01). https://doi.org/10.1016/j.neuint.2013.01.027 Wolf, F. A. et al. SCANPY: large-scale single-cell gene expression data analysis. Genome Biology 2018 19:1 19 (2018–02–06). https://doi.org/10.1186/s13059-017-1382-0 Wolock, S. L., Lopez, R. & Klein, A. M. Scrublet: Computational Identification of Cell Doublets in Single-Cell Transcriptomic Data. Cell Systems 8 (2019/04/24). https://doi.org/10.1016/j.cels.2018.11.005 Korsunsky, I. et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nature Methods 2019 16:12 16 (2019–11–18). https://doi.org/10.1038/s41592-019-0619-0 McInnes, L., Healy, J. & Melville, J. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. (2018/02/09). https://doi.org/10.48550/arXiv.1802.03426 Conde, C. D. et al. Cross-tissue immune cell analysis reveals tissue-specific features in humans. Science 376 (2022–05–13). https://doi.org/10.1126/science.abl5197 Villani, A.-C. et al. Single-cell RNA-seq reveals new types of human blood dendritic cells, monocytes, and progenitors. Science 356 (2017–04–21). https://doi.org/10.1126/science.aah4573 Friebel, E. et al. Single-Cell Mapping of Human Brain Cancer Reveals Tumor-Specific Instruction of Tissue-Invading Leukocytes. Cell 181 (2020/06/25). https://doi.org/10.1016/j.cell.2020.04.055 Müller, S. et al. Single-cell profiling of human gliomas reveals macrophage ontogeny as a basis for regional differences in macrophage activation in the tumor microenvironment. Genome Biology 2017 18:1 18 (2017–12–20). https://doi.org/10.1186/s13059-017-1362-4 Zheng, G. X. Y. et al. Massively parallel digital transcriptional profiling of single cells. Nature Communications 2017 8:1 8 (2017–01–16). https://doi.org/10.1038/ncomms14049 Hodge, R. D. et al. Conserved cell types with divergent features in human versus mouse cortex. Nature 2019 573:7772 573 (2019–08–21). https://doi.org/10.1038/s41586-019-1506-7 Habib, N. et al. Disease-associated astrocytes in Alzheimer’s disease and aging. Nature Neuroscience 2020 23:6 23 (2020–04–27). https://doi.org/10.1038/s41593-020-0624-8 Zhang, Y. et al. Purification and Characterization of Progenitor and Mature Human Astrocytes Reveals Transcriptional and Functional Differences with Mouse. Neuron 89 (2016/01/06). https://doi.org/10.1016/j.neuron.2015.11.013 Richards, L. M. et al. Gradient of Developmental and Injury Response transcriptional states defines functional vulnerabilities underpinning glioblastoma heterogeneity. Nature Cancer 2021 2:2 2 (2021–01–04). https://doi.org/10.1038/s43018-020-00154-9 van den Brink, S. C. et al. Single-cell sequencing reveals dissociation-induced gene expression in tissue subpopulations. Nature Methods 2017 14:10 14 (2017–09–29). https://doi.org/10.1038/nmeth.4437 Muzellec, B., Teleńczuk, M., Cabeli, V. & Andreux, M. PyDESeq2: a python package for bulk RNA-seq differential expression analysis. Bioinformatics 39 (2023/09/02). https://doi.org/10.1093/bioinformatics/btad547 Dimitrov, D. et al. Comparison of methods and resources for cell-cell communication inference from single-cell RNA-Seq data. Nature Communications 2022 13:1 13 (2022–06–09). https://doi.org/10.1038/s41467-022-30755-0 Additional Declarations There is NO Competing Interest. Supplementary Files GraphicalAbstract.jpg Graphical Abstract ST01patientdemographics.csv Supplementary Table 1 ST05loosensitivity.csv Supplementary Table 5 ST03acnvpseudobulkscores.csv Supplementary Table 3a ST03bcnvdrivergenes.csv Supplementary Table 4 ST07gbmcyclingvalidation.csv Supplementary Table 7 ST11icicheckpointstats.csv Supplementary Table 11 ST04cellcompositionbyzone.csv Supplementary Table 3b ST02qcmetricspersample.csv Supplementary Table 2 ST12neuronprogrammestats.csv Supplementary Table 12 ST14genelists.csv Supplementary Table 14 ST08spatialvariablegenes.csv Supplementary Table 8 ST09pathwayenrichment.csv Supplementary Table 9 ST10celltypemarkerenrichment.csv Supplementary Table 10 ST13stressedneurondeg.csv Supplementary Table 13 SupplementalFigures.docx Supplementary Figures ST06bt077vssiletti.csv Supplementary Table 6 Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9590848","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Biological Sciences - Article","associatedPublications":[],"authors":[{"id":634039681,"identity":"45e7d118-f4b3-4094-895c-4f345da0e79f","order_by":0,"name":"Teresa Purzner","email":"data:image/png;base64,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","orcid":"","institution":"Queen's University","correspondingAuthor":true,"prefix":"","firstName":"Teresa","middleName":"","lastName":"Purzner","suffix":""},{"id":634039682,"identity":"54951d79-56c5-44ff-a78f-66f3d9ad6a13","order_by":1,"name":"Kaytlin Andrews","email":"","orcid":"https://orcid.org/0009-0007-7673-3621","institution":"Queen's University","correspondingAuthor":false,"prefix":"","firstName":"Kaytlin","middleName":"","lastName":"Andrews","suffix":""},{"id":634039683,"identity":"1ce9f8cd-60c6-4629-98ce-57b8b9f76d46","order_by":2,"name":"Jacob Howran","email":"","orcid":"","institution":"Queen's University","correspondingAuthor":false,"prefix":"","firstName":"Jacob","middleName":"","lastName":"Howran","suffix":""},{"id":634039684,"identity":"cf0c86ad-2a23-491c-b852-9b198dd0c639","order_by":3,"name":"Douglas Quilty","email":"","orcid":"","institution":"Queen's University","correspondingAuthor":false,"prefix":"","firstName":"Douglas","middleName":"","lastName":"Quilty","suffix":""},{"id":634039685,"identity":"1492af19-6f9c-4085-b872-b189d4643ddc","order_by":4,"name":"Gary Bader","email":"","orcid":"https://orcid.org/0000-0003-0185-8861","institution":"University of Toronto","correspondingAuthor":false,"prefix":"","firstName":"Gary","middleName":"","lastName":"Bader","suffix":""},{"id":634039686,"identity":"85a689df-68e7-4a1a-afd7-a393ff5d6986","order_by":5,"name":"Pamela S. Ohashi","email":"","orcid":"","institution":"Princess Margaret Cancer Centre","correspondingAuthor":false,"prefix":"","firstName":"Pamela","middleName":"S.","lastName":"Ohashi","suffix":""},{"id":634039687,"identity":"489cddec-da13-4797-9b7c-12554f6ea53a","order_by":6,"name":"John Rossiter","email":"","orcid":"","institution":"Queen's University","correspondingAuthor":false,"prefix":"","firstName":"John","middleName":"","lastName":"Rossiter","suffix":""},{"id":634039688,"identity":"6ebafc99-8887-40bc-8bf3-579726f4acfe","order_by":7,"name":"Douglas Cook","email":"","orcid":"","institution":"Queen's University","correspondingAuthor":false,"prefix":"","firstName":"Douglas","middleName":"","lastName":"Cook","suffix":""},{"id":634039689,"identity":"44148c18-6021-4ab3-bf9d-0d039e7106a2","order_by":8,"name":"Shervin Taslimi","email":"","orcid":"","institution":"Queen's University","correspondingAuthor":false,"prefix":"","firstName":"Shervin","middleName":"","lastName":"Taslimi","suffix":""},{"id":634039690,"identity":"64085df2-1edb-451a-aea5-7e4bcd84aba7","order_by":9,"name":"Ron Levy","email":"","orcid":"","institution":"Queen's University","correspondingAuthor":false,"prefix":"","firstName":"Ron","middleName":"","lastName":"Levy","suffix":""},{"id":634039691,"identity":"25d2d54e-e31c-47b1-8049-e99cb76a5f40","order_by":10,"name":"James Purzner","email":"","orcid":"","institution":"Queen's University","correspondingAuthor":false,"prefix":"","firstName":"James","middleName":"","lastName":"Purzner","suffix":""}],"badges":[],"createdAt":"2026-05-02 05:30:18","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9590848/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9590848/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":109205739,"identity":"1d54ad25-3c5f-486d-a827-4d3482759e28","added_by":"auto","created_at":"2026-05-13 15:07:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":5908806,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSingle-cell atlas of glioblastoma across the tumour-to-brain axis. (A,B) \u003c/strong\u003ePre-operative T1-weighted post-contrast (A) and T2-FLAIR (B) MRI of a representative en bloc temporal lobectomy patient, with the tumour and lobectomy outline shown.\u003cstrong\u003e (C) \u003c/strong\u003eGross specimen of an en bloc temporal lobectomy with the five anatomical sampling zones marked: tumour Core, Edge, adjacent white matter (AWM), distal white matter (dWM), and distal grey matter (dGM).\u003cstrong\u003e (D) \u003c/strong\u003eUMAP embedding of all 472,089 single cells coloured by tissue zone (n = 7 zones, including non-oncological control white matter and grey matter).\u003cstrong\u003e (E) \u003c/strong\u003eUMAP embedding coloured by cell type, showing 18 transcriptionally distinct populations spanning malignant glioblastoma states (AC-, MES1-, MES2-, NPC1-, NPC2-, OPC-like), normal glia (oligodendrocyte, OPC, astrocyte), neurons (excitatory, inhibitory), and immune populations (microglia, macrophage, T cell, DC, mast).\u003cstrong\u003e(F) \u003c/strong\u003eStacked bar chart of cell type composition by zone, showing progressive change across the spatial axis. Total cell counts per zone are indicated above each bar.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/52a0c63bb06756d9a6a0dace.png"},{"id":109068112,"identity":"1c0d5ed6-dca5-4d6a-b0ff-9a8a8f43300f","added_by":"auto","created_at":"2026-05-12 10:03:42","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":2539166,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSpatial reorganisation of malignant cell states across the tumour-to-brain axis. (A) \u003c/strong\u003eStacked bar chart of malignant subtype composition by zone (AC-, MES1-, MES2-, NPC1-, NPC2-, OPC-like). Mesenchymal subtypes dominate the tumour bulk; NPC2-like cells become enriched in distal grey matter and OPC-like in distal white matter.\u003cstrong\u003e (B) \u003c/strong\u003eHeatmap of malignant subtype marker gene expression across spatial zones (z-scored mean expression per subtype x zone). Markers include canonical Neftel-2019 genes plus data-driven markers.\u003cstrong\u003e (C) \u003c/strong\u003eCycling fraction (% S + G2M) by malignant subtype and zone. Cycling fractions decline monotonically from tumour core to distal zones across all subtypes.\u003cstrong\u003e (D) \u003c/strong\u003eMean programme scores (Hypoxia [MP7], Interferon Response, Invasion/Tumour Microtube, Developmental Stemness) by malignant subtype across spatial zones. Subtype-specific spatial dynamics include Hypoxia decline in MES2-like cells, Interferon peaking at Edge in MES1-like, and Stemness/Invasion peaking at AWM in AC-like cells.\u003cstrong\u003e (E) \u003c/strong\u003eNeural mimicry score by malignant subtype across spatial zones. The signature is near-absent in mesenchymal subtypes throughout and rises sharply in NPC2- and OPC-like cells in distal compartments.\u003cstrong\u003e (F) \u003c/strong\u003eOPC lineage score by malignant subtype across spatial zones, peaking at 0.884 in OPC-like cells at dWM, the highest score of any malignant subtype at any zone.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/964519a2bd183ceb9f485eab.png"},{"id":108840065,"identity":"215d9ffb-71f3-42ed-b288-75832c1fd97f","added_by":"auto","created_at":"2026-05-09 00:52:15","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1401634,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSpatial remodeling of the immune microenvironment. (A) \u003c/strong\u003eButterfly plot of immune cell composition by zone, separated into brain-resident (left: microglia, DC, pDC, B cell, plasma cells) and recruited (right: MDM, neutrophil, CD14+/CD16+ monocyte, CD16+ monocyte, granulocyte, IMC) populations. Bars show percent of CD45+ immune cells per zone.\u003cstrong\u003e (B) \u003c/strong\u003eStacked bar chart of T cell subtype composition by zone (CD4 T, CD8 T, Treg, NK, DN T).\u003cstrong\u003e (C) \u003c/strong\u003eSix-panel summary of T cell checkpoint receptor expression (PD-1, TIM-3, TIGIT, CTLA-4, LAG-3) by T cell subtype across zones, plus PD-L1 (CD274) on producing cell types (Malignant, Microglia, Macrophage). Receptor expression is highest at tumour Core/Edge and declines distally toward control levels.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/8cc67939e43689e90cc03b16.png"},{"id":108840103,"identity":"2ca0e0cc-840f-49bc-8f5c-fba741f7fe96","added_by":"auto","created_at":"2026-05-09 00:52:28","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":2676133,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eNon-malignant cell stress and reactive states extend throughout radiographically normal brain. (A) \u003c/strong\u003eThree-panel UMAP embedding of all neurons (n = 47,262) coloured by zone, neuron subtype, and stress score (Gavish 2023 MP6). Adequate neuron recovery in AWM/dWM/dGM/GM(Ctrl) supports zone-level analyses.\u003cstrong\u003e (B) \u003c/strong\u003eScore difference (dGM tumour patients minus GM(Ctrl)) for six neuron transcriptional programmes. Stress/injury response (Gavish 2023 MP6) shows the largest positive shift (+0.215), with neuroplasticity/LTP and glutamate/excitotoxicity also elevated.\u003cstrong\u003e (C) \u003c/strong\u003eAstrocyte reactive state composition across zones (Homeostatic, Reactive A1 [neurotoxic], Reactive A2, Disease-associated [DAA], Mixed/Low-grade reactive). Reactive states (A1+A2+DAA) dominate at tumour Edge (96.7%) and decline distally. \u003cstrong\u003e(D) \u003c/strong\u003eOligodendrocyte transcriptional programme score differences across all five tumour-patient zones (Core, Edge, AWM, dWM, dGM), shown as five programme lines (Stress MP6, Myelin maturity, Progenitor markers, Remyelination, Inflammatory) with per-zone FDR significance markers above each point and zero-line shading (red = elevated above non-oncological control, blue = reduced).\u003cstrong\u003e (E) \u003c/strong\u003eOPC transcriptional programme score differences across the same five zones in matching line-plot format. Myelin and progenitor programmes are strongly reduced at Core/Edge (myelin −0.50/−0.52, progenitor −0.39/−0.29; FDR\u0026lt;0.05) and recover toward neutral in oligodendroglial-rich distal zones, while OPC stress remains elevated more broadly than in mature oligodendrocytes.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/9730aa78e1fee55d16d0fcc6.png"},{"id":109219672,"identity":"0686d714-59f8-4f4b-9b3d-8a54a2165942","added_by":"auto","created_at":"2026-05-13 19:59:19","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":12160695,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/14593d42-da06-436e-a6c6-46b79f7a96be.pdf"},{"id":108840060,"identity":"4415c5b6-b774-4062-bc74-49e81280b564","added_by":"auto","created_at":"2026-05-09 00:52:15","extension":"jpg","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1638025,"visible":true,"origin":"","legend":"Graphical Abstract","description":"","filename":"GraphicalAbstract.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/76cff675bcdfe232740b5c16.jpg"},{"id":108976645,"identity":"d8b08736-a6a7-49b7-970b-bfeae5667a86","added_by":"auto","created_at":"2026-05-11 11:27:10","extension":"csv","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":912,"visible":true,"origin":"","legend":"Supplementary Table 1","description":"","filename":"ST01patientdemographics.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/6d8444f8976e5f4875de29a4.csv"},{"id":108840062,"identity":"c331ad0f-5973-4f4a-9d67-d8401f1f8e2a","added_by":"auto","created_at":"2026-05-09 00:52:15","extension":"csv","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":3524,"visible":true,"origin":"","legend":"Supplementary Table 5","description":"","filename":"ST05loosensitivity.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/0588f57c3b85dd9d26d86872.csv"},{"id":108840102,"identity":"308c1e75-6b79-43cc-a9eb-ea35cd3f83d8","added_by":"auto","created_at":"2026-05-09 00:52:28","extension":"csv","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":636,"visible":true,"origin":"","legend":"Supplementary Table 3a","description":"","filename":"ST03acnvpseudobulkscores.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/748e2e62e2c1c530b58fbb87.csv"},{"id":108840081,"identity":"79078d2b-6ace-4179-b75e-74d8ae4db60e","added_by":"auto","created_at":"2026-05-09 00:52:16","extension":"csv","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":5941,"visible":true,"origin":"","legend":"Supplementary Table 4","description":"","filename":"ST03bcnvdrivergenes.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/9f434a949538200bbc333aaf.csv"},{"id":108840078,"identity":"946b2d71-aff5-4707-9e64-78d7be211976","added_by":"auto","created_at":"2026-05-09 00:52:15","extension":"csv","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":315,"visible":true,"origin":"","legend":"Supplementary Table 7","description":"","filename":"ST07gbmcyclingvalidation.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/36ddba08826ae4eb913159fd.csv"},{"id":108840079,"identity":"73d5eb30-ef79-401d-b9a2-63240d31ae1c","added_by":"auto","created_at":"2026-05-09 00:52:16","extension":"csv","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":963,"visible":true,"origin":"","legend":"Supplementary Table 11","description":"","filename":"ST11icicheckpointstats.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/54e468ca132aedabb6e55ccf.csv"},{"id":108977649,"identity":"6395d0ee-51f8-4622-bbeb-0674de3d4d72","added_by":"auto","created_at":"2026-05-11 11:32:26","extension":"csv","order_by":8,"title":"","display":"","copyAsset":false,"role":"supplement","size":1829,"visible":true,"origin":"","legend":"Supplementary Table 3b","description":"","filename":"ST04cellcompositionbyzone.