Spatial analysis reveals the evolving organization of IDH-mutant glioma

Summary

Adult diffuse gliomas comprise malignant cell states interwoven with the non-malignant brain microenvironment. Here, we combine spatial transcriptomics and spatial proteomics of isocitrate dehydrogenase (IDH)-mutant gliomas to define organizational principles across histological grades. In low-grade tumors, spatial organization is shaped by underlying brain anatomy. We identify a functional white–grey matter junction that restricts cortical invasion and is associated with marked changes in tumor composition and cellular phenotypes. This junction is preferentially traversed by oligodendrocyte progenitor (OPC)-like malignant cells, suggesting a role in tumor expansion. In contrast, tumors with intermediate histological features are largely disorganized, with few recurring interactions between cancer cell states and microenvironmental cell types. In high-grade tumors, hypoxia-associated structure emerges, resembling IDH-wildtype glioblastoma. Together, these findings reveal two independent axes of spatial organization—from anatomy-driven structure in low-grade tumors to hypoxia-driven organization in high-grade tumors—and establish a framework linking tumor grade to recurrent spatial interactions.

Graphical Abstract

Hoefflin et al. define organizational principles of IDH-mutant gliomas using multimodal spatial profiling. They identify three spatial archetypes across histological grades and reveal that brain anatomy and hypoxia represent distinct spatial organizing forces, with the white–grey matter junction acting as a functional barrier to invasion.

Introduction

IDH-mutant (IDH-mut) gliomas frequently progress from relatively indolent lesions to recurrent, treatment-resistant tumors, underscoring the need to understand the cellular and microenvironmental features that shape their evolution¹. The two major types of IDH-mut glioma, astrocytoma (IDH-A) and oligodendroglioma (IDH-O), share similar malignant cellular states and inferred developmental hierarchies, despite distinct genetic hallmarks (ATRX loss vs. chromosome 1p/19q co-deletion)^2-5 and distinct composition of their tumor microenvironment (TME).

IDH-mut gliomas are further classified into WHO CNS5 grades, which correlate with tumor aggressiveness and prognosis^2,6. Yet, how malignant cell states and TME populations differ across these grades remains incompletely understood^7,8. Even less is known about how these cellular programs are spatially organized during progression, as spatial profiling of IDH-mutant glioma has so far focused primarily on the immune microenvironment^9,10.

Recent work in IDH-wildtype glioblastoma (IDH-wt GBM), which are by definition grade 4, has shown that hypoxia and necrosis are dominant forces shaping tumor architecture^11-13. In GBM, structured regions are formed by recurrent spatial cell clusters resulting in a layered organization around hypoxic and necrotic areas. In contrast, regions far away from hypoxia/necrosis appear disorganized with limited spatial cell state patterns. While necrosis is frequently observed in grade 4 IDH-A and less commonly in grade 3 IDH-O, IDH-mut gliomas lacking necrosis might be expected to exhibit limited spatial organization.

To interrogate the spatial architecture of IDH-mut gliomas, we performed multimodal spatial profiling of IDH-mut gliomas across grades 2-4 using 10x Visium spatial transcriptomics and CODEX spatial proteomics, and integrated these datasets with our previously published spatial GBM study¹¹. We reveal that the patterns and drivers of spatial organization shift across glioma grades, from brain anatomy-associated organization in lower-grade tumors to hypoxia/necrosis-associated architecture in high-grade disease. We further identify the white–grey matter junction (WGJ) as an unexpected anatomical barrier in lower-grade IDH-mut gliomas, marked by abrupt changes in TME composition and malignant cell states.

Results

Spatial proteomics and transcriptomics of IDH-mut glioma

We performed spatial proteomics and transcriptomics on 34 primary IDH-mut glioma samples from 28 patients. CODEX was applied to 30 samples, comprising approximately 2.6 million cells, and 10x Visium to 23 samples, comprising 63,839 spots (Fig. 1A, Table S1). In most cases (n = 20 samples), adjacent tissue sections were profiled by both technologies (Fig. 1B). The cohort included IDH-A (n = 23) and IDH-O (n = 10) across WHO grades 2–4 (Fig. 1B). For joint analysis across glioma types, we defined categorical grades based on WHO CNS5 histological features: low-grade for grade 2 gliomas, high-grade for tumors with necrosis and/or microvascular proliferation (MVP), and mid-grade for all other gliomas. IDH-WT GBM, by definition grade 4, was included as a separate category using data from Ravi et al.¹³, our previously published dataset¹¹, and seven newly profiled GBM samples analyzed by CODEX (Table S1).

Figure 1. Experimental design and metaprogram generation.

(A) Experimental overview and patient cohort. Fresh-frozen tissue sections from IDH-mut gliomas (n = 30) and IDH-WT GBM (n = 7) were profiled by CODEX using a 61-antibody panel. 10x Visium spatial transcriptomics was performed on adjacent sister sections (n = 24). This dataset was integrated with previously published spatial datasets^11,13. Schematic created with BioRender.com. (B) Summary of the patient cohort including standard neuropathological classification, WHO grade based on tumor, and WHO grade based on the profiled section. (C) Similarity matrix based on gene-set overlap (Jaccard index) across all-transcriptional programs derived from NMF and Leiden clustering of the combined IDH-WT GBM (n = 26) and IDH-mut (n = 24) datasets. Programs are annotated by sample, glioma type, categorical grade, and MP; cluster numbers correspond to those in (D). (D) Table of MPs showing representative genes for each numbered cluster in (C). Malignant MPs are prefixed with “c”. (E) Visium spatial maps displaying the spatial distribution of MPs stratified by non-malignant TME and malignant MPs.

Identifying glioma cell states and non-malignant cell types of the TME

To define recurrent glioma cell states and cell types, we applied Leiden clustering and nonnegative matrix factorization to each Visium sample and then clustered the resulting gene programs across samples¹¹. This generated metaprograms (MPs) from IDH-mut samples alone (IDH-mutMP; Fig. S1A,B, Table S2) and from the integrated IDH-mut/IDH-wt GBM cohort (GliomaMP; Fig. 1C,D). MPs with high gene-expression overlap and corresponding to the same cell state or cell type were further collapsed into generalized MPs (GeneralMP; Fig. S1C).

The GliomaMPs captured established malignant states defined by prior scRNA-seq of IDH-mut gliomas³, GBM¹⁴, and spatial profiling¹¹. Malignant MPs are prefixed with “c” (for “cancer”) throughout the manuscript and included undifferentiated (cUndiff), astrocyte-like (cAC), oligodendrocyte progenitor-like (cOPC), neuronal progenitor-like (cNPC), and canonical mesenchymal-like states (canMES), including astrocytic- and hypoxia-enriched subsets (cMESAst, cMESHyp, Fig. S1D). New malignant MPs that were not previously defined from spatial transcriptomics analysis of IDH-wt GBM and that were enriched in IDH-mut gliomas were related to oxidative phosphorylation (cOXPHOS), and to non-canonical mesenchymal programs (ncanMES) enriched with vascular or inflammatory genes (cMESVasc, cMESInf).

Non-malignant MPs reflected normal brain and TME compartments, including neurons, oligodendrocytes, blood–brain barrier-related vasculature (VascBBB), angiogenic vasculature (VascAng), and distinct myeloid programs, including inflammatory microglia (InfMac), tumor-associated macrophages (Mac), and scavenger macrophages (MacScav)¹² (Fig. S1E). Visium spots were annotated by their highest-scoring MP¹¹ (STAR Methods, Fig. 1E, Fig. S1F).

