Diffuse Glioma Heterogeneity and Its Therapeutic Implications

The (Epi)genomic Landscape of Glioma Subtypes

Unbiased clustering of patients with adult glioma based on either transcriptional or DNA-methylation data robustly segregates tumors into two mega-clusters defined by their IDH status, emphasizing the biological importance of this key glioma driver (18). IDH mutation leads to the accumulation of the oncometabolite R(-)-2-hydroxyglutarate (2HG), which competitively inhibits α-ketoglutarate (α-KG)–dependent dioxygenases, a family of chromatin modifiers (23–25). This drives extensive epigenetic rewiring characterized by histone methylation alterations (26–28) and global DNA hypermethylation—a phenomenon known as the cytosine–phosphate–guanine (CpG) island methylator phenotype (G-CIMP; refs; 15, 27, 29). Even within the category of IDH-mutant glioma, there is clinically meaningful heterogeneity; IDH-mutant 1p19q codeleted gliomas are termed oligodendroglioma with frequent TERT promoter mutations and improved prognosis, whereas IDH-mutant 1p19q non-codeleted tumors are called astrocytomas with high incidences of TP53- and ATRX-inactivating mutations (30). Of note, despite their names, neither tumor type yet has a confirmed cell of origin. DNA-methylation profiling of IDH-mutant 1p19q non-codeleted tumors shows that most have the G-CIMP phenotype (G-CIMP^hi subtype), but a fraction can lose some of their DNA methylation and progress into a G-CIMP^lo subtype. These G-CIMP^lo astrocytomas are enriched in the small fraction of severe GBMs that are IDH-mutant (31); so too are IDH-mutant astrocytomas with polysomy of chromosome 1 and/or 19 (32).

IDH–wild-type gliomas usually present as GBMs and typically have complex genomic landscapes with frequent chromosomal rearrangements and multiple genetic amplifications/mutations. However, despite their diversity, genetic aberrations tend to converge on three specific biological pathways: the p53 axis (MDM2, MDM4, and TP53), the Rb tumor-suppressive pathway (CDK4, CDK6, CCND2, CDKN2A/B, and RB1), and the MAPK/PI3K pathway (PIK3CA, PIK3R1, PTEN, EGFR, PDGFRA, and NF1; ref. 16). Mutations within a particular pathway are usually mutually exclusive and occur with varying frequency, with the most common aberrations being in TP53, CDKN2A, and EGFR. Chromosomal rearrangements are also common in GBM, particularly gain of chromosome 7 (encoding EGFR) and loss of chromosome 10 (encoding PTEN), which are thought to represent some of the earliest steps of GBM pathogenesis (33) and may even arise independently in different subclones (34). As with most cancers, telomere maintenance is required for glioma proliferation, and thus TERT promoter mutations are nearly universal in IDH–wild-type glioblastomas (16, 35). Genetic intertumoral heterogeneity of IDH–wild-type gliomas also manifests at the transcriptional level, with three transcriptional subtypes having been identified: classic, proneural, and mesenchymal, which are closely associated with mutations in EGFR, PDGFRA, and NF1, respectively (14, 36). By contrast, the large majority of IDH-mutant gliomas fall under the proneural category, lending this transcriptional subtype an improved prognosis (16).

pHGGs are genomically distinct from adult glioma, sharing few genetic aberrations common in adulthood and instead enriched for mutations in histones H3 variants (6, 7). They are robustly subtyped according to their histone H3 variant into three groups with different anatomic location, age of onset, mutational profiles, and clinical outcomes (37). Briefly, H3.3K27M pHGGs are the most common with the worst clinical outcomes and are found throughout the midline and pons. Less common are H3.1K27M tumors, which are restricted to the pons, present at younger ages, and have slightly improved survival, whereas H3.3G34R/V tumors are almost entirely restricted to the cerebral hemispheres. Those pHGGs without histone mutations have varied anatomic locations and represent a heterogeneous group, some of which resemble low-grade pediatric gliomas with MAPK signaling dysregulation and CDKN2A/CDKN2B deletion, with the remainder having poorer outcomes and differentiated by methylation profiling into EGFR/MYCN/CDK6 amplification or PDGFRA/MET amplification groups (37). pHGGs that appear in infants (<4 years) are enriched for MAPK mutations or gene-fusion driver genes (ALK, NTRK1/2/3, ROS1, and MET), with improved outcomes and responses to targeted therapy (38).

Glioma Subtype–Specific TMEs

The brain represents a relatively immunosuppressive microenvironment for glioma cells, which contains unique immune/regulatory cell types such as microglia and astrocytes, lacks classic lymphoid structures, and is protected from circulating immune cells and/or drugs by the blood–brain barrier (BBB) and cerebrospinal fluid (CSF)–brain barrier (39). Despite these challenges, the widespread success of immunomodulatory therapy across diverse cancer types has inspired significant interest in characterizing the immune fraction of the glioma TME (40). Importantly, TME cells are remodeled by malignant cells into active participants of tumor progression, and as such there is heterogeneity in the nature of the TME of different glioma genetic subtypes. Gliomas are enriched for tumor-associated macrophages (TAM), which can adopt a wide variety of protumorigenic activation states and play a critical role in promoting invasion, angiogenesis, metastasis, and immune suppression (41). Glioma TAMs can originate from two distinct lineages: tissue-resident microglia (CD49^lo MG-TAMs) or monocytes recruited from peripheral circulation (CD49^hi M-TAMs; refs. 42, 43).