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/fcc787a9381f2c2a4ad21f2e.csv"},{"id":108976909,"identity":"29107e58-ac6e-41e6-ba8e-fdeb1d2c5bb3","added_by":"auto","created_at":"2026-05-11 11:29:35","extension":"csv","order_by":9,"title":"","display":"","copyAsset":false,"role":"supplement","size":3910,"visible":true,"origin":"","legend":"Supplementary Table 2","description":"","filename":"ST02qcmetricspersample.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/01237bd98c31a15a66caaa19.csv"},{"id":108840082,"identity":"3ca6b0cd-c254-4da7-914b-d261e76e1b9c","added_by":"auto","created_at":"2026-05-09 00:52:16","extension":"csv","order_by":10,"title":"","display":"","copyAsset":false,"role":"supplement","size":1330,"visible":true,"origin":"","legend":"Supplementary Table 12","description":"","filename":"ST12neuronprogrammestats.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/4d2483f47491a68cbb31b504.csv"},{"id":108840084,"identity":"6a06bcaa-b1ba-4cac-aec5-8235c22a53a8","added_by":"auto","created_at":"2026-05-09 00:52:16","extension":"csv","order_by":11,"title":"","display":"","copyAsset":false,"role":"supplement","size":3872,"visible":true,"origin":"","legend":"Supplementary Table 14","description":"","filename":"ST14genelists.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/69b52383abb602d80df492ed.csv"},{"id":108840086,"identity":"8b444e96-7c13-4298-8350-a4615e0dcdbc","added_by":"auto","created_at":"2026-05-09 00:52:16","extension":"csv","order_by":12,"title":"","display":"","copyAsset":false,"role":"supplement","size":384121,"visible":true,"origin":"","legend":"Supplementary Table 8","description":"","filename":"ST08spatialvariablegenes.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/eec71636601a0cc42c65a181.csv"},{"id":108840087,"identity":"7f96d7cb-5d89-4c05-8a7e-6de660f8d873","added_by":"auto","created_at":"2026-05-09 00:52:16","extension":"csv","order_by":13,"title":"","display":"","copyAsset":false,"role":"supplement","size":186317,"visible":true,"origin":"","legend":"Supplementary Table 9","description":"","filename":"ST09pathwayenrichment.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/fe8972053038de9efaf9c103.csv"},{"id":109203918,"identity":"5908bbc4-dcad-40f2-8a91-2f9359da2798","added_by":"auto","created_at":"2026-05-13 14:50:37","extension":"csv","order_by":14,"title":"","display":"","copyAsset":false,"role":"supplement","size":14636,"visible":true,"origin":"","legend":"Supplementary Table 10","description":"","filename":"ST10celltypemarkerenrichment.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/c4f71c96c117192163fe37a7.csv"},{"id":108977283,"identity":"3b952596-0703-4313-b311-3ed595fe4d80","added_by":"auto","created_at":"2026-05-11 11:31:11","extension":"csv","order_by":15,"title":"","display":"","copyAsset":false,"role":"supplement","size":1750653,"visible":true,"origin":"","legend":"Supplementary Table 13","description":"","filename":"ST13stressedneurondeg.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/b6fa78e01174c92f65846a9f.csv"},{"id":109068119,"identity":"028dd088-5aa5-4c4a-8b67-3a93321a0a24","added_by":"auto","created_at":"2026-05-12 10:03:46","extension":"docx","order_by":16,"title":"","display":"","copyAsset":false,"role":"supplement","size":6589913,"visible":true,"origin":"","legend":"Supplementary Figures","description":"","filename":"SupplementalFigures.docx","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/ad6b3aa2242717ac62a634bd.docx"},{"id":108840085,"identity":"7b32f652-bd92-4ba6-a679-6d0b0b756be6","added_by":"auto","created_at":"2026-05-09 00:52:16","extension":"csv","order_by":17,"title":"","display":"","copyAsset":false,"role":"supplement","size":769,"visible":true,"origin":"","legend":"Supplementary Table 6","description":"","filename":"ST06bt077vssiletti.csv","url":"https://assets-eu.researchsquare.com/files/rs-9590848/v1/c1e78f0e289140c5133b9866.csv"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Single-cell atlas of glioblastoma across the tumour-to-brain axis reveals the distinct cellular landscape of infiltrative disease","fulltext":[{"header":"Introduction","content":"\u003cp\u003eGlioblastoma (GB) is the most common malignant brain tumour in adults, with a median survival of 14-17 months and 5-year survival of 5% despite maximal surgical resection,\u003csup\u003e1-4\u003c/sup\u003e\u0026nbsp; radiotherapy, and temozolomide chemotherapy. Recurrence typically occurs within 7 months and is inevitable irrespective of the extent of resection, historically documented even following hemispherectomy, underscoring that GB is diffusely infiltrative at diagnosis.\u003csup\u003e5-7\u003c/sup\u003e Despite this, previous human GB studies have been limited to primary tumour tissue removed at the time of surgery. As a result, current models of GB are built on what we remove, rather than what remains after surgery. Yet a patient's clinical course, including progression patterns, treatment resistance, and survival, is dictated by this residual disease, which until now has remained beyond the reach of systematic study, obtainable only through rare surgical procedures that preserve the anatomical continuity between tumour and the surrounding brain. This occult infiltrative disease (OID), composed of sparse tumour cells infiltrating normal-appearing brain beyond the resection margin, exists in a microenvironment fundamentally distinct from the densely cellular tumour bulk. Consequently, therapeutic targets identified from resected tissue may not fully apply to the disease driving recurrence.\u003c/p\u003e\n\u003cp\u003eSingle-cell and spatial transcriptomic profiling of resected GB tissue has revealed remarkable heterogeneity, identifying four principal malignant cell states (astrocyte-like, mesenchymal-like, neural progenitor-like, and oligodendrocyte-like),\u003csup\u003e8-10\u003c/sup\u003e that differ in their proliferative capacity, therapy resistance, and immune interactions. These studies have further characterized the immune and stromal microenvironment, and mapped regional variation in cellular composition within the tumour mass. Yet it remains unknown which of these cellular states and microenvironmental features are present beyond the resectable margin, in the tissue where recurrence arises.\u003csup\u003e9,10\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; Here we address this gap by leveraging rare surgical specimens that provide access to the complete GB tissue continuum. En bloc lobectomies, in which the tumour is resected together with the surrounding lobe, provide anatomically intact specimens that preserve spatial relationships spanning tumour core, invasive margin, and distal infiltrated brain. Such specimens are exceptionally rare; to date, only four centres worldwide have published case series of GB lobectomy, collectively reporting fewer than 200 patients at accrual rates of 3–4 cases annually. Our cohort of five IDH-wildtype GB patients, together with non-oncological lobectomy controls, comprises 32 spatially resolved specimens across the tissue continuum. Using multiplexed single-cell RNA sequencing of these spatially matched fresh-frozen biopsies, we profiled 472,089 cells across five anatomically defined spatial zones: two tumour compartments (tumour core and tumour edge) and three OID compartments (adjacent white matter, AWM; distal white matter, dWM; and distal grey matter, dGM), together with corresponding zones from non-oncological controls. We reveal that the OID is distinct from both tumour and normal brain: malignant cells adopt transcriptional programs mirroring the dominant cell type of the region they occupy, the immune landscape transitions from blood-derived myeloid dominance toward brain-resident microglial composition, and non-malignant cells exhibit transcriptional stress with evidence of incomplete compensation. Together, these findings suggest that the biological extent of GB substantially exceeds the radiographically defined tumour margin, and that the tissue in which recurrence arises is already transcriptionally and immunologically distinct from normal brain before any treatment is administered. This spatially resolved characterization provides a framework for therapeutic strategies designed around the biology of residual disease.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eOID compartments are distinct from both tumor and normal brain\u003c/p\u003e\n\u003cp\u003eTo characterize the cellular landscape of GB across the complete tissue continuum, we performed single-cell RNA sequencing on spatially matched biopsies from five IDH-wildtype, treatment-na\u0026iuml;ve GB patients undergoing en bloc temporal lobectomy and one non-oncological lobectomy control (Fig. 1A-C; Supplementary Table 1). For each tract, biopsies were collected along a defined anatomical axis: tumour core, tumour edge, adjacent white matter (AWM), distal white matter (dWM), and distal grey matter (dGM), with the latter two zones located in macroscopically normal-appearing brain parenchyma outside the contrast-enhancing margin. After quality control and doublet removal, 472,089 cells were retained (median 1,532 genes per cell;\u0026nbsp;median 2,375 UMIs per cell). Cells from equivalent anatomical zones co-localized across patients in UMAP space (Fig. 1D), with\u0026nbsp;integration quality assessed by Local Inverse Simpson Index (LISI) scoring: integration LISI improved from 1.48 (pre-Harmony) to 1.64 after batch correction (maximum 2.0), while cluster LISI remained near-optimal at 1.10 (ideal 1.0), confirming that patient-of-origin effects were reduced without obscuring biologically meaningful cell-type boundaries (Supplementary Table 2). Malignant cells were identified by characteristic GB copy-number alterations and reference-based classification (Methods; Supplementary Fig. 1; Supplementary Table 3). Multi-method cell type annotation combining GBmap SCANVI reference mapping, CellTypist classification, CNV inference, and marker panel scoring (Methods) resolved 18 transcriptionally distinct populations spanning malignant GB states, neural populations, and immune infiltrates (Fig. 1E; Supplementary Fig. 2).\u003c/p\u003e\n\u003cp\u003eCell type composition shifted progressively across the spatial axis (Fig. 1F; Supplementary Table 4). Malignant cells comprised 75.2% of cells at the tumour core, declining to 22.1% in AWM, approximately 20% in dWM, and 10% in dGM. Oligodendrocytes showed the inverse pattern, rising from less than 5% at the core to 41.0% in dWM, declining to 22.5% in dGM. Neurons and astrocytes were nearly absent from the tumour core (neurons 0.1%, astrocytes 0.2%), with partial recovery in dGM (neurons 25.1%, astrocytes 12.0%), and reached their highest proportions in GM(Ctrl) (neurons 32.2%, astrocytes 12.6%). Notably, while total immune cell density relative to non-malignant cells returned to near-control levels in the OID (AWM 1.06x, dWM 0.89x, dGM 0.85x of control), the immune compartment was qualitatively remodeled at every zone: monocyte-derived macrophages remained 25- to 33-fold elevated above control in AWM and dWM, while microglia were reciprocally depleted, indicating that tumour-associated immune remodeling extends into the OID compartments even where overall immune density appears normal.\u003c/p\u003e\n\u003cp\u003eBeyond compositional shifts, cells in the OID compartments occupy a transcriptionally intermediate position between tumour and normal brain. Wasserstein distance analysis of zone-level transcriptional profiles confirmed a continuous spatial gradient (Spearman \u0026rho; = 0.94, p = 0.005 across all zones) rather than a binary tumour\u0026ndash;normal boundary (Supplementary Fig. 3). Tumour-patient dWM retained 19.9% of the transcriptional distance from tumour core to WM(Ctrl); dGM retained 7.2% of the distance to GM(Ctrl).\u003c/p\u003e\n\u003cp\u003eInfiltrating malignant cells mirror the transcriptional identity of their host tissue compartment\u003c/p\u003e\n\u003cp\u003eMalignant cell subtype composition varied markedly and non-uniformly across the spatial axis (Fig. 2A). Mesenchymal subtypes were concentrated in the tumour mass: 93.7% of MES1-like cells and 97.2% of MES2-like cells were located at the core or edge, with mesenchymal subtypes together comprising only 7.2% of malignant cells in dGM. By contrast, NPC2-like and OPC-like cells together accounted for 60.1% of all malignant cells at dGM and 21.4% at dWM, comprising the dominant malignant population within the OID compartments. Within the NPC2-like population itself, 61.0% of cells resided in dGM (compared with only 3.5% at the tumour core; Spearman \u0026rho;=0.90, p=0.037); OPC-like cells showed the complementary pattern with 26.3% at AWM and 44.2% at dWM (Kruskal-Wallis H (KW H) = 13.3, p = 0.010). Each subtype was thus enriched in the zone whose dominant normal cell type shares its transcriptional lineage. AC-like cells were present across all zones, consistent with their numerical dominance as the largest subtype rather than specific enrichment in any compartment. This spatial distribution was consistent across all five GB patients and robust to leave-one-out patient exclusion (Supplementary Table 5).\u003c/p\u003e\n\u003cp\u003eBeyond the compositional differences described above, cells of the same malignant subtype show substantial regional transcriptional differences, with the genes most strongly upregulated in each subtype varying systematically across tumour and OID compartments. Within OPC-like cells, oligodendrocyte lineage genes were most highly expressed in white matter zones relative to the tumour core, with MBP and MOBP showing 6-fold and 10-fold increases respectively from the tumour core to dWM. Within NPC2-like cells, neuronal identity genes showed the strongest expression in grey matter, with GRIN1 increasing 74-fold from the tumour core to dGM and RBFOX3 expression absent at the tumour core (mean=0.00) rising to 0.66 in dGM. The small number of MES1-like cells present in distal zones showed progressively lower expression of mesenchymal markers, including CHI3L1 (3.74 at Core, declining to 0.94 in dGM) and VEGFA (3.87 at Core, declining to 0.95 in dGM). AC-like cells showed comparatively modest spatial variation in AC marker expression, with core identity genes CST3 (4.67 to 2.54) and BCAN (2.82 to 1.58) showing gradual decline across zones (Fig. 2B). Notably, concordant patterns were observed when the analysis was repeated using the Neftel et al. (2019)\u003csup\u003e10\u003c/sup\u003e gene sets that originally defined these subtypes, confirming the robustness of these findings to gene set choice (Supplementary Fig. 4). Similarly, malignant subtype cycling fractions within the tumour compartment were closely concordant with those observed when reanalyzing data from Darmanis et al. (2017)\u003csup\u003e8\u003c/sup\u003e using an identical cell cycle scoring pipeline across an independent cohort profiled on a distinct platform (Smart-seq2 vs 10x Flex; AC-like: 35.1% vs 36.0%, n=467 vs n=47,710; MES1-like: 33.5% vs 37.6%, n=403 vs n=42,552; Supplementary Fig. 5, Supplementary Tables 6 and 7), supporting the reproducibility of subtype-specific cell cycle scoring.\u003csup\u003e8,10\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eOf eight published GB transcriptional programs examined here, spatial variation was most prominent in hypoxia (Gavish et al. 2023; MP7)\u003csup\u003e11\u003c/sup\u003e declining distally in MES2-like cells (0.814 at Core to 0.165 in dGM), interferon response peaking at the tumour edge in MES1-like cells (0.263 at Edge, declining to near-zero in OID zones), and both developmental stemness and invasion/tumour microtube score peaking at AWM in AC-like cells (0.423 and 0.351 respectively, declining to 0.137 and 0.174 in dGM; Fig. 2D; Supplementary Fig. 6). HIF1\u0026alpha;-driven paracrine signaling provides a candidate mechanism by which hypoxia-induced stress propagates into non-hypoxic distal compartments, consistent with reports of HIF1\u0026alpha; target genes differentially expressed across the GB core-margin axis.\u003csup\u003e12,13\u003c/sup\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eNPC2-like cells are the dominant malignant cell type in dGM, where they exhibit a markedly reduced proliferative fraction compared with the tumour bulk. Cell cycle phase scoring across all 144,089 malignant cells (Tirosh et al., 2016; Supplementary Fig. 7)\u003csup\u003e14\u003c/sup\u003e revealed that NPC2-like cycling fraction (S+G2M) fell from 55.2% at the tumour core to 19.7% in dGM (Spearman \u0026rho; = -0.90, p = 0.037 across tumour zones; K-W H = 355, p = 1.2 x 10⁻⁷⁵), substantially below the all-malignant mean at every distal zone (Core 40.5%, Edge 35.2%, AWM 29.2%, dWM 31.7%, dGM 27.5%) (Fig. 2C). This quiescent profile was accompanied by progressive enrichment of the neural mimicry transcriptional program. Analysis of all spatially variable genes within NPC2-like cells enriched for chemical synaptic transmission distally (p=3.7x10⁻\u0026sup1;\u0026sup3;) (Supplementary Fig. 8; Supplementary Tables 8 and 9). Consistent with this, genes gained distally in NPC2-like cells showed significant overlap with the top 100 excitatory neuron marker genes in our dataset (87/100 excitatory neuron markers; Fisher\u0026apos;s exact p=1.0x10⁻\u0026sup1;⁴⁶) (Supplementary Fig. 9; Supplementary Table 10). Similarly, the neural mimicry program, defined by neuronal identity genes including NRGN, RBFOX3, NEFL, NEFM, STMN2, and MAPT, showed the highest expression in NPC2-like cells at every distal zone, peaking at dGM (mean score +0.757; Core mean +0.003), while remaining near-absent in mesenchymal subtypes throughout (Fig. 2E).