As an orthogonal single-cell-resolved approach, we profiled samples by CODEX using a 61-marker antibody panel (Fig. 1A, Table S3). This mapped malignant and TME cells (Fig. 2A-C, Fig. S2A), along with Ivy Glioblastoma Atlas Project (IvyGAP) regional classification (Fig. S2B). Cell identities were annotated through a combined supervised and unsupervised approach (STAR Methods). Within the malignant compartment, CODEX resolved cUndiff, cOPC, cAC, cMES, and cMES-Hyp states. While the IDH1 R132H antibody was incompatible with CODEX, malignant cells in IDH-A samples, but not in IDH-O samples, showed the expected ATRX loss, consistent with pathology reports and supporting the discrimination between malignant and non-malignant cells (Fig. 2A, Fig. S2C).

Figure 2. CODEX cell state annotation and compositional changes across glioma types and grades.

(A) Heatmap of relative Nimbus scores (Z-score) per protein and cell type/state. Each row represents the indicated cell type or state in one sample, with varying numbers of rows across categories due to absence of certain populations in some samples. Heatmaps are stratified by glioma type and highlight ATRX-loss (protein name in red) in malignant cell states of IDH-A. (B) CODEX image of a sample showing the indicated markers (left), with corresponding spatial maps of annotated non-malignant TME cells (middle) and malignant cell states (right). (C) Representative CODEX staining illustrating cell segmentation: CODEX images with indicated markers (top row). Nuclear segmentation masks expanded by 2 μm and colored by cell state (bottom row). (D) Bar plot showing relative abundance of Visium MPs per sample (n = 49), stratified by glioma type. Statistics are summarized in panel (E). (E) Bar plot showing mean Visium MP abundance per glioma type and per categorical grade. Differences in metaprogram abundance across tumor grades and between tumor types were assessed separately for each metaprogram using the Kruskal–Wallis test. p values were adjusted across metaprograms using the Benjamini–Hochberg FDR procedure, with adjusted p < 0.05 considered significant; adj p<0.05*, adj p<0.01** ; adj p<0.001***. (F) Heatmap showing significantly enriched cell types/states (black tiles) from CODEX and Visium data, grouped by categorical grade and glioma type. Grade- and type- specific enrichment of each metaprogram was assessed by comparing its abundance in samples of the indicated grade or type with its abundance in all other grades or types combined using a two-sided Welch’s t-test where p values were adjusted across metaprograms for each grade or type using the Benjamini–Hochberg false-discovery-rate procedure, with adjusted p < 0.05 considered significant.

Differences in tumor composition between glioma type and grade

All malignant MPs were detected across IDH-A, IDH-O, and IDH-WT GBM, but their relative abundances differed (Fig. 2D-F). IDH-A was enriched in cAC (p = 0.006), cOXPHOS (p = 0.0001), ncanMES (p = 0.04), and cUndiff (p = 0.0001), whereas GBM was enriched in canMES (p = 0.002). These observed differences were not driven by major imbalances in anatomical sampling (Fig. S2D). CODEX analysis within IvyGAP cellular tumor regions, further revealed distinct TME composition (Fig. 2F, Fig. S2E): IDH-A contained fewer neurons (0.5% vs. 2.5%, p = 0.001) and vascular cells (4.2% vs. 6.7%, p = 0.03), but more myeloid cells (14.7% vs. 4.7%, p = 0.005), than IDH-O. These patterns are consistent with prior transcriptomic studies³ and the preferential localization of IDH-O to cortical grey matter and IDH-A to subcortical white matter^2,15,16.

Compositional analysis across categorical grades revealed systematic changes in malignant and TME states (Fig. 2E,F). Low-grade gliomas were enriched in cAC (p = 0.0003) and cOXPHOS (p = 0.008). Intermediate-grade tumors showed increased cUndiff (p = 0.002), accompanied by expansion of InfMg (p = 0.005), Mac (p = 0.003), and VascBBB (p = 0.008). High-grade IDH-mut tumors and GBM were dominated by hypoxia- and necrosis-associated malignant and non-malignant states, including ncMES, VascAng, cMES, and MacScav. CODEX resolved more granular TME remodeling: in IDH-A, myeloid populations shifted from inflammatory microglia in low-grade tumors toward immunosuppressive macrophage states in high-grade tumors (p = 0.04; Fig. S2F), while the vascular compartment shifted from VascBBB to aberrant VascAng with increased ECM deposition (p = 0.04; Fig. 2F, Fig. S2G).

Together, these findings indicate a grade-associated transition from more differentiated malignant states and relatively physiological microglial and vascular compartments in low-grade IDH-mut glioma, through increasing de-differentiation and TME remodeling at intermediate grade, toward a GBM-like hypoxia-associated ecosystem in high-grade disease including immunosuppressive macrophages and aberrant vasculature.

IDH-mut gliomas show low state-specific clustering compared to GBM

We subsequently quantified glioma spatial organization using two complementary metrics: state-specific coherence (State-Coh) and consensus interactions. State-Coh measures whether spots assigned to a given MP form coherent patches (State-Coh-high) or remain spatially dispersed (State-Coh-low), and was calculated as previously described¹¹ (Methods, Fig. S3A).

IDH-A and IDH-O samples exhibited significantly lower State-Coh than GBM (p = 2.9×10⁻⁵ and p = 2.1×10⁻⁶, respectively; Fig. 3A, Fig. S3B), indicating that IDH-mut gliomas generally reflect a State-Coh-low architecture. Across glioma types, hypoxic samples showed higher State-Coh than non-hypoxic samples (p = 3×10⁻⁵; Fig. S3C), suggesting that hypoxia represents a conserved spatial organizing force, extending our previous findings in GBM¹¹.

Figure 3. Three spatial archetypes of IDH-mut glioma.

(A) Violin and boxplot of mean State-Coherence of Visium samples stratified by glioma type. Significance was calculated by Wilcoxon rank-sum test. Center line, median; box limits, first and third quartiles; whiskers, 1.5 × interquartile range. (B) Line plot showing the number of consensus interactions of Visium samples per glioma type. (C) Heatmap of MP pairs engaged in consensus interactions for each glioma type, color-coded by interaction category. (D) Proportion of interaction categories per glioma type. (E) Violin and boxplot of mean State-Coherence of Visium samples stratified by categorical grade. Significance and boxplot characteristics are as described in (A). (F) Line plot showing the number of consensus interactions of Visium samples per categorical grade. (G) Heatmap of MP pairs engaged in consensus interactions per categorical grade, color-coded by interaction category. (H) Proportion of interaction categories per categorical grade. (I) Model of the three spatial archetypes, defined by (i) State-Coherence, (ii) consensus interactions, and (iii) cell density. Representative Visium spatial maps are shown for three samples per category with annotations for: (i) MPs; (ii) Total organziation; consensus interactions for (iii) TME–TME; (iv) Cancer–TME; and (v) Cancer–Cancer.

We next quantified spatial associations between MP pairs across three scales of spatial resolution: spot-level co-occurrence (colocalization), spot adjacency, and regional co-enrichment within hexagonal windows (regional composition)¹¹ (STAR Methods, Fig. S3D, Data S1). Moreover, we defined consensus interactions as MP associations supported by more than one spatial metric and recurring across multiple samples (Methods). GBM was consensus interaction-rich (n = 20), followed by IDH-A (n = 12), whereas IDH-O was comparatively consensus interaction-poor (n = 6; Fig. 3B).

