Diffuse midline gliomas (DMG) driven by a lysine27-to-methionine (K27M) mutation in histone 3 (H3) are among the most lethal brain tumors1,2,3,4,5. Primarily identified in younger children (<10 years), the same oncohistone mutation is also recurrently observed in midline gliomas in adults6,7,8. In children, the spatiotemporal pattern of H3-K27M DMG incidence, peaking at 6–9 years of age in the brainstem pontine region, has shaped the hypothesis that the cell-intrinsic and -extrinsic context in which the K27M mutation occurs and elicits oncogenic transformation is developmental stage specific9. Indeed, previous studies have hinted at precursor cells in the pons10 and an early neurodevelopmental window11 as spatiotemporal correlates in K27M mutation-mediated gliomagenesis. Cell-intrinsically, the K27M mutation leads to broad epigenetic dysregulation and thus transformation of a developmentally restricted cell to a tumorigenic stem-like state12,13,14,15,16,17,18. The resulting active chromatin landscape reflects an early oligodendroglial lineage19,20. Single-cell RNA-sequencing (scRNA-seq) of pediatric, predominantly pontine H3-K27M tumors, further demonstrated that most glioma cells are stalled in a cancer stem cell-like oligodendrocyte precursor cell (OPC)-like state that is capable of self-renewal and tumor initiation21,22. In contrast, more differentiated noncycling glia-like cells were shown to have lost their tumorigenic capacity21. Together, this indicates OPC-like cells to be at the core of K27M mutation-mediated tumorigenesis, and hence, may present a strategic therapeutic target in pediatric pontine H3-K27M DMGs.

However, it remains incompletely understood whether H3-K27M DMGs of different midline locations—such as thalamus, pons or spinal cord—as well as different age groups and different morphological features at presentation, have similar cellular compositions. In particular, the more recently recognized group of adolescent (10–19 years) and adult (≥20 years) H3-K27M DMGs remains understudied. In addition to cell-intrinsic modes of dysregulation, mounting evidence indicates that microenvironmental factors critically contribute to glioma growth23,24,25,26,27,28, and it has been suggested that the developing brain provides a permissive environment that can be exploited for pediatric brain tumor growth29,30. However, the interplay between age- and region-specific tissue environments and the varying clinico-anatomical characteristics of H3-K27M DMGs, and its contribution to tumor pathology remain unexplored.

To address these questions, we have utilized single-cell multi-omics and spatial transcriptomic approaches to profile an extended cohort of H3-K27M DMGs encompassing a broad range of age groups and anatomical locations. We thereby identify how age- and location-dependent contexts underlie cell-intrinsic and -extrinsic features that together determine variation in glioma spatial and cellular architecture in light of the common K27M mutation.


Cohort of H3-K27M DMGs across age groups and locations

We conducted multi-omic profiling of 50 H3-K27M mutant patient tumors, selected only by criteria of the oncohistone mutation, spanning pontine (n = 27), thalamic (n = 20), lower brainstem (n = 1) and spinal (n = 2) locations (Fig. 1a,b and Supplementary Table 1). The median age was 12 (2.5–68) years, encompassing 36 pediatric (18 early childhood (0–9 years), 18 adolescent (10–19 years)) and 14 adult (20–68 years) tumors. Samples were obtained pre-treatment (n = 30) and post-treatment (n = 20) from 29 female and 21 male patients. We performed deep full-length Smart-seq2 fresh single-cell (n = 18) or frozen single-nucleus (n = 25) RNA-sequencing (scRNA-seq/snRNA-seq) of 43 tumors (Fig. 1a–c). We additionally analyzed the open chromatin profiles of eight tumors utilizing the single-cell/single-nucleus assay for transposase-accessible chromatin using sequencing (scATAC-seq/snATAC-seq), as well as the single-cell spatial transcriptomic architecture of 14 tumors by in situ sequencing (Fig. 1a,b).

Fig. 1: H3-K27M DMG cohort profiled by single-cell multi-omics.
figure 1

a, Schematic of the workflow. b, Clinico-molecular cohort characteristics. The upper legend bars depict the single-cell profiling method by scRNA-seq (n = 18)/snRNA-seq (n = 25), snATAC-seq (n = 8) and/or single-cell in situ sequencing (n = 14). The lower row specifies the method of genetic characterization. Most frequently detected and previously reported co-mutations are shown in the middle for 43 of 50 tumors profiled by whole or targeted exome sequencing. Clinico-anatomical characteristics are shown by the bottom legend bars. c, UMAP of all cells profiled by scRNA-seq/snRNA-seq. The color legend highlights malignant, types of nonmalignant cells detected based on clustering, copy number profiles and expression of canonical marker genes. For this visualization, scRNA-seq/snRNA-seq data were integrated by the Harmony algorithm, while downstream analyses were performed separately on scRNA-seq and snRNA-seq data to control for technical biases. d, Copy number alteration (CNA) profiles inferred from scRNA-seq/snRNA-seq data. Cells are ordered by their original tumors as rows and are clustered by their pattern of CNAs across chromosomal locations (columns). Representative fresh spike-in nonmalignant cells lacking CNAs are shown on top.

