Cancer stem cell hypothesis 2.0 in glioblastoma: Where are we now and where are we going?

Abstract

Over the past 2 decades, the cancer stem cell (CSC) hypothesis has provided insight into many malignant tumors, including glioblastoma (GBM). Cancer stem cells have been identified in patient-derived tumors and in some mouse models, allowing for a deeper understanding of cellular and molecular mechanisms underlying GBM growth and therapeutic resistance. The CSC hypothesis has been the cornerstone of cellular heterogeneity, providing a conceptual and technical framework to explain this longstanding phenotype in GBM. This hypothesis has evolved to fit recent insights into how cellular plasticity drives tumor growth to suggest that CSCs do not represent a distinct population but rather a cellular state with substantial plasticity that can be achieved by non-CSCs under specific conditions. This has further been reinforced by advances in genomics, including single-cell approaches, that have used the CSC hypothesis to identify multiple putative CSC states with unique properties, including specific developmental and metabolic programs. In this review, we provide a historical perspective on the CSC hypothesis and its recent evolution, with a focus on key functional phenotypes, and provide an update on the definition for its use in future genomic studies.

Glioblastoma (GBM) remains one of the most lethal malignancies, and despite progress in other advanced cancers, prognosis of patients with GBM remains poor. GBM has long been appreciated to exhibit a high degree of cellular heterogeneity, characteristics of a developing organ, and tumor cells that appear to be in an undifferentiated state. Nearly 2 decades ago, in addition to progress in developmental biology, advances in leukemia showed the presence of a cancer stem cell (CSC) population¹ and provided a framework to appreciate these features in GBM,² which, along with breast cancer,³ was one of the first solid tumors where CSCs were identified. Over the last 20 years, the CSC hypothesis has revealed key cellular and molecular insights regarding GBM through functional interrogation. Looking back, the brain tumor field has benefited greatly from the CSC hypothesis, including the identification of new molecular regulators that drive GBM growth and therapeutic resistance, as well as from the identification of novel targets that are at various stages of therapeutic development. While the CSC hypothesis has been useful in many regards, it has not been able to reconcile the cell of origin for GBM, which could potentially include more than one cell type, and has led to debates on the cell cycle status of CSCs. This limitation in particular has proven challenging when the CSC hypothesis is applied to genomic/transcriptomic data, which intrinsically depend on expression of stemness or differentiation markers instead of functional properties. Consequently, genomic analyses have further reinforced the high degree of cellular heterogeneity in GBM, but at the same time they have also generated some confusion as to the exact identity of a CSC (or CSCs) due to the frequent inability to functionally validate these observations. Moreover, the CSC hypothesis was originally structured to identify discrete cell populations and has been adjusted to now encompass a state (or series of states) that takes into account the underlying cellular plasticity that has been revealed in many advanced cancers, including GBM. In this review, we seek to provide an overview of the CSC hypothesis in GBM and its evolution over the last 20 years, while proposing a new definition that has been adjusted to fit the next generation of genomic data and the modern state of the field.

Cancer Stem Cells in Brain Tumors: A Longstanding Controversy

Following the discovery and isolation of CSCs in leukemia,¹ the concept of a solid-tumor cellular compartment with CSC characteristics was hotly debated, from both a theoretical and experimental perspective. The theoretical debate centered on the definition of CSCs and the assays required to functionally validate this cell population. It was commonly accepted that solid tumors are often composed of cells with “undifferentiated” characteristics, but was it possible to identify a population of CSCs with a more “undifferentiated” or stem-like phenotype than other cells? Were these cells uniquely responsible for maintaining long-term tumor growth? Early work with solid tumors showed that some malignant cells had the crucial CSC property of self-renewal,⁴ or self-replication,^2,5,6 but their lineage relationships with other cells in the same tumor were not clear. Some confusion arose from the terminology associated with CSCs, as for some it implied that a stem cell was the cell of origin of the malignancy in question. The cell of origin of some tumors, including GBM, was highly debated in the 2000s and early 2010s, with several laboratories arriving at different conclusions. Competing groups identified neural stem cells (NSCs), neural progenitor cells, astrocytes, and oligodendrocyte progenitor cells (OPCs)^7–11 as the putative GBM cells of origin. Therefore, from the developmental biology perspective, the term CSC was inappropriate because the GBM cell of origin was not firmly established. For our purposes, a CSC is not necessarily defined based on the cell of origin of a tumor type but rather through rigorous functional assays that validate functional properties.

