Resolving the lineage relationship between malignant cells and vascular cells in glioblastomas

Abstract

Glioblastoma multiforme (GBM), a highly malignant and heterogeneous brain tumor, contains various types of tumor and non-tumor cells. Whether GBM cells can trans-differentiate into non-neural cell types, including mural cells or endothelial cells (ECs), to support tumor growth and invasion remains controversial. Here we generated two genetic GBM models de novo in immunocompetent mouse brains, mimicking essential pathological and molecular features of human GBMs. Lineage-tracing and transplantation studies demonstrated that, although blood vessels in GBM brains underwent drastic remodeling, evidence of trans-differentiation of GBM cells into vascular cells was barely detected. Intriguingly, GBM cells could promiscuously express markers for mural cells during gliomagenesis. Furthermore, single-cell RNA sequencing showed that patterns of copy number variations (CNVs) of mural cells and ECs were distinct from those of GBM cells, indicating discrete origins of GBM cells and vascular components. Importantly, single-cell CNV analysis of human GBM specimens also suggested that GBM cells and vascular cells are likely separate lineages. Rather than expansion owing to trans-differentiation, vascular cell expanded by proliferation during tumorigenesis. Therefore, cross-lineage trans-differentiation of GBM cells is very unlikely to occur during gliomagenesis. Our findings advance understanding of cell lineage dynamics during gliomagenesis, and have implications for targeted treatment of GBMs.

Graphical abstract

Keywords

glioblastoma / mural cells / endothelial cells / trans-differentiation / lineage tracing / single-cell sequencing / copy number variation

Introduction

The developmental process of multicellular organisms is highly stereotypic, with three germ layers giving rise to distinct cellular progenies that constitute functional tissues and organs. The brain is mostly composed of neural cells that are progenies of the ectoderm, including neurons, astrocytes, ependymal cells, and oligodendrocytes (OLGs; Merkle and Alvarez-Buylla, 2006; Gao et al., 2014). In addition, cellular components of brain vasculature and microglial cells have distinct origins. Mural cells, vascular smooth muscle cells (vSMCs) and pericytes, originate from the neural crest lineage and the mesoderm (Korn et al., 2002; Kurz, 2009; Yamanishi et al., 2012; He and Soriano, 2015), and endothelial cells (ECs) and microglial cells are derivatives of the mesoderm and hematopoietic progenitors in the yolk sac, respectively (Ginhoux et al., 2010). Macrophages, also originating from the yolk sac, can infiltrate into brain parenchyma upon inflammation and gliomagenesis (Poon et al., 2017; Quail and Joyce, 2017).

Glioblastoma multiforme (GBM) are malignant primary brain tumors with properties of high invasiveness and abundant blood supply (Furnari et al., 2007). Deciphering the cellular composition and lineage history of the GBM micro-environment, including ECs, mural cells, and tumor-associated microglia and macrophages (TAMs), has important implications for developing targeted therapies for GBM. GBM cells are thought to originate from mutated neural stem/progenitor cells (Alcantara Llaguno et al., 2009; Liu et al., 2011; Lee et al., 2018; Alcantara Llaguno et al., 2019; Wang et al., 2020). Previous studies indicated that GBM stem cells might give rise to vascular ECs (Ricci-Vitiani et al., 2010; Wang et al., 2010; Soda et al., 2011; Hu et al., 2016; Mei et al., 2017) or mural cells (Scully et al., 2012; Cheng et al., 2013; Zhou et al., 2017) to support tumor growth and invasion. Our earlier studies also found that stem cell markers can label both glioma stem-like cells and mural cells (Zhong et al., 2016; Li et al., 2017), suggesting the possibility of cross-lineage trans-differentiation of glioblastoma cells. Moreover, intra-tumoral trans-differentiation in lung, prostate, brain, and gastric cancers could confer heterogeneity and novel properties on tumors, enabling tumors better adapt to growth and treatment (Murata-Kamiya et al., 2007; Han et al., 2014; Zou et al., 2017; Yao et al., 2020).

Recent advances of single-cell technologies, combined with lineage-tracing analyses, have deep implications for understanding cellular components, cell status and fate transitions during development, tissue homeostasis, regeneration, and pathological conditions (Guimaraes-Camboa et al., 2017; Kester and van Oudenaarden, 2018; Neftel et al., 2019; Bhaduri et al., 2020; Wang et al., 2021). Particularly, genomic signatures such as somatic single-nucleotide variants were recently applied for study lineage history of human cells during development (Bizzotto et al., 2021). Here we combined lineage-tracing approaches with single-cell sequencing to dissect lineage relationships between malignant cells and non-neural cells, especially ECs and mural cells, using both human GBM samples and mouse GBM models. First, genetic lineage-tracing and transplantation studies inferred that GBM cells only extremely rarely gave rise to vascular cells in vivo. Intriguingly, however, GBM cells could promiscuously express mural cell markers, including Tbx18, during GBM tumorigenesis. We further showed almost non-overlapping copy number variation (CNV) signatures between GBM cells and vascular cells in both mouse and human GBMs. Thus, contradictory to previous opinion, cross-lineage trans-differentiation of GBM cells is uncommon, occurring only in extremely rare events that are more the exception than the norm.

Results

Establishing genetic GBM models to trace fates of GBM cells

We first utilized genetic mouse models of GBM to investigate whether glioma cells could give rise to non-neural cells. Embryonic neural progenitor cells (NPCs) were transformed into GBM-initiating cells (GICs) by inactivating Trp53, Pten, and Nf1, the three most mutated tumor suppressors in human GBM (Kwon et al., 2008; Chen et al., 2015; Yu et al., 2020). This GBM model was named as NPC^TKO. Briefly, plasmids co-expressing Cas9, luciferase (Luc) and single-guide RNAs (sgRNAs) targeting Trp53, Pten, and Nf1 (Cancer Genome Atlas Research, 2008; Chen et al., 2012), along with a lineage-labeling cassette that stably expresses enhanced green fluorescent protein (EGFP) or Discosoma red fluorescent protein (DsRed) in progeny of GICs (Cheng et al., 2016), were in utero electroporated into neocortical neural precursors at embryonic day 14.5 (Fig. 1A). Notably, the genome-editing and lineage-labeling cassette was incorporated into embryonic NPCs via the piggyBac transposase system (Fig. S1A). Adult mice were screened for brain tumors by Luc-mediated imaging analysis. The tumor mass was extensively labeled with cytosolic EGFP or DsRed depending on the lineage-labeling cassette used. Intracranial tumors showed typical features of diffusely infiltrating high-grade gliomas: prominent tumor mass with high proliferation evidenced by pleomorphic and mitotic nuclei, widespread Ki67⁺ cells, indistinct tumor borders, pseudopalisading necrosis, and dilated vasculature (Fig. S1B–F). GBM sections were stained with markers for neurons (NeuN), OLGs (MBP), astrocytes (GFAP), oligodendrocyte precursor cells (OPC, OLIG2), and neural stem cells (SOX2). Around 50%–80% of GBM cells expressed OLIG2, and 10% expressed GFAP (Fig. S1G), indicating OPC features for most GBM cells, aligning with cellular properties observed in OPC-like GBMs (Liu et al., 2011; Wang et al., 2020; Yu et al., 2020). Moreover, approximately 60%–80% of tumor cells also expressed SOX2, suggesting their undifferentiated status (Fig. S1G). EGFP⁺ or DsRed⁺ tumor cells were sorted out from four samples for transcriptome analyses. Gene Set Enrichment Analysis (GSEA) was used to examine the relatedness of NPC^TKO GBMs to molecular signatures of human GBMs. Among the four NPC^TKO GBM cells examined, three of them displayed molecular features of mixed The Cancer Genome Atlas proneural (TCGA-PN) and mesenchymal (MES) GBMs, with one bearing predominantly the TCGA-PN feature (Verhaak et al., 2010; Wang et al., 2017; Fig. S1M). Therefore, genetic inactivation of three tumor-suppressor genes in embryonic neural precursor cells (NPC^TKO) leads to tumorigenesis of TCGA-MES and -PN GBMs in adult mice.