We next classified consensus interactions by partner identity: Cancer–Cancer, Cancer–TME, and TME–TME (Fig. 3C,D). GBM was enriched in recurrent Cancer–Cancer couplings (p = 0.024; Fig. S3G,H), whereas IDH-O was enriched in recurrent TME–TME interactions (p = 0.003) and IDH-A in recurrent Cancer–TME interactions (p = 0.0013; Methods), highlighting the prominent role of microenvironmental coupling in IDH-mut gliomas. To assess whether these differences were influenced by cell density, we correlated cell density with interaction-category frequencies. Cell density was not significantly associated with TME–TME interactions, but negatively correlated with Cancer–TME interactions (R = −0.41, p = 0.02) and positively correlated with Cancer–Cancer interactions (R = 0.58, p = 0.0005, Fig. S3I). Thus, higher cell density, as observed in high-grade IDH-mut gliomas and GBM, may favor malignant self-interactions, whereas lower-density tumors exhibit proportionally stronger coupling with the surrounding microenvironment.

Low-grade IDH-mut gliomas are consensus interaction-rich

We then examined how state-specific coherence and consensus interactions vary across tumor grades. Low-grade IDH-mut gliomas were State-Coh-low (Fig. 3E), yet consensus interaction-rich (n = 17; Fig. 3F), with most couplings involving Cancer–TME and TME–TME pairs (p = 0.01; Fig. 3G,H, Fig. S3J,K). This pattern indicates recurrent cell-state pairings, even in the absence of the State-Coh-high organization typical of GBM.

In contrast, mid-grade gliomas (n = 7) exhibited both low State-Coherence and few consensus interactions (Fig. 3F). They also showed a higher proportion of cUndiff within the malignant compartment (“Undiff score,” p = 0.009; Fig. S3E) and increased malignant state diversity relative to other grades (p = 0.05; Fig. S3F), features that may contribute to their reduced spatial organization. In high-grade IDH-mut gliomas and GBM, both State-Coh and consensus interactions increased significantly (State-Coh: p = 0.038 and p = 0.008, respectively; consensus interactions: n = 12 and n = 20; Fig. 3E,F), accompanied by a trend toward Cancer–Cancer couplings (not significant; Fig. 3G,H, Fig. S3J,K). Changes in consensus interaction number across grades were independently validated by high-resolution CODEX colocalization analyses (Fig. S3L,M) and were not explained by Visium cohort size differences (Fig. S3N).

To visualize these two organizational axes, we overlaid State-Coh-high regions and consensus interaction MP pairs onto Visium maps, further classified by interaction category (Fig. 3I, Data S2). Together, these spatial maps reveal three spatial archetypes: low-grade tumors with low State-Coherence but abundant consensus interactions, intermediate-grade tumors with both low State-Coherence and few consensus interactions, and high-grade tumors and GBM with both high State-Coherence and abundant consensus interactions.

The white-grey matter junction (WGJ) is a functional barrier for cell migration

The enrichment of Cancer–TME interactions in low-grade IDH-mut gliomas suggested that residual brain architecture may shape tumor organization. We therefore classified CODEX regions as grey matter, based on positive MAP2 and negative PLP1 staining, or white matter, based on the opposite pattern (Methods, Fig. S4A,B). This assigned 20% of the CODEX data from 25 tumors to grey matter and 40% from 25 tumors to white matter while 40% remained unassigned. The significant enrichment of the neuronal markers NeuN and SNAP25 in the grey vs. white matter provided orthogonal validation for our assignments (Fig. S4C).

We focused on IDH-A (n = 7) in which cellular tumor was primarily located in subcortical white matter (Fig. 4A, Fig. S4D-F). Four IDH-A tumors contained both white and grey matter regions demarcated by a sharp MAP2 and PLP1 transition, indicating preservation of the WGJ during glioma development (Fig. 4A, Fig. S4A,B). In all but one of these cases, the WGJ coincided with the transition from cellular to infiltrating tumor (Fig. S4F). In four additional IDH-A tumors, the cellular-to-infiltrating transition occurred within white matter, defining a “white–white matter junction” (WWJ) that we compared to WGJ (Fig. S4E,F).

Figure 4. Barrier function of the white–grey matter junction.

(A) Representative CODEX images of two samples showing the indicated markers and IvyGAP annotation. White arrows indicating the transition from the cellular tumor of white matter to infiltrating of grey matter (White-grey junction, top) and from the cellular tumor of white matter to infiltrating of white matter (White-white junction, bottom). (B) Line plot showing the relative changes of cell density and purity (malignant cell state fraction) in CODEX samples at the WGJ (top) and the WWJ (bottom). The black line represents the mean values of individual samples plotted in the background for 1mm bins. The red dashed line represents the mean of the two central bins with p values calculated using a paired t-test. Consensus was labeled TRUE if the directionality was identical in > 50% of the samples. (C) Area plots of the same samples showing relative changes in cell-state fractions, cell density, and tumor purity across the WGJ (top) and WWJ (bottom). (D) Representative CODEX images showing the indicated markers, highlighting cell-state composition differences at the WGJ (top) versus the WWJ (bottom). (E) Line plots showing relative changes in cell-type fractions across the WGJ (top two rows) and the WWJ (bottom two rows). Black lines represent mean values of individual samples (1-mm bins); the red dashed line indicates the mean of the two central bins with p values calculated using binomial tests on both sides of the junction, and directionality was required to be identical in > 50 % of samples. (F) Line plots showing relative changes in marker expression across the WGJ (top) and the WWJ (bottom). Black lines represent mean values of individual samples (1-mm bins); the red dashed line indicates the mean of the two central bins with p values calculated using two-sample Welch t-tests on both sides of the junction, with directionality consistent in > 50 % of samples. (G) Schematic summary illustrating differences in invasive behavior across the white–grey junction compared with the white–white junction.

To quantify cell distributions across these junctions, we annotated cells by distance to the junction (Methods, Fig. S4F). At the WGJ, cell density and malignant cell fraction (tumor purity) decreased sharply, in contrast to the more gradual, approximately linear decrease across the WWJ (Fig. 4B). This pattern suggests that the WGJ acts as a functional barrier to glioma infiltration, whereas the WWJ simply reflects the expected progressive decrease in cell density with increasing distance from the cellular tumor.

Changes in cell composition across the white-grey matter junction

Consistent with a barrier function, the WGJ was associated with abrupt and consistent shifts in cellular composition (Fig. 4C-E, Fig. S4F). Grey matter beyond the WGJ was enriched in neurons (0.8% vs. 28%, p < 0.001), astrocytes (0.1% vs. 4.4%, p < 0.001), and vascular cells (3.5% vs. 10.7%, p < 0.001), while macrophages were strongly depleted (23.1% vs. 1.2%, p < 0.001; Fig. 4E).

Among malignant states, cOPC cells were enriched after the WGJ (10.6% vs. 21.5%, p < 0.001), whereas cAC cells were excluded (AC1: 20.1% vs. 5.4%, p < 0.001, AC2: 15.2% vs. 0.8%, p < 0.001), suggesting that cOPC may selectively cross the WGJ barrier. No such effect was observed in GBM (n = 7, Fig. S5A,B), consistent with the tendency of normal brain structures to erode in high grade gliomas (Fig. S5C).

cOPCs beyond the WGJ showed reduced SOX2, loss of PDGFRA, and a trend toward increased SOX10, without changes in ATRX expression (Fig. 4F, Fig. S5D), consistent with a more differentiated Oligo-lineage phenotype after crossing. In the grey matter of these samples, we observed spatial clusters of tumor cells around neuronal cell bodies, consistent with neuronal satellitosis. Using neuropathology-guided H&E annotations, we identified 119 satellitosis clusters across the four WGJ samples (Fig. S5E,F). Most satellite cells were malignant (78.5%) and predominantly in the cOPC-state (49.4%, Fig. S5G), again showing a shift toward a more differentiated, Oligo-lineage phenotype (Fig. S5H).