Most theoretical controversies surrounding CSCs were shaped by the inability to achieve the strict experimental criteria established by leukemia researchers to functionally define this cell population.^12,13 In the case of GBM, there was evidence that a subset of tumor cells could self-renew both in vitro and in vivo.^6–12 However, experimental platforms to infer stemness have been strongly debated. As stemness is a functional property of a cell, it should only be tested with functional assays.^14,15

A continuing challenge for CSC studies in solid tumors, including GBM, has been the lack of reliable cell-surface markers for prospective CSC enrichment. The lack of such markers has significantly contributed to both the theoretical and experimental debate surrounding the existence of CSCs in tumors including GBM. The first experimental proof of principle of the existence of GBM CSC populations identified CD133 as a marker of tumor-initiating and self-renewing cells, as shown by serial transplantation experiments.⁶ A second cell-surface marker that was used in early studies was SSEA-1/CD15.¹⁶ In both cases, these markers could enrich cells that had properties consistent with those of CSCs; with both CD15 and CD133 cells marking actively cycling cells^17,18 Over the years, several other markers have been associated with stemness, including CD44 and ID1,¹⁹ CD49f,²⁰ and CD109.²¹ However, these markers—including CD133 and SSEA-1/CD15—are expressed in highly variable fractions of cells in GBM specimens and exhibit large inter-sample variability.^6,16,22 It is therefore clear that no single-cell-surface protein considered to date can be used as a universal marker for the enrichment of CSCs in GBM samples. Additionally, it is important to extensively validate the association of a cell marker with a specific functional cellular property for each tumor sample. It is not sufficient to isolate CD133⁺ and CD133⁻ fractions, for instance, and call the former “CSCs” and the latter “non-CSCs,” because these markers are not expected to be universally applicable to all specimens. Although some markers may have wide applicability, experimental approaches must be employed to determine the information content of any marker or combination of markers for each GBM specimen. Unfortunately, many laboratories did adopt these markers at face value and used them to prospectively sort marker-positive and marker-negative cells from primary GBM samples and even from cell lines. The positive and negative fractions were then compared in functional and genomic assays, leading to confusing and often contradictory results. This has consequently confounded the field and undermined acceptance of the existence of CSCs in brain tumors. Additionally, the adoption of more sensitive transplantation models has eroded the notion that universal markers of stemness exist even in leukemia.²³

What Did We Learn About GBM Using CSC-Enriched Models?

The characterization of CSC phenotypes in GBM has strongly relied on patient-derived cell lines and xenografts, which maintain salient characteristics of the tumor from which they were derived.²⁴ It is important to note that although some patient-derived cultures and cell lines are labeled as “GBM stem cell lines,” these cultures are in fact heterogeneous populations of cells enriched for stem-like properties, but these properties are not uniform. Nevertheless, these models have been pivotal in advancing knowledge of CSC behavior in GBM. The enrichment of stem-like cells using culture conditions enabled experimental assays independently of a prior marker selection. Therefore, patient-derived models have been the workhorse of the GBM CSC field, replacing, in some instances, high-passage serum-cultured GBM cell lines that were abandoned due to the lack of cellular heterogeneity and accumulation of chromosomal anomalies.

Over the past 2 decades, a monumental amount of work using patient-derived models has highlighted key characteristics of GBM CSCs. CSCs possess potent tumor-initiating properties and are responsible for treatment resistance, invasion, cancer progression and, ultimately, cancer recurrence.^25–28 In GBM, CSC abundance is associated with poorer prognosis, treatment resistance, and more aggressive malignancies.^28,29 Currently, standard of care treatment for GBM includes targeted ionizing radiation (IR) with concurrent chemotherapy (temozolomide).³⁰ This modality leads to damaged DNA in the tumor cells, which hinders their genetic stability, proliferation, and survival. Functional studies have established a solid understanding that multiple factors contribute to CSC complexity and allow them to resist conventional treatments. CSCs can move in and out of a quiescent state,³¹ and many other factors also contribute to their complexity, including cell lineage, tumor–microenvironment interactions, and the spatial location of the CSCs. Additionally, CSCs have increased activation of the DNA damage response, which allows them to repair single- or double-stranded DNA breaks generated by conventional treatment regimens.²⁹ Further contributing to this complexity, many glioma CSC populations upregulate antiapoptotic genes such as BCL2 and MCL1, along with downregulation of proapoptotic genes such as BAX and caspase-3, which promote resistance to IR.^32–35 Another feature of CSCs that results in treatment resistance is the acquisition of genetic mutations, increasing the heterogeneity of CSCs and leading to the selection of mutations that convey resistance.³⁶ Overall, these findings indicate that CSCs represent an important clinical target, and future studies should prioritize translational efforts focused on targeting and eradicating CSCs.

The Problem With CSC Assays: Is Tumor Initiation Sufficient to Test for CSC Properties?