In parallel, we established another GBM genetic model by transforming adult neural precursors into GICs in situ (Fig. 1B). Lentiviruses co-expressing Trp53 short hairpin RNAs (shRNAs) and human constitutively active HRAS^G12V oncogene were injected into the ventricular-subventricular zone of adult C57BL/6 mice (NSC^HRas-shp53), which developed high-grade glioma with typical histopathological features of GBM within 3–4 weeks after lentiviral injection (Fig. S1I; Marumoto et al., 2009; Neftel et al., 2019). Single-cell RNA sequencing (scRNA-seq) analysis of mouse NSC^HRas-shp53 GBMs showed that each tumor contained three cellular states identified in human GBM, i.e. OPC-like, astrocyte-like (AC-like), and MES-like (Neftel et al., 2019). Interestingly, NSC^HRas-shp53 GBMs contained significantly fewer OLIG2-expressing cells than NPC^TKO GBMs (Fig. S1J–L). RNA-seq transcriptome analyses indicated all three NSC^HRas-shp53 GBMs molecularly recapitulate the TCGA-MES subtype of human GBMs (Fig. S1M). Moreover, like other genetic GBM models, both NPC^TKO and NSC^HRas-shp53 tumors displayed mixed transcriptome signatures of OLGs, astrocytes, and OPCs but not neurons (Fig. S1N; Cahoy et al., 2008); and were associated with the OLG lineage-associated GBM subtype (Type 2) (Fig. S1O; Alcantara Llaguno et al., 2015; Wang et al., 2020).

GBM tumorigenesis induces widespread alterations of brain vasculature

We next examined how the two types of GBMs affected vasculature structures in the brain. NPC^TKO and NSC^HRas-shp53 GBMs were generated de novo in Tbx18^CreERT2/+;Ai14 brains in which mural cells were genetically labeled upon tamoxifen administration. In line with previous studies, 99.8% of pericytes (CD146⁺PDGFRB⁺) and 99.6% of ACTA2-expressing vSMCs were co-labeled with Tbx18-driven tdTomato in wild-type Tbx18^CreERT2/+;Ai14 brains, reconfirming the reliability of Tbx18 in labeling mural cells (Fig. S2A). Consistently, mural cells of NPC^TKO and NSC^HRas-shp53 GBMs were extensively labeled with Tbx18-driven tdTomato and overlaid CD31-labeled ECs (Fig. 1G and 1H). In both models, the CD31 signals of ECs were enhanced, and many tumor vessels were dilated, stiffened, concentrated, or highly disorganized (Fig. 1C–H). To our surprise, in NPC^TKO GBMs there were fewer vessels and decreased vessel coverage in tumor regions relative to non-tumor sites, with vessels exhibiting glomeruli-like structures. In contrast, in NSC^HRas-shp53 GBMs, vessels were greatly dilated and interwoven into net-like structures, with increased vessel density and vessel area when compared to vessels in NPC^TKO GBMs (Fig. 1E–J). Since NPC^TKO GBMs contained significantly more OLIG2⁺ cells than NSC^HRas-shp53 GBMs (Fig. S1L), these observations are in line with a previous study showing that OLIG2-positive GBMs had almost normal vasculature, whereas OLIG2-negative GBMs promote angiogenesis and blood–brain barrier breakdown (Griveau et al., 2018).

Minimal trans-differentiation of GBM cells into mural cells

We next explored the possibility of trans-differentiation of GBM cells into mural cells. Pregnant Tbx18^H2B-GFP/+ mice were electroporated to allow us to examine fate choices of DsRed-labeled malignant cells in NPC^TKO GBMs. In line with previous studies, almost all mural cells, i.e., 99.4% of vSMCs and 99.7% of pericytes, were genetically labeled by Tbx18-driven nuclear GFP (H2B-GFP) in Tbx18^H2B-GFP/+ brains (Fig. S2A; Guimaraes-Camboa et al., 2017). If GICs could differentiate into mural cells, we expected to observe DsRed⁺/GFP⁺ double-positive cells, because most GBM cells stably expressed DsRed. Flow cytometry analyses showed that DsRed⁺ GBM cells and Tbx18::H2B-GFP⁺ mural cells were separate populations, with only very few double-labeled cells detected (Fig. 2A), the latter of which could be rare trans-differentiation events or misexpression of Tbx18 by some GBM cells.

In parallel studies, tumor sections were immunostained for Tbx18::H2B-GFP and smooth muscle actin alpha 2 (ACTA2), a marker for most vSMCs. Confocal imaging and 3D reconstruction demonstrated that, although many DsRed⁺ GBM cells were in close proximity to blood vessels, no overlap between GBM cells and the ACTA2 signal or Tbx18::H2B-GFP could be detected (Figs. 2B and S2B; Video S1).

Negligible differentiation of GBM stem-like cells into mural cells following intracranial grafting

Previous studies suggested that GICs with stem-like features could mostly recapitulate pathological and molecular features of primary GBMs upon intracranial transplantation. In addition, mouse GICs derived from the two GBM models could propagate tumors in both immunocompromised and immunocompetent environments, which facilitated additional lineage-tracing studies. Tumor cells derived from NPC^TKO(Tbx18^H2B-GFP/+) GBMs were in vitro cultured in serum-free conditions to propagate GICs (Ignatova et al., 2002; Pollard et al., 2009) followed by fluorescence-activated cell sorting (FACS) to enrich DsRed⁺ GICs. We next grafted DsRed⁺ GICs intracranially into nude and ICR mice to produce allograft tumors. No DsRed⁺/Tbx18::H2B-GFP⁺ or DsRed⁺/ACTA2⁺ double-positive cells were observed in sections of intracranial tumors by IF staining (Figs. 2E and S2C). By flow cytometry, although no Tbx18::H2B-GFP⁺ double-positive cells were detected in 9 nude brain grafts or in 11 outbred ICR grafts, two ICR grafts contained a very small portion of Tbx18::H2B-GFP cells: of 2.8 million cells analyzed, only 365 and 3523 Tbx18::H2B-GFP cells were detected, respectively, in the two ICR grafts (Fig. 2C and 2D). However, these Tbx18::H2B-GFP cells did not express ACTA2 and located away from ACTA2-labeled vessels (Fig. S2F). We then stained sections of these two ICR brains to address whether these Tbx18::H2B-GFP cells represented fully trans-differentiated mural cells or whether they were GBM cells that misexpressed Tbx18. IF data showed that ~98% of Tbx18::H2B-GFP cells co-expressed GBM markers SOX2 and/or OLIG2, and exhibited a clustered distribution consistent with clonal propagation, suggesting that Tbx18 was being misexpressed by a subset of GBM cells that retained expression of GBM markers (Fig. S2D and S2E).

In addition, GICs derived from an NSC^HRas-shp53(Tbx18^CreERT2/+;Ai14) GBM sample were implanted into brains of BALB/c nude and ICR mice. Tamoxifen was injected once per week for the first 2 weeks and three times in the third week (Fig. S2F). Therefore, cells that expressed Tbx18 would be labeled by tdTomato. Flow cytometry analyses did not detect tdTomato⁺ cells (progeny of Tbx18-expressing cells) in either nude or ICR GBM grafts (Fig. 2F and 2G). Moreover, IF staining showed that the few rarely detected tdTomato⁺ cells in these grafts were located distantly from CD31⁺ vessels and co-expressed SOX2 and/or OLIG2, indicating their GBM cell properties (Fig. S2G and S2H). Next, we transplanted human GBM stem-like cells labeled with ZsGreen into nude mouse brains (Li et al., 2017), but did not observe ZsGreen⁺ACTA2⁺ cells in all brain sections (Fig. S3). Altogether, these findings demonstrated that GICs were barely able to differentiate into mural cells in allografted or xenografted animals, although rarely a small percentage of GBM cells expressed markers for mural cells during gliomagenesis.

GBM cells failed to trans-differentiate into ECs

To examine the potential trans-differentiation of GBM cells into hematopoietic or endothelial lineages, we performed lineage-tracing studies by generating NPC^TKO and NSC^HRas-shp53 GBMs in Tek-Cre;Ai14 brains, in which 99.8% of ECs (Kisanuki et al., 2001) and 90.1% of TAMs (Gomez Perdiguero et al., 2015) were genetically labeled with cytosolic tdTomato (Fig. S4A). TAMs are derived from brain-resident microglia and infiltrated macrophages and express Iba1 (Poon et al., 2017; Quail and Joyce, 2017).

At postnatal days 60–150, GBM brains were minced and digested into single-cell suspensions for flow cytometry analyses. Akin to findings showed in Fig. 2A, EGFP⁺ GBM cells and tdTomato⁺ cells labeled by Tek-Cre were distinct populations, with only a small fraction of double-labeled cells detected (Fig. 3A), likely owing to misexpression of Tek by a few GBM cells during tumorigenesis. Interestingly, vessel walls were largely intact and did not show evidence of GBM cells invading or being incorporated into vascular structures, with numerous Iba1⁺ TAMs aggregated around tumor vessels (Fig. S4D). Many tumor vessels were fully or partially covered with ACTA2-expressing mural cells (Fig. 3B). Importantly, although many EGFP⁺ tumor cells of NPC^TKO GBMs were in close proximity to blood vessels, no overlap of EGFP with the tdTomato signal was observed, arguing against cross-lineage differentiation of EGFP-expressing GBM cells into tdTomato-labeled endothelial or microglia/macrophage lineages (Figs. 3B and S4E; Video S2).