Recently, two complementary studies used flow cytometry, CyTOF, RNA-seq, and protein arrays to comprehensively profile the immune landscape of adult gliomas, revealing subtype-specific TMEs dictated by IDH status. Both confirmed that gliomas are immunogenically “cold” with an abundance of TAMs but low levels of infiltrating T cells, particularly in IDH-mutant glioma (44, 45). Intriguingly, the ratio of TAMs derived from microglia or monocytes varies according to IDH status, with IDH-mutant gliomas enriched for MG-TAMs and IDH–wild-type tumors instead enriched for M-TAMs (44, 45). Detailed analysis of glioma TAMs revealed a diversity of activation states well beyond canonical M1 (antitumorigenic) and M2 (protumorigenic) polarization (46). MG-TAMs and M-TAMs had distinct genomic profiles that are additionally influenced by a glioma's IDH status, with TAMs in IDH–wild-type expressing higher levels of “reactive” markers CD14 and CD64 (44, 45). Pseudotime analysis of monocyte-TAM maturation trajectories showed that macrophage maturation is specified by tumor type, with CD163⁺ CX3CR1⁺ CADM1⁺ M-TAMs found exclusively in IDH–wild-type glioma (and not brain metastases), whereas those few cells present in IDH-mutant tumors did not mature and remained as monocytes (45). Moreover, TAMs from both lineages upregulated genes associated with innate anti–PD-1 resistance (47), and enrichment of an M-TAM gene signature consisting of macrophage activation markers, chemokine receptors, and cytokines could predict lower survival in IDH-mutant gliomas (44).

Interestingly, the ratio between infiltrating immature natural killer (NK) cells and mature cytotoxic NK cells—another cell type involved in innate tumor immunity (48)—is also skewed according to IDH status, with accumulation of immature NK cells in IDH–wild-type tumors (45). In contrast to adult gliomas, pHGGs appear relatively immunologically inert with an absence of macrophage or T-cell infiltration with no correlation to survival (49, 50). Furthermore, unlike adult GBM, pHGG cells do not repolarize cocultured macrophages to an immunosuppressive phenotype and are susceptible to NK cell–mediated destruction (50). The functional differences between TAMs from different lineages, and how the unique immune TME of each glioma subtype is best harnessed by immunotherapy, represent important areas for future research.

Nonimmune, neuroglial cell types undoubtedly also play an important role in glioma biology, but to date have been relatively understudied. Reactive astrocytosis is a common histologic feature of glioma, and GBM-associated astrocytes are more proliferative and marked by JAK–STAT pathway activation and CD274 expression (51, 52). In cooperation with microglia, they secrete anti-inflammatory cytokines such as TGFB, IL, and G-CSF, which contribute to an immunosuppressive environment (51). Even neurons can have protumorigenic functions, by either paracrine or autocrine mechanisms (53), as well as through functional synaptic integration into the calcium signaling networks of glioma cells joined by tumor microtubules (54–56). Oligodendrocytes are detected in relatively high numbers in scRNA-seq of glioma clinical samples, but their potential role in glioma pathology has yet to be explored. Systemic analysis of how neuro/glial cells vary between the TME of different types of glioma has not yet been performed, but will likely reveal important differences and perhaps even opportunities for targeted therapies.

Subclonal Genetic Heterogeneity in Glioma

Cancer is associated with progressive genomic instability, and the interaction of stochastically acquired somatic mutations and environmental selection pressures drives branched evolutionary trajectories and the emergence of multiple genetically distinct subclones (57). In recent years, technological innovations such as laser-capture microdissection, computational deconvolution of bulk genomics data, and single-cell sequencing techniques have led to a growing appreciation for glioma's intratumoral genetic heterogeneity. For IDH–wild-type glioma, this was first demonstrated by fluorescent in situ hybridization, revealing nonoverlapping amplification of EGFR, PDGFRA, and MET glioma driver genes (58, 59), and later confirmed using regional sequencing (60, 61). More recently, inferred copy-number alteration analysis from scRNA-seq data has revealed the presence of multiple subclones with branched evolutionary trajectories for IDH–wild-type GBMs (62, 63), as well as for IDH-mutant gliomas (64, 65) and also pHGG (66). Computational deconvolution of bulk genetics can detect multiple genetically distinct populations, with 85% of pHGG tumors harboring 3 to 10 subclones (67). Although not yet a widely adopted technique, early single-cell DNA sequencing analysis indicates the presence of mutations below the detection threshold of bulk sequencing, suggesting genetic heterogeneity likely extends even beyond current estimates (68).

EGFR, a clinically targetable receptor tyrosine kinase (RTK), presents a case study in the complexity introduced by the various forms of glioma heterogeneity (69). First, different EGFR variants have been detected mutually exclusively within the same tumor by regional sequencing (60, 61) and also by single-nucleus DNA sequencing (70). Furthermore, EGFRvIII (a common glioma EGFR rearrangement variant; ref. 71) may be harbored on double-minute fragments—highly transcribed fragments of extra chromosomal DNA (ecDNA) that segregate unequally upon cell division, further contributing to heterogeneity (72). Heterogeneous EGFR mutation and amplifications can be further compounded by variable expression, and scRNA-seq of GBM clinical samples revealed largely mutually exclusive transcript expression of wild-type EGFR, EGFRvIII, or EGFR with an exon 4 deletion (73). It may even be the case that heterogeneity in EGFR mutations between subclones can promote tumorigenesis cooperatively such that one GBM study found that even a small number of GBM cells expressing EGFRvIII can contribute to the growth of wild-type EGFR cells through a novel IL6-dependent pathway (74). Similar instances of cooperative behavior between genetically distinct subclones have been shown for pHGG, in which rare subclones can confer increased migratory capacity on their neighbors through paracrine mechanisms (67). Indeed, certain subclones appear enriched in particular glioma niches (see below) indicative of a degree of specialization, a hallmark of cooperativity (34, 60, 61).