If CSCs can only be defined based on functional attributes, what are the experiments necessary to identify them and validate stemness? Much of the literature on functional CSC assays falls within the context of the hematopoietic system and leukemia, and most of these experimental techniques rely on the definition of “colony-forming” frequency, which is a quintessential term forged in the early studies that unraveled stem cell hierarchy in the hematopoietic system.³⁷ Based on this concept, the most widely used test of stemness is the limiting-dilution assay (LDA).⁴ Many GBM studies have tested self-renewal of cell populations using in vitro LDAs but rarely with in vivo LDAs, which are considered the gold standard in the leukemia field. In vivo LDAs rely on the concept that CSCs, by definition, must be able to initiate tumor formation, so these experiments are crucial to assess the frequency of CSCs in a sample. To perform this assay, decreasing doses of cells are transplanted into immunocompromised mice, and the cell doses that successfully engraft and produce a tumor within a given time frame are scored. This assay enables comparisons of estimated CSC frequencies between different tumor specimens, cell populations, and cells that have been genetically engineered to mutate, knockdown, or overexpress genes, for example. These experiments are time-intensive and technically onerous because each transplantation requires performing brain surgery on a mouse. However, in vivo LDAs are absolutely crucial to test CSC frequency in a specimen or cell fraction, and their adoption by the GBM community should become more widespread.

Some laboratories have used in vivo tumor initiation as a test of stemness. This led to the use of the term “tumor-initiating cells,” which was even used as an alias for CSCs by some groups. Is tumor initiation sufficient to define a cell as a CSC? Strictly speaking, no. Work in leukemia has shown that progenitor-like cells with self-renewal potential lower than that of CSCs can still engraft in immunocompromised animal models and lead to malignancies. Researchers studying leukemia have historically distinguished between CSCs and progenitor-like cells by performing serial transplantation assays.¹ Cell fractions that can sustain tumor engraftment in immunocompromised hosts over a few (usually 2 or 3) serial transplantations are deemed to possess long-term self-renewal and are therefore considered true CSCs. Cell fractions that can sustain engraftment only in the primary hosts, often with decreasing efficiency, are considered to be progenitor-like because they only possess short-term self-renewal. Serial transplantations with GBM cells are experimentally laborious but technically feasible.³⁸ Although rarely performed in the literature, this experimental paradigm would allow GBM research to move from merely assessing tumor initiation and general engraftment potential to more rigorously determining the short- or long-term self-renewal of specific cell populations. Ultimately, as performing experiments at the highest standards is the only way to provide incontrovertible data on the stemness properties of GBM cells, serial transplantations should become the standard in CSC-related GBM literature.

The Need to Revisit the Hierarchical CSC Model

The GBM field has also operated under the assumption that this tumor type is organized along a strict hierarchical model. This type of cellular organization was supported by mathematical models generated using data from single-cell viral barcoding in patient-derived xenograft settings.³⁸ This hierarchical model was consistent with CSCs producing fast-proliferating progenitor-like cells—that is, cells that are not CSCs but maintain a certain degree of self-renewal—which, in turn, produced differentiated cells destined to die. It must be noted, however, that these results did not unambiguously determine whether all the relationships in this hierarchy are unidirectional. The possibility that some cells with limited self-renewal revert to cell states that have higher degrees of self-renewal cannot be excluded based on current literature and has in fact been suggested.³⁹

Recent high-profile papers using single-cell transcriptomic (scRNA-seq) approaches to profile surgical specimens of GBM have provided a conceptual alternative to the hierarchical model. Different groups proposed their own version of the underlying organizational map of GBM, although these versions can easily be reconciled^40–48 (Figure 1). Building on previous work using spatial sampling from the same patient to assess the extent of mutational heterogeneity, groups led by Suva and Tirosh proposed that isocitrate dehydrogenase (IDH)-wildtype GBM cells exist in 4 broad cellular states: OPC-like, neural-progenitor-like (NPC-like), astrocyte-like (AC-like) and mesenchymal-like (MES-like).⁴¹ It is important to note that their original studies included both GBM and pediatric high-grade glioma specimens, which at the time were still called “pediatric GBM.” Once this distinction is made, over 30% of adult GBM cells were classified as AC-like, whereas over 30% of pediatric glioma cells were classified as OPC-like. This classification is largely consistent with other studies. The first, from the Diaz laboratory,⁴⁸ showed that GBM cells exist along a single axis of variation that spans proneural and mesenchymal states. Similarly, the Pugh and Dirks laboratories⁴⁶ found that GBM cells exist along a single axis that they called neurodevelopment-injury response. Finally, the Petrecca group⁴⁷ reported an axis that includes progenitor and differentiated cell types. The developmental signature-based classifications proposed by all these groups are remarkably congruent and highlight continuous cell states that are consistent with proneural/OPC-like states and MES/AC-like states in GBM samples.

Figure 1.

Summary of nomenclatures derived from scRNA-seq studies of GBM. Each panel represents a graphical summary of transcriptional states proposed by different research groups. The circles represent individual cells arranged in an idealized 2-dimensional representation. Proposed transcriptional states are represented in different colors. All transcriptional states are here arranged along an axis that has neural progenitor-like characteristics at one end and mesenchymal-like characteristics at the other.