In summary, we identified an unexpected selective barrier function of the WGJ, in which the few malignant cells that cross this boundary are enriched for cOPC-like states, frequently form neuronal satellitosis clusters, and adopt a more differentiated Oligo-lineage phenotype (Fig. 4G, Fig. S5I).

Discussion

We evaluated IDH-mut glioma organization across grades using two complementary metrics: State-Coh, which captures spatial clustering of individual cell types or states, and consensus interactions, which identify recurrent cell–cell couplings across spatial scales and samples. State-Coh-high implies the appearance of tumor regions (i.e. patches) that are each dominated by a single cell type/state, while a high number of consensus interactions (consensus interaction-rich) implies that specific pairs of cell types/states tend to be proximal and thereby have a larger impact on tumor organization.

In our previous work on GBM, State-Coh-high regions surrounding hypoxia and necrosis were also consensus interaction-rich, suggesting that the two metrics may be related¹¹. Here, low-grade IDH-mut gliomas decouple these dimensions: they are consensus interaction-rich despite being State-Coh-low. Together, the spatial metrics define three grade-associated spatial archetypes: (i) low-grade tumors are consensus interaction-rich and State-Coh-low; (ii) intermediate-grade tumors are consensus interaction-poor and State-Coh-low; and (iii) high-grade tumors are consensus interaction-rich and State-Coh-high. The high-grade archetype mirrors the layered, hypoxia-anchored organization of GBM, whereas the low-grade archetype appears to reflect organization driven by underlying brain anatomy. In contrast, the intermediate archetype appears to reflect the absence of strong influences from either hypoxia/necrosis or parenchymal brain structures, in which case cellular plasticity and perhaps cellular mobility may give rise to largely disorganized tumors.

Because IDH-mut gliomas typically progress in a stepwise manner, comparison across categorical grades can reveal organizational changes associated with progression. Early in the progression of IDH-mut gliomas, malignant cells expand within the intrinsically structured brain parenchyma. Consistent with this, low-grade IDH-mut gliomas show recurring short-range couplings between malignant and non-malignant states, including cOPC–Oligo, cOXPHOS–Neuron, and cMES–Myeloid interactions. Such local interactions occur without clustering of individual states, producing an organization that is State-Coh-low but consensus interaction-rich.

As tumors advance to intermediate grade, malignant cells increase proliferation and de-differentiation^3,8, while consensus interactions largely collapse. We interpret this as a transition in which physiological brain architecture is progressively disrupted, before hypoxia and necrosis impose a new organizing principle. This model is consistent with increasing white-matter tract disruption across IDH-mut glioma grades¹⁷ and experimental evidence that glioma cells induce axonal injury and Wallerian-like degeneration¹⁸.

In high-grade tumors, hypoxia and necrosis emerge as dominant features associated with global organization. State-Coh increases across malignant and TME compartments, and Cancer–Cancer consensus interactions tend to increase. The spatially coherent cell clusters are anchored by pseudopalisading necrosis, recapitulating the hypoxia-driven layers of GBM^2,11,12. Thus, IDH-mut gliomas appear to transition from brain anatomy-associated organization, through a relatively disorganized intermediate state, toward hypoxia/necrosis-driven organization in high-grade disease.

The WGJ emerged as an important anatomical region shaping both malignant and non-malignant compartments in lower-grade IDH-mut gliomas. Myeloid inflammation, a characteristic feature of glioma tissue^12,19,20, was largely confined to white matter, with few myeloid cells detected in grey matter. This may reflect limited recruitment or retention within this compartment, rather than active exclusion at the junction itself. Most malignant states showed a similar restriction, except cOPC-like cells, which were comparatively enriched among tumor cells crossing into grey matter.

This apparent restriction of infiltration at the WGJ was unexpected given the absence of a true anatomical barrier and given the diffuse growth of IDH-mut gliomas. The WGJ may act as a functional barrier to cortical invasion; alternatively, white matter may provide a more permissive niche, such that grey-matter infiltration requires additional adaptation. In either scenario, cOPC-like malignant cells appear relatively advantaged in traversing this interface. At the junction, cOPCs showed phenotypic shifts consistent with adaptation, including reduced SOX2 expression. Together with prior evidence that cOPC-like cells spatially couple to vasculature¹¹ or are associated with axonal invasion²¹, these findings raise the possibility that they exploit an OPC-like migratory program, potentially using perivascular and axonal routes, to infiltrate surrounding brain tissue²².

Several features of the WGJ may contribute to its barrier-like behavior. Grey-matter extracellular matrix is denser and enriched in sulfated proteoglycans, including neuroglycan C (CSPG5)²³. In parallel, tissue architecture shifts from aligned white-matter tracts to a more isotropic cortical organization, altering topography and mechanical context²⁴. These constraints may hinder infiltration by most malignant states, whereas perivascular migration could provide cOPC-like cells with a route across the junction²⁵.

In summary, we provide a comprehensive spatial description of IDH-mutant glioma and define distinct organizational principles across histological grades. More broadly, our quantitative framework provides a foundation for investigating how tissue architecture constrains tumor organization and can be applied across cancer types to study how spatial organization changes during disease progression and in response to therapy.

Limitations of the study

This study has several limitations. First, despite the overall cohort size, the number of tumors within individual molecular and histological subgroups remains modest, limiting power for subgroup analyses and genotype–phenotype associations. In particular, IDH-O tumors were underrepresented (n = 10 samples). We prioritized the analysis of large, intact tissue sections rather than tissue microarrays to preserve the spatial context required to investigate higher-order tumor organization. Second, the spatial resolution of spot-based Visium limits true single-cell discrimination. We partially mitigated this limitation through orthogonal validation with high-resolution CODEX imaging. Finally, our study is based on observational analyses of human tumor specimens. Thus, the mechanistic models proposed here are inferred from spatial associations rather than experimental perturbation.

STAR Methods

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Glioma samples for spatial was collected from patients undergoing primary glioma resection at the Hospital St. Gallen, Switzerland, Massachusetts General Hospital, Boston, MA, Brigham and Women’s Hospital, Boston, MA, and at University Clinic Freiburg, Freiburg im Breisgau, Germany. All specimens were obtained with written informed consent and in accordance with institutional ethical approvals (KEK-ZH-Nr. 2015-0163; University Hospital Zurich; IRB #10-417; Dana-Farber Cancer Institute; University Hospital Freiburg IRB# 24-1100-S1; IRB #1360-1 at the Weizmann Institute of Science). Clinical characteristics of the cohort are summarized in Table S1.

METHOD DETAILS

Visium sample preparation

Fresh tumor samples were embedded in OCT at the time of freezing (Scigen OCT Compound, #4586). RNA integrity from multiple cryosectioned regions was assessed by TapeStation (Tapestation RNA ScreenTape, Agilent) following extraction with the Zymo Quick RNA MicroPrep Kit (#ZR-R1051). Only samples with RIN >7 proceeded to spatial transcriptomics profiling. Frozen tissue blocks were cryosectioned at 10 μm and placed onto the capture areas of 10x Genomics Visium Gene Expression slides following the manufacturer’s instructions (Slide & Reagent Kit PN-1000184). Sections mounted on Visium slides were stored at −80 °C and processed within two weeks.