\u003csup\u003e14\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eIn contrast to the quiescent NPC2-like population in grey matter, OPC-like cells in white matter zones maintained high proliferative activity. Cycling fraction for OPC-like cells remained at 58.4% in dWM, 29.7 percentage points above the all-malignant mean of 28.7% in the same zone (OPC-like \u0026gt; non-OPC malignant in all five tumour patients; one-sided paired Wilcoxon p=0.031), and comparable to levels observed at the tumour core (65.7%) and edge (54.0%). OPC lineage score showed a systematic increase in white matter zones, rising from 0.086 at the tumour core to 0.884 at dWM, the highest score of any malignant subtype at any zone (Fig. 2F), while invasion/tumour microtube program score remained near-absent (mean 0.018 at AWM). This contrasts with AC-like cells, which carried the highest invasion/tumour microtube program score at AWM (mean 0.351).\u003csup\u003e15\u003c/sup\u003e Consistent with this, the top spatially variable genes within OPC-like cells in white matter zones included canonical myelinating oligodendrocyte genes MBP (6-fold increase from Core to dWM), MOBP (10-fold) and MYRF (7-fold). In keeping, genes gained distally in OPC-like cells showed significant overlap with the top 100 oligodendrocyte marker genes in our dataset (92/100; Fisher\u0026apos;s exact p = 2.0x10⁻\u0026sup1;⁹\u0026sup3;) (Supplementary Fig. 9; Supplementary Table 10).\u003csup\u003e15\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eOPC-like cells in white matter and NPC2-like cells in grey matter thus present distinct transcriptional profiles: proliferative and oligodendrocyte-like in one compartment, quiescent and neuron-like in the other.\u003c/p\u003e\n\u003cp\u003eThe immune microenvironment of GB demonstrates spatial variation\u003c/p\u003e\n\u003cp\u003eImmune populations were identified across all spatial zones, including monocyte-derived macrophages (MDMs), microglia, T cells, neutrophils and granulocytes (Fig. 3A). While all major immune populations were detectable throughout the tissue axis, their composition and density underwent striking spatial remodeling (Fig. 3A; Supplementary Fig. 10).\u003c/p\u003e\n\u003cp\u003eMDMs were the dominant immune population at the tumour core and edge, reaching 109.9 and 176.5 cells per 1,000 total cells respectively. MDM density declined progressively across the OID but remained substantially elevated relative to control at every zone: \u003cstrong\u003e71.8\u0026nbsp;cells per 1,000 in AWM and 56.1 in dWM (33- and 26-fold above WM(Ctrl))\u003c/strong\u003e, and\u0026nbsp;10.2 in dGM (3.7-fold above GM(Ctrl)). Microglia showed the inverse pattern: virtually absent from the tumour core and edge (\u0026lt;1 cell per 1,000 total cells, compared with 255.6 in white matter control), with partial recovery across the OID, reaching\u0026nbsp;57% of WM(Ctrl) levels in AWM, 49% in dWM, and 57% in dGM where it is the dominant immune cell type. The microglia:MDM ratio reflected this spatial inversion, with a ratio of 9:1 in dGM (compared with 58:1 in control grey matter), 2.2:1 in dWM, 2.0:1 in AWM, \u0026lt;0.1:1 at the tumour edge and \u0026lt;0.1:1 at the tumour core. These data reveal that MDM-driven myeloid remodeling extends substantially beyond the tumour mass, with the OID compartments representing a transitional zone between tumour-associated MDM dominance and the microglial-dominant composition of normal brain.\u003c/p\u003e\n\u003cp\u003eT cell abundance demonstrated spatial variance. T cells at the tumour core comprised 23.7% of immune cells (1.6-fold above white matter control), then decreased to levels comparable to white matter control at the tumour edge (16.3% vs 14.9%) and\u0026nbsp;14.2%, 18.4%, and 32.3% in AWM, dWM, and dGM respectively (Supplementary Fig. 10; Supplementary Table 4). T cell composition exhibited a marked shift across spatial zones (Fig. 3B). CD8+ cytotoxic T cells comprised\u0026nbsp;14.8% of T cells in the tumour core, rising to 17.5% at the tumour edge before declining to\u0026nbsp;11.6% and 9.9% in AWM and dWM respectively, and only 4.4% in dGM. Regulatory T cells showed the inverse pattern, rising from 9.7% at the tumour edge to\u0026nbsp;18.2% and 9.3% in AWM and dWM respectively, and up to 39.8% in dGM, approaching the fraction seen in grey matter control (40.7%). The CD8:Treg ratio thus inverted across the tissue axis, from\u0026nbsp;1.9:1 at the tumour edge to\u0026nbsp;0.1:1 in dGM (0.2:1 in GM(Ctrl)). These data reveal that the OID compartments harbor a T cell population that is not merely depleted but compositionally shifted toward regulatory dominance. Expressed as a proportion of CD45+ immune cells (0.9\u0026ndash;1.1% across all zones, KW p=0.71) or as absolute density per 1,000 total cells (1.3\u0026ndash;2.9, KW p=0.47), Treg abundance was not significantly different between zones or between tumour-patient dGM and GM(Ctrl). The apparent Treg enrichment in dGM as a fraction of the T cell pool thus reflects local depletion of conventional CD4+ and CD8+ T cells rather than expansion of the Treg compartment itself.\u003c/p\u003e\n\u003cp\u003eAmong non-conventional lymphocyte populations, NK cells and double-negative T cells (DNTs) were detected at low but consistent frequencies across all spatial zones, comprising 21.8% and 22.7% of T-lineage cells in dWM and dGM respectively (Fig. 3B). NK cells and DNTs, when cytotoxic, engage targets through MHC-unrestricted mechanisms, though functional activity cannot be inferred from transcriptional profiling alone.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eT cell checkpoint receptor expression followed a consistent spatial gradient (Fig. 3C). The three principal checkpoint receptor targets, PDCD1 (PD-1), HAVCR2 (TIM-3), and TIGIT, all showed significant spatial variation (K-W H = 701, 392, and 865 respectively; all p \u0026lt; 10⁻⁸\u0026sup1;), peaking at the tumour edge (7.4%, 7.9%, and 9.2% of T cells respectively) and declining progressively across the OID: 2.9%, 4.4%, and 4.7% in AWM; 2.9%, 3.6%, and 5.7% in dWM; and 0.4%, 1.8%, and 0.5% in dGM, approaching levels observed in GM(Ctrl) (0.3%, 1.4%, 0.3%). Co-expression of two or more checkpoint receptors showed a similar spatial pattern: 4.1% of T cells at the tumour edge expressed multiple checkpoint markers, declining to 1.5% in AWM, 1.8% in dWM, and 0.1% in dGM (Supplementary Table 11). PD-L1 (CD274), the principal ligand for PD-1, was correspondingly elevated on malignant cells at the tumour edge (14.3% positive) and core (7.5%), declining sharply to 1.9% in dWM and 1.2% in dGM (KW H = 3,524, p \u0026asymp; 0), establishing spatial co-localization of checkpoint ligand and receptor expression at the tumour margin. These data reveal that the molecular signatures of T cell exhaustion and immune evasion are concentrated at the tumour margin and progressively diminish across the OID, a spatial gradient observed consistently across all five GB patients (Supplementary Table 11).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eNon-malignant cells exhibit spatially graded transcriptional stress\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNeurons were present at very low frequency in the tumour core and edge (n=28 and n=77 total respectively) precluding reliable zone-level analysis, but were abundant in AWM (n=1,949), dWM (n=3,077), dGM (n=19,455) and GM(Ctrl) (n=18,844), providing adequate statistical power for spatial comparisons across AWM through GM(Ctrl) (Fig. 4A). In macroscopically normal-appearing dGM, 39.6% of neurons scored above zero for the stress program,\u003csup\u003e11\u003c/sup\u003e compared with 6.7% of GM(Ctrl). The same pattern held in AWM (81.8%) and dWM (50.7%), and in both excitatory (40.2% in dGM) and inhibitory subtypes (38.3%; Supplementary Table 12), consistent across all leave-one-out patient-exclusion scenarios. The stress program was significantly elevated in OID compartments relative to zone-matched non-oncological control (KW H=12,817; AWM FDR \u0026asymp; 0, dWM FDR = 2.6x10⁻\u0026sup1;⁶\u0026sup3;, dGM FDR \u0026asymp; 0; Supplementary Fig. 11; Supplementary Table 12), with mean scores positive across all OID compartments (dGM mean +0.023) and negative in GM(Ctrl) (mean -0.107) (Supplementary Table 5; Supplementary Table 12). Stress scores peaked at AWM (mean +0.286) and declined through dWM (mean +0.103) and dGM (mean +0.107) (Fig. 4B), yet remained above zero in every GB patient zone. In contrast, synaptic program expression in dGM neurons was comparable to GM(Ctrl) (99.6% vs 96.5% respectively; Supplementary Fig. 11). The GB-associated field effect thus selectively activates a stress and injury transcriptional response while leaving core synaptic identity intact. Beyond the six scored programs, agnostic spatial variable gene analysis identified 3,776 genes with significant expression variation across spatial zones in neurons (Kruskal-Wallis H, FDR \u0026lt; 0.05, effect size \u0026gt; 0.3), of which 1,306 were elevated in tumour-patient zones and 881 depleted relative to non-oncological control, confirming that the neuronal field effect encompasses broad transcriptional reprogramming beyond the stress response alone (Supplementary Fig. 12). Additionally, established diffusible inducers of cell stress including VEGF, HIF1\u0026alpha;, IL-6, IL-8 (CXCL8), CCL-2, and TGF-\u0026beta;1 are upregulated in the tumour mass relative to non-oncological control, with residual elevation persisting through dWM and dGM, consistent with previous reports of these factors as components of the established GB secretome.\u003csup\u003e13,16-19\u003c/sup\u003e TNF-\u0026alpha; showed the inverse spatial pattern, preferentially expressed by macrophages and microglia in distal compartments rather than at the tumour mass.\u003csup\u003e20\u003c/sup\u003e Each of these inducers engages stress signaling via the kinase arms profiled above.\u003csup\u003e12,21,22\u003c/sup\u003e\u003c/em\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe neuronal stress response was not driven by a single upstream pathway. Scoring neurons across five integrated stress response (ISR) module programs revealed co-activation of the PERK/UPR arm (BiP/HSPA5 elevated in 18.9% of stressed neurons), the HRI/mitochondrial UPR arm (ATF5, HMOX1), and an excitotoxicity/immediate-early gene signature (NPAS4, NR4A1) that was the dominant signature across all cell types (Supplementary Fig. 13). The PKR/interferon arm was not differentially activated in MP6-classified stressed neurons (\u0026Delta; \u0026asymp; 0; Mann-Whitney p = n.s.), arguing against type-I interferon signalling as a major driver. Although the strongest ISR-arm activation was observed in neurons, scoring the same kinase-arm modules across all cell types showed detectable activation across all four profiled cell types (neurons, OPCs, oligodendrocytes, astrocytes; Supplementary Figure 13C), supporting a shared upstream paracrine driver rather than a neuron-restricted mechanism. The CHOP:GADD34 ratio in stressed neurons indicated an active resolution phase rather than chronic terminal stress. The skewed ratio toward GADD34 over CHOP, together with the predominance of immediate-early and excitotoxicity transcripts over committed pro-apoptotic effectors (e.g., BAX, BAK1, BIM all near baseline), is consistent with a pro-survival ISR signature rather than a commitment to apoptosis. Ligand-receptor analysis identified tumour-associated macrophages (MDM) and AC-like tumour cells as the principal candidate sources of upstream ISR-activating signals, expressing TGFB1, CXCL12, IL1B, and CSF1 at highest levels in Core and Edge zones with expression declining across the spatial axis (Supplementary Fig. 14).\u003c/p\u003e\n\u003cp\u003eTo identify a discrete stressed neuron population, neurons were classified as stressed if their MP6 stress score exceeded a zone-matched threshold (non-oncological control neuron mean + 2 SD, computed separately for white matter and grey matter zones to account for baseline transcriptional differences; Methods). This identified 8,401 stressed neurons (17.8% of all 47,262 neurons). Comparing stressed neurons to unstressed neurons by pseudobulk differential expression (DESeq2; n=5 tumour patients) identified 169 genes significantly upregulated and 24 downregulated in the stressed population (adjusted p \u0026lt; 0.05, |log2 fold-change| \u0026gt; 0.5), with pathway enrichment confirming activation of TNF\u0026alpha; Signaling via NF-\u0026kappa;B (FDR = 1.6x10⁻⁷\u0026sup1;), Hypoxia (FDR = 6.8x10⁻\u0026sup1;⁶), p53 Pathway (FDR = 2.4x10⁻⁸), and Apoptosis (FDR = 1.4x10⁻⁷) programs (Supplementary Table 13; Supplementary Fig. 15). The proportion of stressed neurons was elevated in AWM relative to WM(Ctrl) (39.6% vs 4.5%), in dGM relative to GM(Ctrl) (32.3% vs 3.9%), and in dWM relative to WM(Ctrl) (14.1% vs 4.5%) (Supplementary Fig. 16), suggesting that tumour-associated neuronal stress extends throughout macroscopically normal brain parenchyma in GB patients.\u003c/p\u003e\n\u003cp\u003eAstrocytes demonstrated a progressive reactive gradient that tracked distance from the tumour mass (Fig. 4C). At the tumour edge, 96.7% of astrocytes occupied a reactive state (Reactive A1, A2, or DAA), declining to 88.5% in AWM, 70.5% in dWM, and 61.5% in dGM (Spearman \u0026rho; = -0.900, p = 0.037 across tumour zones), approaching but remaining above the 56.9% observed in non-oncological GM(Ctrl) (Core 96.2%; n=80, interpreted with caution given limited recovery). This gradient represents the steepest and most spatially coherent cellular transition in the dataset, robust to leave-one-out exclusion with a minimum AWM-to-dGM difference of 24.0 percentage points (Supplementary Table 5). The A1 neurotoxic state, associated with complement pathway activation and neuronal injury,\u003csup\u003e23\u003c/sup\u003e was the dominant reactive state in AWM (42.8% of astrocytes) and remained elevated above control in both dWM (27.9%) and dGM (13.8% vs 12.0% in GM(Ctrl)). The disease-associated astrocyte (DAA) state, characterized by upregulation of SERPINA3 (logFC 7.63) and SPHK1 (logFC 6.38) relative to homeostatic astrocytes within the dataset, was detectable across all GB zones (26.2% in dGM vs 23.2% in GM(Ctrl)). Astrocytic reactivity, including neurotoxic and disease-associated states, is therefore a region-wide property of the GB-affected brain that persists into macroscopically normal tissue. Consistent with this, agnostic spatial variable gene analysis identified 2,615 genes with significant spatial variation in astrocytes, the second-highest of any non-malignant cell type, of which 1,659 were tumour-elevated, enriching for regulation of apoptotic processes (FDR = 4.4 x 10^-6) and transcriptional repression (FDR = 6.7 x 10^-7), indicating that reactive astrocytes in the OID are broadly transcriptionally reprogrammed beyond the scored reactive state markers (Supplementary Fig. 12).\u003c/p\u003e\n\u003cp\u003eOligolineage cells demonstrated widespread stress (Fig. 4D; Supplementary Fig. 17). Stress program scores were elevated in oligodendrocytes at the Core (+0.262) and remained significantly elevated at AWM (mean +0.083; MWU FDR = 3.2x10⁻\u0026sup2;⁴⁷) and dWM (mean +0.031; MWU FDR = 4.6x10⁻⁴⁵), but did not extend robustly to dGM (mean -0.004; KW H = 10,741 across all zones, p \u0026asymp; 0). OPCs showed a spatially broader stress response, elevated across tumour zones (Core +0.235, AWM +0.117, dWM +0.165, dGM +0.051; KW H = 6,170, p \u0026asymp; 0) (Fig. 4E), well above WM(Ctrl) in white matter zones and above GM(Ctrl) in dGM; notably, OPC stress scores in dWM (+0.165) exceeded those of mature oligodendrocytes in the same zone (+0.031) (MWU p approximately 0), suggesting the progenitor compartment is more affected than the mature population. Beyond the stress program, three additional findings were consistent across all leave-one-out patient-exclusion scenarios (Supplementary Table 5): myelin and maturity marker scores (MBP, PLP1, MOG) were elevated rather than suppressed in GB patient zones relative to control (AWM difference +0.28; dGM difference +0.16; LOO-robust at all zones); remyelination signaling was elevated vs matched control in white matter zones (AWM +0.08, dWM +0.05; LOO-robust) but reduced specifically at dGM (-0.03; LOO-robust); and GB patient oligodendrocytes retained higher progenitor marker scores than non-oncological control oligodendrocytes in all five zones (AWM difference +0.10, dGM difference +0.11; LOO-robust at all zones). An inflammatory response programme was modestly elevated in tumour-proximal zones in both oligodendrocytes (Core +0.09, Edge +0.14) and OPCs (Core +0.08, Edge +0.08), recovering toward non-oncological control levels in distal compartments and consistent with the spatially restricted myeloid signal at the tumour mass (Fig. 4D, 4E). Agnostic spatial variable gene analysis corroborated these findings, identifying 1,303 spatially variable genes in oligodendrocytes and 2,133 in OPCs. Among OPC tumour-elevated genes, pathway enrichment identified transcriptional regulation (FDR = 5.3 x 10^-7), while tumour-depleted OPC genes enriched for synapse organization and nervous system development (FDR = 1.5 x 10^-5), suggesting that OPCs in tumour-proximal zones lose normal supportive functions alongside their stress activation. Oligodendrocyte tumour-elevated genes enriched for sterol biosynthesis (FDR = 2.3 x 10^-4), consistent with the paradoxical upregulation of myelin maturity markers (Supplementary Fig. 12).