GBM cell states are associated with specific genetic landscapes. For instance, OPC-like states are enriched in cells with PDGFRA amplifications, whereas the AC-like state is enriched in EGFR-amplified cells.⁴¹PDGFRA and EGFR amplifications are mutually exclusive in GBM cells,⁴⁹ and subclones with these amplifications coexist in the same tumor. Nevertheless, the abovementioned transcriptional states are not static. Some evidence from xenograft models suggests that cells of one transcriptional subtype can produce transcriptionally heterogeneous tumors in mice.⁴¹ Neftel et al. showed that tumor initiation in immunocompromised mice is a property shared by both NPC-like (CD24-high) cells and MES-like (CD44-high) cells. This work also showed that engraftment of cells from either cell state could recapitulate the distribution of cell states found in the patient’s tumor. A potential confounder in this experiment was that CD24-high cells were mostly NPC-like but also had a small fraction of MES-like cells, and vice versa for the CD44-high cells. Nonetheless, this pivotal work highlights the ability of cells to generate progeny in different transcriptional states. Parenthetically, the ability of a cell to produce progeny with different expression and functional states is postulated by the traditional CSC theory. Ideally, future work should define the ability of transcriptionally homogeneous populations of cells to generate diverse progeny. However, cellular heterogeneity is found in patient-derived cell lines even after they have been cultured for several passages, and therefore, the goal of obtaining transcriptionally homogeneous cell lines might require deriving them from individual cells.⁵⁰ Alternatively, the community should consider transplantations of single cells into immunocompromised mice. Although this is a laborious experiment, the effort would pay dividends in improving our understanding of cellular dynamics in GBM and defining the behavior of CSC populations and their properties, as was done for melanoma.⁵¹

Computational studies from the Diaz lab have indicated that MES-like cells are predicted to give rise to NPC-like cells in GBM, opening the possibility that different cell states with tumor-initiation potential can interconvert.⁴⁸ Neftel et al. also showed that approximately 15% of GBM cells display expression patterns that are intermediate between 2 cell states,⁴¹ findings supported by earlier work.⁴⁰ For example, some cells could be classified as AC-like/MES-like, NPC-like/OPC-like, or AC-like/OPC-like. The first 2 cell hybrids could be the consequence of overclustering the data, given the similarities between the AC and MES (mesenchymal) states and between NPC and OPC (proneural) states. However, the AC/OPC hybrid is particularly interesting because it could represent an intermediate between significantly different states, potentially with different stem-like characteristics; thus, these hybrids could represent cells transitioning from a stem-like state (OPC-like) to a more differentiated state (AC-like).

This idea aligns well with the findings from the Pugh and Dirks groups, who emphasized the existence of a transcriptional gradient between cell types⁴⁶ instead of markedly different groups of cells. Second, there is also good alignment with a report from the Petrecca group showing that progenitor-like GBM cells have stronger tumorigenic potential in vivo than AC-MES cells.⁴⁷ When comparing proneural and mesenchymal signatures using scRNA-seq data from developmental data sets, Wang et al. (2019) found that proneural signature genes were mostly enriched in OPCs, and the mesenchymal genes were enriched in astrocytes, as expected, but also in NSCs.⁴⁸ Both cell states therefore appear to have some transcriptional similarities with stem and progenitor cells of the normal developing human brain. Epigenetic differences between transcriptionally defined GBM cell types have been described. Single-cell DNA methylation (scDNAme) approaches found that NPC- and OPC-like cells tend to have DNA hypomethylation at known PRC2 targets compared to AC- and MES-like cells.⁵² Among the hypomethylated targets were HOX genes and other developmentally relevant transcription factors like LHX2. Integration of scRNA-seq and scDNAme data showed that NPC- and OPC-like cells tend to have signatures consistent with high activity of stem cell regulators, including SOX2 and OLIG2, whereas AC- and MES-like cells had high activity of SOX9, a transcription factor associated with astrocyte differentiation, and members of the AP-1 complex.⁵³ These observations highlight the power of single-cell genomic and informatics tools to dissect molecular underpinnings of cellular states that coexist in a tumor. However, future experiments will need to functionally test the degree by which experimental perturbation of epigenetic landscapes can shift cells from NPC/OPC-like states to AC/MES-like states and vice versa. Interestingly, Chaligne et al.⁵² found that hypomethylated promoters in NPC/OPC-like cells were associated with bivalent histone marks (H3K27me3⁺H3K4me3⁺), whereas hypomethylated promoters in MES/AC-like cells had histone marks associated with active transcription (eg H3K4me3). These results indicate that large-scale differences in chromatin structure could lead to cell state transitions, as recently demonstrated in the context of chromatin “reprogramming” induced by modulation of the levels of the histone variant macroH2A2.⁵⁴ Similarly, Johnson et al.⁵³ found general principles of DNA methylation associated with NPC/OPC-like states. They applied the concept of DNAme “disorder,” defined as the fraction of sequencing reads discordant for DNA methylation status per site or per cell, to gliomas. They found that NPC/OPC-like cells had increased promoter DNA methylation disorder than MES/AC-like cells. These findings collectively highlight large-scale epigenetic variation among classes of GBM cells. Future studies that functionally investigate whether manipulation of these epigenetic states can force cell state transitions are required, because blocking the proposed fluid nature of GBM transcriptional states could be an important weapon against therapy resistance.