Epigenetic Heterogeneity and Transcriptional States in Glioma

Epigenetic heterogeneity is also a critical determinant of tumor biology, with malignant cells co-opting developmental signaling pathways and adopting pseudo-developmental cellular hierarchies. The discovery of glioma stem cells (GSC) was a paradigm-shifting moment in glioma research, and GSCs are now the central focus in our understanding of glioma initiation, progression, and therapy resistance. GSCs represent a population of stem-like tumor cells maintained by self-renewal, and with the capacity to generate more differentiated progenies that compose the bulk of the tumor mass. Fundamentally the term GSC is a phenotypic/functional descriptor, which can only be defined experimentally by their tumor-propagating potential in vivo and sustained self-renewal potential (75, 76). Compelling evidence that glioma tumor cell stemness is a “property” shared by a heterogeneous group of cells, rather than a cellular “identity,” is provided by the large array of nonoverlapping, intermittently, and variably expressed cell-surface markers previously used for their isolation (e.g., CD133, CD44, SSEA1, L1CAM, A2B5, PDGFRA, and EGFR; refs. 77–81). Despite their nonoverlapping marker gene expression, GSCs' shared properties are likely sustained by a common network of master transcription factors and chromatin regulators defining their epigenetic/transcriptional status (82). Indeed, forced expression of core neurodevelopmental transcription factors (POU3F2, SOX2, SALL2, and OLIG2) is sufficient to transform differentiated cells into GSCs with tumor-initiating potential (83).

A profoundly important consequence of GSCs' capacity to produce more differentiated progeny is that it superimposes a layer of phenotypic heterogeneity—in the form of epigenetically defined transcriptional states—over a patchwork of intratumoral genomic heterogeneity. A series of papers leveraging scRNA-seq of clinical glioma samples have now begun to catalog the diversity of glioma phenotypic heterogeneity, and enabled reevaluation of the GSC model in the context of reconstructed cellular hierarchies (Fig. 2; refs. 84, 85). Analysis of IDH-mutant gliomas reveals a branched pseudo-developmental trajectory with cells falling into three subpopulations: at the apex, a group of cells transcriptionally similar to neural progenitor cells (NPC) and enriched for proliferation, with the remaining cells falling along a continuous trajectory toward more differentiated astrocytic (AC)-like or oligodendrocyte (OC)-like cells, with reduced levels of proliferation (64, 65). This finding has important consequences, as it implies oligodendroglioma (IDH-mut, 1p/19q codeleted) and astrocytoma (IDH-mut, 1p/19 non-codeleted) have a shared cell of origin, contradicting earlier studies (86, 87).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

Glioma epigenetic-mediated intratumoral heterogeneity and tumor cell pseudo-developmental hierarchies. A, Canonical developmental pathway represented on a Waddington epigenetic landscape adapted from ref. 85. Stem/progenitor cells pass down “valleys/creodes” (developmental trajectories) to “troughs/attractors” (differentiated cell types). B, Inferred pseudo-developmental hierarchy for IDH-mutant glioma. Cells reside on a branched continuous trajectory from a highly proliferative NPC-like population to nonproliferative, more differentiated astrocyte (AC)-like or oligodendrocyte (OC)-like cells. C, Inferred pseudo-developmental hierarchy for HK27M pHGG. Cells reside on a branched continuous trajectory from a highly proliferative OPC-like population to nonproliferative, more differentiated AC-like or OC-like cells. OPC-like cells represent the large majority of cells (80%) and differentiation is biased toward AC-like cells. D, IDH–wild-type glioma is characterized by four cell states, with tumor-specific cell state ratio at least partially influenced by genetics; examples include EGFR amplification/mutation associated with AC-like cells, PDGFR amplification associated with OPC-like cells, CDK mutation associated with NPC-like cells, and NF1 mutation associated with mesenchymal (MES)-like cells. Evidence supports a nonhierarchical model with high levels of cell state plasticity; cells in all states can proliferate and function as GSCs.

As IDH-mutant gliomas progress and recur, there is a redistribution of transcriptional states, with a reduction in the fraction of OC/AC-like cells and an increase in the fraction of proliferative GSCs (65, 88, 89). This phenomenon, often referred to as “differentiation block,” is even more evident in H3K27M pHGGs, perhaps explaining in part their increased aggressiveness (66). These pHGG tumors show a similar pseudo-developmental hierarchy but with a markedly increased proportion of GSCs (∼80%), which, instead of NPCs, resemble oligodendrocyte-progenitor cells (OPC), likely reflecting an alternate cell of origin (66). Somewhat surprisingly, there is a bias in differentiation of pHGG OPC-like cells toward AC-like rather than OC-like cells. This may be explained by H3K27M-mediated inhibition of the PRC2 complex, which plays an important role in oligodendrocyte differentiation (90–92), though one should not necessarily presume malignant cells to follow developmental trajectories with fidelity.