To identify other mutations, we performed whole or targeted exome sequencing in 43 of 50 tumors (Fig. 1b). Recurrent mutations in TP53, PDGFRA and PIK3CA were broadly observed across all clinico-anatomical groups stratified by age and location, while alterations in HIST1H3B and BRAF were only rarely detected in childhood tumors, which is in line with previous reports of H3-K27M DMGs1,5,8,31,32.

Overall, our cohort covers a representative clinico-molecular range of H3-K27M DMGs. Interestingly, we did not detect significant differences in co-mutational profiles between different groups, and next set out to investigate non-genetic features and heterogeneity of H3-K27M DMGs across different spatiotemporal contexts.

H3-K27M DMG cell composition across age and location

We aimed at delineating and comparing transcriptional heterogeneity within our cohort stratified by age and location (Fig. 1c,d and Extended Data Fig. 1a–d). Complementary approaches assessing inter- and intratumoral heterogeneity concordantly identified tumor cells differentially expressing actively cycling, OPC-like, ‘astrocyte-likeʼ (AC-like), ‘oligodendrocyte-likeʼ (OC-like) and ‘mesenchymal-likeʼ (MES-like) signatures (Fig. 2a–d, Extended Data Fig. 2a–g and Supplementary Table 2). OPC-like cells were further resolved into three subpopulations (OPC-like-1, OPC-like-2 and OPC-like-3) (Fig. 2c,d and Extended Data Fig. 2b,c). Interestingly, the MES-like signature, which has been described in glioblastoma (GBM)33,34, has not been identified in H3-K27M DMGs before, hinting at unique properties uncovered from previously understudied clinico-anatomical groups within our extended cohort.

Fig. 2: Intratumoral transcriptional heterogeneity of H3-K27M DMGs.
figure 2

a, UMAP of all fresh tumor cells, highlighting identified clusters. b, Marker genes (y axis) of identified fresh tumor cell clusters, grouped and annotated on the x axis. Dot sizes represent the percentage of cells expressing the gene in the given cluster, and the color scale shows scaled average relative expression. c, Heatmap representing the relative expression (color bar) of the top 30 marker genes (rows) for the tumor metaprograms identified by NMF across all fresh tumor cells (columns). d, Proportions (y axis) of fresh tumor-derived NMF metaprograms (color legend) in tumor cells for each fresh sample (x axis). e, Cell type-specific TF regulatory networks (regulon, x axis) derived by SCENIC, plotted against their normalized specificity score (y axis). f, Boxplots representing relative frequencies of metaprograms in all fresh and frozen tumors in adult (n = 10) versus pediatric (n = 23) age groups. The median is marked by the thick line within the boxplot, the first and third quartiles by the upper and lower limits, and the 1.5 times interquartile range by the whiskers. Three asterisks denote credible statistical changes as assessed by a Bayesian scCODA model with FDR < 0.05 and without multiple test corrections. g, Boxplots representing relative frequencies of metaprograms in all fresh and frozen tumors grouped by pontine (n = 19) versus thalamic (n = 14) locations. The median is marked by the thick line within the boxplot, the first and third quartiles by the upper and lower limits, and the 1.5 times interquartile range by the whiskers. Three asterisks denote credible statistical changes as assessed by a Bayesian scCODA model with FDR < 0.05 and without multiple test corrections. h, RNA in situ hybridization for MES-like (CD44) and macrophage (CD14) markers in two adult and two pediatric H3-K27M DMGs. Two to three slides were stained for each sample with 10–15 fields of view taken per slide. i, Two-dimensional representations of the OC-like versus AC-like (x axis) and OPC-like (y axis) scores for adult and pediatric H3-K27M DMGs, respectively.

OPC-like cells were ubiquitously present in all tumors independent of age or location (Fig. 2a,d and Extended Data Fig. 2e,f). Interestingly, even in this expanded cohort, we did not detect any neuronal lineage tumor cells, placing this in contrast to all other high-grade glioma types and isocitrate dehydrogenase (IDH)-mutant glioma33,35,36,37. To investigate whether this may be a phenomenon specific to the midline location, we single-cell profiled two location- and age-matched IDH-mutant midline gliomas (Supplementary Table 1), revealing that neuronal lineage programs are present within rare midline IDH-mutant tumors (Extended Data Fig. 2h). Hence, this comparison of primary gliomas of the same location and age groups, but different genotypes, supports a direct cell-intrinsic effect of the K27M mutation to skew tumor cells toward a glial/OPC-like instead of a neuron-like identity.

We next reconstructed networks of active transcription factors (TFs) and their downstream gene targets (gene regulatory networks (GRNs)) (Fig. 2e and Supplementary Table 3) by single-cell regulatory network inference and analysis (SCENIC)38. We indeed found key GRNs known from normal glial specification (for example, SOX10 in OPC-like cells, TFEB in OC-like cells, SOX9 in AC-like cells) to be likewise active in respective H3-K27M DMG tumor cell counterparts, highlighting parallels between normal developmental and glioma cell fate determination. Moreover, we identify GRNs (for example, GLI2 and NFATC4) that have not yet been implicated in normal development and may hence present glioma-specific regulatory aberrations.