\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eSingle-cell atlases of resected glioblastoma tissue have defined the malignant cellular states, immune composition, and transcriptional heterogeneity of the tumour mass in considerable detail.\u003csup\u003e8-10,24,25\u003c/sup\u003e Yet recurrence is driven not by the tissue removed at surgery but by the occult infiltrative disease (OID) that persists beyond the resection margin. Here we reveal that this OID is biologically distinct from both tumour bulk and normal brain. Infiltrating GB cells adopt transcriptional programs that mirror the dominant cell type of the region they occupy: OPC-like cells in white matter, NPC2-like cells in grey matter. This transcriptional resemblance is supported at the gene level by significant overlap between spatially gained genes and independently-derived normal cell type markers (OPC-like: 92/100 oligodendrocyte markers, p=2.0x10⁻¹⁹³; NPC2-like: 87/100 excitatory neuron markers, p=1.0x10⁻¹⁴⁶). Concurrently, the surrounding parenchyma acquires attributes of the tumour microenvironment, with myeloid remodeling, neuronal stress, and astrocytic reactivity extending well beyond the contrast-enhancing margin, spatial patterns observed consistently across all five profiled patients and robust to leave-one-out exclusion. The chimeric tissue identity of the OID, with malignant cells assuming host phenotypes while host tissue assumes tumour-associated features, has direct therapeutic implications which we address in turn.\u003c/p\u003e\n\u003cp\u003eNPC2-like cells are enriched in distal grey matter, where they adopt a progressively quiescent state, with cycling fraction declining monotonically from\u0026nbsp;tumour core to dGM (Fig. 2C). This spatial gradient of quiescence has important therapeutic implications. Slow-cycling glioma cells survive temozolomide and radiation, persist through treatment, and can regenerate the tumour mass\u003csup\u003e26,27\u003c/sup\u003e; a non-cycling population residing outside the standard radiation field therefore represents a reservoir from which recurrence would be expected to arise. The therapeutic challenge is compounded by the neural mimicry program carried by NPC2-like cells,\u0026nbsp;a transcriptional resemblance to neurons so deep that 87% of excitatory neuron markers are gained distally (Fisher's exact p=1.0x10⁻¹⁴⁶). This expression pattern increases with distance from the\u0026nbsp;tumour, peaking in dGM where NPC2-like cells and neurons co-occupy the same spatial compartment and where Treg frequency is highest. Central self-tolerance is established in part through AIRE-mediated thymic expression of tissue-restricted antigens, including neuronal proteins, leading to deletion of T cells reactive against these self-antigens.\u003csup\u003e28,29\u003c/sup\u003e NPC2-like cells expressing these neuronal identity genes in distal grey matter may therefore escape T cell-mediated recognition through the same mechanism that protects normal neurons from autoimmune attack, rendering them resistant to both neoantigen-based and T cell-redirecting immunotherapies while\u0026nbsp;their quiescent state also renders them resistant to standard cytotoxic therapy.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;The white matter OID compartments are characterized by enrichment of OPC-like cells which, unlike the quiescent NPC2-like population in grey matter, maintain a high proliferative index rendering them, in principle, both radiosensitive and chemosensitive. In the white matter, then, the therapeutic challenge is not\u0026nbsp;tumour cell biology but drug delivery and radiation targeting: these zones lie outside the standard radiation field and while temozolomide crosses the blood-brain barrier it achieves only approximately 20% of plasma concentrations in normal brain interstitium,\u003csup\u003e30,31\u003c/sup\u003e substantially below the levels reached in the\u0026nbsp;tumour mass where BBB integrity is compromised.\u003csup\u003e32\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eTogether, the two-phenotype model predicts that effective treatment of the OID will require combination approaches: anti-proliferative strategies directed at the cycling OPC-like population, and non-cell-cycle-dependent approaches directed at quiescent NPC2-like cells. For the latter, differentiation therapy represents a rational strategy, with BMP4 and retinoic acid both shown to induce terminal differentiation of glioma stem cells, reduce self-renewal capacity, and sensitize cells to cytotoxic therapy.\u003csup\u003e33-35\u003c/sup\u003e Alternatively, strategies that force quiescent cells into cycle, such as checkpoint kinase inhibition or metabolic modulation, could render them susceptible to radiation and temozolomide.\u003csup\u003e36-38\u003c/sup\u003e CHK1 inhibition represents a well-characterized approach in this category: glioma stem cells survive radiation through preferential activation of the CHK1-mediated DNA damage checkpoint, and pharmacological CHK1 inhibition reverses this radioresistance by abrogating the protective G2/M arrest; CHK1 inhibition similarly potentiates temozolomide cytotoxicity 5-fold by preventing the G2 arrest that otherwise allows DNA damage repair.\u003csup\u003e27,39\u003c/sup\u003e Metabolic modulation offers a complementary strategy: slow-cycling glioma cells depend on oxidative phosphorylation, and metformin-induced AMPK activation has been shown to promote differentiation of stem-like cells into non-tumorigenic populations.\u003csup\u003e40,41\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eTumour-associated macrophages and microglia are the dominant immune population in GB, and recent single-cell studies have begun to distinguish their distinct functional states and spatial distributions within the tumour.42-45 Our data extend these observations across the complete tissue axis, revealing a striking spatial inversion of the myeloid compartment: at the tumour core MDMs outnumber microglia by more than 100-fold, while in dGM the composition inverts to 90% microglia and 10% MDM (microglia:MDM ratio 9:1). T cell composition shifts in parallel, with CD8+ cytotoxic T cells declining from 17.5% at the tumour edge to 4.4% in dGM while regulatory T cells rise from 9.0% to 35.4%. Tumour-infiltrating Tregs maintained CTLA-4, TIGIT, and LAG-3 expression across core, edge, and AWM zones, consistent with the activated Treg phenotype described in human cancer, while PD-1 remained minimal on this population and was instead carried predominantly by CD8+ T cells. The principal checkpoint receptor targets (PD-1, TIM-3, and TIGIT) peak at the tumour edge and decline to levels approaching normal brain in dGM (0.4-1.8%), a spatial gradient observed consistently across all five patients. These findings have direct implications for immunotherapy in GB, which we address in turn.\u003csup\u003e42-45\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;The three checkpoint receptors described above are the principal targets of checkpoint inhibitor trials in glioblastoma, selected on the basis of expression profiling in resected tumour. To date, however, every phase III checkpoint inhibitor trial in glioblastoma has failed to demonstrate a survival benefit, a finding confirmed by a recent meta-analysis of 12 randomized controlled trials.\u003csup\u003e46-49\u003c/sup\u003e While the failure of checkpoint blockade in glioblastoma is undoubtedly multifactorial, our data provide some spatial context. Checkpoint therapies are designed around the biology of the tumour mass and its immediate margin, where our data confirm high checkpoint expression, exhausted T cells, and MDM-dominated immunosuppression, providing a rational biological substrate for blockade. Consistent with this, Nivolumab (anti-PD-1) reaches saturating levels on intratumoural T cells in glioblastoma and induces local T cell activation and proliferation, and neoadjuvant PD-1 blockade promotes intratumoural T cell clonal expansion and cDC1 activation.\u003csup\u003e50-53\u003c/sup\u003e In the OID, however, PD-1, TIM-3, and TIGIT are near-absent (0.4-1.8% in dGM, approaching normal brain levels), and PD-L1 on malignant cells declines from 14.3% at the tumour edge to 1.2% in dGM, indicating that the molecular substrate for local checkpoint reinvigoration is largely confined to the tumour mass and its immediate margin.\u003csup\u003e46-53\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eImportantly, the principal mechanism of PD-1 blockade is thought to operate not locally but through progenitor exhausted T cells residing in tumour-draining lymph nodes, which proliferate, differentiate, and traffic to the tumour after systemic blockade.\u003csup\u003e54-59\u003c/sup\u003e This pathway remains intact irrespective of local checkpoint expression in the OID. Yet even if checkpoint blockade successfully generates tumour-reactive effector T cells, those cells must traffic to and function within the OID to eliminate residual disease, and our data suggest that this effector phase is substantially limited. Antigen-reactive T cells are concentrated at the tumour edge rather than in the OID compartments, consistent with 5-ALA-guided spatial profiling of glioblastoma showing preferential enrichment of PD-1⁺CD103⁺ tissue-resident T cells in the intermediate and marginal tumour layers rather than in distal tissue.\u003csup\u003e60,61\u003c/sup\u003e The OID thus presents compounding effector-phase barriers: Treg dominance (CD8:Treg ratio of 0.1:1 in dGM), persistent MDM-mediated immunosuppression (25- to 33-fold above control in AWM and dWM), and the neural mimicry program of NPC2-like cells, which may exploit central tolerance mechanisms to evade MHC-restricted recognition.\u003csup\u003e15,62-64\u003c/sup\u003e As such, patients with subtotal resections, in whom checkpoint-high, MDM-dominated tumour bulk remains in situ, may represent a more biologically appropriate population for checkpoint inhibitor trials.\u003c/p\u003e\n\u003cp\u003eThese effector-phase barriers are not unique to checkpoint blockade. Neoantigen vaccines, bispecific T cell engagers, and other T cell-redirecting strategies face the same compounding limitations in the OID, as each ultimately depends on endogenous T cells recognizing and eliminating tumour cells. MHC-unrestricted approaches such as CAR-T platforms targeting glioblastoma-specific surface antigens circumvent the antigen visibility problem, although antigen loss and intratumoural heterogeneity remain well-documented challenges.\u003csup\u003e65-69\u003c/sup\u003e Regardless of modality, the pronounced Treg dominance in the OID suggests that any immunotherapy strategy directed at residual disease may benefit from concurrent efforts to shift the CD8:Treg ratio toward cytotoxic predominance.\u003csup\u003e63\u003c/sup\u003e Beyond the malignant and immune compartments, non-malignant cells are transcriptionally altered in spatially structured patterns extending from the\u0026nbsp;tumour margin throughout the OID. Neurons, astrocytes, and oligodendrocytes each exhibit transcriptional changes in tissue that is macroscopically and radiographically normal-appearing, constituting a region-wide field effect that was consistent across all five patients and robust to leave-one-out exclusion in each cell type examined.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Neurons exhibit a stress/injury transcriptional response that is elevated across all OID compartments relative to non-oncological controls, peaking at AWM and persisting into macroscopically normal dGM. This extends prior immunohistochemical observations of stress-response protein upregulation in tumour-infiltrated cortex to well outside regions of direct tumour contact.70 Critically, synaptic program expression in dGM neurons is preserved relative to controls, suggesting that core synaptic identity remains intact despite widespread stress activation. This dissociation between stress activation and preserved synaptic identity is consistent with a sublethal injury state in which neurons are functionally compromised but not yet irreversibly damaged, and defines a potential therapeutic window. The character of this response implicates diffusible signals from the tumour and the microenvironment rather than direct hypoxic injury, with immediate-early transcription factors and molecular chaperones dominating over canonical HIF1A-driven hypoxia genes. The pattern of ISR kinase arm co-activation is consistent with specific tumour-derived diffusible signals known to be elevated in the GB microenvironment: VEGF and HIF1α engaging the PERK/UPR arm, IL-6 and IL-8 activating NF-κB signaling, and TGF-β1 contributing to GCN2-arm metabolic stress.\u003csup\u003e70,71\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Astrocytes exhibit a progressive reactive gradient extending throughout the OID, from 88.5% in the AWM to 61.5% in the dGM, where the total reactive fraction remains above non-oncological controls. The disease-associated astrocyte (DAA) and A1 neurotoxic states are the dominant reactive subtypes across OID compartments, characterized respectively by upregulation of SERPINA3 and complement pathway activation. These findings are consistent with emerging evidence that tumour-associated reactive astrocytes are not passive bystanders but active participants in GB pathogenesis: astrocyte-derived CCL2 and CSF1 govern tumour-associated macrophage recruitment, astrocyte-derived cholesterol supports glioma cell survival, and a distinct astrocyte subset limits tumour immunity by inducing T cell apoptosis through TRAIL driven by tumour-derived IL-11 signaling through STAT3.\u003csup\u003e72,73\u003c/sup\u003e DAA and A1 astrocytes in the OID therefore represent co-existing mechanisms of neuronal injury and immunosuppression that persist independently of direct malignant cell contact. These reactive glia, along with microglia and macrophages, are also primary sources of reactive oxygen species: the ROS response program is upregulated in a gradient-dependent manner, peaking at AWM in microglia and persisting through the OID in astrocytes, providing a spatially graded cell-extrinsic inducer of stress in surrounding tissue.\u003csup\u003e74,75\u003c/sup\u003e Together, these glial changes contribute to the region-wide transcriptional stress response compounding the neuronal injury described above.\u003csup\u003e72-75\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eOligodendrocyte lineage cells exhibit a spatially restricted stress response concentrated in white matter zones, but the more striking finding is a pattern of active but failing compensation. More revealing is the combination of paradoxically upregulated myelin maturity markers alongside reduced remyelination signaling and elevated progenitor marker scores, a pattern consistent with active but failing compensation: mature oligodendrocytes are attempting to maintain existing myelin at the transcriptional level while OPCs fail to differentiate, impairing the capacity for white matter repair.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTogether, these findings describe a brain under sustained transcriptional stress across multiple cell types and spatial zones, in tissue that appears normal by conventional macroscopic and radiographic assessment. Current neuroprotective strategies in GB have focused predominantly on mitigating treatment-induced toxicity, including radioprotectant agents, hippocampal-sparing radiation techniques, and anti-glutamatergic combinations, and were not designed around the pre-treatment molecular landscape of the affected parenchyma.\u003csup\u003e76\u003c/sup\u003e Our data reveal that neuronal stress is already established before any treatment is administered, is driven by multiple parallel mechanisms including co-activation of distinct ISR kinase arms (PERK/UPR, HRI/mitochondrial), A1 astrocytic neurotoxicity, and oligolineage compensation failure, and extends into tissue outside any treatment field.\u0026nbsp;Glutamate-mediated excitotoxicity through system is well-established as one contributor to peritumoral neuronal injury,\u003csup\u003e77-79\u003c/sup\u003e but the co-activation of multiple ISR kinase arms identified here suggests this is one of several parallel mechanisms, and that single-agent neuroprotective strategies targeting individual pathways may be insufficient. That affected neurons retain synaptic identity despite widespread stress activation suggests a therapeutic window may exist. Neuroprotective intervention informed by the specific injury mechanisms identified here could preserve cognitive and functional outcomes in a patient population for whom quality of life during the highest-functioning period after diagnosis remains an unmet clinical priority.