Numerous recent studies of IDH–wild-type GBM have revealed similar heterogeneity of transcriptional states, but much less clear evidence of consistent pseudo-developmental hierarchies. Instead, these tumors appear to be characterized by high levels of cellular state plasticity (62, 63, 93–97). Patel and colleagues' pioneering study was the first to apply scRNA-seq to IDH–wild-type clinical GBM samples, and showed that individual single cells from within the same tumor can occupy all three of the (bulk-defined) GBM transcriptional subtypes, mesenchymal, classic, and proneural (73). Since then, numerous groups have defined IDH–wild-type GBM transcriptional states based on clustering approaches to scRNA-seq data. These cellular states vary slightly according to computational methodology—however, a mesenchymal (MES)-like state unique to IDH–wild-type GBM is consistently detected. Reports of other transcriptional states resembling neuroglial cell types are more variable, but there is consensus that (i) a variety of transcriptional states exists and (ii) averaging of malignant transcriptional states, which occur in tumor-specific proportions, accounts for GBM's bulk transcriptomic subtypes (proneural, classic, and mesenchymal). The ratio of cells occupying each transcriptional state is partially influenced by tumor genetics, with EGFR amplification being associated with high a frequency of AC-like cells and the classic GBM transcriptional subtype, whereas PDGFRA/CDK4 amplification is associated with OPC-like/NPC-like cells and the proneural GBM transcriptional subtype. Finally, NF1 mutations are associated with MES-like cells and the mesenchymal GBM transcriptional subtype (14, 63). However, the majority of genetic aberrations found in IDH–wild-type GBM do not appear to have strong associations with the relative enrichment of a particular state, implying a greater influence by alternative factors such as epigenetics and microenvironmental interactions.

Controversy remains over the existence of consistent cellular hierarchies in IDH–wild-type GBM. For instance, Wang and colleagues used RNA velocity—a technique using the time derivative of gene expression to predict the future transcriptional state of cells (98)—and mitochondrial mutational phylogenies to support a hierarchical and continuous axis of variation from mesenchymal GSCs to proneural GSCs in clinical samples (95). Simultaneously, Neftel and colleagues published a study reporting four transcriptional subtypes (AC-like, NPC-like, OPC-like, and MES-like) but no evidence of consistent unidirectional cellular hierarchies (63). Definitive resolution of the debate over cellular hierarchies will require further experimentation (99). To that end, using various experimental GBM models, we found transcriptional subtypes analogous to Neftel and colleagues and also did not detect consistent cellular hierarchies by RNA velocity (97). We contend that, unlike IDH-mutant glioma and pHGGs, IDH–wild-type GBMs do not conform to rigid, unidirectional cellular hierarchies characteristic of development but instead are defined by their plasticity and spontaneous transitions between malignant cellular states.

FAC-sorted populations of cells from NPC-like (CD24^hi) and MES-like (CD44^hi) states within the same tumor both robustly induce tumorigenesis in orthotopic PDX models, with the resulting tumors recapitulating the full transcriptional diversity of the parental tumor. In addition, lentiviral barcoding revealed that cells sharing the same barcode (i.e., that proliferated from the same original NPC-like or MES-like cell during tumorigenesis) could occupy all four transcriptional states—implying state transitions (63). Bhaduri and colleagues also used scRNA-seq of clinical samples to show that multiple cell populations within individual tumors have GSC characteristics (i.e., marker gene expression) and that FACS-sorted populations of PTPRZ1 high or low cells (a newly described GSC marker typically expressed in outer radial glia cells) were both able to form heterogeneous tumors in an ex vivo organoid model. Again, tumors derived from both populations recapitulated the full transcriptional diversity of the parental tumor (62). These dual findings of (i) heterogeneous GSCs with tumorigenic potential and (ii) high levels of plasticity were emulated in an elegant study by Dirkse and colleagues, who used the combinatorial expression of four GSC cell-surface markers (CD113, CD44, A2B5, and CD15) to phenotype GBM cells into 16 subgroups, all of which were tumorigenic and, given time, could reconstitute all subgroups when grown in vitro or in vivo. Mathematical modeling suggested that phenotypic heterogeneity was regained through stochastic reversible state transitions rather than unidirectional cellular hierarchies. Intriguingly, differences in subgroup tumorigenicity and mouse survival were linked to plasticity, with the most inherently plastic populations producing the most aggressive tumors (100). This link between plasticity and malignancy could be explained by cooperative behavior between malignant transcriptional states and/or adaptability to different tumor niches (101).

Another important distinguishing feature of IDH–wild-type GBM is that cells with proliferative signatures are found in all transcriptional states, albeit biased toward more “neural” NPC/OPC-like cells. Furthermore, well-established GSC markers are distributed across GBM transcriptional subtypes with enrichment of CD24 in NPC-like cells, CD133 in OPC-like cells, Nestin in AC-like cells, and CD44 in MES-like cells, reinforcing the notion that IDH–wild-type GBM stemness is a property not restricted to a particular transcriptional state (84). Thus, in each subtype of glioma, tumors with a wide range of genetic backgrounds appear to converge on a limited set of transcriptional states, which loosely parallel healthy wild-type cell populations. The occupancy of different transcriptional states may be biased by certain mutations, but the transcriptional states themselves appear epigenetically encoded by as yet undefined gene regulatory networks, which may vary according to cellular hierarchies and a tumor's cell of origin. As technologies for single-cell multiomics progress, we anticipate the linkage between genetics and transcriptional phenotypes at the cellular level to be an important area of study.