We next compared cellular compositions across tumor locations and age groups (Fig. 2f,g and Extended Data Fig. 2i,j). Interestingly, the MES-like metaprogram was substantially enriched in adult tumors (Fig. 2f), which persisted when we controlled for location as a potential confounding factor (Extended Data Fig. 2i). This was validated by RNA in situ hybridization (Fig. 2h). Except for one NF1-mutated pediatric tumor (Fig. 1b), which was associated with a stronger MES-like signature as previously reported33,39, we did not detect any additional recurring genetic mutations in coding gene regions in tumors enriched for MES-like cells, suggesting either non-coding mutations and/or non-genetic determinants may underlie the observed age-specificity. As such, this age-related difference points toward the emerging role of the tumor microenvironment in shaping the MES-like signature, as has been illustrated in recent studies27,28,40.

Together, we demonstrate that H3-K27M DMGs are biased toward an OPC-like cell identity independent of age or midline location, which suggests cell-intrinsic effects of the K27M oncohistone mutation itself rather than environmental determinants to underlie this cellular state. Contrastingly, an association with age is observed for the MES-like signature (Fig. 2i), potentially linking this cellular state to cell-extrinsic/environmental drivers.

Location specificity of OPC-like subpopulations

We next examined the three OPC-like subpopulations uniquely detected in our extended scRNA-seq dataset, termed OPC-like-1, OPC-like-2 and OPC-like-3 (Figs. 2c,d and 3a). While all OPC-like subpopulations were defined by high expression of canonical OPC markers (for example, PDGFRA, SOX10 and OLIG1/2), these markers together with other known marker genes of committed OPCs (for example, CSPG4, GPR17 and EPN2) were most highly expressed by OPC-like-1 cells (Fig. 3a–c)41,42. In contrast, OPC-like-2 and −3 cells depicted higher expression of marker genes linked to more immature oligodendrocyte precursors of the developing brain, also termed pre-OPCs—a state of oligodendroglial lineage differentiation between less differentiated neural stem cell and more differentiated OPC (for example, ASCL1, HES6, BTG2, DLL1 and EGFR) (Fig. 3c)41,42,43. Additionally, OPC-like-2 cells highly expressed genes encoding ribosomal proteins (for example, RPL17 and RPS18), and OPC-like-3 cells exhibited higher expression of immediate early response genes (for example, JUNB and EGR1) (Fig. 3a), which have been previously described as markers of different normal (pre-)OPC subpopulations44,45. When we projected these OPC-like subpopulations onto scRNA-seq atlases of the human telencephalon and mouse cortex41,43,46, the OPC-like-1 subpopulation indeed mapped to committed/maturing OPCs, whereas OPC-like-2 and OPC-like-3 cells were more similar to pre-OPCs (Fig. 3d–f and Extended Data Fig. 3a–c). Comparison with cell populations from other glioma types and trajectory analyses (Extended Data Fig. 2c; 3d,e) also pointed toward a more immature state of OPC-like-2 and OPC-like-3 cells, and stronger lineage commitment of OPC-like-1 cells.

Fig. 3: Region-specific states of OPC-like cells.
figure 3

a, Heatmap representing the relative expression (color scale) of the top 30 marker genes (rows) for the different OPC metaprograms across all fresh tumor cells (columns). b, Violin plots depicting log normalized absolute expressions of canonical OPC marker genes in OPC-like-1, OPC-like-2 and OPC-like-3 subpopulations. Expressions in AC-like cells (orange) are shown for comparison. c, Heatmap representing the relative expression (color scale) of canonical pre-OPC and OPC marker genes (rows) in tumor OPC-like-3, OPC-like-2 and OPC-like-1 populations (columns). d, Projection of OPC-like-1, OPC-like-2 and OPC-like-3 populations (x axis) onto normal pre-OPC and OPC (y axis) from a scRNA-seq dataset of the human hippocampus46. Color scale presents expression scores of normal cell signatures in tumor cells, while dot sizes depict expression scores of tumor cell signatures in normal cells. e, Projection of OPC-like-1, OPC-like-2 and OPC-like-3 populations (x axis) onto normal pre-OPC, OPC and OAPC (HOPX+SPARCL1+ glial progenitor cell) (y axis) from a scRNA-seq dataset of the human developing cortex41. Color scale presents expression scores of normal cell signatures in tumor cells, while dot sizes depict expression scores of tumor cell signatures in normal cells. f, Projection of OPC-like-1, OPC-like-2 and OPC-like-3 populations (x axis) onto different normal OPCs of varying maturation stages (y axis) from a scRNA-seq dataset of the neonatal mouse cortex43. Color scale presents expression scores of normal cell signatures in tumor cells, while dot sizes depict expression scores of tumor cell signatures in normal cells. g, TF regulatory networks (regulon, x axis) derived by SCENIC for each tumor OPC-like subpopulation, plotted against their normalized specificity score (y axis). h, Dotplots representing the distribution (mean ±2 × s.e.m.) of the proportions of different OPC-like tumor states across all fresh tumors grouped by pontine (n = 11) and thalamic (n = 6) locations. Three asterisks denote credible statistical changes as assessed by a Bayesian scCODA model, with FDR < 0.05 and without multiple test corrections.