Absence of EC differentiation by GBM stem-like cells following intracranial grafting

Next, EGFP⁺ GICs derived from three NPC^TKO(Tek-Cre;Ai14) GBMs were grafted into brains of immunocompromised BALB/c nude mice or immunocompetent ICR mice to produce allograft tumors. If GICs could differentiate into ECs, we expected to observe EGFP⁺/tdTomato⁺ double-positive cells. However, in flow cytometry experiments, we did not detect EGFP⁺/tdTomato⁺ cells from grafted tumors (Fig. 3C and 3D). Consistent with these results, only a few EGFP⁺/tdTomato⁺ cells were visualized in sections of intracranial tumors; however, these double-labeled cells lacked EC morphology and did not express CD31 (Figs. 3E and S4F).

To ask whether GBM cells derived from NSC^HRas-shp53 tumors could produce ECs, we also implanted GICs enriched from NSC^HRas-shp53(Tek-Cre;Ai14) GBMs into brains of BALB/c nude and ICR mice. In line with findings using NPC^TKO GICs, by flow cytometry we did not detect tdTomato⁺ cells (progeny of Tek-expressing cells) from grafted tumors derived from NSC^HRas-shp53 GICs (Fig. 3F and 3G). Rare tdTomato⁺ cells could be seen in grafted brains; however, these cells were located distantly from vessels and did not express CD31 (Fig. S4G), indicating misexpression of Tek rather than trans-differentiation. Thus, GICs did not give rise to ECs in vivo under either immunocompromised or immunocompetent conditions in our experimental settings.

scRNA-seq of GBM cells distinguishes tumor from non-tumor cell types

We next deployed scRNA-seq to characterize distinct cell types within GBM samples followed by analyses of lineage relationships between malignant cells and vascular cells. Two GBM specimens from the NPC^TKO model, one of each being generated in either Tek-Cre;Ai14 or Tbx18^CreERT2/+;Ai14 brains, were dissected and enzymatically digested into single-cell suspensions, then subjected to 10× Genomics scRNA-seq. A total of 25,667 single cells were successfully sequenced (Fig. 4A).

Uniform Manifold Approximation and Projection (UMAP) was employed for dimensionality reduction to cluster cells, which stratified cells into 18 principal cell clusters (Figs. 4B and S5A; Table S1). Well-established marker genes were then mapped onto the UMAP representation to designate cell types. Specifically, vascular cells (Pecam1, Cldn5, and Tek for ECs; Pdgfrb, Kcnj8, Acta2, and Tbx18 for mural cells) (Fig. S5B and S5C), TAMs (Cd14, Aif1, and Tmem119), OLGs (Klk6, Mag, Mbp, and Cldn11), and T lymphocytes (Cd2, Cd3d, Cd3e, and Cd3g) were respectively clustered into distinct groups. Notably, clusters of ECs and mural cells were either non-separable or close to each other, indicating their lineage kinship. Putative tumor cells (Fig. 4B and 4C) widely expressed neural stem cell (NSC) markers (Sox2, Nestin, and Olig2) and Pdpn, a marker indicating the mesenchymal-like state of GBM cells (Hara et al., 2021; Fig. S5E). As expected, expression of Luciferase and Egfp could be mostly detected in cells showing NSC features (Fig. S5D). However, interestingly, expression of markers for ECs or mural cells was not confined to vascular cell clusters, but could be also detected, although at comparably moderate levels, in other cell types. This was particularly true for Cspg4 (Ng2), a marker also labeling OPCs, which was highly expressed in GBM cells. Consistently, although expression of tdTomato was predominant in vascular cells and TAMs, it could be also detected in other cell clusters, demonstrating aberrant expression of vascular cell markers during GBM tumorigenesis (Fig. S5D). On the other hand, expression of NSC markers could be detected at relatively moderate levels in non-neural cells (Fig. S5E), echoing a recent scRNA-seq study of human GBMs (Bhaduri et al., 2020).

In addition, a substantial portion of tumor cells and non-tumor cells expressed Mki67, indicative of their proliferative status (Fig. S5E). In summary, analyses of scRNA-seq data of NPC^TKO GBMs supported our lineage-tracing analyses by showing that although vascular cells and TAMs clustered in groups clearly distinct from those of malignant cells, some GBM cells promiscuously expressed markers of other lineages, thus emphasizing the need to exercise extreme caution before assigning cell-type identity by using a few individual marker(s).

CNVs of vascular cells are distinct from those of GBM cells

Cells with common ancestors bear similar genetic variations such as CNVs, and recent studies have analyzed lineage relationships according to CNVs (Patel et al., 2014; Venteicher et al., 2017). To further explore the lineage relationship between tumor cells and non-tumor cells, we therefore utilized scRNA-seq data to compare CNV signatures of GBM cells with those of vascular cells. Our assumption was that if ECs and mural cells arise by trans-differentiation from GBM cells, they should share common CNVs with tumor cells. We performed CNV analyses in NPC^TKO GBM samples, utilizing CNVs of T cells as the reference (Fig. 4D and 4E). Using the inferCNV algorithm, CNVs were estimated by sorting the analyzed genes by their chromosomal location and applying a moving average to the relative expression values, with a sliding window of 100 genes within each chromosome. Hierarchical clustering grouped the rest single cells into multiple cell groups based on their CNV properties. Three groups of malignant cells (Group A/B/C) could be identified from each of the two NPC^TKO GBM specimens based on their prominent CNV characteristics (Tables S2 and S3), including genomic deletions and amplifications, which showed intra-tumoral and inter-tumoral heterogeneity. We found that CNV patterns of vascular cells were distinct from those of malignant cells. Cells recognized as malignant were then assigned into cellular clusters previously grouped according to their transcriptome signatures (Fig. 4B and 4C). Not surprisingly, clusters 0, 3, 4, 5, 6, 7, 8, and 11 of the two NPC^TKO GBMs contained significant portions of malignant cells. More than 90% of vascular cells from the two NPC^TKO GBM samples (cluster 17: 84/92 of NPC^TKO #1 and 80/81 of NPC^TKO #2) were designated as non-malignant cells (Fig. 4F and 4G). For the few malignant cells that were clustered into vascular cells, we reasoned that these cells might display some transcriptome features of vascular cells. Alternatively, some vascular cells could acquire a certain CNV characteristics of malignant cells due to clonal expansion (Zhou et al., 2020). In addition, the copyKAT (copynumber karyotyping of tumors) algorithm was applied to assign CNVs for individual cells (Gao et al., 2021), which also designated most vascular cells as diploid non-malignant cells: 85/92 for NPC^TKO #1 and 80/81 for NPC^TKO #2.

Next, we superimposed CNV signals and correlations between vascular cells and T cells on those of each batch of malignant cells. “CNV signal” reflects the overall extent of CNVs of each cell, defined as the mean of the squares of CNV values across the genome. “CNV correlation” refers to the correlation between the CNV profile of each cell and the average CNV profile of all malignant cells from the corresponding tumor cluster. Of the three clusters of malignant cells in the two NPC^TKO GBM samples, CNV properties for vascular cells and T cells were largely distinct from malignant cells, supporting distinct lineage origins for these cell types (Fig. S6A–L). Of note, CNVs of NSC^HRas-shp53 GBM cells were less drastic than those of NPC^TKO GBM cells, probably due to different onco-drivers and/or the shorter gliomagenesis period for NSC^HRas-shp53 GBMs (data not shown).