Anatomic Variations within the TME

The TME represents the final piece of the puzzle: interacting with genetic and epigenetic intratumoral heterogeneity to influence the individual transcriptional phenotype and biological behavior of each malignant cell (Fig. 3). At least three different TMEs—or niches—have been commonly described in glioma: the hypoxic–necrotic core, the perivascular niche, and the invasive edge (102). In each of these different niches, tumor cells are exposed to different combinations of extrinsic cues in the form of mechanophysical forces, nutrient gradients, and interactions with different TME cell populations. The role of the TME in shaping malignant cell phenotype is emphasized by the comparison of different glioma models; ex vivo organoid or PDX-based models which partially recapitulate the glioma TME transcriptionally phenocopy parental tumors to a much greater degree than glioma cells simply grown in 2-D or as tumor organoids (97).

• Download figure
• Open in new tab
• Download powerpoint

Figure 3.

Glioma anatomic intratumoral heterogeneity. Schematic representation of glioma and its TME highlighting important cell types and four key anatomic regions/niches. (1) The SVZ is a vascularized stem-cell niche consisting of epidermal cells lining the CSF-filled ventricles, neighbored by adult NSCs and their more differentiated progeny. SVZ stem/progenitor cells are the glioma putative cell of origin. (2) The tumor–brain interface is termed the invasive edge and harbors invasive glioma cells that migrate along blood vessels as well as white matter tracts to infiltrate the brain. (3) Proangiogenic, vascularized regions of the tumor leading to leaky, dysfunctional vessels forming in part the perivascular niche. (4) Poorly vascularized regions of the tumor form the hypoxic core, with necrosis and hypoxic stress signaling. Glioma cells can project tumor microtubules to form a tumor signaling network, with functional connections to TME neuro/glial cells.

GBMs have high levels of abnormal angiogenesis forming leaky, dysfunctional vessels with microvascular proliferating structures (a GBM histologic hallmark), forming an environment known as the perivascular niche, which is enriched for GSCs (103). Together, endothelial cells and perivascular immune cells promote glioma stemness and growth through secreted growth factors and cytokines such as TGFB or CXCL12 and the NOTCH, sonic hedgehog, and nitric oxide (NO) signaling pathways (77, 104–109). The interactions are bidirectional, with GSCs promoting maintenance and expansion of the niche by secretion of VEGF and other angiogenic factors (110), and potentially even differentiating into endothelial- or pericyte-like cells themselves, although the extent to which this phenomenon occurs in human GBM is highly controversial (111–113). The vasculature also plays an integral role at the invasive edge—the brain–tumor interface—as GSCs migrate along preexisting blood vessels and white matter tracts to invade the healthy brain (114–116). To facilitate invasion within this tight environment, GSCs shrink by shedding cytoplasmic water, upregulate ephrinB2 to override vascular repulsion, and lift up astrocytic endfeet and pericytes remodeling basal lamina as they progress along the abluminal surface of endothelial cells (117, 118). Importantly, the reciprocal interactions between GSCs and the TME cells of the invasive niche reshape the phenotype of the GSCs, inducing a more mesenchymal transcriptional state that favors migration over proliferation (119, 120). Although these more invasive mesenchymal GSCs may be relatively less tumorigenic in PDX experiments than GSCs from the tumor core (121), they are perhaps the most clinically important because it is this population that persists after surgical resection and is thought to drive recurrence (122).

In contrast to the other regions of the glioma TME, the hypoxic–necrotic core is depleted of vasculature and contributes to GSC maintenance and gliomagenesis via alternate mechanisms—predominantly mediated by hypoxia-inducible factor 1 (HIF1α) and HIF2α (102). HIF2α is associated with chronic hypoxia and promotes expression of stemness genes such as KLF4, SOX2, and OCT4, whereas HIF1α is more associated with acute hypoxia and mediates metabolic adaptation, promotion of a mesenchymal transcriptional state, and expression of prosurvival factors (123–127). The nature of the TME at any particular anatomic region is not static, but changes dynamically in accordance with tumor growth and/or response to therapy. For example, HIF1α-driven VEGF secretion in regions of hypoxia can prime cells for regeneration of the perivascular niche (128) and GSCs migrating along blood vessels disrupt the homeostasis of the BBB, facilitating release of blood-borne cytokines and immune cells, leading to vasculature remodeling, which, as the tumor grows, leads to regression of the co-opted vessels in favor of angiogenesis—thus generating the perivascular niche (129).

Finally, a less commonly recognized but important fourth TME niche is the subventricular zone (SVZ). The SVZ is a unique microanatomic niche within the brain consisting of ependymal cells, astrocytes with extensions to the ependymal surface, and adult neural stem cells (NSC) of various differentiated stages all interacting with one another and all sampling the unique environment of the CSF-filled ventricular system (130). Not only have SVZ-based NSCs frequently been implicated as the cell of origin for IDH–wild-type GBM (131, 132), but glioma contact with this niche predicts worse prognosis, and if gliomas access the subependymal space and the CSF, it allows largely unrestricted access to almost anywhere in the brain through diffuse CSF spreading with dramatic therapeutic implications (133). GSCs may also gain access to the CSF via the glymphatic system—a glial-dependent perivascular network with pseudolymphatic function (134)—and thereby spread along the cranial spinal axis and rarely metastasize extracranially (135). The function of the glymphatic system in the context of glioma is only now being explored, but roles regulating tumor drainage and tumor immunity via dendritic cell trafficking to the cervical lymph nodes are beginning to emerge (136).