Analysis of OPC-like subpopulation-specific GRNs using our scRNA-seq dataset identified TFs such as SHOX2 and OTX2 to be most specifically active in OPC-like-1 cells (Fig. 3g and Supplementary Table 3). GRNs specific to OPC-like-2 cells included Notch signaling regulator HES6 as well as multiple patterning TFs of the HOX family, and GRN characteristics of OPC-like-3 cells were linked to the AP-1 TF family (Fig. 3g). Of note, HOX patterning TFs have been demonstrated to be expressed in mice embryonal pre-OPCs while being downregulated in postnatal OPCs44. Moreover, immediate early response regulators have been implicated as specific to human pre-OPCs compared to committed OPCs45, further hinting at a more immature and pre-OPC-like state of DMG OPC-like-2 and OPC-like-3 cells.

We next compared proportions of these OPC-like subpopulations across our spatiotemporally stratified cohort and observed a remarkable enrichment of pre-OPC-like (OPC-like-2 and OPC-like-3) cells in pontine compared to thalamic tumors. Conversely, OPC-like-1 cells were enriched in thalamic tumors (Fig. 3h). These differences remained when stratifying for age groups as potential confounders (Extended Data Fig. 2j).

Therefore, we identify tumor location as a contextual determinant of OPC-like states, with immature pre-OPC-like progenitors enriched in pontine, and more committed OPC-like cells enriched in thalamic tumors.

The open chromatin landscape of H3-K27M DMG cell populations

To resolve how H3-K27M DMG cellular heterogeneity is governed at the chromatin level, we probed single-nucleus accessible chromatin profiles by snATAC-seq of eight tumors complementing their single-cell transcriptomes. De novo annotation of malignant cell clusters proved largely concordant with scRNA-seq-derived cell populations and included an additional group of AC-like (AC-like-alternative) cells with increased gene activity scores for synaptic marker genes (for example, GABBR2, GRIA1 and CAMK2B) (Fig. 4a,b, Extended Data Fig. 4a–f and Supplementary Table 4; Supplementary Note). Cross-modality integration with scRNA-seq data further demonstrated overall congruence between chromatin- and transcriptome-defined cell states (Extended Data Fig. 4g). Notably, this also revealed distinct clusters of OPC-like-1, OPC-like-2 and OPC-like-3 cells in snATAC-seq space (Extended Data Fig. 4h). Concordant with our scRNA-seq findings, OPC-like-2 and OPC-like-3 cells also exhibited similarities with pre-OPCs at open chromatin level, whereas OPC-like-1 cells depicted higher chromatin accessibility for genes also described in healthy committed OPCs45 (Extended Data Fig. 4h–j). Thus, our finding of different OPC-like subpopulations is represented at both transcriptome and accessible chromatin levels.

Fig. 4: Characteristic chromatin profiles of H3-K27M DMG cell populations.
figure 4

a, UMAP of all snATAC-seq derived tumor nuclei after batch effect correction, highlighting de novo assigned clusters. b, Dotplot representation of top marker genes with differential gene activities (color scale) and proportion of nuclei accessible (dot size) within snATAC-seq derived cell states. c, Heatmap showing normalized chromatin accessibility and gene expressions of 13,632 substantially linked CRE-gene pairs (left rows, chromatin accessibility; right rows, linked gene expressions). Rows were clustered using hierarchical clustering. For visualization, 5,000 rows were randomly selected. d, Barplot representing distribution of numbers of linked CREs per gene. Red dashed line denotes the top 5% threshold of numbers of linked CREs that define GPC. e, Ranking of genes (x axis) by numbers of linked CREs (y axis) highlighting genes with top 20 linked CREs in color. Genes differentially expressed in a tumor cell state or identified as a cell state-specific TF regulon by SCENIC are colored according to the legend. f, Venn diagram representing overlap of GPCs with H3-K27M DMG super-enhancer associated genes, identified by Nagaraja et al.19. P value of a two-sided hypergeometric test is shown. g, Dotplot of integrative TF analysis representing the top cell state (columns)-specific TFs (rows). Average relative expression level assessed by scRNA-seq is depicted by dot size, and relative activity inferred by SCENIC analysis is presented by color scale. h,i, Integrative representation of gene loci of the h, OPC-like cell-specific SEZ6L gene, and i, AC-like cell-specific ITPKB gene. At the top, pseudobulk chromatin accessibility track plots are shown colored by cell type. In the middle row, bars depict the locus of putative CREs. In the bottom row, loops denote the correlation between chromatin accessibility of each peak and expression of its linked gene, representing putative CREs that are enriched for the OPC-like cell-specific SOX8 (h), AC-like cell-specific SOX9 (i), TF motifs, respectively.