Visium H&E Staining and imaging

Tissue sections on Visium slides were fixed in methanol (Millipore Sigma #34860), followed by an aqueous eosin-based H&E protocol according to manufacturer’s instructions (10x Visium Methanol Fixation, H&E Staining, and Imaging Protocol CG000160). Brightfield images were acquired on a Leica DMIi8 inverted microscope equipped with a DFC310FX color camera using a 10×/0.25 NA dry objective. Images were stitched using Leica Application Suite X and further processed in Fiji (v2.3.1).

Visium cDNA synthesis and library generation and sequencing

After imaging, tissue permeabilization was performed directly on the slide to release mRNA. Optimal permeabilization (9 minutes) was established using a Visium tissue-optimization time course. cDNA synthesis and library generation were performed with the Visium Spatial Gene Expression Slide and Reagent Kit (10x Genomics). Reverse transcription was performed by a template switch oligo protocol to generate a second strand, and cDNA synthesis was carried out according to qPCR results (KAPA FAST SYBR qPCR master mix, KAPA Biosystems). Amplified cDNA was fragmented, end-repaired, adaptor-ligated, and size-selected using SPRIselect beads (Beckman Coulter) at each step. Library quality was assessed using the Agilent TapeStation and NEBNext Library Quant Kit for Illumina (New England Biolabs). Indexed libraries were diluted to 1.8 nM, pooled, denatured, and sequenced on an Illumina NovaSeq SP flow cell (100-cycle kit) with 1% PhiX spike-in. Run configuration: Read 1 = 28 cycles; Read 2 = 90 cycles; Index 1 = 10 cycles; Index 2 = 10 cycles.

PhenoCyler Fusion

The antibody panel was curated to identify major cancer cell states and non-malignant cell types of the TME. We added markers to our previously published panel⁹ to increase phenotypic granularity in the myeloid compartment and the extracellular matrix. A detailed summary of all antibodies can be found in Table S3. Every antibody has been validated with conventional immunofluorescence staining using adequate control tissues from human duodenum and glioma, of which adjacent Visium data was available. PhenoCycler Fusion 2.0 (formerly CODEX; Akoya Biosciences) runs have been performed following the manufacturer’s instructions. In short, fresh frozen sections were fixed with 4% PFA for 20 min and blocked with 4% BSA and 0.25% TritonX-100 for 30 min. Primary antibodies were incubated overnight at 4 °C and secondary antibodies incubated for 2h at room temperature. All antibodies (including pre-conjugated antibodies from Akoya) were additionally validated on the respective glioma control samples using multiple CODEX runs to adjust antibody concentration and exposure times to optimize signal to noise ratio. After the run, hematoxylin and eosin staining was performed after flow cell removal with Xylene.

Spatial transcriptomic analysis

Alignment, processing, and QC of Visium data

FASTQ alignment to GRCh38, UMI counting, and initial spot filtering were performed with Space Ranger (versions v1.2–v2.0, 10x Genomics). Spot-to-image alignment and spatial coordinate registration were carried out in Loupe Browser (v5.0.1).

Gene expression values were defined as $E_{i,j}={\log }_2(1+{\text{CPM}}_{i,j}∕10)$, in which counts per million ${(\text{CPM})}_{i,j}$ refers to ${10}^6\times {\text{UMI}}_{i,j}∕\text{sum}({\text{UMI}}_{1\dots n,j})$, for gene $i$ in sample $j$, with $n$ being the total number of analyzed genes. The average number of UMIs detected per spot was less than 100,000; thus, CPM values were divided by 10 to avoid inflating the differences between detected ($E_{i,j}>0$) and undetected ($E_{i,j}=0$) genes as previously described (Tirosh, 2016)

For each spot, the number of counts was used as a proxy for sample quality. Spots with fewer than 1000 counts and/or expressing more than 20% mitochondrial genes, another proxy for low quality, were filtered out. For each sample, the 7,000 most highly expressed genes were retained, and expression was centered per sample to define relative expression values by subtracting the average expression of each gene $i$ across all $k$ spots: ${\text{Er}}_{i,j}=E_{i,j}-\text{average}(E_{i,1\dots k})$, where $Er$ represents relative expression values.

Per sample clustering

Unsupervised clustering of spots was performed on each sample separately using two complementary approaches: Leiden clustering to capture discrete patterns of variation and NMF to capture continuous patterns of variation. For each sample, after PCA, Leiden clustering was performed on the SNN graph (Seurat v4.3.0). For Leiden clustering, cluster-specific gene programs were defined by differential expression analysis based on the top 50 most differentially expressed genes by the Wilcoxon Rank Sum Test with a p < 0.005.

For NMF, negative values were transformed to zero in each centered expression matrix. To minimize the influence of selecting an individual k parameter, we performed NMF over multiple k values (2-10), generating 54 NMF programs per sample. Each NMF program was defined by the top 50 genes based on NMF score.

Generating consensus metaprograms

Gene programs from all samples (Leiden and NMF) were clustered based on their Jaccard overlap to identify consensus metaprograms (MPs). MPs were first defined for IDH-mutant gliomas alone, and a second set was subsequently defined that included both IDH-mutant and IDH-wild-type glioblastomas (including datasets from^11,13) to enable cross-tumor-type comparisons.

Due to the high number of gene programs generated by NMF, programs were included in the consensus clustering only if they were (i) robust within a sample (i.e., 20/50 gene overlap between at least two programs across more than one value of k), (ii) non-redundant (i.e., for a group of similar programs, defined as at least 10% overlap with another program within a sample, only one is retained), and (iii) shared ≥10% gene overlap with programs from another sample. Clustering followed the procedures described in Gavish et al.²⁶ and Greenwald et al.¹¹.

Gene-set enrichment analysis

Visium-derived MPs generally map to a single cell type or cell state (see Greenwald et al. 2024 for validation of this approach). We assessed the enrichment of MP signatures with Gene Ontology terms (MSigDB modules H, C2, C5, C8) as well as published signatures for glioma cell states^{3,11,14,27-29}, and non-malignant brain cell types^30-36 by hypergeometric test (FDR adjusted p < 0.01 was considered significant).

Spot scoring and assignment of spots to MPs

Spot-wise MP scores were computed using the scalop R package (https://github.com/jlaffy/scalop). Given a set of genes (Gj) reflecting an MP corresponding to a cell state or cell type, we calculate for each spot i, a score, SCj(i), quantifying the relative expression of Gj in spot i, as the average relative expression (Er) of the genes in Gj, compared to the average relative expression of a control gene set $(Gj\underset{¯}{count}):SCj(i)=\text{average}[Er(Gj,i)]-\text{average}[Er(Gjcount,i)]$. The control gene-set is defined by first binning all analyzed genes into 20 bins of aggregate expression levels (Ea) and then, for each gene in the gene set Gj, randomly selecting 75 genes from the same expression bin. The control gene set has a comparable distribution of expression levels to that of Gj, and the control gene set is 100-fold larger, such that its average expression is analogous to averaging over 100 randomly selected gene sets of the same size as the considered gene set. Each spot was assigned to the MP with the highest score.