The Role of Tumor Heterogeneity in Glioma Therapy

Intratumoral genetic heterogeneity is generally thought to play a major role in tumor recurrence with secondary drug resistance through clonal selection of preexisting resistant clones (Fig. 4). This was intuited from the early days of chemotherapy whereupon initially treatment-responsive tumors showed dramatic cytoreduction only to ultimately recur with drug-resistant cells. Improved longitudinal molecular/genetic profiling and the advent of targeted therapies with singular mechanisms of action have now convincingly validated this theory (57). For example, in initially EGFR inhibitor–responsive lung cancers, tumor recurrence occurs driven by outgrowth of clones harboring the T790M EGFR mutation or MET amplification, both of which can be detected at low levels in pretreatment samples and are associated with shorter progression-free survival (137–141).

• Download figure
• Open in new tab
• Download powerpoint

Figure 4.

Models of intratumoral heterogeneity–mediated resistance to therapeutic stress (e.g., EGFR inhibition) in glioma. A, Resistance to EGFR inhibition could be driven by intratumoral genetic heterogeneity and clonal selection of a preexisting resistant clone or by acquisition of a secondary mutation that confers resistance as commonly seen in non–small cell lung carcinoma. B, Hypothetically, resistance of an EGFR-amplified IDH–wild-type GBM, with a high proportion of AC-like cells, to EGFR inhibition could be epigenetically driven by (1) depletion of vulnerable stable epigenetic state(s) and outgrowth of stable resistant epigenetic state(s) (i.e., epigenetic ‘clonal’ selection) or (2) by a plastic cell state transition from a vulnerable epigenetic state to a resistant epigenetic state. (3) If epigenetic resistance is driven by plastic state transitions, then removal of EGFR inhibition may result in a redistribution of cell states to their steady-state proportions.

Longitudinal sequencing studies show that although glioma progression following primary therapy is certainly accompanied by the appearance of novel mutations, only rarely is there evidence of clonal selection. More generally, the original driver genes persist, and the strongest selective pressures appear to occur early in gliomagenesis prior to therapy, with subsequent clonal evolution being largely stochastic (68, 88, 142–149). A fairly large proportion of recurrent gliomas (16%), particularly IDH-mutant and/or MGMT promoter methylated gliomas, present with extremely high intratumoral genetic heterogeneity referred to as the “hypermutator phenotype.” However, there is no survival difference to suggest that hypermutation leads to secondary mutations driving acquired resistance (142). This phenotype can arise de novo due to defects in DNA mismatch repair (MMR) genes and DNA polymerase but is more commonly caused by temozolomide treatment (68). In other cancers, a hypermutator phenotype can confer sensitivity to PD-1 blockade (150, 151), fueling some optimism for glioma (152). Unfortunately, hypermutated gliomas do not appear to have increased PD-1 blockade sensitivity, likely due to the late onset of acquired MMR deficiency in glioma, leading to subclonal neoantigen expression (68).

The depletion of glioma cells expressing EGFR in the context of EGFR small molecular inhibition or anti-EGFRvIII immunotherapy could be used to argue the importance of clonal selection in glioma recurrence (153). However, these treatments rarely if ever result in tumor reduction or stabilization, raising the question of whether genetic EGFR heterogeneity represents a true mechanism of drug resistance versus the mere elimination of a subpopulation of cells expressing a nonessential molecular target. This reinforces the importance of identifying and targeting truncal mutations that arise early during tumorigenesis, unlike EGFR mutations, which arise during the later stages of glioma evolution (154). Unfortunately, genes with truncal glioma mutations—IDH, PIK3CA, and H3—appear relatively scarce and have not yet been successfully targeted (149, 154, 155).

In a similar manner, a fair number of gliomas harboring the BRAF^V600E mutation show good initial responsiveness to BRAF inhibitors that cross the BBB but later recur with BRAF inhibitor resistance, despite continued expression of the Braf^V600E mutation. Although clonal selection could explain this phenomenon, it has not yet been experimentally shown, whereas resistance driven by a secondary BRAF mutation not present in the original tumor or by nonclonal phenotypic changes both have (156, 157).

It may be the case that inheritably stable epigenetic subpopulations of glioma cells operate as the selected “unit” in a process akin to clonal selection. For instance, it is common for treated patients with temozolomide/nitrosourea-sensitive gliomas with hypermethylated MGMT promoters to ultimately recur with a hypomethylated MGMT promoter chemotherapy-resistant tumor (158). Given the intratumoral cell-to-cell heterogeneity of MGMT promoter methylation, it is presumed that treatment relapse occurs as a result of survival and eventual outgrowth of previously existing MGMT-hypomethylated populations of cells. However, this has never been rigorously demonstrated and it is possible that the hypomethylated recurrent tumor merely reflects a nonclonal phenotypic adaptation to an otherwise less favored epigenetic state in the surviving fraction of cells. Similarly, global DNA-methylation profiling indicates that recurrent IDH-mutant gliomas deregulate their cell cycle epigenetically (159) and demethylation occurs as IDH-mutant G-CIMP^hi astrocytomas progress into more aggressive recurrent IDH-mutant G-CIMP^lo GBMs (31). Again, whether these changes represent outgrowth of stable epigenetic subpopulations or adaptive phenotypic changes remains to be determined.