As snATAC-seq resolves gene-distal and intragenic accessible chromatin regions containing potential cis-regulatory DNA elements (CREs) that underlie gene expression, we next inferred putative CREs integrating snATAC-seq and scRNA-seq modalities. By correlating snATAC-seq-derived accessible chromatin regions/peaks to scRNA-seq measured expression levels of their nearest associated gene (Fig. 4c)47,48, we identified 13,632 potential peak-gene links of CREs and their target genes (Supplementary Table 4 and Extended Data Fig. 4k). Among these, 287 genes exhibited more than eight (top 5%) linked CREs, denoting high regulatory locus complexity that has been described as ‘predictiveʼ chromatin and thereby a determinant of key lineage marker genes (Fig. 4d,e)47,48. We identified a higher number of genes linked with predictive chromatin (termed ‘GPCsʼ) specific to OPC/OC-like as compared to AC-like/MES-like cells, indicating highly cooperative regulation of the oligodendroglial lineage pervasively underlying H3-K27M DMGs (Fig. 4e; Methods). Because large groups of CREs are related to the concept of ‘super-enhancersʼ47,48, we overlayed our candidate GPCs with H3-K27ac ChIP-seq derived super-enhancer profiles of H3-K27M primary tumors19. This demonstrated a significant overlap of GPCs with H3-K27M DMG super-enhancer regulated genes (Fig. 4f), and further points toward a key role of these multimodally derived marker genes in orchestrating H3-K27M tumor cell identities.

We next sought to reconstruct and refine interdependent circuits of gene regulation by integrating expressions and activities of TFs inferred from scRNA-seq and enrichment of TF binding motifs in CREs derived from snATAC-seq (Methods). We identified 65 putative cell state-specific TFs that our analysis indicated to be (1) expressed at sufficient levels, (2) binding to characteristic motifs substantially enriched in CREs and (3) altering expressions of downstream target genes in a cell-type-specific manner (Fig. 4g). Moreover, we examined which TFs potentially regulate GPCs, focusing on TFs predicted to regulate expressions of GPCs and having binding sites detected within GPC-linked CREs. For example, the OPC-like marker gene SEZ6L is differentially expressed and accessible in OPC-like cells, and is linked to 16 CREs containing TF binding sites of SOX8, which is again predicted to be differentially active in OPC-like cells (Fig. 4h). We describe the same interdependencies between gene expression, chromatin accessibility and enrichment of cell state-specific TFs in CREs for GPCs of all tumor cell states, such as for AC-like marker gene ITPKB (Fig. 4i), which is linked to 11 CREs that harbor TF binding sites for SOX9, NFATC4 and RFX3, whose regulons are predicted to govern the expression of ITPKB. Together, our data further corroborate the closely interwoven and cell state-specific loops of chromatin regulation and gene expression identified at multiple levels.

In summary, we show that single-cell chromatin accessibility independently recapitulates the main cellular lineages identified in corresponding single-cell transcriptomes of H3-K27M DMG tumors. Our multimodal analysis reveals putative cell state-specific CREs as building blocks of larger GPC-associated regulatory complexes. These GPCs are enriched in OPC-like/OC-like cells, reinforcing the central role of the oligodendroglial lineage in H3-K27M DMGs. These results can be leveraged to more deeply investigate select key intrinsic regulators of H3-K27M DMG cell identities.

The age-specific myeloid cell landscape in H3-K27M DMGs

Various cellular and structural components constitute the glioma microenvironment and extrinsically influence glioma cell identities49,50. It remains to be elucidated whether these components are characteristic of their respective location or age-related brain environments. Here our age- and location-stratified H3-K27M glioma cohort uniquely lends itself to dissecting such context-specific differences largely independent of tumor subtype and genetic drivers. Because glioma- or tumor-associated myeloid cells (GAMs/TAMs) presented the largest proportion of nonmalignant cells within our scRNA-seq dataset (Fig. 1c), we focused on characterizing and comparing this microenvironmental component across our clinico-anatomical patient groups.

We classified TAMs into brain-resident microglia or monocyte-derived macrophages using reported sets of canonical marker genes35 (Fig. 5a–c). Overall TAM proportions were not different between adult and pediatric samples (Extended Data Fig. 5a). However, comparison of microglia versus macrophage proportions across age groups revealed a higher rate of microglia in pediatric DMGs, while adult DMGs contained higher rates of macrophages (Fig. 5d). Tumor location did not seem to influence these proportions (Extended Data Fig. 5b).

