Glioblastoma remains a cancer of high unmet clinical need. Since 2005, the standard of care (SOC) treatment for younger, more physically fit patients has been surgery followed by radiotherapy and chemotherapy with TMZ. Unfortunately, surgical intervention is unable to prevent GBM progress due to the difficulty in obtaining the maximal resection of all tumour material whilst minimising surgical morbidity of essential brain regions. There is some evidence that the higher the level of tumor resection, aided by the development of image-guided surgical applications such as 5-aminolevulinic acid (5-ALA) or BLZ-100 to more readily define the tumour/brain tissue border, thus improving the amount of tumour surgically removed (1, 2), the better the resulting patient outcome (3). Coupled with surgical treatment is chemoradiation that combines six weeks of IR, delivering 60 Gy in 30 fractions to residual disease with daily TMZ (4), followed by six months of adjuvant TMZ using a 5-day over 28-day schedule (5). This standard therapeutic combination has not changed for almost 20 years but inevitably tumour recurrence occurs with a median survival for paitents treated with this schedule of under two years (6, 7).

Since the adoption of TMZ, there have been considerable efforts to broaden the number and range of therapeutics available to treat GBM ( Table 1 ). However, the vast majority of systemic treatments, including all trials involving small molecules and biologics, have failed to improve patient outcomes. There are several factors prevalent in GBM driving this poor record of drug development and, consequently, the absence of improvement in patient outcomes; these include the striking range of molecular and cellular heterogeneity found in GBM tumours, the presence of both chemotherapeutic and radioresistant sub-populations of glioblastoma stem cells (GSC) in the tumour, the development and existence of protectant tumour microenvironments (TME) and of course, the existence of the blood-brain barrier which impedes the movement of small molecule and biologic therapeutics into brain tissue.

In this article, we summarize the current state of clinical trials in GBM and highlight the fact that poor clinical translation of potential GBM therapeutics could, in part, be attributed to the failure of drug development campaigns to incorporate many of the key pathophysiological features and sources of heterogeneity of human GBM into the earliest stages of the drug development pipeline. We review the steps academic research scientists are taking to address these failures and argue that high throughput screening to identify novel GBM combinatorial targets for drug development must incorporate the use of a diverse set of patient-derived GSC lines from as wide a cross-section of the GBM disease spectrum as possible, should include chemotherapeutic and radio-resistant cells, and be screened in physiologically relevant conditions including hypoxia in combination with radiotherapy +/- TMZ. Next, we discuss the rationale for current GBM combination trials is not always clear and we wish to emphasise the need for more systematic and transparent strategies for identification, validation and prioritization of combinations that lead to clinical trials. Finally, we make some specific recommendations to the preclinical, small compound screening paradigm that we feel could increase the likelihood of identifying tractable, combinatorial, small molecule inhibitors and better drug targets specific to GBM.

The degree of intra-/inter-heterogeneity of GBM tumor profiles is particularly extensive ( Figure 1 ) and has proven to be an important factor in determining long term cancer survival (67). Whole-genome sequencing of tumours has revealed extensive molecular heterogeneity within (68) and between (69, 70) patients, presenting substantial challenges to the identification and prediction of targeted (precision) therapeutic strategies. Tumours exhibit many somatic mutations in both coding (69) and non-coding regions (71) affecting both gene copy number and coding mutations, as well as regulatory elements for specific genes. However, clearly not all are driver mutations. Instead, the majority are ‘passenger’ mutations that may confer additional survival advantages under the selective pressures of drug or radiation treatment and promote clonal evolution during therapy, providing opportunities for drug resistance to occur (67).

Tumour heterogeneity and the implications of mutations on GBM pathology are seen at the genomic level when the WHO re-classification of GBM based on the mutational status of isocitrate dehydrogenase (IDH) (72) and recently completed a further re-classification to include one additional feature such as of loss chr 10 or gain chr 7, EGFR ampli, or TERT (73). IDH mutant status in GBM has a significant impact on prognosis. Although IDH-mutated glioma generally exhibit a better disease outcome (74), the incidence of IDH mutations in secondary tumour suggests that lower-grade glioma with IDH mutation often recur after having undergone malignant transformation to a higher grade. In addition, IDH-mutated glioma is more likely to develop a hypermutation phenotype (75). Ongoing research has generated a growing list of molecular and protein differences found in GBM tumours ( Figure 1 ) relating to treatment response and patient outcome (76). These mutations have been well catalogued with strong correlation across testing platforms (69, 77, 78).