After quantifying malignant state composition per sample based on spot-based metaprogram annotations, we derived a per sample stemness score. For each sample, we calculated the proportion of each malignant state within the malignant compartment. The “stemness” score is defined as the proportion of “cUndiff”.

State specific coherence score

State-Coh was quantified independently for each metaprogram (MP) within each sample. For a given MP, we computed the mean number of immediately adjacent spots that shared the same MP assignment, averaged across all spots assigned to that MP. This value reflects the local self-association of the MP in space. To normalize this metric, we estimated a lower and upper bound on spatial coherence. The lower bound (expected minimum) was obtained by randomly permuting spot positions 100 times per sample and recalculating the mean same-MP neighbor count for each permutation; the shuffled results were then averaged to yield the expected value under spatial randomness. The upper bound (expected maximum) was defined analytically by assuming that all spots assigned to the MP were organized in a regular hexagon pattern, and computing the corresponding average number of same-MP neighbors under this configuration. The final spatial coherence score was computed by linearly scaling the observed value between these bounds:

$$\text{State-}\mathrm{Coh}=\frac{\text{observed}-{\text{expected}}_{\min }}{{\text{expected}}_{\max }-{\text{expected}}_{\min }}$$

This normalization yields a score that facilitates comparison of State-Coh across MPs and samples.

Spot colocalization

For each sample, we first performed scoring-based spot deconvolution by assigning metaprogram (MP) scores to all spatial spots as described in Spot scoring and assignment of spots to MPs. MP scores above 0.1 were interpreted as indicating MP presence, and deconvolution matrices were binarized such that each MP was encoded as present (1) or absent (0) per spot. For any given MP pair (a, b) within a sample, we calculated the number of spots in which both MPs were present (N), as well as the total number of spots containing MP a (Ta) or MP b (Tb). Colocalization was quantified as:

$$\text{colocalization}=\frac{N}{T_a+T_b}$$

This observed value was averaged across spots to obtain a per-sample colocalization score for each MP pair. To estimate the expected colocalization, we repeated the same calculation on 500 shuffled versions of the deconvolution matrix per sample. Statistical significance was assessed using Fisher’s exact test by comparing observed and shuffled co-occurrence frequencies. An effect size was defined for each MP pair as the ratio of expected to observed colocalization. MP pairs were considered significantly colocalized within a sample if the effect size was ≥1.3 and the associated p was ≤0.01. Robustness across samples was evaluated by computing the fraction of samples in which each MP pair met these criteria. For downstream analyses, we used both the mean colocalization score per MP pair and the proportion of samples exhibiting significant colocalization.

Spot adjacency

Spatial adjacency was calculated independently for each sample. For a given ordered MP pair (A->B), we defined the observed asymmetric adjacency as the fraction of immediate neighbors of MP A spots that were assigned to MP B, normalized by all immediate neighbors of MP A:

$$\frac{\#\text{immediate MP B neighbors of MP A}}{\#\text{immediate neighbors of MP A}}$$

To remove the spatial coherence effect of MP B on its adjacency, we normalized this value by the relative adjacency capacity of MP B. Specifically, we calculated the proportion of MP B spots that had at least one neighboring spot not assigned to MP B, divided by the total number of such non-self adjacency opportunities across all MPs except MP A. To assess statistical significance, we generated a null distribution by randomly permuting spot positions 10,000 times per sample and recalculating adjacency scores. The p value is determined by how many times the observed score was higher or lower than the shuffled scores (the smaller of the two).

Regional composition

Regional composition was calculated for each sample. For each radius (defined as the number of hexagons encircling a spot) ranging from 1 to 15, we defined a window around each spot. We then measured the abundance of all MPs within the window and computed correlations in abundance between MP pairs across all windows of the same size, excluding windows containing only one MP. This process was repeated over 500 shuffled permutations of the spot positions to generate a random distribution, from which we assessed the significance of the correlation using Fisher’s exact test.

Consensus interactions

To integrate spatial relationships across multiple scales, we rescaled the outputs of each spatial analysis (colocalization, adjacency, and regional composition) using min–max normalization, mapping all scores to a common range of −1 to 1. For each analysis, MP pairs with a mean scaled score greater than 0.35 across samples and that were statistically significant in at least 20% of samples within a given grade or tumor type were classified as strongly connected. An MP pair was defined as a consensus interaction if it satisfied the strong-connection criterion in at least two of the five spatial analyses (colocalization, adjacency, and regional composition at radii r = 5, 8, and 15), with the additional requirement that at least one measure reflected direct spatial coupling (colocalization or adjacency). MP pairs exhibiting consistent spatial avoidance were identified using the same criteria and classified as repelled interactions.

In order to validate that the overall low number of consensus interactions in mid-grade tumors is robust and to control for the differing number of samples per grade, we performed resampling and permutation controls for robustness using the proximity interaction statistics as computed from the spot adjacency measure per sample. To test whether differences in the number of recurrent interactions across grades could be explained by unequal group sizes, we down sampled each categorical grade to a common sample size (n=6) and repeated the procedure for six independent trials (random subsampling without replacement in each grade). For each down sampled group and trial, we computed the number of consistent enriched interactions as the number of metaprogram pairs with enrichment observed in at least three samples within the subsampled set and we computed the number of positive consensus-like interactions per trial. As a label-agnostic control, we generated six trials of random 4-way partitions. In each trial, we randomly selected six samples from each grade, permuted the pool samples, and split them into four groups of six samples each. For each group we computed proximity interactions and quantified the number of recurrent interactions. This analysis estimates the expected variability in recurrent interactions under random grouping when sample size and grade composition are controlled

Interaction category pairs

Consensus interactions and repulsions were further categorized by their interaction type (i.e., TME-TME; cancer-TME; and cancer-cancer) depending on the annotation of each MP as cancer (cOPC1, cOPC2, cAC, cUndiff, cNPC, cOXPHOS, cProlifMetab, cMESInf, cMESVasc, cMES, cMESAst, cMESHyp) or TME (Oligo, Neuron, Synaptic, Myeloid, Vasc). In order to compute the statistical significance of interaction category types within samples by grade and tumor type, we first quantified the relative abundance of each metaprogram pair participating in a consensus interaction across any grade or type (i.e., union of consensus interactions – global consensus pair) per sample. In order to quantify the relative abundance of each global consensus pair, we performed per-spot annotation of consensus interactions by checking whether the neighbor edges of each spot generated a consensus interaction with that spot. These per-spot annotations were stored both by metaprogram pair identity and by three category-level labels: Cancer-Cancer, Cancer-TME, and TME-TME. To test whether global consensus pair prevalence differs across groups and to generate type- and grade-specific consensus interaction annotations independent of our original definition of consensus interactions, we performed a Welch’s paired t-test on per-sample prevalence values by tumor grade and type. Grade comparisons included low/med/high, while type comparisons included IDH-A, IDH-O, and IDH-wt (GBM). We identified pairs whose prevalence differed significantly across grades or tumor types (p < 0.05). For pairs that were not significantly associated with a group, we retained them globally in all samples, and for pairs significantly associated with a given group, we retained them only for that group (collectively referred to as ‘allowed pairs’).

Allowed pairs were divided into interaction categories (Cancer-Cancer, Cancer-TME, TME-TME) using the pair-family annotation described above, and the relative abundance of each consensus interaction category among allowed pairs was calculated per sample. To test for differences in interaction composition across groups, we performed pair-wise Wilcoxon rank-sum tests on per-sample proportions (with Benjamini-Hochberg correction within each set of comparisons) and adjusted p < 0.05 were considered significant.