Fig. 5: The myeloid cell landscape of H3-K27M DMGs.
figure 5

a, UMAP of TAMs analyzed by scRNA-seq, color scaled by expression scores for microglia and macrophage gene sets. b, UMAP of TAMs colored by classification as macrophage or microglia cell type. c, Violin plot depicting log normalized expression levels of representative microglia and macrophage marker genes across TAMs scored as either microglia or macrophage. d, Dotplots representing the distribution (mean ± 2 × s.e.m.) of assigned macrophage versus microglia proportions across adult and pediatric tumors (N = 16 biologically independent samples). Three asterisks denote credible statistical changes determined by a Bayesian scCODA model with FDR < 0.05 and without multiple test corrections. e, Violin plots of log normalized expression levels of OSM gene in adult and pediatric TAMs. Three asterisks denote P = 0.003 (two-sided Kolmogorov–Smirnov test). Three asterisks in light green represent comparisons between adult and pediatric tumors for macrophages. f, Violin plots of log normalized expression levels of OSMR gene in adult and pediatric tumor cells. Three asterisks denote P = 0 (two-sided Kolmogorov–Smirnov test). g, Violin plots of log normalized expression levels of MES-like marker genes in adult and pediatric TAMs. P values from different comparisons are shown (two-sided Kolmogorov–Smirnov tests; black: within age-group comparisons between macrophages and microglia; light green: adult versus pediatric macrophages; dark green: adult versus pediatric microglia). h, Heatmap representation of scaled relative expressions (color scale) of MES-like state-associated ligands and marker genes (rows) in a single-cell atlas of normal mice microglia and brain myeloid cells across different age groups (E14.5, P7, P60)52 (columns).

Mounting evidence suggests a causal role of TAMs in establishing a mesenchymal cell state in GBM through TAM-secreted ligands binding to receptors on glioma cells, such as between ligand-receptor pair OSM-OSMR, or via chemokine signaling27,28,40,51. Given the significant enrichment of MES-like cells in adults compared to pediatric H3-K27M DMGs in our cohort, we hypothesized that this may be driven by differences in such tumor–immune interactions. We indeed detected higher expression of OSM in adult TAMs, and the corresponding receptor OSMR in adult tumor cells (Fig. 5e,f), indicating immune-mediated engagement of a previously validated pathway27 in inducing the MES-like phenotype in adult tumors. Moreover, we observed increased expression of MES-like marker genes in adults compared to pediatric TAMs, which were shown to be increased in mesenchymally enriched gliomas27 (Fig. 5g). To assess whether these transcriptional differences of MES-like state marker genes and inducing ligands may be inherent to normal brain myeloid cells during temporal development and aging, we analyzed gene expressions across age in a normal mouse brain myeloid cell atlas. Indeed, we observed an increase of ligands such as OSM and of mesenchymal marker genes with age (Fig. 5h)52, supporting that the increase of the H3-K27M DMG tumor MES-like state with age is linked to changes of the brain myeloid compartment that also occur during normal development and aging processes.

Last, we interrogated receptor–ligand interactions between TAMs and OPC-like subpopulations, revealing shared OPC-wide (for example, SEMA3E-PLXND1) and subpopulation-specific interactions (Extended Data Fig. 5c–f). This may point toward a harnessing of microenvironmental factors in reinforcing the OPC-like lineage and further determining their varying maturation, which provides the basis for follow-up investigations to better understand the contributions of cell-extrinsic regulators to the different OPC-like states.

In summary, we reveal that adult H3-K27M DMGs harbor higher proportions of monocyte-derived macrophages, while pediatric tumors are enriched for brain-resident microglia. We also show that H3-K27M DMG-associated TAMs upregulate ligands and marker genes that can induce tumor cell MES-like states with increasing age, thereby linking the age-specific tumor immune microenvironment to the observed increase of MES-like tumor cells in adult H3-K27M DMGs. This illustrates how age-related microenvironmental factors can differentially shape tumor cellular states.

Charting the single-cell spatial architecture of H3-K27M DMG

To map our scRNA-seq/snATAC-seq derived cell populations to their spatial positions within intact H3-K27M DMG tissues, we performed hybridization-based in-situ sequencing (HybISS)53 in 16 patient H3-K27M DMG tissue sections (14 different tumors, 2 tumors with multi-region sampling), using a panel of 116 cell-type-specific combinatorial marker genes curated from our scRNA-seq dataset (Fig. 1b, 6a–c, Extended Data Fig. 6a–e and Supplementary Table 5).

Fig. 6: The single-cell spatial transcriptomic architecture of H3-K27M DMGs.
figure 6