An example of tumour heterogeneity at the epigenetic level occurs with the methylation status of O6-methylguanine-DNA methyltransferase (MGMT). When MGMT is silenced through methylation, it is unable to repair the DNA damage induced by TMZ (79) which has been associated with improved survival. Protein arginine methyltransferase 5 (PRMT5) gene expression in GBM is inversely correlated with survival (80). Recent studies show that pharmacological inhibition of PRMT5 suppresses the growth of GSC cultures and significantly prolongs the survival of mice with orthotopic patient-derived GBM xenografts (81). These emerging studies, together with observations that mutations in genes linked to chromatin organisation, are a common feature of GBM (69), implying that epigenetic processes may be a common driver of GBM across heterogeneous subtypes.

Another confounding factor in the resistance of GBM to chemotherapeutics is the developmental state of GBM cells within the tumour. GBM hijacks mechanisms of neural development to produce subcompartments of GSCs that exhibit resistance to radiotherapy and chemotherapies contributing to tumour-propagating potential and recurrence following treatment (82–85). In a recent study, an integrative approach incorporating single-cell RNA-sequencing, bulk genetic and expression analysis of multiple patient tumours, functional assays and single-cell lineage tracing was employed to derive a unified model of cellular states and genetic diversity in GBM (86). The authors postulate that malignant cells in GBM exist in four main cellular states that recapitulate distinct neural cell type features; Astrocyte like, Oligodendrocyte precursor cells, Neural progenitor cells, and Mesenchymal like.

Along with genetic/epigenetic inputs and GSC, the TME drives cellular heterogeneity and subsequently treatment outcomes in GBM. A major structural component of the TME is established by a complex mixture of >15 glycosylated extracellular matrix (ECM) proteins, which establishes a physical and biochemical niche which impacts the interactions between tumour cells and the ECM, including regulation of cell fate, differentiation and migration (87) and are critical to the invasiveness and malignancy of GBM. Importantly, the TME in GBM helps to create avascular regions within the tumour, which leads to decreased tumour O₂ tension and the development of a necrotic core in the tumour, a hallmark of GBM. These hypoxic conditions play a crucial role in protection against chemotherapy and radiation, likely through low O₂ content in the tumour as radiation treatment relies on O₂ and upregulation of numerous proliferative and metastatic pathways attributed to the hypoxia-inducible factors (HIF1α and HIF2α) (88). HIF has been implicated in the dedifferentiation of GBM cells driving the proneural-to-mesenchymal transition, believed to be a hallmark of resistance in recurrent tumours. Another important characteristic of the GBM microenvironment is the significant infiltration of resident microglia and peripheral macrophages. Recent research has implicated these tumour-associated microglia and macrophages in central aspects of tumour development in GBM, including proliferation, angiogenesis and immunosuppression (89), which contributes to the chemoradiation resistance and rapid progression of the disease.

The need for a combination strategy can be rationalized by the well-established cellular heterogeneity of brain tumours. There is evidence to suggest that brain tumours may arise from a stem cell origin (90). The proportion of stem cells in tumours, and the proteins they express (such as CD133 and Ki67), often determine the aggressiveness of recurrent GBM and can inversely correlate with patient outcomes (91, 92). Additionally, from a collection of experimental approaches (68, 86, 93), this stem cell origin gives rise to a proliferative hierarchy in primary GBM, leading to the presence of a range of cell states, from stem cell to differentiated cells within a tumour. Developmental and lineage subtypes invariably generate therapy-resistant populations due to higher cellular entropy (94), providing various genetic and epigenetic mechanisms stem cells use to evade death from radiation and chemotherapy treatment (95, 96).

Stem cells are increasingly becoming the target cell for further research into understanding GBM and treatment responses. The variable responses of tumour cell populations to standard GBM treatment implies that it is sensible to consider the combination of more than one drug; one of these may target stem cells and the other target non-stem cells in the bulk tumour. Data increasingly show the importance of targeting both subpopulations. For example, using patient-derived cells from biopsies, it has been reported that for greatest patient benefit, both stem cells and ‘tumour bulk’ cells need to be killed by >40% and >55%, respectively (97). Additional evidence that stem cells and bulk cell populations respond differently to drugs has been gathered from the development of a chemotherapeutics assay termed ChemoID (97). The assay involves culturing GBM stem cells and tumour bulk cells from patient biopsy material and assessing the in vitro killing ability of a panel of chemotherapeutic drugs. Significant differences in sensitivity to TMZ between stem cells and bulk cells within the same patient tumour sample ( Figure 2 ) have been observed.