Are There Multiple CSC Compartments in GBM?

Just as the GBM field sought a CSC at the apex of GBM, neural stem cell biologists worked to understand the neural stem cell at the apex of olfactory (and hippocampal) neurogenesis. To understand how this parallel field influenced GBM biology, we must take a slight tangent to briefly describe the olfactory neurogenesis that originates in the subventricular zone (SVZ). New neurons are generated throughout life within the unique microenvironment of the SVZ, a compartment that lines the dorsolateral edge of the lateral ventricles. Relatively quiescent astrocytic NSCs generate a pool of rapidly cycling progenitors that give rise to all the neuron subtypes that maintain and refine the olfactory system in the adult brain. For nearly 10 years, developmental neurobiologists assumed that this system was built on a single neural stem cell community at the top of the neurogenic hierarchy. In 2007, Merkle and colleagues used spatially restricted, anatomically precise adenoviral labeling at different points along the dorsal–ventral axis of the SVZ and discovered not one but multiple NSCs at the apex of their own, more limited neuronal lineages.⁵⁵ For example, one apex NSC generated dopaminergic neurons, while a second separate and noninterconvertible NSC generated calretinin-expressing neurons.⁵⁵ It should be noted that some controversy exists regarding the degree to which adult neurogenesis in the human brain resembles the system in the rodent brain. For example, Sanai and colleagues reported that most new neurons generated within the human SVZ migrate to the prefrontal cortex rather than the olfactory bulb.⁵⁶ However, Curtis et al. refuted this claim, demonstrating that the human rostral migratory stream presents multiple collaterals that feed different brain regions but, like the rodent system, terminates in the olfactory bulb.⁵⁷ Nonetheless, Merkle’s observations represented a seismic shift in our understanding of solid tissue stem cells. It became clear that multiple apex stem cells reside within a single growth zone and give rise to separate lineages of cells, which in turn cooperate to maintain proper function of the larger system. If multiple NSCs exist during development, is it also possible that GBM may replicate this fundamental system of cooperative but independent cellular lineages born of and maintained by separate apex CSCs? The first evidence supporting a multi-CSC lineage model of GBM was presented nearly 10 years ago when Patel and colleagues⁴⁰ interrogated GBM using single-cell RNA sequencing. These authors established 2 divergent transcriptional signatures of GBM cells. The first signature was generated using tumor cells cultured in conditions optimized for the growth and maintenance of NSCs. The second signature was generated using forcibly differentiated matched cells. The authors presented these signatures at opposite ends of a linear axis of gene transcription. They then captured the transcriptional profiles of several hundred cells from the tumors of 5 GBM patients and positioned them according to their resemblance to the CSC signature. If a single, apex CSC population was responsible for generating all the malignant cell types present within these tumors, we would expect to observe a bimodal distribution of cells along this stem-to-non-stem axis. We should find a relatively small fraction of cells that bear strong likeness to the CSC signature in contrast to most cells that would resemble the “differentiated” signature. Instead, the authors⁴⁰ discovered a gaussian distribution of transcriptional signatures along the stem-to-non-stem axis, indicating that the majority of GBM cells are neither purely stem nor purely “differentiated.” These data can be interpreted in 1 of 2 possible ways. (1) There are no stem or non-stem tumor cells present in GBM, rather all malignant cells display varying degrees of a stem cell phenotype. This interpretation does not fit well with our understanding of solid tissue stem cells, especially the NSCs of the SVZ.⁴⁶ (2) The second possibility, which was informed by our contemporary view of the subventricular neurogenic system suggested the presence of an unknowable number of cellular lineages. In this interpretation, a fate-restricted, parental CSC would give rise to a given cellular lineage, largely populated by cells of diminished, but not absent stem cell character. These findings presented a provocative and challenging view of GBM. Clearly, the binary model of stem versus non-stem was grossly oversimplified. Over the following decade, our current view of the cellular composition of these tumors would be revealed as well as the connection between the CSC phenotype, cell cycle rate, and cellular quiescence (Figure 2).

Figure 2.