a, Schematic of HybISS experimental approach. Briefly, mRNA is amplified in situ by RT, and the product cDNA is hybridized with a custom complementary padlock probe. Next, RCA reaction is run to generate a blob of DNA that can then be barcoded with individualized gene bridge probes and fluorescently barcoded. After imaging, the sample is stripped of bridge probes, and the cycle is repeated five times with different fluorophores for decoding and identification of gene signals based on their decoding sequence. b, Representative image of malignant and nonmalignant cell type/state assignments in one primary human H3-K27M DMG section (UMPED65_A2; 1 experiment over the entire tumor section with N = 22,813 cells assigned), outlining the distribution of malignant and nonmalignant cell populations within the sample. c, Proportions (x axis) of scRNA-seq derived tumor cell states (color legend) identified by pciSeq across 16 human H3-K27M DMG samples (y axis). d, Violin plot representing the distribution of MES-like cell proportions in adult compared to pediatric H3-K27M DMGs (N = 7,004 MES-like cells across 16 biologically independent samples) profiled by spatial transcriptomics. Whiskers show minimum/maximum proportions. An asterisk denotes P = 0.024 (two-sided t-test). e, Heatmap representations of neighborhood enrichment analysis between malignant cell populations, identified at 50 μm, across all samples. The color scale denotes the probability of finding a cell when a second cell type is presently divided by the probability of finding the second cell type. f, Representative multiplexed IF CODEX images from three of four primary human H3-K27M DMGs, showing spatially distinct subpopulations of malignant (marker: H3-K27M) OPC-like (marker: PDGFRA), OC-like (marker: BCAS1), AC-like (marker: GFAP) and proliferating cells (marker: Ki67). For each tumor, one experiment was performed with ~70,000 to 1.2 million individual cells profiled per sample over the entire tumor section. g, Sample-wide scatter plot representing each cell population’s tendency to cluster with other cell populations (degree of centrality, y axis) or to cluster with themselves (clustering coefficient, x axis).

We analyzed spatial cell state compositions by probabilistic cell typing by in situ sequencing (pciSeq). Here we interestingly observed AC-like cells to constitute the major malignant cell compartment (Fig. 6c), which is in contrast to the predominance of OPC-like cancer cells observed by scRNA-seq. This held true across tumor sections of different sizes, cell densities and qualities. Our spatial analysis also identified larger numbers and diversity of nonmalignant cell types, that were either not detected or showed only low representation in scRNA-seq (Extended Data Fig. 6e). Because larger numbers of cells are assessed on average, and processing-associated biases are reduced in intact tissues, spatial transcriptomics is likely more representative of true cell state compositions than conventional scRNA-seq. Stratification within our spatially profiled cohort again revealed that adult H3-K27M DMG sections harbor substantially higher proportions of MES-like tumor cells relative to pediatric tumors (Fig. 6d), orthogonally underscoring the association of age with the MES-like state.

We next performed neighborhood enrichment analyses to investigate spatial relationships between individual cell populations. Here we observed marked variability in neighborhood structures, highlighting overall intertumoral spatial heterogeneity (Supplementary Fig. 3). Global analysis of malignant cell neighborhoods indicated higher colocalization of OPC-like/cycling and OC-like cells (Fig. 6e). We validated these findings on the protein level by multiplexed immunofluorescence (IF) imaging (codetection by indexing (CODEX) system) in four H3-K27M gliomas (Fig. 6f and Extended Data Fig. 6f–h). Concordantly, this approach indicated a preferred mitotic niche of proliferating OPC-like and OC-like cells, encircled by more differentiated, nonproliferating AC-like cells (Fig. 6f).

Neighborhood analysis between cancer and noncancer cells revealed closer proximities between vascular cells and MES-like tumor cells (Extended Data Fig. 6i), pointing toward increased vascularization that has been associated with the mesenchymal state54. Within a subset of samples (7 of 12 with >1,000 cells profiled), we also observe increased colocalization of microglia/macrophages with MES-like, OC-like and AC-like cancer cells (Supplementary Fig. 3).

Further, we assessed the tendencies of each cell population to either form their own homogeneous cluster, by calculating their clustering coefficient (that is, degree to which members of a cell population favor clustering together), or to cluster heterogeneously with other populations, as represented by their degree centrality (that is, ratio of nonmembers connected to members of a cell population). Here we observed that AC-like cells, nonmalignant astrocytes and TAMs depicted the highest tendency to cluster with other cell types/states, hinting at their more diffuse distribution rather than localization within a restricted spatial compartment. In comparison, vascular cells, neurons, and cycling OPC-like cells exhibited a higher tendency to cluster with members of the same cell population, which is further indicative of a propensity to form specific structures/niches (Fig. 6g).

In summary, we resolved the spatial architecture of scRNA-seq–defined H3-K27M DMG cell populations directly within the native tumor tissue. Our results shed light on global and heterogeneous cellular relationships and neighborhoods, notably suggesting the presence of mitotic stem-like niches in which H3-K27M tumor cells of oligodendroglial lineage (OPC-like and OC-like cancer cells) colocalize. These findings lend themselves to further investigation of potential therapeutic avenues directed at regional and temporal perturbation of H3-K27M DMG tumor cell populations and their associated niches.


We previously demonstrated the preponderance of OPC-like tumor cells in seven pediatric H3-K27M DMGs through scRNA-seq. However, it remained unknown whether the same cellular composition—proposed to arise as a function of early pontine development—holds true across multiple spatiotemporal environments in which these tumors occur. To address these questions, we generated a multi-omic single-cell atlas of H3-K27M DMGs, comprising various midline locations and ages ranging from 2 to 68 years. Our data shed light on understudied thalamic locations and adolescent/adult age groups and provide a blueprint for the spatiotemporal context-specificity of tumor cell-intrinsic properties, spatial tissue architectures, and microenvironmental interactions that co-orchestrate cellular identity against the backdrop of the shared K27M driver mutation.