\u003c/p\u003e\n\u003cp\u003eThese\u0026nbsp;findings provide a spatially informed blueprint for combination therapies designed around the biology of each compartment: anti-proliferative strategies for cycling OPC-like cells in white matter, non-cell-cycle-dependent approaches including MHC-unrestricted immunotherapies for quiescent NPC2-like cells in grey matter and neuroprotective intervention directed at the specific injury mechanisms identified in the surrounding parenchyma. Additionally, therapies should reflect the disease present at the time of treatment. Gross total resection leaves predominantly OID with quiescent invaders and low checkpoint target expression, while subtotal resection retains mesenchymal-dominant, checkpoint-high\u0026nbsp;tumour bulk. Current protocols often apply uniform adjuvant treatment approaches regardless of what remains, and clinical trials that select for patients with gross total resections may paradoxically select for the population whose residual disease is least amenable to the therapies under investigation. Stratification by extent of resection may be necessary to match therapeutic strategy to the biology actually present. Together these findings recontextualize the cellular states, immune biology, and therapeutic vulnerabilities previously characterized in resected\u0026nbsp;tumour, providing a framework for GB therapeutics designed not in response to the tumour removed, but in anticipation of recurrence driven by the disease that remains.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eStudy design and patient cohort\u003c/p\u003e\n\u003cp\u003eTissue samples were obtained from patients who underwent en bloc temporal lobectomy at [Institution] and consented to tissue donation (REB [withheld for blinding]). Five patients with histologically confirmed glioblastoma (IDH-wildtype, WHO Grade 4; Supplementary Table 1) and one non-oncological control patient (BT077), who underwent temporal lobectomy for medically refractory epilepsy, were enrolled under the same protocol. Written informed consent was obtained from all participants. For each GB patient, five spatially matched biopsies were collected along a defined anatomical axis: tumour core, tumour edge, adjacent white matter (AWM; peri-tumoural white matter immediately bordering the tumour mass, resected en bloc with the specimen), distal white matter (dWM), and distal grey matter (dGM). dWM and dGM were sampled from macroscopically normal-appearing brain parenchyma outside the contrast-enhancing region on pre-operative MRI. One patient (BT074) contributed two independent biopsy series collected from spatially distinct regions of the same tumour (BT074-A and BT074-C), providing a deliberate within-patient assessment of sampling reproducibility across the spatial axis. The non-oncological control patient provided biopsies from white matter and grey matter locations. All tissue was snap-frozen immediately following resection and stored at -80\u0026deg;C.\u003c/p\u003e\n\u003cp\u003eSingle-cell RNA sequencing\u003c/p\u003e\n\u003cp\u003eSingle-cell RNA sequencing was performed using the 10x Genomics Chromium Flex probe-based library preparation kit, which enables high-quality capture from fresh-frozen tissue and multiplexing of up to 16 samples per sequencing run. All 32 samples were distributed across two 16-plex pools (Pool_152 and Pool_153), reducing batch effects introduced by sequential sample processing. Libraries were sequenced on an Illumina NovaSeq X Plus to a target depth of 30,000 reads per cell. Raw sequencing data were processed using Cell Ranger 10.0.0 with the human GRCh38 reference genome and the 10x Flex probe set (v1.0, 18,108 probes). Sample demultiplexing was performed using the Flex multiplexing algorithm within Cell Ranger.\u003c/p\u003e\n\u003cp\u003eQuality control and doublet removal\u003c/p\u003e\n\u003cp\u003ePer-sample quality control was performed using Scanpy 1.12.80 Cells were filtered to retain those with 205\u0026ndash;13,700 genes/cell detected and fewer than 25% mitochondrial probe counts. Mitochondrial content was low throughout (cohort median 0.17%), consistent with the probe-based capture chemistry of the Flex kit, which includes only 12 mitochondrial probe targets. Doublets were identified using Scrublet 0.2.3 with per-sample threshold optimization, followed by manual review of doublet score distributions.81 Cells flagged as doublets by both automatic classification and manual review were excluded. 4 of the 32 samples with low cell yield were retained after confirming that their inclusion did not alter the direction or magnitude of central findings in leave-one-out sensitivity analysis (Supplementary Table 5). After quality control, 472,089 cells were retained.\u003csup\u003e80,81\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eNormalization, dimensionality reduction, and batch correction\u003c/p\u003e\n\u003cp\u003eCount matrices were normalized to a target sum of 10,000 counts per cell and log-transformed (log1p) using Scanpy. Highly variable genes were identified using the Seurat flavour dispersion method (n=3,000 HVGs), restricted to genes that were highly variable in at least one sample (batch_key=sample_id; 32 samples). Principal component analysis was performed using randomized SVD (50 components). Batch correction was performed using Harmony 0.2.0,\u003csup\u003e82\u003c/sup\u003e with pool identity (Pool_152, Pool_153; 2 pools) as the batch covariate. Pool identity rather than patient identity was used as the Harmony batch covariate to preserve inter-patient biological variation, which was subsequently assessed through leave-one-out sensitivity analysis. Integration quality was assessed using the Local Inverse Simpson\u0026apos;s Index (LISI): integration LISI improved from 1.48 to 1.64 post-correction, while cluster LISI remained at a median of 1.0, indicating effective batch mixing without loss of biological signal. UMAP embedding was computed on the Harmony-corrected neighbour graph (n_neighbors=15, min_dist=0.3) using UMAP 0.5.11.\u003csup\u003e83\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eCell clustering\u003c/p\u003e\n\u003cp\u003eLeiden community detection was applied to the Harmony-corrected neighbour graph at four resolutions (r=0.2, 0.5, 1.0, 2.0), yielding 14, 22, 41, and 73 clusters respectively. The r=1.0 resolution (41 clusters) was used as the primary clustering for cell type annotation. The r=0.5 resolution was used for CellTypist label propagation and the r=2.0 resolution was used for differential expression analyses requiring finer cluster granularity.\u003c/p\u003e\n\u003cp\u003eCell type annotation\u003c/p\u003e\n\u003cp\u003eA multi-method annotation strategy was used, leveraging each tool for the cell populations where it performs most reliably.\u003c/p\u003e\n\u003cp\u003eMalignant cell states.\u003c/p\u003e\n\u003cp\u003eMalignant GB cells were annotated using the GBmap SCANVI reference model (Ruiz-Moreno et al., 2025), trained on a curated glioblastoma single-cell reference atlas.\u003csup\u003e24\u003c/sup\u003e GBmap assigns cells to six malignant states based on the Neftel et al. (2019) classification: AC-like, MES1-like, MES2-like, NPC1-like, NPC2-like, and OPC-like.\u003csup\u003e10\u003c/sup\u003e Cells with GBmap posterior probability below 0.5 were flagged as low confidence. Malignant cell classification was further refined using copy-number variation analysis and post-hoc revision, as detailed below.\u003c/p\u003e\n\u003cp\u003eNormal neural populations.\u003c/p\u003e\n\u003cp\u003eNeurons, oligodendrocytes, OPCs, astrocytes, endothelial cells, and vascular leptomeningeal cells were annotated using CellTypist 1.7.1 with the Adult Human MTG model,\u003csup\u003e84\u003c/sup\u003e applied with cluster majority-vote label propagation at Leiden r=0.5 resolution. CellTypist was used in preference to GBmap for these populations because the GBmap reference is tumour-biased and underperforms on normal brain cell types, for which distal white matter and grey matter samples are enriched.\u003c/p\u003e\n\u003cp\u003eImmune populations.\u003c/p\u003e\n\u003cp\u003eMajor immune populations (microglia, macrophages, T cells, dendritic cells) were initially assigned by GBmap. Fine immune subpopulation annotation (18 populations) was performed using custom marker panel scoring with sc.tl.score_genes, followed by Leiden cluster modal assignment to resolve individual cells to the most frequent population within their cluster. Marker panels were derived from Villani et al. (2017)\u003csup\u003e85\u003c/sup\u003e for dendritic cell subtypes, Friebel et al. (2020)\u003csup\u003e86\u003c/sup\u003e for myeloid populations, M\u0026uuml;ller et al. (2017)\u003csup\u003e87\u003c/sup\u003e for monocyte subtypes, and Zheng et al. (2017)\u003csup\u003e88\u003c/sup\u003e for lymphocyte subtypes. Scoring-based annotation assigns a subset of activated T cells to dendritic cell populations due to overlapping marker expression; for T cell composition analyses, cluster-level modal assignment was therefore used, which correctly separates CD4 T, CD8 T, regulatory T, double-negative T, and NK cell populations. Regulatory T cells were defined by co-expression of FOXP3, IL2RA (CD25), CTLA4, and IKZF2 (Helios) within CD4-expressing clusters; double-negative T cells by CD3D/CD3E/TRAC positivity with neither CD4 nor CD8A/CD8B; NK cells by NKG7, KLRD1, and GNLY in the absence of CD3 transcripts.\u003c/p\u003e\n\u003cp\u003eT cell composition statistics and per-subtype marker scoring.\u003c/p\u003e\n\u003cp\u003eCell composition per zone was computed at three denominators where specified: (i) absolute density per 1,000 total cells, (ii) percentage of CD45+ immune cells (sum of all annotated immune populations), and (iii) percentage of the T cell pool (CD4_T + CD8_T + Treg + DNT + NK_cell). For T cell composition statistical comparisons across zones, biopsy-level (sample_id-aggregated) fractions were used (n = 5 GB patients x 5 tumour zones; 4 BT077 biopsies for WM(Ctrl); 3 BT077 biopsies for GM(Ctrl)). Kruskal-Wallis H tests compared the five tumour zones for each subtype and for total T cell %CD45+; pairwise Mann-Whitney U tests compared each tumour zone to its matched non-oncological control (WM(Ctrl) for AWM and dWM; GM(Ctrl) for dGM; pooled WM+GM control for Core and Edge), with Benjamini-Hochberg correction within each subtype\u0026apos;s pairwise test family. Treg abundance was additionally tested under six denominators (% of T cell pool, % of CD4 lineage [Treg / (CD4_T + Treg)], % of CD45+, per 1,000 total cells, % of total cells, % of non-malignant cells) to distinguish absolute population changes from denominator effects driven by changes in the surrounding T cell compartment.\u003c/p\u003e\n\u003cp\u003ePer-subtype checkpoint receptor expression (CD4 T, CD8 T, Treg) was computed as the fraction of cells with normalised expression \u0026gt; 0 for each receptor (PDCD1/PD-1, HAVCR2/TIM-3, TIGIT, CTLA4, LAG3) per zone; cells per subtype x zone with n \u0026lt; 10 were flagged as low-confidence and shown as open markers in figures. PD-L1 (CD274) was analysed on its source populations (malignant cells, microglia, macrophages) rather than on T cells, since PD-L1 is the ligand and is expressed by antigen-presenting / target cells. CD4+ T cells were further characterised by helper versus cytotoxic gene-set scoring: a helper score (IL7R, CCR7, LEF1, TCF7, SELL, CD40LG, CXCR5) and a cytotoxic score (GZMB, GZMA, GZMK, PRF1, GNLY, IFNG, NKG7, CCL5), each computed via sc.tl.score_genes against the same expression-stable background as other program scores.\u003c/p\u003e\n\u003cp\u003eCopy-number variation analysis\u003c/p\u003e\n\u003cp\u003ePatient-level malignancy was first validated by pseudo-bulk CNV analysis: per-patient counts were aggregated across malignant and matched-normal cells, normalised to CPM, and the per-gene log₂(tumour/reference) ratio smoothed by rolling median (window=100 genes). Six lineage-matched reference panels were used (AC- and MES1/2-like vs. oligodendrocyte; NPC-like vs. neuron; OPC-like vs. OPC). Significant events (|Z|\u0026gt;2 vs the control patient) confirmed hallmark GB alterations (chr7 gain, chr10 loss, focal\u0026nbsp;EGFR\u0026nbsp;amplification) in all five patients (Supplementary Fig. 18; Supplementary Table 3).\u003c/p\u003e\n\u003cp\u003eper-cell CNV calling and tumour classification\u003c/p\u003e\n\u003cp\u003eHigher-resolution per-cell CNV analysis used cnv_group_flex. Each GB subtype was compared to an empirically-matched BT077 reference cell type (AC and MES vs astrocytes; OPC vs oligodendrocytes; NPC1 vs OPCs; NPC2 vs excitatory neurons), with mappings selected by minimising the residual SD of the genome-wide log₂ ratio after drift correction. Three bias-correction layers were applied: sex chromosomes and chrM were excluded; a 138-gene empirical blacklist removed cell-type markers that mimic CNV signal (e.g.\u0026nbsp;GAD2,\u0026nbsp;HTR2A,\u0026nbsp;CNTNAP3B); and a genome-wide drift correction subtracted the median log₂ ratio over expression-stable genes to mitigate the \u0026minus;0.10 to \u0026minus;0.25 baseline shifts seen in tumour-adjacent neurons.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eStage 1.\u0026nbsp;\u003c/em\u003eCells were partitioned into pseudobulk groups of\u0026nbsp;k\u0026nbsp;= 50 within each (patient, cell type, zone) stratum. Per-group regional scores were tested against bootstrap nulls (2,000 iterations) of matched BT077 cells, with BH-corrected two-sided p-values, at three resolutions: chromosome, arm (39 autosomal arms), and focal (15 GB-relevant GISTIC regions \u0026plusmn;1 Mb).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eMarker libraries.\u003c/em\u003e For each patient, markers were derived from group-level penetrance in the GBM_AC core/edge cohort, combining drift-corrected arm markers with pre-drift focal markers (drift correction attenuates focal-amp signal). Library sizes ranged from 8 to 17 markers per patient.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eStage 2.\u003c/em\u003e To recover per-cell resolution at \u003cem\u003ek\u003c/em\u003e = 50, an iterative sort-and-group algorithm (50 seed rounds, 30 ranked rounds with 10% jitter) accumulated per-marker firing as an exponentially-weighted moving average (\u0026alpha; = 0.3). Per-marker scores were averaged within four directional tiers (arm gain/loss, focal gain/loss) and combined as mean_score = 0.7\u0026middot;max(tier means) + 0.3\u0026middot;mean(remaining), bounded [0, 1]. The directional split prevents asymmetric profiles (e.g. NPC2-like loss-only) from being diluted by silent opposite-sign markers.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eClassification.\u003c/em\u003e Cell-type-specific p50 and p99 thresholds were derived per patient by applying the marker library to BT077 controls. Cells were classified normal (\u0026lt; p50), tumour (\u0026ge; p99), or indeterminate; cell types with p99 \u0026ge; 0.90 were flagged unreliable. For subtypes lacking a matched BT077 reference (e.g. GBM_NPC2), fallback thresholds (p50 = 0.10, p99 = 0.50) were used (Supplementary Fig. 18).\u003c/p\u003e\n\u003cp\u003ePost-hoc cell type revision.\u003c/p\u003e\n\u003cp\u003eGBmap classified 92,657 cells as AC-like; however, a subset of these were non-malignant astrocytes misclassified due to shared transcriptional features between reactive astrocytes and the AC-like malignant state. A two-tier revision was applied exclusively to AC-like cells to identify and reclassify non-malignant astrocytes.\u003c/p\u003e\n\u003cp\u003eIn Tier 1, AC-like cells were reclassified as astrocytes if k-nearest neighbor label transfer from non-malignant cells assigned an astrocyte identity with transfer confidence \u0026ge;0.50 (n=16,367 cells; 3.5% of total; mean confidence 0.89). In Tier 2, an additional 33 cells were reclassified based on homeostatic astrocyte marker scoring above the non-oncological control median combined with low malignancy marker expression.\u003c/p\u003e\n\u003cp\u003eAll remaining AC-like cells (n=76,290) were retained as malignant. The revised annotation was used throughout all downstream analyses (Supplementary Table 2).\u003c/p\u003e\n\u003cp\u003eSub-compartment annotation\u003c/p\u003e\n\u003cp\u003eNeuron subtypes.\u003c/p\u003e\n\u003cp\u003eOf 47,262 neurons identified by CellTypist, 8,931 with pan-neuronal marker expression (mean of SNAP25, SYN1, SYN2, SYT1, STMN2, TUBB3, MAP2) below the non-oncological control 25th percentile were reclassified as non-neuronal, leaving 38,331 neurons for subtype analysis. These were subtyped into six classes, excitatory layer subtypes (Ex_L2/3, Ex_L4, Ex_L5/6) and inhibitory subtypes (Inh_PV, Inh_SST, Inh_VIP/LAMP5), using marker panels from Hodge et al. (2019)\u003csup\u003e89\u003c/sup\u003e and the Allen Brain Cell Atlas. Assignment was performed by maximum marker panel score with Leiden cluster modal propagation.\u003c/p\u003e\n\u003cp\u003eAstrocyte states.\u003c/p\u003e\n\u003cp\u003eAstrocytes (n = 25,799) were further classified into reactive states using published marker gene panels scored with sc.tl.score_genes: A1 neurotoxic (C3, SERPING1, GBP2, PSMB8, SRGN, AMIGO2)\u003csup\u003e23\u003c/sup\u003e, A2 neuroprotective (EMP1, S100A10, SPHK1, CD109, PTGS2, TM4SF1)\u003csup\u003e23\u003c/sup\u003e, Disease-Associated Astrocyte (DAA; SERPINA3, SPHK1, CD44, VIM, GFAP)\u003csup\u003e90\u003c/sup\u003e, and Homeostatic (AQP4, GJA1, SLC1A3)\u003csup\u003e91\u003c/sup\u003e. Initial assignment (v2) classified each astrocyte to the state with the highest positive score, with cells co-expressing multiple reactive markers labeled \u0026quot;Mixed\u0026quot; (n = 2,430; 9.4%) and cells with no score above threshold labeled \u0026quot;Uncertain\u0026quot; (n = 7,362; 28.5%). A refinement step (v3) resolved these categories: Mixed cells were assigned to their single highest-scoring reactive state (redistributing to A1 n = +876, A2 n = +1,407, DAA n = +147); Uncertain cells with any reactive score \u0026gt; 0 were assigned to the highest reactive state (n = 2,195; mostly DAA), those with homeostatic score \u0026gt; 0.5 were assigned Homeostatic (n = 2,109), and the remainder were assigned \u0026quot;Astrocyte_Low\u0026quot; (n = 3,058), representing cells with no clear reactive or homeostatic signature. All astrocyte reactive percentages reported in the main text use v3 annotation. The v3 reclassification does not alter any biological conclusion; the reactive gradient is preserved and the relative ordering of states is unchanged (Supplementary Table 5).