CNVs of vascular cells are distinct from those of malignant cells in human GBM samples

To investigate the possibility that ECs and mural cells of human GBMs could derive from malignant cells, we analyzed CNV features of human GBM samples. To this end, high-quality scRNA-seq data of two GBM samples—SF11247 and SF11285—generated by Bhaduri et al. was analyzed (Bhaduri et al., 2020). For sample SF11247, 12 398 cells were grouped into 18 clusters, with ECs and mural cells in close-related but distinct clusters (Fig. S7A, S7C, and S7E; Table S4). For sample SF11285, 11 433 cells were grouped into 14 clusters, with ECs and mural cells in the same cluster (Fig. S7B, S7D, and S7F; Table S5). Similar to findings with mouse GBM samples, expression of most vascular cell markers could be also detected in other cell types (Fig. S7E and S7F). CNV analyses were performed utilizing scRNA-seq data, designating CNVs of T cells, TAMs, and OLGs as the reference. We empirically split GBM cells of each sample into two groups according to their CNV properties (Fig. 5A and 5B). For example, Group A cells of SF11247 showed GBM hallmarks of chromosome 7 gain and chromosome 10 loss; whereas Group B cells of SF11247 displayed gain of chromosome 3 and 10, as well as loss of chromosome 13 (Fig. 5A). As to SF11285, Group A cells contained widespread amplifications and deletions; but Group B cells had prominent amplifications of chromosome 7, 8, 12, 18, and 19 (Fig. 5B). Consistent with data acquired from mouse genetic models, CNV patterns of vascular cells were largely distinct from those of tumor cells. Intriguingly, although some vascular cells shared a few common CNV signatures with malignant cells, including amplification of chromosome 16, CNVs unique to vascular cells such as the amplification of chromosome 13 could be detected in SF11285 (Fig. 5B), suggesting potential clonal expansion following this chromosomal abnormality, similar to findings in a recent report which found that some stromal cells, including immune cells, fibroblasts, and ECs, of human colorectal cancer carried genomic alterations with distinct features probably due to clonal expansion (Zhou et al., 2020).

We then calculated and compared CNV signals and CNV correlations with malignant cells (Neftel et al., 2019). In both tumors, similar to non-tumor TAMs, OLGs, and T cells, most vascular cells had lower CNV signals and were less correlated with GBM cells (Fig. 5C–F), thus forming distinct groups. Consistently, the copyKAT algorithm also assigned the majority of vascular cells as diploid non-malignant cells: for GBM SF11247, ECs—49/51, mural cells—63/68; for GBM SF11285, 126/129. Notably, a small ratio of T cells, TAMs, and OLGs were also assigned as aneuploid cells (Fig. 5G and 5H; Table S6). Thus, CNV properties of vascular cells were distinct from malignant cells in both mouse and human GBMs, strongly suggesting that vascular cells were derived from lineage(s) other than GBM cells.

ECs and mural cells propagate upon GBM tumorigenesis

Single-cell sequencing had demonstrated the presence of proliferating cells, as marked by expression of Ki67, in both tumor cells and vascular cells (Figs S5D, S7E, and S7F). To further investigate cell types that were proliferating, tumor sections of NPC^TKO and NSC^HRas-shp53 brains from genetically labeled animals were co-immunostained with Ki67. Results demonstrated that in NPC^TKO GBM brains, ~7% of tdTomato⁺ cells in Tek-Cre;Ai14 mice, and ~7% of Tbx18::H2B-GFP⁺ cells were Ki67⁺ proliferative cells, whereas matched brain regions on the contralateral side had no marker⁺/Ki67⁺ cells (Fig. 6A–D). NSC^HRas-shp53 GBM brains contained ~9% of tdTomato⁺Ki67⁺ cells in both Tek-Cre;Ai14 mice and Tbx18^CreERT2/+;Ai14 mice (Fig. 6B, 6E, and 6F). Next, resident mural cells and their progeny were labeled by injecting tamoxifen into Tbx18^CreERT2/+;Ai14 mice at P31–P35, when gliomagenesis of NPC^TKO GBMs is at an early stage. Brain tumors were acquired from 2- to 5-month-old mice for immunofluorescence analyses, with BrdU administrated for four consecutive days prior to sacrifice (Fig. 6G). Results showed that ~10% of genetically labeled ECs and mural cells in NPC^TKO GBM brains had incorporated BrdU (Figs. 6H, S8A and S8C). Similar results were seen in NSC^HRas-shp53 GBM brains (Fig. S8D and S8E). Therefore, resident mural cells and ECs were activated to proliferate upon gliomagenesis. In concert with these findings, we found ~7.5% of CD31⁺ ECs and ~5% of ACTA2⁺ vSMCs were proliferative in human GBM specimens, but not in non-tumor sites (Fig. S8F–I). Finally, all ACTA2⁺ mural cells in GBM mouse brains were progeny of Tbx18-CreERT2;Ai14 cells in NPC^TKO GBM brains (Fig. 6I). Altogether, our findings indicate that amplification of major vascular cell types in GBM tissue is a result of self-propagation rather than their being derived by trans-differentiation of GBM cells.

Discussion

In the current study, we attempted to resolve the lineage relationship between GBM cells and non-neural cells including ECs, mural cells, and TAMs by combining genetic lineage-tracing studies with CNV comparisons. The human and mouse GBM samples analyzed in current study showed that ECs and mural cells likely belong to lineages separate from GBM cells, and that reprogramming of GBM cells into vascular cells is very unlikely, echoing a previous study indicating absence of trans-differentiation of hEGFRvIII⁺ GBM cells into ECs (Griveau et al., 2018).

Two genetic GBM models were generated de novo in immunocompetent mouse brains, which recapitulate key pathological and molecular features of human GBMs. We then applied lineage-tracing studies of NPC^TKO and NSC^HRas-shp53 GBM cells in three genetic models that, respectively, labeled ECs and TAMs, or mural cells, which included de novo tumorigenesis and intracranial transplantation of GBM stem-like cells in both immunocompromised and immunocompetent mice. Minimal co-labeled GBM cells and ECs/hematopoietic cells were observed either by flow cytometry or by extensive examination of tissue sections. Intriguingly, a few Tbx18-expressing NPC^TKO GBM cells were detected in a few primary and implanted tumors, which might support trans-differentiation of GBM cells into mural cells. However, in the two transplanted NPC^TKO GBMs where some malignant cells co-expressed Tbx18-driving GFP, these cells mostly aggregated as clones and expressed SOX2 and/or OLIG2, but did not express ACTA2, indicating their GBM cell identity. We did not observe marker-misexpressing cells that showed mural cell morphology and/or had incorporated into tumor vasculature, suggesting that these marker-misexpressing cells were unlikely to be at a transitional state between GBM cells and mural cells.

We further compared CNV patterns between malignant cells and vascular cells in NPC^TKO GBMs. Clusters of vascular cells (ECs and mural cells) were well separated from malignant cells according to their transcriptome characteristics. Importantly, vascular cells in NPC^TKO GBMs exhibit CNV features distinct from malignant cells, arguing against trans-differentiation of GBM cells into ECs and mural cells. Further analyses of single-cell sequencing data derived from two human GBM samples also showed that CNV characteristics of ECs and mural cells were quite different from those of malignant GBM cells, supporting a non-GBM cell-of-origin for vascular cells.

Single-cell transcriptome studies also suggested that many markers that have been used to label mural cells in past studies, including Acta2, Pdgfrb, Tbx18, and Cspg4, could be expressed by cells beyond vascular cell clusters. The promiscuous expression of Tbx18 during gliomagenesis was further supported by dispersed tdTomato expression in Tbx18^CreERT2;Ai14 GBMs. Therefore, misexpression of mural cell markers by some GBM cells, probably due to disorganized epigenomes and signaling cascades, could lead to the conclusion that the cross-lineage trans-differentiation of GBM cells might occur. However, these “misexpressing” cells do not meet transcriptome and functional criteria for ECs and mural cells, as they also expressed markers of NSCs without displaying typical morphology of vascular cells. Interestingly, some vascular cells carried common CNV signatures, indicating their clonal expansion during gliomagenesis. A recent study found that significantly more stromal cells, including fibroblasts, ECs, and immune cells, of colorectal cancer carried CNVs than those in adjacent normal tissues (Zhou et al., 2020). We also showed that a significant portion of vascular cells were actively cycling in GBM tissues but not in non-tumor tissues. Therefore, tumorigenesis could promote proliferation of stromal cells, and stromal cells that bear CNV features might acquire clonal advantage over those without CNVs. Future studies could apply extensive bioinformatic analyses including CellPhoneDB followed by experimental validations to dissect how GBM cells and vascular cells communicate with each other (Efremova et al., 2020).

Our studies further propose potential explanations as to why previous studies have suggested cross-lineage trans-differentiation of GBM cells. Human GBMs might contain immunocompetent environment analogous to our tumorigenesis and allograft models, which could facilitate propagation of “marker-misexpressing” cells. Additionally, previous studies have largely relied on immunofluorescence staining using molecular markers for mural cells including NG2 and PDGFRB, which are also expressed in both neural and non-neural cell types and might lead to false conclusions (Ishii et al., 2008; Cheng et al., 2013; Dimou and Gallo, 2015). Finally, many GBM cells are tightly associated with blood vessels; thus, GBM cells that misexpress a subset of vascular lineage markers could be mistakenly identified as trans-differentiated cells.