In the majority of samples tested, bulk cells were more sensitive to TMZ than stem cells; thus, TMZ-treatment will not likely have significantly killed the stem cell population. Although not the ideal chemotherapeutic drug, due to insensitivity in a large proportion of patients, TMZ may be of greater value for use against bulk tumour cells, opening up the possibility of using TMZ combined with additional drugs to attack the heterogeneous tumour cell complement within the tumour mass.

The treatment regimen of surgery followed by TMZ and IR can slow the progression of GBM tumours, but it is not generally curative. The simplest solution to increasing the effectiveness of treatment would be to incorporate an additional therapeutic agent into the SOC treatment regimen to target one or more specific GBM tumour cell survival pathways to yield a combinatorial or synergistic effect. An example of a hypothesis-driven combinatorial approach has been to target tumour angiogenesis, increased tumour vascularisation supports GBM growth and survival in the face of radiochemotherapy. Vascular Endothelial Growth Factor (VEGF) is over-expressed in a variety of metastatic tumours and bevacizumab (Avastin), an anti-VEGF antibody, was first approved for treatment in metastatic colorectal cancer in 2004 (98) and later for GBM treatment in 2009 (6). Trials in in newly diagnosed GBM, such as the Avaglio trial or RTOG 0825 trial (53, 99) showed no increase in overall survival although progression-free survival times were extended by 3-4 months with bevacizumab (53, 100). To date, it remains the only approved (not in EU) antibody treatment for recurrent GBM (101). The difficulty in obtaining significant clinical benefits for anti-angiogenic therapies in newly diagnosed GBM has been summarized in retrospective meta-analyses reviews (102, 103).

Another promising approach that has had significant success in many advanced malignancies is the use of immune checkpoint inhibitors (ICI) targeting key checkpoint pathways including the cytotoxic T-lymphocyte-associated antigen 4 (CTLA-4)/B7 and programmed death 1 (PD-1)/programmed cell death ligand 1 (PD-L1), reviewed by Zhang et al., 2021 (104). Despite their successes, ICIs are yet to improve survival in GBM. By addressing the restrictive nature of the TME, the timing of checkpoint inhibition and employing novel combinatorial strategies of ICIs this strategy could still provide effective therapeutic options in the future for GBM (105).

Additional small molecules have been tested as potential treatments based on GBM driver pathways. A recent review listed 90 clinical trials with 85 different compounds either as monotherapy or in combination with standard treatment, the majority are tyrosine kinase inhibitors (6), and alternative alkylating agents (Clinicaltrials.gov). Unfortunately, so far, none of these has delivered a beneficial treatment ( Table 1 ).

Data and clinical trials from hypothesis-driven multiple drug combination therapies are less well documented than those seen for single or double additions to SOC treatment. Here, the strategy is to target multiple molecules or pathways implicated in tumour development and survival. One such hypothesis is the Co-ordinated Undermining of Survival Pathways (CUSP9) in GBM, which utilizes nine clinically approved growth-factor inhibiting drugs alongside TMZ treatment for recurrent GBM. (now called CUSP9 with slightly altered panel composition - Table 2 ) (131).

A recent study in patient-derived GSC based on the CUSP9 drug panel, highlights individually these drugs have little effect, but in combination the tumour killing effect can become significant ( Figures 3A, B ) (132).

Figure 3 : CUSP9 with TMZ. A, B Individually, the drugs in the CUSP9 or TMZ did not reduce cell survival, evaluated by the cell viability or cytotoxicity assay; a significant effect was observed when applied as a drug combination in CPCs (both p < 0.0001, one-way ANOVA). C Percentage of growth inhibition of human GAMG cells following the addition of temozolomide (TMZ), ritonavir (RITON) and aprepitant (APREP) in monotherapy; temozolomide + ritonavir, temozolomide + aprepitant and aprepitant + ritonavir in combination therapy. Adapted from Kast RE,. et al., 2015 (131) and Skaga E., et al., 2019 (132)

A further study highlights two of the CUSP9 drugs, aprepitant and ritonavir (133), tumour cell killing is either poor (aprepitant 6%) or equivalent to TMZ alone (ritonavir 14%). However, when used in combination, aprepitant and ritonavir show significantly increased synergistic inhibition which was increased when TMZ was also added for a triple combination ( Figure 3 ). The CUSP idea and initial promising results using tumour cell lines imply that successful combination therapies for GBM may well require more than two or even three drugs.