Spatial proteomics analysis

Segmentation of spatial proteomics data

Raw images generated by the CODEX run were background subtracted as per the automated experimental post-processing. Pyramidal images were then loaded into QuPath³⁷ and DAPI based segmentation with nuclear expansion by 2um was performed using CellPose³⁸. In addition, an independent CellPose segmentation strategy based on DAPI and Vimentin was applied for the segmentation of gemistocytes. Mean expression values per segmentation mask were exported.

Nimbus scoring of spatial proteomics data

Next, multiplex images and segmentation masks were transformed into appropriate formats to be inputted into the Nimbus workflow³⁹. Nimbus is a deep learning–based classifier trained on the Pan-Multiplex dataset, which scores marker positivity from multiplex imaging data. This approach avoids arbitrary intensity thresholds by leveraging pixel-level features to optimize positive/negative calls and is specifically suited to control for signal spillover of adjacent cells. We calculated Nimbus scores for every cell and every marker following the standard workflow.

Cell phenotyping of spatial proteomics data

Cell phenotyping was done in a two-step approach. First Nimbus scores of the most discriminatory markers were used to assign major cell types using a supervised and conservative gating strategy. Discriminatory markers were selected and ranked based on signal-to-noise ratio with an emphasis on nuclear markers as they are associated with less spillover artifact. We used the following gating strategy: Cells expressing OLIG2 > 0.4 were classified as cOPC. Oligo were defined by SOX10 > 0.6 in the absence of OLIG2 expression (< 0.4). Neurons were defined by NeuN > 0.6. T cells were subdivided into CD8⁺ (CD3 > 0.7 and CD8 > 0.6) and CD4⁺ (CD3 > 0.7 and CD4 > 0.6) subsets., astrocytes by SPARC > 0.6 or S100B > 0.5, and vascular cells by CD31 > 0.6 or HSPG2 > 0.5. Myeloid cells were identified by high expression of at least one of the following markers CD11c, CD163, CD206, or P2Y12 (> 0.6). Neutrophils and B cells were identified by CD45 > 0.7 co-expression with CD15 > 0.7 or CD19 > 0.7, respectively.

In a second step, we applied an unsupervised clustering workflow to the remaining unassigned cells. We used self-organizing maps (SOMs)⁴⁰ using mean nuclear intensity values of representative state-specific marker genes (e.g., SOX2, ATRX, GFAP, VIM, S100B, CA9, PDPN). The resulting protein–cell matrices were transposed and used as input for SOM training on a two-dimensional rectangular grid (typically 3×3 or 3×2 units). Training was performed using Euclidean distance as the similarity metric over 10 iterations with learning rates decreasing from 0.05 to 0.01. Each cell was mapped to its closest SOM unit, effectively defining proteomically coherent subclusters. Cluster-level mean expression values and differential expression profiles were computed relative to all other clusters to identify the top marker genes for each SOM unit. Heatmaps of mean and z-scored differential expression were generated to visualize the expression patterns underlying cluster distinctions. To further assess intercluster relationships and subtype hierarchies, we performed hierarchical clustering of average expression profiles using Ward’s linkage. The resulting SOM-derived subclusters were then interpreted based on dominant marker gene expression patterns, and manually annotated as distinct proteomic states (e.g., cAC1, cAC2, Oligo, cUndiff) or filtered out as low-expression.

In a third step we used the same unsupervised SOM-based clustering approach to subtype T cells, Myeloid and Vasc using Nimbus scores of cell type specific markers, resulting in the final cell subtype annotations (Fig. 2A).

Colocalization analysis

We constructed neighborhood matrices for each sample using the imcRtools package in R. Neighborhoods were defined by a radius of 27.5 μm (corresponding to the diameter of one Visium spot). For each cell, the neighborhood matrix contained the counts of all neighboring cell types, with rows representing individual cells and columns the cell-type composition of their local neighborhood. Interaction frequencies between every possible pair of cell types were then summarized by normalizing counts to the abundance of the reference cell type within a sample.

To assess significance, we performed label-shuffling in which cell-type identities were permuted 500 times. Null distributions of interaction counts were derived from the shuffled data, and observed values were converted into z-scores relative to the mean and standard deviation of the null. Empirical p values were calculated by comparing observed interaction strengths to the shuffled distributions, and significant cell–cell relationships were classified as either interaction or repulsion. Pairwise interaction profiles were then aggregated across samples and conditions to identify recurrent and condition-specific spatial relationships.

Junction analysis

White and grey matter classification

Tissue sections stained with the CODEX proteomic panel were segmented into 50 μm tiles in QuPath, exporting mean marker intensities per tile. Expression matrices were log2(x+1)-transformed, followed by median centering and scaling across samples. To remove empty or artifact regions, we applied an unsupervised clustering workflow using a Self-Organizing Map and hierarchical clustering based on normalized marker expressions. Marker intensities were then averaged into 50 μm tiles, retaining those containing ≥ 10 small tiles.

Bimodal thresholds for PLP1 and MAP2 expression were determined using Gaussian mixture models (2-component) with the threshold defined as the midpoint between the two component means. Input values were log2-transformed, then median-centered and scaled across all tiles. Each tile was classified as white matter (PLP1 > threshold AND MAP2 < threshold), grey matter (MAP2 > threshold AND PLP1 < threshold), or ambiguous (all remaining tiles). Across 30 IDH-mutant samples, white matter comprised 41.5%, grey matter 19.0%, and ambiguous tiles 39.5% of total tissue area. Classification was validated using both IvyGAP annotations and expression of additional neuronal markers (NeuN, SNAP25), which showed significantly higher expression in grey matter relative to white matter tiles (Wilcoxon rank-sum test).

Junction cell density and purity analysis

Within each distance bin, the bin area was estimated from the bounding box of cell centroids and a fixed bin thickness of 1mm: $\text{area}({\text{mm}}^2)=(\text{length}\_\mu \mathrm{m}\times \text{thickness})∕{10}^6$. Cell density was computed as counts per mm². Purity was defined as the fraction of malignant cells. Distances were binned into four 1-mm intervals spanning −2 to +2 mm (Bin1 to Bin4). Bin2–Bin3 was defined as mid-zone baseline. Density values were expressed as relative change to this mid-zone baseline (Δ% Cells per mm²), while purity values were expressed as absolute deviation from baseline (Δ Purity). Per-sample behavior was visualized by connecting the bin-wise estimates with straight lines. Sample-level summaries (mean ± s.e. across samples) were plotted at each bin and linked with a black LOESS curve. A vertical reference line at 0 mm and a dashed horizontal reference line equal to the mean of the central bins (Bin2 and Bin3) were added. Paired t-tests comparing Bin1 vs Bin4 were used to assess significance, complemented by a Consensus metric indicating whether >50% of samples changed in a consistent direction (Consensus: TRUE).

Junction cell type composition analysis

The same distance binning was applied. For each sample, cell-type fractions were calculated by dividing the number of cells of a given type by the total cells within that bin, and relative changes were visualized as a per sample line plot. Sample summaries were obtained by averaging cell-type fractions across samples within each bin (mean ± s.e.), and these bin-wise means were connected with a black LOESS curve. A dashed red horizontal reference equals the average of the central bins (Bin2 and Bin3) for each facet. For each sample and cell type we estimated a per-sample null rate from the central region as the fraction in Bin2+Bin3. We then tested Bin1 and Bin4 against this null using binomial tests on counts (successes = cell-type counts; trials = total cells in the bin). For each bin we recorded the direction (above/below null) and p value. If > 50% of evaluable samples agree in direction, we label significance by the fraction of samples with p < 0.05 and matching that majority direction: ≥75% = “**”, >50% = “*”, otherwise “NS”.