Forced expression of so-called master transcriptional regulators (MRs) along with proper micro-environmental cues is required for trans-differentiation events to occur. For example, short-term ETV2 expression and TGFβ inhibition with constitutive ERG1/FLI1 co-expression reprograms mature amniotic cells into vascular ECs with clinical-scale expansion potential (Ginsberg et al., 2012). Similarly, expression of MYOCD reprograms human skin fibroblasts into vSMCs (Bersini et al., 2020). Nonetheless, in vivo reprogramming, even among the neural lineage, by manipulating MRs remains controversial, probably due to leaky expression of reporter genes and lack of stringent lineage-tracing and single-cell sequencing methods (Wang et al., 2021). From this perspective, reprogramming of GBM cells into vascular cells is likely a rare, if not totally absent, event.

Interestingly, we noticed a few key disparities between the two genetic models of GBMs. The NPC^TKO GBMs have mixed molecular features of TCGA-MES and -PN, whereas NSC^HRas-shp53 GBMs molecularly recapitulate the TCGA-MES subtype of human GBMs. NPC^TKO GBM cells contain more prominent genetic alterations compared to NSC^HRas-shp53 GBMs (data not shown), probably because embryonic neural precursors are at the apex of neurodevelopment hierarchy and longer tumorigenic period for NPC^TKO GBMs. We would like to point out that although loss-of-function mutations of three tumor-suppressor genes were introduced at embryonic stages to generate NPC^TKO GBMs, those tumors by no means resemble human pediatric high-grade gliomas, which often carry mutations distinct from adult GBMs (Schwartzentruber et al., 2012; Wu et al., 2012). In addition, there was substantial vascular remodeling and increased proliferation of mural cells and ECs in GBM brains, which could be either direct or adaptive responses to GBM formation. Intriguingly, the two genetic GBM models displayed distinct vascular patterns, possibly reflecting their cellular composition and properties such as the relative proportion of OLIG2⁺ cells. Indeed, there was less vessel coverage and fewer vessels in NPC^TKO GBMs than NSC^HRas-shp53 GBMs, mirroring a previous study showing OLIG2⁺ gliomas had largely normal vasculature, but that vessel density, vessel size, and the vascular and luminal areas within OLIG2⁻ gliomas were significantly increased (Griveau et al., 2018).

Here we did not observe cross-lineage trans-differentiation of GBM cells into vascular cells through careful lineage-tracing and single-cell analyses. We cannot exclude, however, given the highly inter-tumoral and intra-tumoral heterogeneity of GBMs, the possibility that a rare number of GBM cells might gain some molecular and functional features of vascular cells. Lineage-tracing studies indeed showed a very small portion of GBM cells misexpressed markers for vascular cells; and a few vascular cells of the NPC^TKO GBM were designated as malignant cells based on their CNV properties. Therefore, complementing genetic lineage-tracing experiments with single-cell sequencing could more stringently discern lineage relationships. Future studies should apply more lineage-tracing studies using other markers of mural cells or ECs. Moreover, single-cell sequencing and CNV analyses followed by vascular cell enrichment using additional GBM samples could further improve the resolution in differentiating lineages.

In summary, our findings advance understanding of the lineage potential of GBM cells and might provide insights into targeted treatments for GBM patients.

Materials and methods

Key resources table

Vector construction

Oligonucleotides coding for guide RNAs targeting Trp53, Pten, and Nf1 have been described previously (Zuckermann et al., 2015). These oligonucleotides were constructed and cloned into the pPB-sgRNA-SpCas9 (Cheng et al., 2016). For the Bioluminescence detection, the EGFP cassette in the pPB-sgRNA-SpCas9 described above was replaced with Luc cassette from pGL3 (Promega). pPB-EGFP and pPB-DsRed were constructed from the pPB-mU6pro vector. The mU6-EGFP cassette in the pPB-mU6pro vector was replaced with CAG-DsRed (from pCALNL-DsRed; #13769) to construct the pPB-DsRed. Then, the DsRed cassette in pPB-DsRed was replaced with EGFP cassette (from pEGFP-C2) to construct pPB-EGFP.

Animals

Tbx18^H2B-GFP/+ mice (Cai et al., 2008; Guimaraes-Camboa et al., 2017) were maintained on mixed C57BL/6 and CD-1 background. Tek-Cre, Tbx18^CreERT2/+ (Guimaraes-Camboa et al., 2017) and Ai14 mice were maintained on the C57BL/6 background. The Ai14 reporter mice (007914, The Jackson Laboratory) were crossed with Tek-Cre (008863, The Jackson Laboratory) and Tbx18^CreERT2/+ to genetically lineage trace the progenies of ECs and TAMs (Tek-Cre), and mural cells (Tbx18^CreERT2/+), respectively. Genotypes of all mice were determined using PCR analyses of tail or toe genomic DNA with appropriate primers. Both male and female mice were used for all experiments without preference. Mice were housed in certified specific-pathogen-free facilities. All animal procedures were approved by the Laboratory Animal Welfare and Ethics Committee of the College of Life Science and Medical Research Institute, Wuhan University.

In utero electroporation

All pregnant mice (E14.5) were deeply anesthetized with pentobarbital sodium (70 mg/kg) for in utero electroporation. Surgical operation and electroporation-based gene transfer was performed as follows. After injecting 1–2 µL supercoiled plasmids (5 μg/μL in ddH₂O) into telencephalic vesicles of E14.5 embryos, electric square pulses were generated using CUY21VIVO-SQ (BEX) and delivered into dorsal forebrains using forceps-like electrodes (35 mv, 50 ms-on, 950 ms-off, 5 pulses). The uteri were then carefully put back into the abdominal cavity and incisions were sutured. The whole procedure was complete within 30 min. Mice were warmed on heating pad until they woke up and given analgesia treatment (Ibuprofen) in drinking water.

Intracranial injection of lentivirus

Lentiviruses were stereotactically injected into the subependymal ventricular zone of 8-week-old Tek-Cre;Ai14 and Tbx18^CreERT2/+;Ai14 mice. Lentiviruses (1 × 10⁶ IU) suspended in 1.5 µL PBS were loaded on the syringe, and injected slowly (0.2 µL/min) using the following coordinates: 1.0 mm anterior, 1.0 mm lateral, and 2.3 mm dorsal to the bregma. Upon completing injection, the needle was left in place for 5 min, then withdrawn slowly to reduce virus reflux in 2 min.

Bioluminescence imaging

The in utero electroporation-operated embryos were allowed to survive until P21–P30 and were subjected to bioluminescence imaging of Luc activity. Luc signals were captured using the IVIS Lumina II (Caliper Life Sciences). Mice were injected with 30 mg of D-Luciferin, Potassium Salt (Yeasen Biotech) intraperitoneally 8 min before imaging. Images were acquired and radiance was determined within mouse heads. Only mice whose brain carried Luc signals were selected for further analysis. At postnatal days 60–150, all selected mice were subjected to bioluminescence imaging again to make sure tumors were big enough for experiments.

Tamoxifen administration

To efficiently induce CreER-dependent recombination, tamoxifen (150 mg/kg of body weight, T5648, Sigma) dissolved in sunflower seed oil (Sigma) was intraperitoneally administered according to the schematic illustration described in the paper.

Establishment of GICs with stem-like features

Single-cell tumor suspensions were plated to grow neurospheres at a density of roughly 5 × 10⁵ cells/well in 6-well Ultra-Low Attachment Multiple Well Plate (CLS3471, CORNING Costar) in the NBE medium (neurobasal medium containing N-2 and B-27 supplements, EGF, bFGF, and penicillin-streptomycin) as described. We used short-term passage cells (less than five passages in culture) in all experiments.

Intracranial allograft and xenograft implantation of GICs

Experiments were carried out using 6-week-old male Balb/c athymic nude mice or CD-1 mice (HNSJA). A total of 1 × 10⁵ GICs were injected intracranially using a stereotactic device (RWD) and a Hamilton syringe at a depth of 2.5 mm into the right cerebral hemisphere (2 mm posterior and 1.5 mm lateral to the bregma). Animals were sacrificed 4–8 weeks post-surgery.

Tissue processing

Mice were deeply anesthetized with 0.7% pentobarbital sodium and transcardially perfused with 4% paraformaldehyde in PBS. Brains were collected and post fixed in 4% paraformaldehyde in PBS overnight, followed by dehydration in 20% sucrose in PBS until submerging to the bottom (24–48 h). Brains were embedded in optimal cutting temperature compound (O.C.T. Compound, SAKURA) and then sliced coronally at 20 μm equidistant intervals using a cryostat microtome (CM1950, Leica).