Proposed mechanisms for cell state transitions. (A) In one model, mesenchymal-like and neural-like CSCs produce independent lineages. (B) Alternatively, mesenchymal-like CSC are slow cycling and the most self-renewing cell state. They can give rise to neural-like CSCs and therefore can recapitulate the full extent of cellular heterogeneity in GBM. (C) In the third model, mesenchymal-like and neural-like CSCs are functionally equivalent and can interconvert depending on their niche, that is, their state is specified by their specific microenvironment.

The field advanced just 3 years later when teams from the Universities of Florida and Bonn demonstrated the functional implications of the multilineage interpretation of Patel’s bioinformatic findings. This team, led by Reinartz and colleagues,⁵⁸ isolated and tested 33, morphologically, transcriptionally, and functionally distinct, CSC-organized malignant lineages generated from 5 to 9 morphologically unique cell types from 5 GBM patients. This work began by dissociating biopsy specimens from 5 GBM patients and plating them sparsely across large surface areas. The authors then isolated and expanded 5––9 morphologically unique cell types from each patient, resulting in 33 presumptive CSC cultures. Each of the clonal lineages demonstrated CSC functionality in vitro. However, when tested against an exhaustive panel of chemotherapeutics, the authors noted a striking degree of variability in the drug–response patterns from each CSC lineage both within and across individual patients. The situation in vivo was even more interesting and revealed additional levels of complexity. We are accustomed to thinking about GBM as a heterogeneous disease that presents numerous cellular morphologies. When individual CSC lineages were transplanted into the brains of immune-compromised mice, only a portion of histological features present in the original patient biopsy was replicated. For example, one CSC lineage generated a tumor, densely packed with small, uniformly round cells. A second lineage generated a lesion full of massive, multinucleated, giant cells. Both of these morphologies were present in the original patient biopsy, but clearly, they originated from different stem cell parents. Just as an individual neural stem cell generates a single neuron subtype, an individual CSC generated a single malignant cell type. When we examine GBM tumors, we are actually seeing an amalgam of CSC lineages, thoroughly intermixed with one another in the same anatomical space. More importantly, these authors demonstrated how these lineages are dynamically linked when the system is put under pressure by chemotherapeutic or radiation treatment. Because each lineage displays a unique drug-resistance profile, a course of treatment merely results in selection for the set of lineages with the greatest resistance to the prescribed drug (or combination of drugs). Similarly, work from the Dirks lab showed that genetically and functionally diverse CSCs coexist in a single GBM.⁵⁰

Single-cell genomic analyses have investigated how these different cell lineages interconnect, and the relationships were found to be dynamic and apparently in flux, with cells able to switch to a different cell state and then back again. This view, at least at first look, appears radically different from the CSC hierarchy proposed by earlier work in the field. However, we would like to emphasize that most of these conclusions were derived from computational analyses. Cell state transitions at steady state need to be further studied using patient-derived xenograft models that faithfully recapitulate the biological properties of patient tumors. Work from the Phillips laboratory showed that multiple CSC populations likely exist in GBM and that these cells might interconvert.³⁹ Although this original work focused on CD133⁺ and CD133⁻ cells, its findings are strongly consistent with some of the recent computational findings relying on single-cell omics technologies. These results are also consistent with data suggesting the coexistence of 2 types of CSCs in GBM: one associated with a perivascular niche and one associated with hypoxic regions of the tumor.^59,60 It is also consistent with a fundamentally new view of how GBM grows and invades the normal brain. In a recent paper, the Winkler group showed that GBM cells in the tumor core form physical networks among each other and with astrocytes, whereas GBM cells that are unconnected to other cells function as “pioneers” and invade the brain.⁶¹ The connected cells in the GBM core are predominantly of the mesenchymal type. On the other hand, the majority of pioneer cells at the tumor outskirts are of the proneural type. Are these pioneer cells the most primitive type of CSC? A recent paper from the Winkler lab suggests that proneural-type unconnected cells behave like CSCs by migrating out from the tumor and seeding new “colonies” of tumor expansion in the normal brain.^61,62 These cells then generate progeny with reduced self-renewal that acquire the mesenchymal-type cell state and form connected networks with each other and astrocytes, effectively enlarging the tumor core.⁶³ The hierarchical CSC model is not sufficient to account for the emerging data from genomic and functional experiments. New CSC models must therefore incorporate the concepts of plasticity and cell state transitions.