With the failure of clinical trials resulting from hypothesis-driven drug combination studies, non-hypothesis driven, high throughput screening (HTS) strategies are now coming to the forefront of academic and pharmaceutical industry research ( Table 3 ).

Traditionally in HTS, very large collections of small compound libraries with diverse chemical structures have been employed by the pharmaceutical industry to identify starting compounds for drug development programs. However, this process is generally less successful at the academic level due to the prohibitive infrastructural costs required to develop and produce larger screens, as well as the cost to develop a small compound hit into a clinical drug candidate. As a result, many academic screening facilities have instead focussed on developing more physiologically relevant phenotypic assays which better recapitulate key areas of disease biology (145) and then screen smaller, focussed libraries such as “drug repurposing” (146–148), probes for target discovery (149–151) or pharmacologically active small compound sets (152). This focussed approach has the advantage of providing a better understanding of the molecular basis of the disease while simultaneously providing the opportunity to exploit existing therapeutics and compounds with known safety profiles, small compounds that possess drug-like properties and compounds with known protein targets, and can provide a more direct path to clinical translation (153).

Academic drug screens have proven to be successful at identifying novel repurposing opportunities in a variety of other cancers (154), so are good candidates as potential additions for GBM therapy (137, 155, 156). While a number of HTS and drug repurposing screens have been performed recently in GBM, with several hits being tested in early-phase clinical trials ( Table 3 ), such screening hits have yet to translate into robust randomized phase II/III testing ( Table 1 ).

Screens using smaller but diverse small molecule libraries are also becoming more commonplace, including uncharacterized molecules that could lead to a novel therapeutic discovery program rather than just a repurposing opportunity. One example study using approximately 31,000 small molecules identified 8 compounds showing greater inhibitory effect in GSC cultures compared with non-stem cell-enriched GBM cultures (144). In addition, another two compounds inhibited four GBM hub genes ASPM, MELK, FOXM1b and TOP2a. From these xenotransplants of GSC enriched tumour cells pre-treated with four candidate compounds (Emetine, #5560509, #5256360 or OLDA) displayed significantly reduced tumour mass volume suggesting a targeted effect on the tumour initiating stem cells. However, no further studies have been done on these compounds.

For larger cancer focussed HTS employing pharmacologically active compound libraries (157) or collections of probe libraries targeting, for example, kinases (158) or epigenetic modifiers (159), the size of some of the available drug libraries is in the hundreds/thousands, resulting in an exponentially increasing the number of possible drug combinations. Therefore, screening all possible pairwise or higher-order combinations would be a huge task. For time and financial reasons, combinatorial drug screening will, therefore, realistically need to be constrained. If it is likely that one or more drugs would be required in addition to TMZ, what are the criteria that would identify a sole molecule as advantageous to use in combination? This may be an important consideration in the light of, for example, the CUSP9 data showing little activity as sole agents but significant effect in combination. The simplest and quickest screening format would be to identify candidates that already show a significant amount of tumour killing ability by themselves, and the studies described above show that potential drugs can be selected. In contrast, however, that format would not have selected any of the CUSP9 drugs which look potentially promising.

The development of clustered, regularly interspaced short palindromic repeats (CRISPR)/Cas gene-editing technologies can be deployed in the earliest stages of drug development to enable unbiased identification of new molecular targets through genome-wide CRISPR screens in cells or small model organism-based models of disease (160–162). A study using genome-wide pooled CRISPR screening across a panel of patient-derived GCS cultures was recently reported to define the molecular determinants governing GSC growth and survival (163). The screen was performed in the presence of a lethal dose of TMZ to identify genes modulating TMZ sensitivity revealing mechanisms of therapeutic resistance and new strategies for combinatorial therapy, including GSC self-renewal, GSC stemness and proliferation, as well as regulators of Cell Stress Response Pathways which all appear essential for GSC viability and fitness. Positive selection screens performed in two patient-derived GBM cell cultures treated with a lethal dose of TMZ revealed core members of the mismatch repair (MMR) pathway. To identify mechanisms of intrinsic resistance of the GSC population to TMZ, the authors performed negative selection screens in the presence of a sublethal dose of TMZ. These studies identified the Fanconi anemia pathway’ (163)