\u003c/p\u003e\n\u003cp\u003eOligodendrocyte and OPC program scoring.\u003c/p\u003e\n\u003cp\u003eTranscriptional programs were scored across oligodendrocytes (n=104,516) and OPCs (n=26,769) using sc.tl.score_genes. Five programs were evaluated: stress (Gavish et al. 2023, MP6; 49 genes)\u003csup\u003e11\u003c/sup\u003e, myelin maturity (MBP, PLP1, MOG), progenitor markers (PDGFRA, CSPG4, OLIG2), remyelination signaling, and inflammatory response. Unlike astrocytes, which exhibit discrete reactive states, oligodendrocytes and OPCs express stress and maturation markers along a continuous gradient; program scores were therefore used to capture this continuum rather than imposing categorical state boundaries. Between-zone differences were assessed using Kruskal-Wallis H test with Benjamini-Hochberg FDR correction.\u003c/p\u003e\n\u003cp\u003eImmune fine populations.\u003c/p\u003e\n\u003cp\u003eFine immune annotation (18 populations; n=119,164 cells) was performed as described above. Rescue rate was 100%; all cells were assigned to a population via Leiden cluster modal annotation.\u003c/p\u003e\n\u003cp\u003eCell cycle scoring\u003c/p\u003e\n\u003cp\u003eCell cycle phase was scored for all 472,089 cells using the Tirosh et al. (2016) S-phase and G2M-phase gene lists via sc.tl.score_genes_cell_cycle in Scanpy.\u003csup\u003e14\u003c/sup\u003e Cells were classified as S, G2M, or G1 phase based on the maximum of S-score and G2M-score relative to a threshold of 0. Cycling fraction per subtype per zone was defined as the percentage of cells in S or G2M phase.\u003c/p\u003e\n\u003cp\u003eGB malignant state program scoring\u003c/p\u003e\n\u003cp\u003eEight transcriptional programs were scored across all 144,089 malignant cells using sc.tl.score_genes (Scanpy 1.12). Gene lists were obtained from published sources: Hypoxia (MP7; 27 genes)\u003csup\u003e11\u003c/sup\u003e, EMT-I Mesenchymal (MP15; 30 genes)\u003csup\u003e11\u003c/sup\u003e, EMT-III S100/Annexin (MP17; 30 genes)\u003csup\u003e11\u003c/sup\u003e, Heat-Shock Stress (MP9; 29 genes)\u003csup\u003e11\u003c/sup\u003e, Interferon Response (14 genes)\u003csup\u003e11\u003c/sup\u003e, Invasion/Tumor Microtube (17 genes)\u003csup\u003e15\u003c/sup\u003e, Developmental Stemness (14 genes)\u003csup\u003e92\u003c/sup\u003e, and Neural Crest-Like identity (12 confirmed NCC markers)\u003csup\u003e64\u003c/sup\u003e. Gavish meta-program gene lists were obtained from the 3CA Supplementary Table (top 30 genes per meta-program, restricted to genes present in the 10x Flex probe panel). The Hamed NCL gene list comprises the 12 NCC identity markers confirmed in the original publication; injury-associated genes were excluded as they derive from a mouse injury model and include immediate-early genes susceptible to dissociation artifacts.\u003csup\u003e93\u003c/sup\u003e Full gene lists are provided in Supplementary Table 14. Between-zone differences were assessed using Kruskal-Wallis H test; spatial gradients were confirmed at the patient level using Friedman\u0026apos;s test on pseudobulk means (n=5 patients x 5 tumour zones; Methods, Sensitivity analyses).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTwo additional cell-state-specific programmes were scored on the same 144,089 malignant cells to capture the lineage-resemblance signatures referenced in the spatial mimicry analysis: a neural mimicry panel comprising 12 neuronal identity markers (NRGN, RBFOX3, NEFL, NEFM, STMN2, MAPT, CNTNAP2, NEFH, SYT1, SNAP25, JPH3, MAP2) capturing the NPC2-like neuronal transcriptional resemblance, and an OPC lineage panel comprising 9 oligodendrocyte and OPC identity markers (MBP, PLP1, MOG, PDGFRA, CSPG4, OLIG2, SOX10, SOX8, NKX2-2) capturing the OPC-like oligodendroglial resemblance. Both panels were scored using sc.tl.score_genes with the same expression-stable background as the eight published programmes above.\u003c/p\u003e\n\u003cp\u003eGB malignant subtype marker gene identification.\u003c/p\u003e\n\u003cp\u003eMarker genes for each GB malignant subtype were identified by Wilcoxon rank-sum test (sc.tl.rank_genes_groups, method=\u0026quot;wilcoxon\u0026quot;), comparing each subtype against all other malignant cells (n=144,089 total). Tests were performed separately for upregulated and downregulated genes. The top 5 upregulated and top 5 downregulated genes per subtype (ranked by log2 fold-change, filtered to adjusted p\u0026lt;0.05 and minimum detection rate 10%) were selected for visualization. Mean log-normalized expression per subtype per spatial zone was z-scored across zones and clipped at \u0026plusmn;3 for display.\u003c/p\u003e\n\u003cp\u003eSpatially variable gene analysis.\u003c/p\u003e\n\u003cp\u003eTo identify genes whose expression varies systematically across spatial zones, Kruskal-Wallis H tests were applied independently to each malignant subtype and non-malignant cell population. For each cell type, genes expressed in at least 3% of cells were tested for expression variation across spatial zones (5 tumour zones for malignant subtypes; 7 zones including controls for non-malignant populations). Multi-test correction was performed using Benjamini-Hochberg FDR (FDR\u0026lt;0.05, effect size \u0026gt;0.3). Genes were classified as upregulated or downregulated at the infiltration front based on Spearman rank correlation between expression and spatial zone order. Malignant and non-malignant cells were analyzed separately because they have fundamentally different spatial distributions: malignant cell density decreases from Core to distal grey matter, while non-malignant cell density increases; a gene positively correlated with zone in one compartment may show the opposite pattern in the other.\u003c/p\u003e\n\u003cp\u003eCell type marker overlap test.\u003c/p\u003e\n\u003cp\u003eTo assess which normal cell type each malignant subtype most closely resembles transcriptionally, the spatially variable genes of each malignant subtype were compared to normal cell type marker genes. Normal cell type markers were independently derived from the 328,000 non-malignant cells using Wilcoxon rank-sum tests (sc.tl.rank_genes_groups) for each of 8 normal cell types. The overlap between each malignant subtype\u0026apos;s spatially variable genes and each normal cell type\u0026apos;s markers was assessed using Fisher\u0026apos;s exact test, with odds ratios and FDR-corrected p-values reported (Supplementary Fig. 9; Supplementary Table 10).\u003c/p\u003e\n\u003cp\u003eNeuron transcriptional program scoring.\u003c/p\u003e\n\u003cp\u003eCellular stress was scored across all non-malignant cell types using the Gavish et al. (2023) Stress 1 meta-program (MP6; 49 genes from the 3CA Supplementary Table, of which 49 were present in the 10x Flex probe panel).\u003csup\u003e11\u003c/sup\u003e The same\u0026nbsp;program was applied to neurons, oligodendrocytes, OPCs, and astrocytes for consistency. Five additional neuron programs were scored using published gene sets: synaptic function (SYP, SYT1, SNAP25, VAMP2, STX1A, DLG4, NRXN1, BSN, PCLO, SHANK2), neuroplasticity/LTP (BDNF, NTRK2, ARC, HOMER1, CAMK2A, CAMK2B, CREB1, MAPK3, FMR1, SYNGAP1), oxidative stress (SOD1, SOD2, HMOX1, NQO1, GPX4, PRDX1, PRDX2, CAT, TXNRD1), neurodegeneration (APP, PSEN1, MAPT, SNCA, SQSTM1, TARDBP, CLU, VCP), and glutamate/excitotoxicity (GRIA1, GRIA2, GRIN1, GRIN2A, GRIN2B, SLC1A2, SLC1A3, GLS, GLUL). Neurons were classified as stressed if their MP6 score exceeded a zone-matched threshold defined as the non-oncological control neuron mean + 2 standard deviations, computed separately for white matter zones (threshold = 0.241) and grey matter zones (threshold = 0.102) to account for baseline differences in neuronal transcriptional state between tissue compartments.\u003c/p\u003e\n\u003cp\u003eIntegrated stress response (ISR) module scoring.\u003c/p\u003e\n\u003cp\u003eTo dissect the upstream pathways of MP6 stress activation in neurons, five ISR kinase-arm and effector modules were independently scored on the 47,262 CellTypist-identified neurons (reduced to 38,331 after pan-neuronal marker filtering): the PERK/UPR arm (HSPA5/BiP, DNAJB9, HERPUD1, ERO1A, EDEM1, HYOU1, SEC61B, DERL1, DERL2, P4HB, PDIA3, PDIA4, PDIA6, CALR, XBP1); the GCN2/kynurenine arm (ASNS, PHGDH, PSAT1, PSPH, SLC7A5, SLC1A5, CBS, CTH, MTHFD2, CYP1A1, CYP1B1, AHRR); the PKR/interferon arm (IFIT1, IFIT2, IFIT3, OAS1, OAS2, OAS3, MX1, MX2, ISG15, RSAD2, DDX58, HERC5, ISG20, GBP1, GBP2); the HRI/mitochondrial-UPR arm (ATF5, HSPD1, HSPE1, LONP1, CLPP, BNIP3L, PINK1, HMOX1, NQO1, GCLC, GCLM, TXNRD1, SRXN1); and an excitotoxicity / immediate-early gene module (NPAS4, ARC, FOSB, FOSL1, EGR3, BDNF, NR4A1, NR4A2, NR4A3). Module scores were computed using sc.tl.score_genes (use_raw=False); between-group comparisons (MP6-classified stressed vs non-stressed neurons) used the two-sided Mann-Whitney U test. Per-marker positivity rates within stressed neurons (e.g. BiP/HSPA5 firing in 18.9% of stressed neurons) were computed as the fraction with normalised expression \u0026gt; 0.\u003c/p\u003e\n\u003cp\u003eNon-malignant cell spatial transcriptional analysis.\u003c/p\u003e\n\u003cp\u003eAgnostic spatial variable gene analysis was applied to four non-malignant cell populations: neurons (n = 47,262), oligodendrocytes (n = 104,516), OPCs (n = 26,769), and astrocytes (n = 25,799). For each cell type, genes expressed in at least 3% of cells were tested for expression variation across all 7 spatial zones (including non-oncological control) using Kruskal-Wallis H test with BH FDR correction (FDR \u0026lt; 0.05, effect size \u0026gt; 0.3). Genes were classified as tumour-elevated (higher expression in tumour-proximal zones; Spearman rho \u0026lt; -0.3 with zone rank) or tumour-depleted (higher in control zones; rho \u0026gt; 0.3). Pathway enrichment on each direction-specific gene list was performed using gseapy enrichr against GO Biological Process 2023 and MSigDB Hallmark 2020 gene sets.\u003c/p\u003e\n\u003cp\u003ePseudobulk differential expression\u003c/p\u003e\n\u003cp\u003ePseudobulk differential expression analysis was performed using PyDESeq2 0.5.4. For each comparison, raw counts were aggregated per patient per zone per cell type to generate pseudobulk profiles.\u003csup\u003e94\u003c/sup\u003e Comparisons included: (1) distal non-malignant cells vs. the non-oncological control, (2) adjacent white matter non-malignant vs. distal non-malignant, (3) adjacent white matter malignant vs. tumour edge malignant, (4) glioblastoma microglia vs. non-oncological control microglia, and (5) stressed vs. non-stressed neurons. Genes with adjusted p\u0026lt;0.05 and |log2 fold-change|\u0026gt;0.5 were considered differentially expressed.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWasserstein distance analysis\u003c/p\u003e\n\u003cp\u003eTranscriptional intermediacy of spatial zones was quantified using the Wasserstein distance (earth mover\u0026apos;s distance) between zone-level UMAP distributions. For each spatial zone, 5,000 cells were subsampled from the 2D UMAP embedding. Pairwise Wasserstein distances were computed using scipy.stats.wasserstein_distance, averaged across both UMAP dimensions (Supplementary Fig. 3).\u003c/p\u003e\n\u003cp\u003eCell-cell communication analysis\u003c/p\u003e\n\u003cp\u003eLigand-receptor interaction analysis was performed using LIANA 1.7.1 on the tumour microenvironment subset (Core + Edge cells) and the infiltration front subset (Edge + AWM cells).\u003csup\u003e95\u003c/sup\u003e The full 472,089-cell dataset was not processed due to memory constraints. LIANA was run with the consensus resource and five scoring methods (CellChat, NATMI, SingleCellSignalR, Connectome, log2FC; n_perms=50); interactions reported are those significant across at least three methods (FDR\u0026lt;0.05).\u003c/p\u003e\n\u003cp\u003eSensitivity and robustness analyses\u003c/p\u003e\n\u003cp\u003eLeave-one-out (LOO) sensitivity analysis was performed to assess whether key quantitative findings were robust to the exclusion of individual patients. Each of the five glioblastoma patients was excluded in turn, and the relevant statistic was recomputed on the remaining four patients. BT074-A and BT074-C (biological replicates) were additionally excluded together as a sixth scenario. A finding was considered robust if the direction of effect was preserved across all LOO iterations. Pseudobulk sensitivity analysis was performed by aggregating program scores to patient-level means and testing with Friedman\u0026apos;s test (non-parametric repeated-measures ANOVA; n=5 patients x 5 tumour zones). Pairwise comparisons used Wilcoxon signed-rank tests paired by patient (Supplementary Table 5).\u003c/p\u003e\n\u003cp\u003eSoftware Used\u003c/p\u003e\n\u003cp\u003eAll statistical analyses were performed in Python 3.10 using scipy 1.16.3, numpy 2.3.5, and statsmodels 0.14.6. Unless otherwise stated: between-group comparisons used the Mann-Whitney U test (two-sided); multiple comparison correction used the Benjamini-Hochberg false discovery rate procedure; significance threshold was FDR\u0026lt;0.05. Kruskal-Wallis H test was used to assess overall between-zone differences for program score analyses. Spearman rank correlation was used for continuous variable associations. Cohen\u0026apos;s d effect sizes for the power calculation were computed as mean delta / SD of within-patient deltas (paired design). All figures were generated using matplotlib 3.10.8 and seaborn 0.13.2. Generative AI assistance: Claude Code (Anthropic; model: Claude Opus 4.6) was used for copy-editing of manuscript prose, assistance with analysis-code refactoring, and debugging. All AI-generated text and code were reviewed, edited, and verified by the authors, who take full responsibility for the accuracy and integrity of the work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cbr clear=\"all\"\u003e \u003c/strong\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eEthics statement\u003c/p\u003e\n\u003cp\u003eAll tissue samples were obtained under written informed consent following protocols approved by the [Institution] Research Ethics Board (REB IDs withheld for double-blind review; provided to editorial office). Procedures conformed to the Declaration of Helsinki and applicable national guidelines.\u003c/p\u003e\n\u003cp\u003eData availability\u003c/p\u003e\n\u003cp\u003eRaw and processed single-cell RNA-sequencing data have been deposited in the Gene Expression Omnibus (GEO) under accession number [withheld for double-blind review; reviewer access provided to editorial office]. The annotated AnnData object (06_publication.h5ad; 472,089 cells x 18,108 genes) and per-figure summary tables are available at [withheld for review]. All publicly available reference datasets (Darmanis et al. 2017; Siletti et al. 2023) are cited in the references and accessed via their original deposition.\u003c/p\u003e\n\u003cp\u003eCode availability\u003c/p\u003e\n\u003cp\u003eAll analysis code, including notebooks for figure generation and the per-cell CNV pipeline (cnv_group_flex), is available at [link provided to editorial office]. The repository includes a Conda environment specification (scrna.yml) and per-script README files. Software dependencies and exact versions are documented in the Methods.\u003c/p\u003e\n\u003cp\u003eAuthor contributions\u003c/p\u003e\n\u003cp\u003e[Author contributions removed for double-blind review; provided in cover letter to editorial office.]\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eStupp, R.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Radiotherapy plus concomitant and adjuvant temozolomide for glioblastoma. \u003cem\u003eNew England Journal of Medicine\u003c/em\u003e \u003cstrong\u003e352\u003c/strong\u003e, 987\u0026ndash;996 (2005). https://doi.org/10.1056/NEJMoa043330\u003c/li\u003e\n \u003cli\u003ePrice, M.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e CBTRUS Statistical Report: Primary Brain and Other Central Nervous System Tumors Diagnosed in the United States in 2018\u0026ndash;2022. \u003cem\u003eNeuro-Oncology\u003c/em\u003e \u003cstrong\u003e27\u003c/strong\u003e, iv1\u0026ndash;iv66 (2025). https://doi.org/10.1093/neuonc/noaf194\u003c/li\u003e\n \u003cli\u003eWen, P. Y.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Glioblastoma in adults: A Society for Neuro-Oncology (SNO) and European Society of Neuro-Oncology (EANO) consensus review on current management and future directions. \u003cem\u003eNeuro-Oncology\u003c/em\u003e (2025). https://doi.org/10.1093/neuonc/noaf177\u003c/li\u003e\n \u003cli\u003eLouis, D. N.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e The 2021 WHO Classification of Tumors of the Central Nervous System: a summary. \u003cem\u003eNeuro-Oncology\u003c/em\u003e \u003cstrong\u003e23\u003c/strong\u003e, 1231\u0026ndash;1251 (2021). https://doi.org/10.1093/neuonc/noab106\u003c/li\u003e\n \u003cli\u003eJr., E. B. \u0026amp; Karnosh, L. J. Cerebral Hemispherectomy. \u003cem\u003eJournal of Neurosurgery\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e (1949\u0026ndash;07\u0026ndash;01). https://doi.org/10.3171/jns.1949.6.4.0285\u003c/li\u003e\n \u003cli\u003eGardner, W. J. Removal of the Right Cerebral Hemisphere for Infiltrating Glioma. \u003cem\u003eJournal of the American Medical Association\u003c/em\u003e \u003cstrong\u003e101\u003c/strong\u003e (1933/09/09). https://doi.org/10.1001/jama.1933.02740360003002\u003c/li\u003e\n \u003cli\u003eDelgado-L\u0026oacute;pez, P. D. \u0026amp; Corrales-Garc\u0026iacute;a, E. M. Survival in glioblastoma: a review on the impact of treatment modalities. \u003cem\u003eClinical and Translational Oncology 2016 18:11\u003c/em\u003e \u003cstrong\u003e18\u003c/strong\u003e (2016\u0026ndash;03\u0026ndash;10). https://doi.org/10.1007/s12094-016-1497-x\u003c/li\u003e\n \u003cli\u003eDarmanis, S.