Junction marker expression analysis

We quantified relative expression of markers within a specific cell state (e.g. cOPC). For every sample and distance bin, relative marker expression scores were calculated and visualized as connected lines. Sample summaries were derived by averaging across samples within each bin (mean ± s.e.), and these bin-wise means were linked using a LOESS curve to display the overall trend. A per-marker “null” level was defined as the average of Bin2 and Bin3 means. For each marker, we formed paired per-sample comparisons between the extreme bins (Bin1 and Bin4). Let Δ = Bin4 – Bin1 be the within-sample change in relative Nimbus mean score. We report a two-sided paired t-test comparing Bin1 vs Bin4. Tests are computed only when ≥3 paired samples are available; otherwise, statistics are reported as NA. In addition, for each sample and marker we compared Bin1 and Bin4 to the per-marker null (Bin2 and Bin3) using two-sample Welch t-tests. For each bin we recorded direction (above/below the null) and p value. Marker-level “votes” summarize consistency across samples: if a majority direction exists, we label significance by the fraction of samples with p < 0.05 and matching that direction: ≥75% = “**”, > 50% = “*”, otherwise “NS”.

Neuronal satellitosis analysis

A board-certified neuropathologist annotated neuronal satellitosis clusters on H&E-stained sections from samples containing the WGJ. Annotations were transferred to matched same-slide CODEX images, and the composition of non-neuronal cell types within each cluster was quantified. Expression of cOPC-relevant markers, including SOX2, PDGFRA, CD24, EGFR, SOX9, SOX10, and OLIG2, was compared between satellite cOPCs and an equal number of randomly selected cOPCs from cellular tumor regions of the same samples.

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistical analyses were performed in R (version 4.3.1). Statistical tests are described in the corresponding Methods sections and figure legends. Depending on the analysis, statistical significance was assessed using Wilcoxon rank-sum tests, paired or unpaired Welch's t-tests, Fisher's exact tests, binomial tests, ore ANOVA. For spatial analyses, empirical null distributions were generated by permutation or label-shuffling approaches as described in the respective Methods sections. Multiple testing correction was performed using the Benjamini–Hochberg procedure where appropriate. Unless otherwise stated, all statistical tests were two-sided, and adjusted p values < 0.05 were considered statistically significant.

Supplementary Material

1

Data S1. Spatial relationship summaries

Data S2. Interaction spatial maps

Table S1. Patient and sample metadata

Table S2. Gene signatures for Visium-derived metaprograms

Table S3. CODEX panel and antibody specifications

KEY RESOURCE TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
See Table S3 for list of all primary antibodies used in CODEX panel
Biological samples
Human IDH-mutant glioma samples	Massachusetts General Hospital/Brigham and Women’s Hospital, Cantonal Hospital St. Gallen, Medical Centre–University of Freiburg	N/A
Chemicals, peptides, and recombinant proteins
OCT Compound	Scigen	Cat #4586
Aqueous Eosin Y solution	Millipore Sigma	Cat #HT110216
Mayer’s Hematoxylin	Millipore Sigma	Cat #MHS16
Bluing Buffer	Dako	Cat #CS70230-2
SPRI Select Reagent	Beckman Coulter	Cat #B23318
DMEM/F-12	ThermoFisher Scientific	Cat #11320033
MEM non-essential amino acids	ThermoFisher Scientific	Cat #11140050
Neurobasal medium	ThermoFisher Scientific	Cat # 21103049
Critical commercial assays
Quick RNA Microprep Kit	Zymo	Cat #ZR-R1051
Tapestation RNA Screen Tape	Agilent	Cat #5067-5576
Tapestation High Sensitivity D5000 Screen Tape	Agilent	Cat #5067-5592
Tapestation High Sensitivity D1000 Screen Tape	Agilent	Cat #5067-5584
Visium Spatial Tissue Optimization Kit	10x Genomics	Cat # PN-1000193
Visium Accessory Kit	10x Genomics	Cat #PN-1000194
Dual Index Kit TT Set A	10x Genomics	Cat #PN-1000215
Novaseq SP 100 Cycles Sequencing Kit	Illumina	Cat #20028401
Visium Spatial Gene Expression Slide and Reagent Kit	10x Genomics	Cat # PN-1000184
FAST SYBR qPCR Master Mix	KAPA Biosystems	Cat # KK4600
NEBNext Library Quanit Kit for Illumina	New England Biolabs	Cat # E7630L
Conjugation Kit	Akoya Biosciences	Cat # 7000009
Staining Kit	Akoya Biosciences	Cat # 7000008
CODEX buffer	Akoya Biosciences	Cat # 7000001
Nuclear stain	Akoya Biosciences	Cat # 7000003
Assay reagents	Akoya Biosciences	Cat # 7000002
Software and algorithms
10x Genomics SpaceRanger 1.0 and 1.1	10x Genomics	N/A
10x Genomics Loupe Browser 5.0.1	10x Genomics	N/A
Seurat 4.3.0	Hao et al.	https://github.com/cran/Seurat
Scalop	J. Laffy, Tirosh Lab	https://github.com/jlaffy/scalop
QuPath	Bankhead et al.	https://qupath.github.io/
CellPose	Stringer et al.	https://github.com/mouseland/cellpose
imcRtools	Eiling et al.	https://github.com/BodenmillerGroup/imcRtools
R 4.4.1		https://cran.r-project.org/bin/windows/base/old/4.4.1/
NMF	Gaujoux, Seoighe et al. 2010	https://cran.r-project.org/web/packages/NMF/index.html
MASS	Venables WN, Ripley BD. 2002	https://www.stats.ox.ac.uk/pub/MASS4/
inferCNA	Patel, Tirosh et al., 2014	https://github.com/jlaffy/infercna
MSigDB	Subramanian, Tamayo, et al., 2005	https://www.gsea-msigdb.org/gsea/msigdb
SpatialExperiment	Righelli et al., 2022	https://github.com/drighelli/SpatialExperiment
Fusion 1.0.5	Akoya Biosciences
Deposited Data
IDH-mutant glioma 10x Visium spatial transcriptomics processed data	This study	https://doi.org/10.5281/zenodo.18380571
IDH-mutant glioma CODEX imaging data	This study	https://doi.org/10.6019/S-BIAD2840
IDH-mutant glioma CODEX processed data	This study	https://doi.org/10.5281/zenodo.21335411
GBM 10x Visium spatial transcriptomics processed data	Greenwald et al.11	https://doi.org/10.5281/zenodo.8105466
GBM 10x Visium spatial transcriptomics raw data	Greenwald et al.11	GEO: GSE237183
GBM CODEX imaging and processed data	Greenwald et al.11	https://doi.org/10.5281/zenodo.8105466
Glioblastoma 10X Visium spatial transcriptomics	Ravi et al.13	https://doi.org/10.5061/dryad.h70rxwdmj
Glioblastoma single-cell RNA sequencing (Smart Seq2)	Neftel et al.14	GEO: GSE131928

Highlights:

• IDH-mutant gliomas exhibit three distinct spatial archetypes

• Brain anatomy organizes low-grade, while hypoxia organizes high-grade gliomas

• The white–grey matter junction acts as a functional barrier to invasion