Other CRISPR-Cas9 screenings have identified further potential targets that sensitize GBM cells to TMZ, including suppressing the NF-κB/E2F6 signalling axis (164). In another study, the authors applied a pooled CRISPR-Cas9 screening approach to the U138 human glioma line in a transwell invasion assay (165). Isolation and sequencing of cells that display accelerated invasion through the transwell filter identified inhibition of mitogen-activated protein kinase 4 (MAP4K4) as a potential therapeutic target for GBM invasion (165). A further study used CRISPR interference (CRISPRi) to screen 5689 Long non-coding RNAs (lncRNAs) loci in human GBM cells, identified 467 hits that modify cell growth in the presence of clinically relevant doses of fractionated radiation and 33 of these lncRNA hits sensitize GBM cells to radiation (166). Follow up studies demonstrated that antisense oligonucleotides targeting lncGRS-1 selectively decrease tumour growth and sensitize glioma cells to radiation therapy (166). The evolution of CRISPR/Cas 9 screening, including arrayed screening strategies across more definitive GBM phenotypic assay endpoints and pooled screens to identify sensitizers of additional drug classes are well placed to identify further therapeutic targets and drug combination hypotheses in GBM.

Tumour heterogeneity and the presence of GSC alone does not explain the lack of successful clinical translation of promising preclinical discoveries in GBM ( Table 1 ) and may in fact be due to experimental reasons; chief among them is the growing realisation that currently, the in vitro and in vivo models of GBM used in preclinical drug discovery do not completely recapitulate tumour biology within patients. GBM cells and their progenitor GSC dynamically respond to their local microenvironment through a multitude of bidirectional communications resulting in the permissive conditions for tumour growth and invasion. In vitro models employed to date have sought to encompass, to varying degrees, some of this environmental complexity in the high throughput setting.

In initial experiments isolating and characterising GSCs, a neurosphere culture paradigm was used successfully with GSCs (167, 168). Further advances have seen the field introduce adherent culture grown on different base layer coatings of ECM components, including collagen and laminin. These methods, previously established for fetal and human neural stem cells, overcame limitations of neurospheres, whilst generating pure and expandable populations of GSCs, are (169) readily amenable to an HTS format and have been extensively used in drug discovery for GBM since their establishment (144, 170).

A major hallmark of GBM and a vital component of the TME that has not featured in drug discovery efforts to date is the hypoxic core of tumours. These regions result from increased cell proliferation overcoming oxygen supply (171), in turn generating an environment that supports tumour progression through GSC maintenance, proliferation, and drug resistance (172–174). In current GBM research the majority of in vitro cell models do not account for the effects of the hypoxic TME with cells cultured in atmospheric oxygen with a partial oxygen pressure of 159 mmHg. Studies quantitating pO₂ in gliomas recorded a highly significant number of pO₂ readings in patients with GBM less than 2.5 mmHg (175, 176), levels known to reduce the efficacy of radiotherapy where cellular sensitivity is halved when pO₂ is below 3 mmHg (177). Recent research generated a novel 3D in vitro model with common GBM cell lines, U87, U251, and SNB19, to study hypoxic effects in the context of TMZ resistance (178), highlighting the significant role of hypoxia on drug resistance.

The role of non-neoplastic cells in GBM is another vital consideration when attempting to model TME conditions in vitro. These consist of immune cells, including invading macrophages and resident microglia, stromal cells including astrocytes, vascular endothelial cells and pericytes (88). Microglia and astrocytes, both of which are found in abundance in the brain, accounting for approximately 30-50% and 50% of the volumes in the tumour and brain tissue, respectively (89, 179) have been implicated in glioma invasion and survival (88) highlighting the importance of their inclusion in comprehensive studies of the effects of TME on GBM. For a more indepth review refer to Decordova et al., 2020 (180). Throughput is a significant limitation on the extent to which the effects of microglia and astrocytes can be utilized; however, several academic laboratories have generated 3D co-culture models to elucidate the roles of these cell types that offer potential for adaption to a higher throughput format. One such study has generated luciferase reporter GBM cell lines with a cell viability readout on 3D floating co-culture spheroids of bioluminescent GBM tumour cells and non-luminescent rat astrocytic cells. The research demonstrated an astrocyte-mediated increase in TMZ resistance in four of a panel of six GBM tumour cell lines (181). This in vitro model was also successfully employed along with stem cells and in vivo models to demonstrate growth-supportive effects of astrocytes (182). The protective effects of astrocytes in co-culture against the main chemotherapeutics, TMZ and vincristine, have also been observed in subsequent research in 2D and 3D co-culture and demonstrated to be contact-dependent (183, 184) where drug resistance might be due to the transfer of mitochondria through tunnelling nanotubes connecting with and rescuing the tumour cells (183). Colleagues from this laboratory utilized the same 3D HA-based model with co-cultured microglia to demonstrate chemoresistance that was not present in monocultures of GBM cell lines (185).