Our study reveals a ubiquitous presence of OPC-like and more differentiated glia-like cells across all clinico-anatomical groups. Concomitantly, neuronal-like tumor cells are absent, which is independent of age and location and stands in contrast to other glioma types. Thus, this likely presents direct consequences of the K27M mutation universally skewing tumor cells toward an OPC-like and away from a neuronal-like state, decoupled from spatiotemporal influences.

We identify two major variable features as a function of regional or temporal context, respectively (Fig. 7):

Fig. 7: Schematic summary of the spatiotemporal context-specific composition of H3-K27M DMGs.
figure 7

Comparisons are between pediatric versus adult patient groups (x axis) and pontine versus thalamic midline locations (y axis) and a representative model image of tumor cell composition is depicted, respectively. All tumor groups are abundant in OPC-like cells and also harbor more differentiated AC-like, OC-like, MES-like and nonmalignant microenvironmental cells, but lack tumor cells of the NPC/neuronal lineage, as delineated by single-cell multi-omics (color legend). MES-like cells increase with age, as indicated by the green arrow, which is associated with age-related changes in the tumor immune microenvironment; in particular, higher proportions of microglia in pediatric tumors as opposed to increased proportions of macrophages in adult tumors. Location specificity exists for varying maturation stages of OPC-like cells—pontine tumors harbor less mature pre-OPC-like cells, while thalamic tumors are enriched for more mature lineage-committed OPC-like cells, either as a result of region-specific cell-intrinsic features or due to location-related diversification driven through interactions within the local environmental niche.

First, we resolve pontine H3-K27M DMGs to harbor more immature pre-OPC-like tumor cells than their thalamic counterparts. This raises the question of whether this diversity reflects region-specific cell-intrinsic features or it is driven by local environmental interactions. While normal murine OPCs have been shown to lack heterogeneity across different brain regions44,55, it is possible that region-specific microenvironments provide distinct cues to differentially foster OPC differentiation. This has been observed in the gray matter where OPC differentiation takes place more slowly compared to white matter56,57. In glioblastoma, the white matter has likewise been suggested as a prodifferentiative niche for oligodendroglial lineage stem-like cells58. It will be of interest to explore in future studies what extrinsic factors in the pons relative to the thalamus may contribute to preserving healthy and aberrant OPC(-like cell)s in a less committed pre-OPC(-like) state and how these specific microenvironmental contexts could be perturbed by targeting such factors.

The finding of a more immature precursor-like cell is accordant with previous modeling studies postulating embryonic neural stem/progenitor cells instead of OPCs as the H3-K27M DMG cell of origin11,59,60,61,62. While the K27M mutation could occur in such an earlier state, it subsequently induces a cellular arrest in a self-renewing OPC-like state59, and the hypothesized original cell of mutation may become diluted and eliminated from fully transformed tumors9. Taken together, the literature supports the idea that the cell state of transformation is an oligodendroglial lineage precursor, whose precise state may vary from pre-OPC to more mature OPC with different histone variants19, anatomical locations and ages.

Second, we observe the mesenchymal signature to increase with higher age, which we link to age-related differences in TAMs that have been illustrated to induce this myeloid-affiliated tumor signature27,28,40. As the mesenchymal state has been associated with a more aggressive phenotype in a broad range of solid tumors63,64, and mesenchymal- and myeloid-directed therapies are under active investigation, it will be of interest to investigate such an age and outcome association in H3-K27M gliomas and other tumors.

Lastly, we reconstructed the single-cell spatial architecture of patient H3-K27M tumors, identifying a niche of proliferating OPC-like/OC-like tumor cells, surrounded by AC-like cells, which constitute the major tumor cell population in situ. This finding contrasts the predominance of OPC-like cells observed by conventional and especially fresh scRNA-seq and may arise due to technical and biological reasons. As AC-like glioma cells have been shown to be interconnected through tumor microtubes19,25,65, we speculate that they may be less viable and more sensitive to tumor dissociation, thereby biasing toward capturing more aggressive OPC-like cells in scRNA-seq. By contrast, AC-like cells may be better preserved in frozen snRNA-seq and spatial approaches. Such a potential predominance of AC-like cells instead of OPC-like cells does not stand in contrast to the proposed role of OPC-like cells as the stem-like drivers of H3-K27M DMGs and would align with a more traditional model in which cancer stem cells present the minority of tumor cells66. With the emergence of spatial technologies, it will be relevant to assess whether similar differences are observed throughout other tumor types and biological systems, pinpointing the importance of multimodal profiling to further refine models derived primarily through the lens of a single modality.

Altogether, we provide an extensive resource of H3-K27M DMG cellular heterogeneity across space and time that lends itself to delineating the multi-faceted interplay between spatiotemporal context-specific cellular properties and microenvironmental niches for the design of rational modeling studies and therapeutic frameworks tailored to the different clinico-anatomical groups of this lethal glioma.