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Single-Cell RNA-Seq Analysis of Infiltrating Neoplastic Cells at the Migrating Front of Human Glioblastoma. \u003cem\u003eCell Reports\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e (2017/10/31). https://doi.org/10.1016/j.celrep.2017.10.030\u003c/li\u003e\n \u003cli\u003eCouturier, C. P.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Single-cell RNA-seq reveals that glioblastoma recapitulates a normal neurodevelopmental hierarchy. \u003cem\u003eNature Communications\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e (2020). https://doi.org/10.1038/s41467-020-17186-5\u003c/li\u003e\n \u003cli\u003eNeftel, C.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e An Integrative Model of Cellular States, Plasticity, and Genetics for Glioblastoma. \u003cem\u003eCell\u003c/em\u003e \u003cstrong\u003e178\u003c/strong\u003e (2019/08/08). https://doi.org/10.1016/j.cell.2019.06.024\u003c/li\u003e\n \u003cli\u003eGavish, A.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Hallmarks of transcriptional intratumour heterogeneity across a thousand tumours. \u003cem\u003eNature 2023 618:7965\u003c/em\u003e \u003cstrong\u003e618\u003c/strong\u003e (2023\u0026ndash;05\u0026ndash;31). https://doi.org/10.1038/s41586-023-06130-4\u003c/li\u003e\n \u003cli\u003eCh\u0026eacute;deville, A. L.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Investigating Glioblastoma Response to Hypoxia. \u003cem\u003eBiomedicines 2020, Vol. 8, Page 310\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e (2020\u0026ndash;08\u0026ndash;27). https://doi.org/10.3390/biomedicines8090310\u003c/li\u003e\n \u003cli\u003eFeldman, L. Hypoxia within the glioblastoma tumor microenvironment: a master saboteur of novel treatments. \u003cem\u003eFrontiers in Immunology\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e (2024/06/26). https://doi.org/10.3389/fimmu.2024.1384249\u003c/li\u003e\n \u003cli\u003eTirosh, I.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Single-cell RNA-seq supports a developmental hierarchy in human oligodendroglioma. \u003cem\u003eNature 2016 539:7628\u003c/em\u003e \u003cstrong\u003e539\u003c/strong\u003e (2016\u0026ndash;11\u0026ndash;02). https://doi.org/10.1038/nature20123\u003c/li\u003e\n \u003cli\u003eVenkataramani, V.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Glioblastoma hijacks neuronal mechanisms for brain invasion. \u003cem\u003eCell\u003c/em\u003e \u003cstrong\u003e185\u003c/strong\u003e, 2899\u0026ndash;2917.e2831 (2022). https://doi.org/10.1016/j.cell.2022.06.054\u003c/li\u003e\n \u003cli\u003eD\u0026apos;Alessio, A.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Analysis of angiogenesis related factors in glioblastoma peritumoral tissue and their derived cancer stem cells. \u003cem\u003eOncotarget\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e (2016\u0026ndash;10\u0026ndash;01). https://doi.org/10.18632/oncotarget.12398\u003c/li\u003e\n \u003cli\u003eJarmuzek, P.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Cytokine Profile in Development of Glioblastoma in Relation to Healthy Individuals. \u003cem\u003eInternational Journal of Molecular Sciences 2023, Vol. 24, Page 16206\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e (2023\u0026ndash;11\u0026ndash;11). https://doi.org/10.3390/ijms242216206\u003c/li\u003e\n \u003cli\u003eHassel, B., Niehusmann, P., Halvorsen, B. \u0026amp; Dahlberg, D. Pro-inflammatory cytokines in cystic glioblastoma: A quantitative study with a comparison with bacterial brain abscesses. With an MRI investigation of displacement and destruction of the brain tissue surrounding a glioblastoma. \u003cem\u003eFrontiers in Oncology\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e (2022/07/29). https://doi.org/10.3389/fonc.2022.846674\u003c/li\u003e\n \u003cli\u003eSasaki, A.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Secretion of Transforming Growth Factor-\u0026beta;1 and -\u0026beta;2 by Malignant Glioma Cells. \u003cem\u003eNeurologia medico-chirurgica\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e (1995). https://doi.org/10.2176/nmc.35.423\u003c/li\u003e\n \u003cli\u003eWei, Q.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e TNF\u0026alpha; secreted by glioma associated macrophages promotes endothelial activation and resistance against anti-angiogenic therapy. \u003cem\u003eActa Neuropathologica Communications 2021 9:1\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e (2021\u0026ndash;04\u0026ndash;14). https://doi.org/10.1186/s40478-021-01163-0\u003c/li\u003e\n \u003cli\u003eSoni, H.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e PERK-mediated expression of peptidylglycine \u0026alpha;-amidating monooxygenase supports angiogenesis in glioblastoma. \u003cem\u003eOncogenesis 2020 9:2\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e (2020\u0026ndash;02\u0026ndash;13). https://doi.org/10.1038/s41389-020-0201-8\u003c/li\u003e\n \u003cli\u003eColwell, N.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Hypoxia in the glioblastoma microenvironment: shaping the phenotype of cancer stem-like cells. \u003cem\u003eNeuro-Oncology\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e (2017/07/01). https://doi.org/10.1093/neuonc/now258\u003c/li\u003e\n \u003cli\u003eLiddelow, S. A.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Neurotoxic reactive astrocytes are induced by activated microglia. \u003cem\u003eNature 2017 541:7638\u003c/em\u003e \u003cstrong\u003e541\u003c/strong\u003e (2017\u0026ndash;01\u0026ndash;18). https://doi.org/10.1038/nature21029\u003c/li\u003e\n \u003cli\u003eRuiz-Moreno, C.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Charting the single-cell and spatial landscape of IDH-wild-type glioblastoma with GBmap. \u003cem\u003eNeuro-Oncology\u003c/em\u003e \u003cstrong\u003e27\u003c/strong\u003e (2025/10/14). https://doi.org/10.1093/neuonc/noaf113\u003c/li\u003e\n \u003cli\u003ePatel, A. P.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Single-cell RNA-seq highlights intratumoral heterogeneity in primary glioblastoma. \u003cem\u003eScience\u003c/em\u003e \u003cstrong\u003e344\u003c/strong\u003e (2014\u0026ndash;06\u0026ndash;20). https://doi.org/10.1126/science.1254257\u003c/li\u003e\n \u003cli\u003eChen, J.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e A restricted cell population propagates glioblastoma growth after chemotherapy. \u003cem\u003eNature 2012 488:7412\u003c/em\u003e \u003cstrong\u003e488\u003c/strong\u003e (2012\u0026ndash;08\u0026ndash;01). https://doi.org/10.1038/nature11287\u003c/li\u003e\n \u003cli\u003eBao, S.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Glioma stem cells promote radioresistance by preferential activation of the DNA damage response. \u003cem\u003eNature 2006 444:7120\u003c/em\u003e \u003cstrong\u003e444\u003c/strong\u003e (2006\u0026ndash;10\u0026ndash;18). https://doi.org/10.1038/nature05236\u003c/li\u003e\n \u003cli\u003eBluestone, J. A. \u0026amp; Anderson, M. Tolerance in the Age of Immunotherapy. \u003cem\u003eNew England Journal of Medicine\u003c/em\u003e \u003cstrong\u003e383\u003c/strong\u003e (2020\u0026ndash;09\u0026ndash;17). https://doi.org/10.1056/NEJMra1911109\u003c/li\u003e\n \u003cli\u003eHandel, A. E.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e The role of thymic tolerance in CNS autoimmune disease. \u003cem\u003eNature Reviews Neurology 2018 14:12\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e (2018\u0026ndash;11\u0026ndash;19). https://doi.org/10.1038/s41582-018-0095-7\u003c/li\u003e\n \u003cli\u003eOstermann, S.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Plasma and Cerebrospinal Fluid Population Pharmacokinetics of Temozolomide in Malignant Glioma Patients. \u003cem\u003eClinical Cancer Research\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e (2004/06/01). https://doi.org/10.1158/1078-0432.CCR-03-0807\u003c/li\u003e\n \u003cli\u003ePortnow, J.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e The Neuropharmacokinetics of Temozolomide in Patients with Resectable Brain Tumors: Potential Implications for the Current Approach to Chemoradiation. \u003cem\u003eClinical Cancer Research\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e (2009/11/15). https://doi.org/10.1158/1078-0432.CCR-09-1349\u003c/li\u003e\n \u003cli\u003eRosso, L.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e A New Model for Prediction of Drug Distribution in Tumor and Normal Tissues: Pharmacokinetics of Temozolomide in Glioma Patients. \u003cem\u003eCancer Research\u003c/em\u003e \u003cstrong\u003e69\u003c/strong\u003e (2009/01/01). https://doi.org/10.1158/0008-5472.CAN-08-2356\u003c/li\u003e\n \u003cli\u003ePiccirillo, S. G. M.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Bone morphogenetic proteins inhibit the tumorigenic potential of human brain tumour-initiating cells. \u003cem\u003eNature 2006 444:7120\u003c/em\u003e \u003cstrong\u003e444\u003c/strong\u003e (2006/12). https://doi.org/10.1038/nature05349\u003c/li\u003e\n \u003cli\u003eYing, M.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Regulation of glioblastoma stem cells by retinoic acid: role for Notch pathway inhibition. \u003cem\u003eOncogene 2011 30:31\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e (2011\u0026ndash;03\u0026ndash;07). https://doi.org/10.1038/onc.2011.58\u003c/li\u003e\n \u003cli\u003eCampos, B.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Differentiation Therapy Exerts Antitumor Effects on Stem-like Glioma Cells. \u003cem\u003eClinical Cancer Research\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e (2010/05/15). https://doi.org/10.1158/1078-0432.CCR-09-1800\u003c/li\u003e\n \u003cli\u003eTallman, M. M.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Improving Localized Radiotherapy for Glioblastoma via Small Molecule Inhibition of KIF11. \u003cem\u003eCancers 2023, Vol. 15, Page 3173\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e (2023\u0026ndash;06\u0026ndash;13). https://doi.org/10.3390/cancers15123173\u003c/li\u003e\n \u003cli\u003eShen, H., Hau, E., Joshi, S., Dilda, P. J. \u0026amp; McDonald, K. L. Sensitization of Glioblastoma Cells to Irradiation by Modulating the Glucose Metabolism. \u003cem\u003eMolecular Cancer Therapeutics\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e (2015/08/01). https://doi.org/10.1158/1535-7163.MCT-15-0247\u003c/li\u003e\n \u003cli\u003eAhmed, S. U.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Selective Inhibition of Parallel DNA Damage Response Pathways Optimizes Radiosensitization of Glioblastoma Stem-like Cells. \u003cem\u003eCancer Research\u003c/em\u003e \u003cstrong\u003e75\u003c/strong\u003e (2015/10/15). https://doi.org/10.1158/0008-5472.CAN-14-3790\u003c/li\u003e\n \u003cli\u003eHirose, Y., Berger, M. S. \u0026amp; Pieper, R. O. Abrogation of the Chk1-mediated G2 Checkpoint Pathway Potentiates Temozolomide-induced Toxicity in a p53-independent Manner in Human Glioblastoma Cells1. \u003cem\u003eCancer Research\u003c/em\u003e \u003cstrong\u003e61\u003c/strong\u003e, 5843\u0026ndash;5849 (2001).\u003c/li\u003e\n \u003cli\u003eHoang‐Minh, L. B.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Infiltrative and drug‐resistant slow‐cycling cells support metabolic heterogeneity in glioblastoma. \u003cem\u003eThe EMBO Journal 2018 37:23\u003c/em\u003e \u003cstrong\u003e37\u003c/strong\u003e (2018\u0026ndash;10\u0026ndash;15). https://doi.org/10.15252/embj.201798772\u003c/li\u003e\n \u003cli\u003eSato, A.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Glioma-Initiating Cell Elimination by Metformin Activation of FOXO3 via AMPK. \u003cem\u003eStem Cells Translational Medicine\u003c/em\u003e \u003cstrong\u003e1\u003c/strong\u003e (2012/11/01). https://doi.org/10.5966/sctm.2012-0058\u003c/li\u003e\n \u003cli\u003eZhang, P.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Therapeutic targeting of tumor-associated myeloid cells synergizes with radiation therapy for glioblastoma. \u003cem\u003eProceedings of the National Academy of Sciences\u003c/em\u003e \u003cstrong\u003e116\u003c/strong\u003e (2019\u0026ndash;11\u0026ndash;19). https://doi.org/10.1073/pnas.1906346116\u003c/li\u003e\n \u003cli\u003eKhan, F.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Macrophages and microglia in glioblastoma: heterogeneity, plasticity, and therapy. \u003cem\u003eThe Journal of Clinical Investigation\u003c/em\u003e \u003cstrong\u003e133\u003c/strong\u003e (2023/01/03). https://doi.org/10.1172/JCI163446\u003c/li\u003e\n \u003cli\u003eKumar, S.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Molecular heterogeneity of Glioblastoma-associated microglia and macrophages and myeloid-derived suppressor cells: Insights from single-cell omics and therapeutic implications. \u003cem\u003eCritical Reviews in Oncology/Hematology\u003c/em\u003e \u003cstrong\u003e220\u003c/strong\u003e (2026/04/01). https://doi.org/10.1016/j.critrevonc.2026.105190\u003c/li\u003e\n \u003cli\u003eAkay, F. \u0026amp; Saleh, M. Rational therapeutic targeting of myeloid cells in glioblastoma: challenges and perspectives. \u003cem\u003eFrontiers in Immunology\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e (2025/06/26). https://doi.org/10.3389/fimmu.2025.1472710\u003c/li\u003e\n \u003cli\u003eG\u0026oacute;mez, E. G.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Efficacy and safety of immune checkpoint inhibitors and mTOR inhibitors as targeted therapy for glioblastoma: A systematic review and meta-analysis of randomized clinical trials. \u003cem\u003eNeurosurgical Review 2026 49:1\u003c/em\u003e \u003cstrong\u003e49\u003c/strong\u003e (2026\u0026ndash;01\u0026ndash;03). https://doi.org/10.1007/s10143-025-04043-6\u003c/li\u003e\n \u003cli\u003eMelo, S. M. d.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Anti‐PD‐1 and anti‐PD‐L1 antibodies for glioma - de Melo, SM - 2025 | Cochrane Library. \u003cem\u003eCochrane Database of Systematic Reviews\u003c/em\u003e \u003cstrong\u003e2025\u003c/strong\u003e (2025). https://doi.org/10.1002/14651858.CD012532.pub2\u003c/li\u003e\n \u003cli\u003eZeng, Y.-F.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e The efficacy and safety of anti-PD-1/PD-L1 in treatment of glioma: a single-arm meta-analysis. \u003cem\u003eFrontiers in Immunology\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e (2023/04/14). https://doi.org/10.3389/fimmu.2023.1168244\u003c/li\u003e\n \u003cli\u003eTykocki, T. Immunotherapy failure in glioblastoma: A systematic review and meta-analysis of randomized controlled trials. \u003cem\u003eCritical Reviews in Oncology/Hematology\u003c/em\u003e \u003cstrong\u003e222\u003c/strong\u003e (2026/06/01). https://doi.org/10.1016/j.critrevonc.2026.105279\u003c/li\u003e\n \u003cli\u003eSkadborg, S. K.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Nivolumab Reaches Brain Lesions in Patients with Recurrent Glioblastoma and Induces T-cell Activity and Upregulation of Checkpoint Pathways. \u003cem\u003eCancer Immunology Research\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e (2024/09/01). https://doi.org/10.1158/2326-6066.CIR-23-0959\u003c/li\u003e\n \u003cli\u003eLee, A. H.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Neoadjuvant PD-1 blockade induces T cell and cDC1 activation but fails to overcome the immunosuppressive tumor associated macrophages in recurrent glioblastoma. \u003cem\u003eNature Communications 2021 12:1\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e (2021\u0026ndash;11\u0026ndash;26). https://doi.org/10.1038/s41467-021-26940-2\u003c/li\u003e\n \u003cli\u003eCloughesy, T. F.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Neoadjuvant anti-PD-1 immunotherapy promotes a survival benefit with intratumoral and systemic immune responses in recurrent glioblastoma. \u003cem\u003eNature Medicine 2019 25:3\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e (2019\u0026ndash;02\u0026ndash;11). https://doi.org/10.1038/s41591-018-0337-7\u003c/li\u003e\n \u003cli\u003eSchalper, K. A.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Neoadjuvant nivolumab modifies the tumor immune microenvironment in resectable glioblastoma. \u003cem\u003eNature Medicine 2019 25:3\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e (2019\u0026ndash;02\u0026ndash;11). https://doi.org/10.1038/s41591-018-0339-5\u003c/li\u003e\n \u003cli\u003eNose, Y.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e PD-1 antibody-bound progenitor-exhausted CD8+ T cells in lymph nodes boost PD-1-blockade anti-tumor immunity in gastrointestinal cancer. \u003cem\u003eNature Communications 2026\u003c/em\u003e (2026\u0026ndash;04\u0026ndash;08). https://doi.org/10.1038/s41467-026-70751-2\u003c/li\u003e\n \u003cli\u003eTsui, C.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Lymph nodes fuel KLF2-dependent effector CD8+ T cell differentiation during chronic infection and checkpoint blockade. \u003cem\u003eNature Immunology 2025 26:10\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e (2025\u0026ndash;09\u0026ndash;15). https://doi.org/10.1038/s41590-025-02276-7\u003c/li\u003e\n \u003cli\u003eWijesinghe, S. K. M.