Thus, the accumulating data, from studies of up to 10,000 patients (68), indicate that although clonal selection can occur, it is a relatively rare occurrence. This strongly implies that intratumoral “genetic” heterogeneity and the selection of treatment-resistant genetic subclones are not the predominant driving forces behind glioma recurrence in most cases. If not genetics, then perhaps it is glioma's profound “phenotypic” intratumoral heterogeneity—a product of not only genetics but also epigenetic cell states and the microenvironment—that underlies their intrinsic treatment resistance? It is accepted that specific phenotypic tumor cell subpopulations, such as GSCs, have enhanced therapeutic stress resistance (160). It follows then that—akin to genotypic diversity and clonal selection—increased phenotypic heterogeneity would be linked to increased potential for therapeutic resistance.

It is notable then that, although both remain lethal, IDH-mutated gliomas tend to have the least intratumoral heterogeneity of the gliomas and are the most responsive to both cytotoxic and targeted treatment (BRAF inhibitors and possibly IDH1/2 inhibitors; ref. 161), whereas IDH–wild-type gliomas are the most diverse and treatment-refractory. Malignant glioma epigenetic states are not fixed but dynamic, and it is common for IDH–wild-type GBMs to change between bulk transcriptional subtypes (proneural, mesenchymal, and classic) on recurrence (36, 143, 162). Because bulk transcriptomic subtypes are a product of averaging of single-cell states, this switching could plausibly result from (i) depletion of a vulnerable cellular state and expansion of a more resistant state or (ii) plastic transitions from one cellular state to another.

For instance, many believe that the high genetic intratumoral heterogeneity, resulting in aberrant expression of various RTKs in different glioma cells (e.g., EGFR, PDGFR, and MET) within a given tumor, accounts for the overall clinical failure of individual RTK inhibitors. This has led to a clinical trial strategy designed to evaluate combination RTK inhibitors (RTKi). To date, however, these trials have not proved successful, with these failures variously attributed to enhanced toxicity of combination therapy, inappropriate selection of targeted agents, and/or non–biomarker-driven trial designs (163–166). Given the inherent challenges of CNS pharmacokinetics and BBB penetration, it cannot be excluded that one of more of the combined therapeutics did not reach effective concentration (167–169). However, this approach is predicated on the assumption that intratumoral genetic heterogeneity is the primary, fixed mechanism for intrinsic drug resistance. It ignores the possibility that glioma cells might transition phenotype to a predominant cellular state no longer requiring RTK signaling.

Similarly, clinical trials evaluating combinations of antiangiogenic agents designed to target the heterogeneous mechanisms that individual glioma cells use to mediate angiogenesis (e.g., VEGF, PDGF, and TIE2 overexpression) have proved ineffective (NCT00458731 and NCT00667394). As in the case of combination RTKis, whether the failure to find an effective combination of antiangiogenic agents is secondary to the ineffectiveness of the specific agents evaluated, lack of biomarker-guided therapies, and/or the possibility that antiangiogenic therapy may be an overall ineffective strategy to combat the profound intratumoral heterogeneity in gliomas remains to be seen. Could glioma cell phenotype transition to a predominant invasive cellular state, not requiring angiogenic support, account for failure of antiangiogenic therapy? Indeed, it has been well documented that tumor progression on anti-VEGF therapy (e.g., bevacizumab) is clinically associated with a more infiltrative/invasive tumor growth pattern and experimentally shown to be associated with a more invasive mesenchymal phenotype (170).

Microenvironmental and anatomic heterogeneity may also be contributing to primary treatment resistance with areas of tumor hypoxia, reducing the effectiveness of radiation and alkylating agents. For instance, areas of hypoxia are associated with reduced vascular perfusion and an acidic microenvironment that will result in diminished drug delivery to areas of the tumor, diminished responsiveness to antiangiogenic agents, and inhibition of immune effector cell infiltration and antitumor cytotoxicity. Not surprisingly, clinical trials attempting to address these issues by using strategies such as hypoxia- activated drugs or hypoxia-correcting therapies (hyperbaric oxygen) have been unsuccessful alone, but may have some use as adjuvant therapies (171). Similarly, the variability of tumor-mediated BBB disruption spanning the spectrum from a fully disrupted barrier within the core of the glioma mass to the fully intact BBB at the infiltrating edge of the tumor mediates anatomically variable drug delivery and thus could account for treatment failure. Nevertheless, if this anatomic/environmental heterogeneity was primarily responsible for treatment failure, one would expect at least a “partial response” in areas of the chemotherapy/radiation-treated glioma not affected by this hypoxia or BBB disruption, a phenomenon not reported in the clinic (albeit regional partial responses are not typically assessed). Relative to neuroanatomic and spatial intratumoral heterogeneity, the recent discovery of interconnecting tumor nanotubes extending between glioma cells and to normal brain cells has been implicated as a therapy-resistance mechanism in glioma models (54–56). However, the true relevance of tumor–brain glioma networks in a clinical setting remains to be determined.