Flow cytometry

GBM regions from Luc positive mice were acutely dissected under fluorescent microscope and enzymatically dissociated into single-cell suspensions using the Papain Dissociation system (Worthington Biochemical). The same areas from the control mice were parallelly dissociated into single-cell suspensions to set up voltage parameters and all gates. Flow cytometry was performed on a BD LSRFortessa X-20 Cell Analyzer or a Beckman Cytoflex Cell Analyzer. Cell sorting was performed on a BD FACSAria III Cell Sorter. Every experimental group used 3 million cells for flow cytometry analyses.

Human GBM specimens

Human GBM samples and derived GBM stem-like cells were previously reported (Zhong et al., 2016; Li et al., 2017). Their acquisition was in accordance with ethic guidelines of Wuhan Union Hospital and Shanghai Tenth People’s Hospital in accordance with institution-approved protocols. Specimens were examined by neuropathologists to verify tumor type and grade.

Immunohistochemistry, immunofluorescence, and Western blot analysis

Tissue sections or adherent cultured cells were incubated for one hour at room temperature in blocking solution (50 mmol/L Tris-HCl, pH 8.0, 0.1 mol/L NaCl, 0.1% Triton X-100, 3% NGS, 0.1% BSA) prior to overnight primary antibody incubation (4°C). Primary antibodies for immunofluorescence studies were: SOX2 (1:1000 dilution, Ab5603, Millipore), Ki67 (1:500 dilution, ab15580, Abcam), BrdU (1:500 dilution, ab6326, Abcam), NeuN (1:500 dilution, ab177487, Abcam), TUJ1 (1:200 dilution, T8660, Sigma-Aldrich), GFAP (1:500 dilution, Z0334, Agilent), MBP (1:300 dilution, SMI-99P, Covance), OLIG2 (1:200 dilution, AB9610, Millipore), OLIG2 (1:50 dilution, MABN50, Millipore), IBA1 (1:200 dilution, ab5076, Abcam), CD31 (1:100 dilution, ab24590, Abcam), ACTA2 (1:400 dilution, ab5694, Abcam), GFP (1:2000 dilution, ab13970, Abcam), CD31 (1:300 dilution, AF3628-SP, R&D), CD146 (1:100 dilution, 134701, BioLegend). We used immunofluorescence staining with Alexa Fluor 488, 555, or 647 (Life Technologies) and biotin-streptavidin-Alexa Fluor-conjugated secondary antibodies (Jackson ImmunoResearch), as well as horseradish peroxidase-based Vectastain ABC Kit (Vector Laboratories). Image acquisitions were performed using a Zeiss LSM880 confocal microscope and image editing done using ZEN, Photoshop or ImageJ.

Bulk RNA isolation and sequencing

RNA isolation was performed using the RNAiso Plus (TAKARA) according to the manufacturer’s instructions. RNA quality was evaluated spectrophotometrically, and the quality was assessed with the Agilent 2100 Bioanalyzer (Agilent Technologies). All samples showed RNA integrity of >7.5. RNA-seq libraries were prepared using the NEBNext Ultra II RNA Library Prep Kit for Illumina (New England Biolabs, no. E7775). Once prepared, indexed complementary DNA (cDNA) libraries were pooled in equimolar amounts and sequenced with paired-end reads on an Illumina NovaSeq 6000 (Illumina). For NPC^TKO tumor samples, GBM bulks were dissociated into single-cell suspensions as described above, EGFP⁺ or DsRed⁺ GBM cells were sorted by FACS followed by RNA isolation and subsequent RNA-seq. For NSC^HRas-shp53 tumor samples, GBM bulks were directly used to isolate RNA for RNA-seq.

Bulk RNA-seq analysis

Quality control of bulk RNA-seq data was performed using FastQC (v0.11.9). Adapter and low-quality bases were trimmed by Cutadapt (v3.2). Clean data were mapped to mouse reference genome (UCSC mm10) by TopHat (v2.1.1) with default params. Cufflinks (v2.2.1) package calculated the gene expression level and normalized by fragments per kilobase of exon model per million mapped fragments. Single-sample GSEA has been described elsewhere (Verhaak et al., 2010). ssGSEA calculated separate enrichment scores between each sample and the specified gene sets that were download from published reference data. An enrichment score calculated by R package GSVA [v1.38.2, gsea (method = “ssgsea”, ssgsea.norm = TRUE)]. A positive score represents a high ranking of up-genes in the signature, and low ranking of down-genes in the signature. A negative value does not indicate the opposite, but rather a lack of effect. For the experiment described in Fig. S1, gene sets for related neural lineage clusters (Fig. S1N; Cahoy et al., 2008), TCGA-GBM subtypes (Fig. S1M; Verhaak et al., 2010), lineage-specific GBM subtypes (Fig. S1O; Wang et al., 2020) were downloaded and processed as described above.

scRNA-seq data generation and processing

Designated cells were sorted into PBS following the 10× Genomics protocol. The cell preparation time before loading onto the 10× Chromium controller was <2 h. Cell viability and counting were evaluated with trypan blue by microscopy, and samples with viabilities >70% were used for sequencing. Libraries were constructed using the Single Cell 3ʹ Library Kit V2 (10× Genomics). Transcriptome profiles of individual cells were determined by 10× Genomics-based droplet sequencing. Once prepared, indexed cDNA libraries were sequenced with paired-end reads on an Illumina NovaSeq 6000 (Illumina). The sequencing depth was 30 M per cell.

scRNA-seq data processing

The quality of sequencing reads was evaluated using FastQC (v0.11.9). Cell Ranger (v4.0.0) was used to align the sequencing reads (fastq) to the mm10 mouse transcriptome and quantify the expression of transcripts in each cell. This pipeline resulted in a gene expression matrix for each sample, which records the number of unique molecular identifiers (UMIs) for each gene associated with each cell barcode. Unless otherwise stated, all downstream analyses were implemented using R 3.6.0 and the package Seurat (v3.1.0). The percentage of mitochondrial and red blood cells were calculated and filtered to retain only higher-quality cells (red blood cells <5%). The feature count matrix was normalized and scaled with NormalizeData (normalization.method = “LogNormalize”, scale.factor = 10,000) and ScaleData function. We performed doublet prediction on the clustered data using Doublet Finder. For the experiment described in Fig. 4, cells from two samples were pooled and analyzed together. After rigorous quality control, in NPC^TKO #1, we obtained 11,722 high-quality cells with a median gene number—4,450 genes, resulting in a total of 19,878 mouse genes detected in all cells. In NPC^TKO #2, we obtained 13,945 high-quality cells with a median gene number—3,402 genes, resulting in a total of 20,038 mouse genes detected in all cells.

Dimension reduction

Dimension reduction was performed at three stages of the analysis: the selection of variable genes, PCA, and UMAP. The FindVariableGenes function was applied to select highly variable genes covering most of the biological information contained in the whole transcriptome. Then, the variable genes were used for PCA implemented with the RunPCA function. Next, we selected principal components 1–20 (for total cells) as input and performed the RunUMAP function to obtain bidimensional coordinates for each cell. We then clustered the cells using FindClusters (resolution = 0.6) function.

Determination of cell-type identity

For each cell type, we used multiple cell type-specific/enriched marker genes that have been previously described in the literature to determine cell-type identity. For ECs: Pecam1, Tek, and Cldn5 (Vanlandewijck et al., 2018). For mural cells: Kcnj8, Pdgfrb, Acta2, and Tbx18 (Vanlandewijck et al., 2018). For T cells: Cd2, Cd3d, Cd3e, and Cd3g. For TAMs: Cd14, Tmem119, and Aif1. For OLGs: Cldn11, Mag, Mbp, and Klk6 (Brown et al., 2016).