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Lymph-node-derived stem-like but not tumor-tissue-resident CD8+ T cells fuel anticancer immunity. \u003cem\u003eNature Immunology 2025 26:8\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e (2025\u0026ndash;07\u0026ndash;29). https://doi.org/10.1038/s41590-025-02219-2\u003c/li\u003e\n \u003cli\u003evan Pul, K. M., Fransen, M. F., van de Ven, R. \u0026amp; de Gruijl, T. D. Immunotherapy Goes Local: The Central Role of Lymph Nodes in Driving Tumor Infiltration and Efficacy. \u003cem\u003eFrontiers in Immunology\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e (2021/03/01). https://doi.org/10.3389/fimmu.2021.643291\u003c/li\u003e\n \u003cli\u003eDammeijer, F.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e The PD-1/PD-L1-Checkpoint Restrains T cell Immunity in Tumor-Draining Lymph Nodes. \u003cem\u003eCancer Cell\u003c/em\u003e \u003cstrong\u003e38\u003c/strong\u003e (2020/11/09). https://doi.org/10.1016/j.ccell.2020.09.001\u003c/li\u003e\n \u003cli\u003eMiller, B. C.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Subsets of exhausted CD8+ T cells differentially mediate tumor control and respond to checkpoint blockade. \u003cem\u003eNature Immunology 2019 20:3\u003c/em\u003e \u003cstrong\u003e20\u003c/strong\u003e (2019\u0026ndash;02\u0026ndash;18). https://doi.org/10.1038/s41590-019-0312-6\u003c/li\u003e\n \u003cli\u003eVanni, A.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e 5‐ALA Assisted Surgery of Human Glioblastoma Samples Reveals an Enrichment of T Cells Expressing PD‐1 and CD103 in the Intermediate and Marginal Layers. \u003cem\u003eEuropean Journal of Immunology\u003c/em\u003e \u003cstrong\u003e55\u003c/strong\u003e (2025/06/01). https://doi.org/10.1002/eji.202451681\u003c/li\u003e\n \u003cli\u003eRomagnoli, G.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e CD8+CD103+PD1+TIM3+ T cells in glioblastoma microenvironment correlate with prognosis. \u003cem\u003eImmunology\u003c/em\u003e \u003cstrong\u003e171\u003c/strong\u003e (2024/02/01). https://doi.org/10.1111/imm.13710\u003c/li\u003e\n \u003cli\u003ePedersen, V.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Tumor and immune cell distribution in the tumor core and outer part of glioblastoma, IDH wildtype. \u003cem\u003eJournal of Neuro-Oncology 2025 175:3\u003c/em\u003e \u003cstrong\u003e175\u003c/strong\u003e (2025\u0026ndash;09\u0026ndash;25). https://doi.org/10.1007/s11060-025-05232-5\u003c/li\u003e\n \u003cli\u003eGalvez-Cancino, F.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Regulatory T cell depletion promotes myeloid cell activation and glioblastoma response to anti-PD1 and tumor-targeting antibodies. \u003cem\u003eImmunity\u003c/em\u003e \u003cstrong\u003e58\u003c/strong\u003e (2025/05/13). https://doi.org/10.1016/j.immuni.2025.03.021\u003c/li\u003e\n \u003cli\u003eHamed, A. A.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Gliomagenesis mimics an injury response orchestrated by neural crest-like cells. \u003cem\u003eNature 2025 638:8050\u003c/em\u003e \u003cstrong\u003e638\u003c/strong\u003e (2025\u0026ndash;01\u0026ndash;01). https://doi.org/10.1038/s41586-024-08356-2\u003c/li\u003e\n \u003cli\u003eKarschnia, P.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Chimeric Antigen Receptor T Cells for Glioblastoma. \u003cem\u003eNeurology\u003c/em\u003e \u003cstrong\u003e97\u003c/strong\u003e (2021\u0026ndash;August\u0026ndash;3). https://doi.org/10.1212/WNL.0000000000012193\u003c/li\u003e\n \u003cli\u003eHayat, M. A., Yu, S., Tao, G., Ghani, M. U. \u0026amp; Hu, J. Chimeric antigen receptor T cell therapy for glioblastoma: overcoming current barriers and strategies to enhance efficacy for therapeutic implications. \u003cem\u003eBiochemical Pharmacology\u003c/em\u003e \u003cstrong\u003e248\u003c/strong\u003e (2026/06/01). https://doi.org/10.1016/j.bcp.2026.117851\u003c/li\u003e\n \u003cli\u003ePant, A., Lim, M., Pant, A. \u0026amp; Lim, M. CAR-T Therapy in GBM: Current Challenges and Avenues for Improvement. \u003cem\u003eCancers 2023, Vol. 15, Page 1249\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e (2023\u0026ndash;02\u0026ndash;16). https://doi.org/10.3390/cancers15041249\u003c/li\u003e\n \u003cli\u003eIsmail, F. S.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Current and Future Roles of Chimeric Antigen Receptor T-Cell Therapy in Neurology. \u003cem\u003eJAMA Neurology\u003c/em\u003e \u003cstrong\u003e82\u003c/strong\u003e (2025/01/01). https://doi.org/10.1001/jamaneurol.2024.3818\u003c/li\u003e\n \u003cli\u003eLuksik, A. S.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e CAR T Cell Therapy in Glioblastoma: Overcoming Challenges Related to Antigen Expression. \u003cem\u003eCancers 2023, Vol. 15, Page 1414\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e (2023\u0026ndash;02\u0026ndash;23). https://doi.org/10.3390/cancers15051414\u003c/li\u003e\n \u003cli\u003eGoel, S., Wharton, S. B., Brett, L. P. \u0026amp; Whittle, I. R. Morphological changes and stress responses in neurons in cerebral cortex infiltrated by diffuse astrocytoma. \u003cem\u003eNeuropathology\u003c/em\u003e \u003cstrong\u003e23\u003c/strong\u003e (2003/12/01). https://doi.org/10.1046/j.1440-1789.2003.00510.x\u003c/li\u003e\n \u003cli\u003ePakos‐Zebrucka, K.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e The integrated stress response. \u003cem\u003eThe EMBO Reports 2016 17:10\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e (2016\u0026ndash;09\u0026ndash;14). https://doi.org/10.15252/embr.201642195\u003c/li\u003e\n \u003cli\u003ePerelroizen, R.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Astrocyte immunometabolic regulation of the tumour microenvironment drives glioblastoma pathogenicity. \u003cem\u003eBrain\u003c/em\u003e \u003cstrong\u003e145\u003c/strong\u003e (2022/09/14). https://doi.org/10.1093/brain/awac222\u003c/li\u003e\n \u003cli\u003eFaust Akl, C.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Glioblastoma-instructed astrocytes suppress tumour-specific T cell immunity. \u003cem\u003eNature 2025 643:8070\u003c/em\u003e \u003cstrong\u003e643\u003c/strong\u003e (2025\u0026ndash;05\u0026ndash;21). https://doi.org/10.1038/s41586-025-08997-x\u003c/li\u003e\n \u003cli\u003eConti, A.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Role of Inflammation and Oxidative Stress Mediators in Gliomas. \u003cem\u003eCancers 2010, Vol. 2, Pages 693-712\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e (2010\u0026ndash;04\u0026ndash;26). https://doi.org/10.3390/cancers2020693\u003c/li\u003e\n \u003cli\u003eOlivier, C., Oliver, L., Lalier, L. \u0026amp; Vallette, F. M. Drug Resistance in Glioblastoma: The Two Faces of Oxidative Stress. \u003cem\u003eFrontiers in Molecular Biosciences\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e (2021/01/27). https://doi.org/10.3389/fmolb.2020.620677\u003c/li\u003e\n \u003cli\u003eKumaria, A. \u0026amp; Ashkan, K. Novel therapeutic strategies in glioma targeting glutamatergic neurotransmission. \u003cem\u003eBrain Research\u003c/em\u003e \u003cstrong\u003e1818\u003c/strong\u003e (2023/11/01). https://doi.org/10.1016/j.brainres.2023.148515\u003c/li\u003e\n \u003cli\u003eSokolov, E., Dietrich, J. \u0026amp; Cole, A. J. The complexities underlying epilepsy in people with glioblastoma. \u003cem\u003eThe Lancet Neurology\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e (2023/06/01). https://doi.org/10.1016/S1474-4422(23)00031-5\u003c/li\u003e\n \u003cli\u003eOh, M. C.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Overexpression of Calcium-Permeable Glutamate Receptors in Glioblastoma Derived Brain Tumor Initiating Cells. \u003cem\u003ePLOS ONE\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e (Oct 23, 2012). https://doi.org/10.1371/journal.pone.0047846\u003c/li\u003e\n \u003cli\u003eBuckingham, S. C. \u0026amp; Robel, S. Glutamate and tumor-associated epilepsy: Glial cell dysfunction in the peritumoral environment. \u003cem\u003eNeurochemistry International\u003c/em\u003e \u003cstrong\u003e63\u003c/strong\u003e (2013/12/01). https://doi.org/10.1016/j.neuint.2013.01.027\u003c/li\u003e\n \u003cli\u003eWolf, F. A.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e SCANPY: large-scale single-cell gene expression data analysis. \u003cem\u003eGenome Biology 2018 19:1\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e (2018\u0026ndash;02\u0026ndash;06). https://doi.org/10.1186/s13059-017-1382-0\u003c/li\u003e\n \u003cli\u003eWolock, S. L., Lopez, R. \u0026amp; Klein, A. M. Scrublet: Computational Identification of Cell Doublets in Single-Cell Transcriptomic Data. \u003cem\u003eCell Systems\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e (2019/04/24). https://doi.org/10.1016/j.cels.2018.11.005\u003c/li\u003e\n \u003cli\u003eKorsunsky, I.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Fast, sensitive and accurate integration of single-cell data with Harmony. \u003cem\u003eNature Methods 2019 16:12\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e (2019\u0026ndash;11\u0026ndash;18). https://doi.org/10.1038/s41592-019-0619-0\u003c/li\u003e\n \u003cli\u003eMcInnes, L., Healy, J. \u0026amp; Melville, J. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. (2018/02/09). https://doi.org/10.48550/arXiv.1802.03426\u003c/li\u003e\n \u003cli\u003eConde, C. D.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Cross-tissue immune cell analysis reveals tissue-specific features in humans. \u003cem\u003eScience\u003c/em\u003e \u003cstrong\u003e376\u003c/strong\u003e (2022\u0026ndash;05\u0026ndash;13). https://doi.org/10.1126/science.abl5197\u003c/li\u003e\n \u003cli\u003eVillani, A.-C.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Single-cell RNA-seq reveals new types of human blood dendritic cells, monocytes, and progenitors. \u003cem\u003eScience\u003c/em\u003e \u003cstrong\u003e356\u003c/strong\u003e (2017\u0026ndash;04\u0026ndash;21). https://doi.org/10.1126/science.aah4573\u003c/li\u003e\n \u003cli\u003eFriebel, E.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Single-Cell Mapping of Human Brain Cancer Reveals Tumor-Specific Instruction of Tissue-Invading Leukocytes. \u003cem\u003eCell\u003c/em\u003e \u003cstrong\u003e181\u003c/strong\u003e (2020/06/25). https://doi.org/10.1016/j.cell.2020.04.055\u003c/li\u003e\n \u003cli\u003eM\u0026uuml;ller, S.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Single-cell profiling of human gliomas reveals macrophage ontogeny as a basis for regional differences in macrophage activation in the tumor microenvironment. \u003cem\u003eGenome Biology 2017 18:1\u003c/em\u003e \u003cstrong\u003e18\u003c/strong\u003e (2017\u0026ndash;12\u0026ndash;20). https://doi.org/10.1186/s13059-017-1362-4\u003c/li\u003e\n \u003cli\u003eZheng, G. X. Y.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Massively parallel digital transcriptional profiling of single cells. \u003cem\u003eNature Communications 2017 8:1\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e (2017\u0026ndash;01\u0026ndash;16). https://doi.org/10.1038/ncomms14049\u003c/li\u003e\n \u003cli\u003eHodge, R. D.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Conserved cell types with divergent features in human versus mouse cortex. \u003cem\u003eNature 2019 573:7772\u003c/em\u003e \u003cstrong\u003e573\u003c/strong\u003e (2019\u0026ndash;08\u0026ndash;21). https://doi.org/10.1038/s41586-019-1506-7\u003c/li\u003e\n \u003cli\u003eHabib, N.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Disease-associated astrocytes in Alzheimer\u0026rsquo;s disease and aging. \u003cem\u003eNature Neuroscience 2020 23:6\u003c/em\u003e \u003cstrong\u003e23\u003c/strong\u003e (2020\u0026ndash;04\u0026ndash;27). https://doi.org/10.1038/s41593-020-0624-8\u003c/li\u003e\n \u003cli\u003eZhang, Y.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Purification and Characterization of Progenitor and Mature Human Astrocytes Reveals Transcriptional and Functional Differences with Mouse. \u003cem\u003eNeuron\u003c/em\u003e \u003cstrong\u003e89\u003c/strong\u003e (2016/01/06). https://doi.org/10.1016/j.neuron.2015.11.013\u003c/li\u003e\n \u003cli\u003eRichards, L. M.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Gradient of Developmental and Injury Response transcriptional states defines functional vulnerabilities underpinning glioblastoma heterogeneity. \u003cem\u003eNature Cancer 2021 2:2\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e (2021\u0026ndash;01\u0026ndash;04). https://doi.org/10.1038/s43018-020-00154-9\u003c/li\u003e\n \u003cli\u003evan den Brink, S. C.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Single-cell sequencing reveals dissociation-induced gene expression in tissue subpopulations. \u003cem\u003eNature Methods 2017 14:10\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e (2017\u0026ndash;09\u0026ndash;29). https://doi.org/10.1038/nmeth.4437\u003c/li\u003e\n \u003cli\u003eMuzellec, B., Teleńczuk, M., Cabeli, V. \u0026amp; Andreux, M. PyDESeq2: a python package for bulk RNA-seq differential expression analysis. \u003cem\u003eBioinformatics\u003c/em\u003e \u003cstrong\u003e39\u003c/strong\u003e (2023/09/02). https://doi.org/10.1093/bioinformatics/btad547\u003c/li\u003e\n \u003cli\u003eDimitrov, D.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Comparison of methods and resources for cell-cell communication inference from single-cell RNA-Seq data. \u003cem\u003eNature Communications 2022 13:1\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e (2022\u0026ndash;06\u0026ndash;09). https://doi.org/10.1038/s41467-022-30755-0\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-9590848/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9590848/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"In most cancers, the tissue removed at surgery forms the basis of our molecular understanding of the disease, yet it is the residual disease left behind that drives recurrence and determines patient outcomes. This disjunction is particularly consequential in glioblastoma, a diffusely infiltrative brain tumour in which recurrence is inevitable even following maximal resection. While single-cell profiling has revealed remarkable heterogeneity within resected glioblastoma, it remains unknown which malignant states, immune populations, and microenvironmental features characterize the occult infiltrative disease beyond the surgical margin that ultimately drives recurrence and treatment failure. Here, leveraging rare en bloc lobectomy specimens that preserve the complete continuum from tumour core to macroscopically normal brain, we profiled 472,089 single cells across 32 spatially matched biopsies from five glioblastoma patients, revealing that this residual disease constitutes a biologically distinct entity from both tumour and normal brain. Malignant cell composition shifts dramatically across the tissue axis: mesenchymal-like cells dominate the tumour bulk, while neural progenitor-like and oligodendrocyte precursor-like cells predominate in distal grey and white matter respectively, each mirroring the transcriptional identity of the dominant normal cell type in their zone and rendering them resistant to different and largely non-overlapping therapeutic approaches. The immune landscape undergoes a spatial inversion from blood-derived macrophage dominance at the tumour core to microglial dominance distally, with progressive regulatory T cell enrichment and near-complete loss of checkpoint target expression, providing a spatial context for the challenges that have been faced with immunotherapy in glioblastoma. Non-malignant cells throughout radiographically normal parenchyma exhibit transcriptional stress before any treatment, suggesting a therapeutic window for neuroprotection. These findings provide a framework for therapeutic strategies matched to the biology of residual disease, and illustrate a principle likely applicable across solid tumours: that meaningful progress against recurrence requires studying not what is removed, but what remains.","manuscriptTitle":"Single-cell atlas of glioblastoma across the tumour-to-brain axis reveals the distinct cellular landscape of infiltrative disease","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-09 00:51:52","doi":"10.21203/rs.3.rs-9590848/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"0e58530d-c99f-4e3d-89ec-9d5dedc5f097","owner":[],"postedDate":"May 9th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Reject before peer review","date":"2026-05-13T13:43:02+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-05-04T10:03:23+00:00","index":"","fulltext":""},{"type":"submitted","content":"Nature","date":"2026-05-02T05:29:27+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":67471817,"name":"Biological sciences/Cancer/CNS cancer"},{"id":67471818,"name":"Biological sciences/Cancer/Tumour heterogeneity"},{"id":67471819,"name":"Biological sciences/Cancer/Cancer microenvironment"},{"id":67471820,"name":"Biological sciences/Neuroscience/Diseases of the nervous system/Cancer in the nervous system"},{"id":67471821,"name":"Health sciences/Diseases/Cancer/CNS cancer"}],"tags":[],"updatedAt":"2026-05-13T13:47:42+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-09 00:51:52","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9590848","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9590848","identity":"rs-9590848","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}