Cell Cycle Properties of CSCs in GBM

Quiescent cells were first identified in a model of primary leukemia stem cells. Human acute myeloid leukemia stem cells are more resistant to AraC chemotherapy-induced apoptosis, with a majority of these cells found to reside in the G₀ phase of the cell cycle.³¹ Quiescence is one piece of a complex array of elements that contribute to the complexity of CSCs. Because many conventional cancer treatments target rapidly dividing cells, the ability of CSCs to remain quiescent for periods of time allows them to escape these treatments and reinitiate tumor growth later. There are several well-established techniques to study quiescence, including nucleotide-based pulse-chase, dye-based lipid and protein labeling, histone 2B (H2B) analysis, cell cycle phase reporters, and gene promoter-driven reporters.⁶⁴ The obvious downside to several of these early methods was the requirement for fixation, which inhibits the identification and collection of live slow-cycling cancer cells for functional readout. The field then transitioned to live-cell imaging dye-based labeling techniques, such as labeling the membrane protein PKH followed by carboxyfluorescein succinimidyl ester⁶⁴ or H2B fluorescent timer (H2B-ft).⁶⁵ Although this latter technique will miss certain populations of quiescent cells, it can be used in vivo and has a longer turnover of 4–6 weeks. These techniques have allowed us to begin to understand the quiescent state in CSCs.

Computational analyses from independent laboratories studying GBM consistently found that actively cycling cells are unevenly distributed between cell states. Wang et al.⁴⁸ estimated that 21%–30% of proneural CSC-like cells and 0.3%–10% of mesenchymal CSC-like cells are actively cycling. In this study, RNA velocity analyses were used to infer cell state transitions from a fraction of the proliferating mesenchymal-like cells to proliferating proneural-like cells. This supports the hypothesis that at least 2 CSC-like populations exist in GBM—one with proneural characteristics and one with mesenchymal characteristics—with different roles in tumor growth, as discussed above.

Activation of cells from quiescence to produce other cell types has been a key tenet of the hierarchical models of GBM and is extensively supported by in vivo experiments using patient-derived xenografts and genetic mouse models.^38,66,67 Data from the Dirks group support the idea that CSCs are slow cycling.³⁸ The authors generated patient-derived xenograft models that were barcoded with a library of lentiviruses at the single-cell level. These xenografts were serially transplanted into immunocompromised mice, and the transcriptome and exome of the resulting tumors were sequenced. The barcodes allowed for the reconstruction of modes of tumor growth and associated clonal dynamics. Importantly, these experiments were agnostic to cell markers associated with stemness. The data were consistent with the most primitive CSC population being quiescent or slow cycling, and the fast-proliferating cells had limited self-renewal. In addition to studies using xenografts, recent work from the Parada group used a genetic mouse model of GBM to assess the cycling properties of CSCs, and reached equivalent conclusions.⁶⁶ The authors demonstrated that bioinformatics approaches can be used to identify subsets of quiescent CSCs, which, upon TMZ administration, exhibited high self-renewal capacity, high levels of proliferation, and potent tumor-initiating properties. CSCs exit quiescence, produce fast-proliferating cells, and then reenter quiescence. Studies such as these lay the blueprint for how functional and bioinformatics approaches can work in tandem to identify subsets of quiescent CSCs and a novel therapies paradigm to target therapeutically resistant quiescent cells. Moreover, these studies can help refine the current functional understanding of CSCs (Figure 3).

Figure 3.

Regulation of the quiescent stem cell state. Patient-derived xenograft and genetic mouse studies have demonstrated that both cell-intrinsic and cell-extrinsic mechanisms, many of which have previously been suggested to regulate the overall CSC phenotype, regulate the quiescent CSC state. Cell-intrinsic regulators include cell lineage, tumor subtype, and cell cycle properties; cell extrinsic regulators include intratumoral spatial location, cell–cell interactions, and the GBM TME.

The exact mechanisms responsible for the activation of proliferation of GBM cells upon treatment are not completely understood, but they might imply global epigenetic alterations. Data from patient-derived PDGFRA-amplified in vitro models indicate that cells that survive treatment with dasatinib are relatively quiescent and display overall decreased levels of H3K27me3.⁶⁸ It is interesting to observe that global reductions in H3K27me3 levels were also observed in treatment-resistant models of triple-negative breast cancer⁶⁹ and treatment-refractory brain malignancies, including diffuse midline gliomas⁷⁰ and PFA ependymoma^71,72 in pediatric cohorts. More in vivo functional studies are required to better map the involvement of epigenetic forces in treatment resistance and in shifting cells between quiescent and proliferating states.

These pivotal findings demonstrate that a subset of CSCs in GBM exist in a quiescent state and have the ability to transition to highly proliferative states. Although cell-tracing and imaging techniques have been sufficient to prove that cells do move in and out of quiescence, a more comprehensive understanding of how cells transition into and out of a quiescent state warrants novel technologies and approaches. Similarly, when studying quiescence, the complexity of CSCs must not be forgotten, and it should be acknowledged that multiple factors contribute to different quiescent signals and states, including the tumor microenvironment (Figure 3). For example, a mesenchymal shift was observed in quiescent CSCs of the proneural and mesenchymal subtype, with hypoxia and tumor microenvironmental factors contributing to quiescent signals in CSCs.⁷³

It is also critical to understand that a quiescent state in vitro is not the same as a quiescent state in vivo, where the (tumor) microenvironment adds an additional level of complexity (Figure 3) and contributes to the CSC state. Ample evidence exists of CSC behavior being affected by other tumor cells, like non-CSCs or “differentiated” cells,⁷⁴ and other cell types that exist in the TME. Recent in vivo work has identified how the crosstalk between intratumoral CSCs and surrounding NSCs contributes to a quiescent cell state capable of tumor initiation.⁷⁵ Further informatics studies of the genetic and epigenetic alterations present pre- and postradiation and chemotherapy in vivo are needed to better understand these quiescent properties of CSCs, while similar functional work is needed to target the specific signaling networks that characterize these cells.