Thus, the data to date suggest that no one aspect of the dramatic genomic, environmental, or anatomic heterogeneity found in gliomas is unto itself responsible for the intrinsic treatment resistance. Rather it is more likely that a combination of many, if not all, of these factors acting individually on a given cell, and in concert on the tumor as a whole, provides numerous mechanisms sufficient for overall resistance to all current therapeutic modalities. Indeed, some of these individual tumor mechanisms may work not only in a cell-autonomous way, but through a paracrine cooperative type of mechanism, with individual glioma cells transporting molecules that aid in stress response both to the local microenvironment and directly to adjacent glioma cells through secretion, exosome and tunneling microtubes. Thus, this profound intratumoral glioma heterogeneity cushions and buffers the tumor as a whole in a web of therapeutic stress resistance.

Potential Therapeutic Strategies to Mitigate Glioma Heterogeneity

The current general approach to address intratumor heterogeneity is to combine various agents with different mechanisms of action. However, based on the discussion above and the protean number of heterogeneity-mediated resistance mechanisms simultaneously operative in glioma at any given time, this combination therapy approach seems overly simplistic, impractical, and likely to fail clinically (as it has done to date). We therefore suggest an alternative strategy whereby rather than trying to address and overcome each of the individual tumor heterogeneity–mediated survival mechanisms, such as treating individual glioma cell genotypes, it may prove far more productive to devise strategies based on glioma cell phenotypes. This strategic approach centers on the concept that despite the myriad of possible individual cell genotypes and environmental factors, glioma cells tend to converge on a fairly limited set of epigenetically encoded cell states. Because each cellular state defines the types of stress response it is capable of mounting, one might then rationally suppose that directing therapies against that “cellular state” will a priori address each of those resistant mechanisms mediated therein by that state. In essence, rather than focusing on frequently mutated genes or pathways, might we instead search for and target factors essential for the stable maintenance of a particular glioma cell state?

The idea of targeting cellular states has very different implications and strategic applications based on the different hierarchical ontogenic structures of the three major types of gliomas (pHGG, IDH-mutant, and IDH–wild-type). pHGG and IDH-mutant gliomas share a more classic developmental hierarchy with tumor growth driven by a highly proliferative stem cell population (OPC-like in pHGG and NPC-like in IDH-mutant gliomas) asymmetrically dividing to generate more differentiated (presumably less tumorigenic) progeny (OC-like and AC-like). This developmental hierarchy suggests that effective therapies targeted against the stem cell and proliferative state should be highly effective for eliminating the capability of the tumor to propagate further. Indeed, this may be why standard radiation and genotoxic chemotherapy (e.g., temozolomide and nitrosoureas) can be so effective in IDH-mutant gliomas compared with IDH–wild-type gliomas. Similarly, it may explain why the ability to target proliferative, stem cell drivers such as the MAPK kinase pathway in some NF1- and BRAF-mutated pHGGs is so effective. This strongly suggests that the development of other genetic and epigenetic modifiers of targets intrinsic to NSC proliferation and survival (e.g., BMI1, SOX2, and PRC2 complex) may ultimately yield highly effective therapies. In addition to killing GSCs, this approach might drive “differentiation,” forcing malignant cells into the more indolent AC-like state that is both less proliferative and less tumorigenic (84).

By contrast, IDH–wild-type gliomas do not conform to a rigid developmental hierarchy but rather represent a limited set of convergent epigenetically encoded transcriptional states (OPC-like, NPC-like, AC-like, and MES-like) between which cells can spontaneously and stochastically transition. If our hypothesis that specific cellular states mediate different mechanisms of stress/therapy resistance holds true, then it is wholly possible that cellular plasticity between states represents the overarching mechanism of IDH wild-type GBM therapy resistance. For instance, could glioma cells adaptively transition away from directly targeted cellular states such that EGFR inhibition drives cells out of the EGFR-dependent AC-like state and anti-PDGFRA therapy depletes the OPC-like state that is associated with PDGFRA amplification? This same principle could also apply indirectly when a particular therapy alters the TME to preferentially stabilize/destabilize a particular state such as the previously mentioned shift toward an invasive MES-like state in response to antiangiogenic treatment. If this can be proved to be the case, then one can envision treatment strategies geared to “trapping” glioma cells through targeted stress into a particular state, which is then selectively vulnerable to a specific therapeutic attack—a strategy we call “State Selective Lethality” (Fig. 5). Such a strategy would be independent of the vast levels of genomic, environmental, and anatomic heterogeneity operative in glioma, having essentially only to deal with a finite number of states rather than a near-infinite number of heterogeneity-mediated combinatory conditions, and would be applicable across a large number of gliomas, making drug discovery and personalized therapy much more practical.

• Download figure
• Open in new tab
• Download powerpoint

Figure 5.

Key principles of a “State Selective Lethality” strategy for glioma therapy. (1) Each cell occupies one of a limited set of states each with a unique combination of intrinsic therapeutic vulnerabilities and potential resistance mechanisms. Successful targeting of a vulnerable cell state will drive plastic state transitions to a resistant cell state. (2) Combination of cell state–targeted therapies will prevent plasticity-driven resistance, but concurrent use of multiple drugs may have high cumulative toxicity. (3) Novel “trapping agents” (i.e., drugs targeting chromatin remodelers) may inhibit plastic transitions between cellular states—this would eliminate a mechanism of resistance and enhance efficacy of existing therapeutics. (4) Such a trapping agent would enable sequential targeting of all tumor cell states present, and may reduce toxicity of combination therapy with multiple drugs.