CNV inference from single-cell data using inferCNV

CNVs were estimated by sorting the analyzed genes by their chromosomal location and applying a moving average to the relative expression values, with a sliding window of 100 genes within each chromosome by inferCNV. For mouse data in Figs. 4 and S6, T cells were used to define the baseline of normal karyotype as the non-malignant cell types, such that their average CNV value was subtracted from all cells. Using hierarchical clustering of the rest single cells, three groups with the most significant and concordant CNV profiles were identified, and were named as Group A/B/C cells, respectively. Percentages of Group A/B/C cells in each Seurat clusters were presented (Fig. 4F and 4G). Then, we applied the “two CNA-based measures” to score each cell as previously reported (Neftel et al., 2019) to compare the lineage relationship of target Seurat clusters with Group A/B/C cells, respectively. “CNV signal” reflects the overall extent of CNVs, defined as the mean of the squares of CNV values across chosen chromosomes containing significant and concordant CNV profiles. In NPC^TKO #1, we selected chromosome 2, 3, 5, 8, 11, 12, 13, 15 as CNV profile for Group A cells; Chr 5, 11, 12, 13 for Group B cells; Chr 6, 13, 18, 19 for Group C cells. In NPC^TKO #2, we selected chromosome 2, 8, 9, 10, 11, 14, 15, 17, 19 as CNV profile for Group A cells; Chr 2, 9, 10, 11, 14, 15, 17, 19 for Group B cells; Chr 7, 18 for Group C cells. “CNV correlation” refers to the correlation between the chosen CNV profile of each cell and the average CNV profile of Group A/B/C cells, respectively. For human data in Figs. 5 and S7, T cells, TAMs, and OLGs were used to define the baseline of normal karyotype as the non-malignant cell types (Neftel et al., 2019). In SF11247, we selected chromosome 7, 10, 13, 19 as CNV profile for Group A cells; Chr 3, 7, 13 for Group B cells. In SF11285, we selected all chromosomes across the genome as CNV profile for Group A cells; Chr 7, 8, 12, 18, 19 for Group B cells.

CopyKAT

To distinguish malignant cells from non-malignant cells in human GBM samples, genome-wide aneuploidy in single cells were identified at 5 MB resolution to separate tumor cells from normal cells by copyKAT (V1.0.8.) using integrative Bayesian approaches. For experiments described in Fig. 5, raw gene expression matrixes of two GBM samples were prepared to run copyKAT, respectively. The parameter “ngene.chr=1” was applied to keep as many cells as possible.

Measurements and statistics

To quantify the percentage of marker-positive tumor cells, numbers of NeuN-, MBP-, GFAP-, OLIG2-, or SOX2-expressing tumor cells were counted in tumor cores, then divided by numbers of total tumor cells (EGFP or DsRed positive cells). To quantify the percentage of BrdU⁺ or Ki67⁺ ECs or mural cells, double-positive cells (BrdU⁺tdTomato⁺ or Ki67⁺tdTomato⁺) were counted in both tumor and non-tumor sites, then divided by numbers of total ECs or mural cells (>1000 cells). Vessel area and numbers were analyzed using glioma brains from three NPC^TKO and three NSC^HRas-shp53 model generated in Tbx18^CreERT2/+;Ai14 mice. Fluorescence sections were scanned with Leica THUNDER Imager Tissue Scanner. For each mouse brain, two images were cropped at the tumor core (without necrosis) and at the matched region of the contralateral non-tumor parenchyma, respectively. Average number of vessels per mm² and average total vascular area per field were measured with the ImageJ and analyzed with Graphpad Prism. OLIG2⁺ cells density was analyzed using glioma brains from eight NPC^TKO and seven NSC^HRas-shp53 model generated in Tek-Cre;Ai14 and Tbx18^CreERT2/+;Ai14 mice. Immunohistochemistry sections were scanned with the Leica Aperio VERSA 8 Brightfield, Fluorescence & FISH Digital Pathology Scanner. The number of OLIG2⁺ cells per mm² in each mouse brain tumor core was measured with the ImageJ and analyzed with Graphpad Prism. All grouped data are presented as mean ± SEM. Statistical tests were calculated on the GraphPad Prism software (version 8.4.3). The statistical significance of a single comparison on continuous data was performed using Student’s two-tailed unpaired t-test with Welch’s correction when required (non-equal variances) or the Mann–Whitney non-parametric test when data did not fit a normal distribution (assessed by Shapiro–Wilk normality test). For multiple comparison of Ki67⁺tdTomato⁺ in Fig. 6B, one-way analysis of variance (ANOVA) followed by Tukey’s post hoc test was performed. For multiple comparison of vascular density in Fig. 1I and 1J, we used two-way ANOVA, after checking that our data fitted to a normal distribution (assessed by Shapiro–Wilk normality test). Significant difference is indicated by a P-value less than 0.05 (*P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001).

References

[1] Alcantara Llaguno S, Chen J, Kwon CH et al. Malignant astrocytomas originate from neural stem/progenitor cells in a somatic tumor suppressor mouse model. Cancer Cell 2009;15:45–56.

[2] Alcantara Llaguno S, Sun D, Pedraza AM et al. Cell-of-origin susceptibility to glioblastoma formation declines with neural lineage restriction. Nat Neurosci 2019;22:545–55.

[3] Alcantara Llaguno SR, Wang Z, Sun D et al. Adult lineage-restricted CNS progenitors specify distinct glioblastoma subtypes. Cancer Cell 2015;28:429–40.

[4] Bersini S, Schulte R, Huang L et al. Direct reprogramming of human smooth muscle and vascular endothelial cells reveals defects associated with aging and Hutchinson-Gilford progeria syndrome. elife 2020;9:e54383.

[5] Bhaduri A, Di Lullo E, Jung D et al. Outer radial glia-like cancer stem cells contribute to heterogeneity of glioblastoma. Cell Stem Cell 2020;26:48–63.e6.

[6] Bizzotto S, Dou Y, Ganz J et al. Landmarks of human embryonic development inscribed in somatic mutations. Science 2021;371:1249–53.

[7] Brown TI, Carr VA, LaRocque KF et al. Oligodendrocyte heterogeneity in the mouse juvenile and adult central nervous system. Science 2016;352:1323–6.

[8] Cahoy JD, Emery B, Kaushal A et al. A transcriptome database for astrocytes, neurons, and oligodendrocytes: a new resource for understanding brain development and function. J Neurosci 2008;28:264–78.

[9] Cai CL, Martin JC, Sun Y et al. A myocardial lineage derives from Tbx18 epicardial cells. Nature 2008;454:104–8.

[10] Cancer Genome Atlas Research, N. Comprehensive genomic characterization defines human glioblastoma genes and core pathways. Nature 2008;455:1061–8.

[11] Chen J, McKay RM, Parada LF. Malignant glioma: lessons from genomics, mouse models, and stem cells. Cell 2012;149:36–47.

[12] Chen F, Rosiene J, Che A et al. Tracking and transforming neocortical progenitors by CRISPR/Cas9 gene targeting and piggyBac transposase lineage labeling. Development 2015;142:3601–11.

[13] Cheng L, Huang Z, Zhou W et al. Glioblastoma stem cells generate vascular pericytes to support vessel function and tumor growth. Cell 2013;153:139–52.

[14] Cheng M, Jin X, Mu L. Combination of the clustered regularly interspaced short palindromic repeats (CRISPR)-associated 9 technique with the piggybac transposon system for mouse in utero electroporation to study cortical development. J Neurosci Res 2016;94:814–24.

[15] Dimou L, Gallo V. NG2-glia and their functions in the central nervous system. Glia 2015;63:1429–51.

[16] Efremova M, Vento-Tormo M, Teichmann SA et al. CellPhoneDB: inferring cell-cell communication from combined expression of multi-subunit ligand-receptor complexes. Nat Protoc 2020;15:1484–506.

[17] Furnari FB, Fenton T, Bachoo RM et al. Malignant astrocytic glioma: genetics, biology, and paths to treatment. Genes Dev 2007;21:2683–710.

[18] Gao R, Bai S, Henderson YC et al. Delineating copy number and clonal substructure in human tumors from single-cell transcriptomes. Nat Biotechnol 2021;39:599–608.

[19] Gao P, Postiglione MP, Krieger TG et al. Deterministic progenitor behavior and unitary production of neurons in the neocortex. Cell 2014;159:775–88.

[20] Ginhoux F, Greter M, Leboeuf M et al. Fate mapping analysis reveals that adult microglia derive from primitive macrophages. Science (New York, NY) 2010;330:841–45.

[21] Ginsberg M, James D, Ding BS et al. Efficient direct reprogramming of mature amniotic cells into endothelial cells by ETS factors and TGF beta suppression. Cell 2012;151:559–75.

[22] Gomez Perdiguero E, Klapproth K, Schulz C et al. Tissue-resident macrophages originate from yolk-sac-derived erythro-myeloid progenitors. Nature 2015;518:547–51.

[23] Griveau A, Seano G, Shelton SJ et al. A glial signature and Wnt7 signaling regulate glioma-vascular interactions and tumor microenvironment. Cancer Cell 2018;33:874–89.e7.