Additionally, given the recent advances in cancer neuroscience, it is now apparent that synaptic-like signaling between neurons or astrocytes and glioma cells influences the proliferation of tumor cells and the growth profiles of tumors. For example, it has recently been demonstrated that a subpopulation of GBM cells has the ability to form multicellular networks; examining molecular profiles of this population revealed these cells express high cell stemness molecular features.⁷⁶ Follow-up studies have demonstrated that subpopulations of glioma cells, that are highly plastic and highly active, display rhythmic Ca2⁺ oscillations that subsequently activate MAPK-NF-κB signaling within the network of cells.⁷⁷ Additionally, GBM cells that hijack traditional neuronal mechanisms to enhance cell invasion mechanisms have been shown to have neuronal and neural-progenitor-like cell states, closely linking CSC phenotypes with neuronal activity.⁶¹ Future experimental work that explores the role of neural signaling between tumor cells and the TME, and elucidation of downstream molecular mechanisms that control entry and exit from quiescent states, would significantly expand the ability for therapeutic development.

An Updated Definition of CSCs

Here, we would like to propose an updated definition for CSCs in GBM. This definition has to be simple, taking into account the new developments in the field, while also incorporating high experimental standards that can be applied to the expanding multiomics landscape of research. CSCs are cells endowed with long-term self-renewal that can produce progeny with a variety of cell states. By a “variety of cell states” we mean that cell progeny needs to have a variety of functional (eg, exhibiting more or less self-renewal) and molecular profiles. Incorporating the idea of a GBM CSC that gives rise to multiple cell states better represents the dynamic cellular organization of GBM and better reflects what we know about intratumoral heterogeneity than does the term “multilineage differentiation,” given that GBM cells are incapable of true differentiation. CSCs are not giving rise to a rigid hierarchical cellular structure in GBM. Rather, cellular plasticity and cell state transitions need to be taken into account when assessing stemness properties. At the same time, we want to stress the need for more widespread adoption of traditional functional assays for the experimental determination of stemness and self-renewal properties. Long-term self-renewal must be tested with serial transplantations in vivo. The ability of a marker to enrich for self-renewal needs to be tested through in vitro and in vivo LDAs.

We envision that future technological and technical advances will improve our understanding of CSC function in appropriate environmental contexts. First, the adoption of improved mouse models may enable more precise measurements of CSC activity in cellular milieus that more closely resemble the human context, thus refining our understanding of this cell population in GBM. For instance, xenotransplantation in immunocompromised mouse models that have been reconstituted with human immune cells could provide significant benefits over the more standard use of NSG or NOD/SCID mice. These humanized mouse models would enable the study of CSCs through functional experiments in mammalian brain environments complemented with human immune cells. Similarly, in vitro functional assays with organoid models that incorporate cells normally found in the neural microenvironment might represent a resource to be more fully explored in the neuro-oncology field. Second, the adoption of spatially defined genomic technologies may be impactful, allowing the study of CSCs in their appropriate environmental contexts in surgical specimens and in xenograft models. These approaches promise to allow the reconstruction of interactions between TME and cancer cells in native environments. Spatial RNA-seq has already been used to explore GBM samples and upcoming approaches to expand these studies to epigenetics and metabolites will be transformative. However, we would like to conclude that irrespective of the advancements in genomic and bioinformatic approaches, ultimately CSC properties must be defined with functional experimental platforms.

Cell plasticity and cell state transitions represent a further challenge that must be overcome to achieve better therapeutic strategies for GBM. Incorporation of this concept in preclinical studies could lead to more faithful models of GBM and significant changes in data interpretation that, hopefully, will alter current clinical paradigms.

Conflict of interest statement

None declared.

Funding

Work in the Lathia laboratory is supported by National Institutes of Health (NIH) [grants P01 CA245705 and R35 NS127083], the Lerner Research Institute, and the Case Comprehensive Cancer Center. D.J.S. is supported by an American Brain Tumor Association Discovery Grant. A.R.S. is supported by a Cancer Biology Training Grant (NIH T32 CA059366) and a Midwest Brain Tumor Foundation Postdoctoral Fellowship award. The Gallo laboratory was supported by a Project Grant and a Project Grant–Priority Announcement in Pediatric Cancer Research from the Canadian Institutes of Health Research.