[24] Guimaraes-Camboa N, Cattaneo P, Sun Y et al. Pericytes of multiple organs do not behave as mesenchymal stem cells in vivo. Cell Stem Cell 2017;20:345–59.e5.

[25] Han X, Li F, Fang Z et al. Transdifferentiation of lung adenocarcinoma in mice with Lkb1 deficiency to squamous cell carcinoma. Nat Commun 2014;5:3261.

[26] Hara T, Chanoch-Myers R, Mathewson ND et al. Interactions between cancer cells and immune cells drive transitions to mesenchymal-like states in glioblastoma. Cancer Cell 2021;39:779–92.e11.

[27] He FL, Soriano P. Sox10ERT2CreERT2 mice enable tracing of distinct neural crest cell populations. Dev Dynam 2015;244:1394–403.

[28] Hu B, Wang Q, Wang YA et al. Epigenetic activation of WNT5A drives glioblastoma stem cell differentiation and invasive growth. Cell 2016;167:1281–95.e18.

[29] Ignatova TN, Kukekov VG, Laywell ED et al. Human cortical glial tumors contain neural stem-like cells expressing astroglial and neuronal markers in vitro. Glia 2002;39:193–206.

[30] Ishii Y, Matsumoto Y, Watanabe R et al. Characterization of neuroprogenitor cells expressing the PDGF β-receptor within the subventricular zone of postnatal mice. Mol Cell Neurosci 2008;37:507–18.

[31] Kester L, van Oudenaarden A. Single-cell transcriptomics meets lineage tracing. Cell Stem Cell 2018;23:166–79.

[32] Kisanuki YY, Hammer RE, Miyazaki J et al. Tie2-Cre transgenic mice: a new model for endothelial cell-lineage analysis in vivo. Dev Biol 2001;230:230–42.

[33] Korn J, Christ B, Kurz H. Neuroectodermal origin of brain pericytes and vascular smooth muscle cells. J Comp Neurol 2002;442:78–88.

[34] Kurz H. Cell lineages and early patterns of embryonic CNS vascularization. Cell Adh Migr 2009;3:205–10.

[35] Kwon C.-H, Zhao D, Chen J et al. Pten haploinsufficiency accelerates formation of high-grade astrocytomas. Cancer Res 2008;68:3286–94.

[36] Lee JH, Lee JE, Kahng JY et al. Human glioblastoma arises from subventricular zone cells with low-level driver mutations. Nature 2018;560:243–7.

[37] Li YM, Wang W, Wang FY et al. Paired related homeobox 1 transactivates dopamine D2 receptor to maintain propagation and tumorigenicity of glioma-initiating cells. J Mol Cell Biol 2017;9:302–14.

[38] Liu C, Sage JC, Miller MR et al. Mosaic analysis with double markers reveals tumor cell of origin in glioma. Cell 2011;146:209–21.

[39] Marumoto T, Tashiro A, Friedmann-Morvinski D et al. Development of a novel mouse glioma model using lentiviral vectors. Nat Med 2009;15:110–6.

[40] Mei X, Chen YS, Chen FR et al. Glioblastoma stem cell differentiation into endothelial cells evidenced through live-cell imaging. Neuro Oncol 2017;19:1109–18.

[41] Merkle FT, Alvarez-Buylla A. Neural stem cells in mammalian development. Curr Opin Cell Biol 2006;18:704–9.

[42] Murata-Kamiya N, Kurashima Y, Teishikata Y et al. Helicobacter pylori CagA interacts with E-cadherin and deregulates the β-catenin signal that promotes intestinal transdifferentiation in gastric epithelial cells. Oncogene 2007;26:4617–26.

[43] Neftel C, Laffy J, Filbin MG et al. An integrative model of cellular states, plasticity, and genetics for glioblastoma. Cell 2019;178:835–49.e21.

[44] Patel AP, Tirosh I, Trombetta JJ et al. Single-cell RNA-seq highlights intratumoral heterogeneity in primary glioblastoma. Science 2014;344:1396–401.

[45] Pollard SM, Yoshikawa K, Clarke ID et al. Glioma stem cell lines expanded in adherent culture have tumor-specific phenotypes and are suitable for chemical and genetic screens. Cell Stem Cell 2009;4:568–80.

[46] Poon CC, Sarkar S, Yong VW et al. Glioblastoma-associated microglia and macrophages: targets for therapies to improve prognosis. Brain 2017;140:1548–60.

[47] Quail DF, Joyce JA. The Microenvironmental landscape of brain tumors. Cancer Cell 2017;31:326–41.

[48] Ricci-Vitiani L, Pallini R, Biffoni M et al. Tumour vascularization via endothelial differentiation of glioblastoma stem-like cells. Nature 2010;468:824–8.

[49] Schwartzentruber J, Korshunov A, Liu X-Y et al. Driver mutations in histone H3.3 and chromatin remodelling genes in paediatric glioblastoma. Nature 2012;482:226–31.

[50] Scully S, Francescone R, Faibish M et al. Transdifferentiation of glioblastoma stem-like cells into mural cells drives vasculogenic mimicry in glioblastomas. J Neurosci 2012;32:12950–60.

[51] Soda Y, Marumoto T, Friedmann-Morvinski D et al. Transdifferentiation of glioblastoma cells into vascular endothelial cells. Proc Natl Acad Sci USA 2011;108:4274–80.

[52] Vanlandewijck M, He L, Mae MA et al. A molecular atlas of cell types and zonation in the brain vasculature. Nature 2018;554:475–80.

[53] Venteicher AS, Tirosh I, Hebert C et al. Decoupling genetics, lineages, and microenvironment in IDH-mutant gliomas by single-cell RNA-seq. Science 2017;355:eaai8478.

[54] Verhaak RG, Hoadley KA, Purdom E et al. Integrated genomic analysis identifies clinically relevant subtypes of glioblastoma characterized by abnormalities in PDGFRA, IDH1, EGFR, and NF1. Cancer Cell 2010;17:98–110.

[55] Wang R, Chadalavada K, Wilshire J et al. Glioblastoma stem-like cells give rise to tumour endothelium. Nature 2010;468:829–33.

[56] Wang Q, Hu B, Hu X et al. Tumor evolution of glioma-intrinsic gene expression subtypes associates with immunological changes in the microenvironment. Cancer Cell 2017;32:42–56.e6.

[57] Wang LL, Serrano C, Zhong X et al. Revisiting astrocyte to neuron conversion with lineage tracing in vivo. Cell 2021;184:5465–81.e16.

[58] Wang Z, Sun D, Chen YJ et al. Cell lineage-based stratification for glioblastoma. Cancer Cell 2020;38:366–79.e8.

[59] Wu G, Broniscer A, McEachron TA et al. Somatic histone H3 alterations in pediatric diffuse intrinsic pontine gliomas and non-brainstem glioblastomas. Nat Genet 2012;44:251–3.

[60] Yamanishi E, Takahashi M, Saga Y et al. Penetration and differentiation of cephalic neural crest-derived cells in the developing mouse telencephalon. Dev Growth Differ 2012;54:785–800.

[61] Yao M, Ventura PB, Jiang Y et al. Astrocytic trans-differentiation completes a multicellular paracrine feedback loop required for medulloblastoma tumor growth. Cell 2020;180:502–20.e19.

[62] Yu K, Lin CJ, Hatcher A et al. PIK3CA variants selectively initiate brain hyperactivity during gliomagenesis. Nature 2020;578:166–71.

[63] Zhong XL, Liu X, Li YM et al. HMGA2 sustains self-renewal and invasiveness of glioma-initiating cells. Oncotarget 2016;7:44365–80.

[64] Zhou Y, Bian S, Zhou X et al. Single-cell multiomics sequencing reveals prevalent genomic alterations in tumor stromal cells of human colorectal cancer. Cancer Cell 2020;38:818–28.e5.

[65] Zhou W, Chen C, Shi Y et al. Targeting glioma stem cell-derived pericytes disrupts the blood-tumor barrier and improves chemotherapeutic efficacy. Cell Stem Cell 2017;21:591–603.e4.

[66] Zou M, Toivanen R, Mitrofanova A et al. Transdifferentiation as a mechanism of treatment resistance in a mouse model of castration-resistant prostate cancer. Cancer Discov 2017;7:736–749.

[67] Zuckermann M, Hovestadt V, Knobbe-Thomsen CB et al. Somatic CRISPR/Cas9-mediated tumour suppressor disruption enables versatile brain tumour modelling. Nat Commun 2015;6:7391.

RIGHTS & PERMISSIONS

©The Author(s) 2022. Published by Oxford University Press on behalf of Higher Education Press.