When designing an optimized platform for high throughput combinatorial screening, balancing the need for larger screening libraries to generate novel drug candidates and physiologically relevant models to better predict drug response is crucial. In order to better model tissue-like features and cellular interactions of the TME, recent HTS efforts have utilized 3D in vitro models of GBM. High content screening against a panel of 83 chemotherapeutics in both 2D and 3D cultures in the U87 cell line highlighted shortcomings in the former’s ability to detect drug sensitivities with only those resulting from 3D cultures aligning with in vivo results (138). These findings are also supported by an independent study demonstrating TMZ sensitivity is decreased in 3D cultures of the common GBM cell lines, U87, U251, and SNB19, when compared with its 2D counterpart (178) Others have employed a 3D in vitro model with cell lines cultured in a HA-gelatin based hydrogel to model the ECM in 3D cultures, with chondroitin sulphate, chitosan and collagen/gelatine have also been utilized (186). Increased drug resistance was attributed to the co-cultured microglia and astrocytes.

Finding the appropriate balance in a GBM model between the complexity of the TME and the practical considerations of HTS is a challenge that has to be addressed for more successful translation of hits in drug discovery. In recent research, this balance may well have been achieved (187) in a study of 461 mostly FDA-approved drugs in 12 patient-derived GSCs, that incorporated both a 2D culture model in the primary HTS followed by a secondary 3D model for hit validations experiments. Results from this study demonstrate significant differences in sensitivity to hits representing multiple drug classes across patient-derived GSC models.

In complex heterogeneous tumours, no single molecular event drives continued proliferation and tumour progression, and redundancy in signalling pathways limits the efficacy of therapies targeting single pathways (188). Network and pathway switching permit rapid tumour evolution and therapeutic evasion; this requires a holistic approach to understand cancer cell signalling networks, ‘driver’ pathways, and how best to collapse the robustness of such networks to address therapeutic resistance. Thus the complexity of GBM has been likened to the “three lock problem”, describing that a door with three locks will not open any better even if keys are found to one or two locks (106). With so many different signalling pathways and cellular heterogeneity involved in tumour growth, combinations of targeted agents may well be most effective in treating rapidly evolving heterogeneous tumours, provided we can identify the network changes that permit cancer cells to subvert single agents.

In addition to the importance of the disease model and optimal drug combination delivery, equally relevant questions are; what is the rationale behind choosing specific drugs for potential combination therapies? How many drugs will be most beneficial to patients? What is the quickest and most successful way to identify these combinations? The rationale behind many drug combinations strategies tested in clinical trials studies is often unclear, and to date, unsuccessful. Even with the automated HTS instruments, exhaustive screening of drug combinations becomes quickly impractical, both in terms of time and cells required, as the number of potential drug and dose combinations increases exponentially with the number of tested drug components and dose levels. This combinatorial explosion requires intelligent computational-experimental strategies to guide the discovery of the most potent and less toxic combinations to be prioritized for further preclinical development before entering into lengthy and costly animal or clinical studies (189).

For rational combinatorial screening, there is a need for computationally efficient and experimentally feasible approaches that can (i) reduce combinatorial search space of potential combinations, (ii) and to identify both synergistic and safe combinatorial therapies for each individual patient and cell-context. For cancer-selectivity, it is important to guarantee that the combinations target stem cells and bulk tumour cells in GBM, while avoiding co-inhibition of non-malignant cells, to avoid treatment resistance and severe toxic effects. Ianevski et al. demonstrated recently the feasibility of using XGBoost machine learning (ML) algorithm together with ex vivo single-agent responses and single-cell transcriptomic profiling to identify patient-specific and cancer-selective pairwise combinations for treatment-refractory AML patients, each with different molecular backgrounds and synergy mechanisms (190). Using flow cytometry drug assay, they demonstrated that the predicted combinations resulted not only in synergistic cancer cell co-inhibition, but were also capable of targeting specific AML cell subpopulations that emerge in differing stages of disease pathogenesis or treatment regimens. Such patient- and cell subpopulation-specific combinatorial approaches may avoid overlapping toxic effects through co-inhibiting mainly malignant cell types, therefore increasing their likelihood for clinical success.

There are many experimental and computational challenges that need to be solved to implement similar translational approaches for GBM combinatorial screening. Machine learning algorithms could help in predicting full dose-response matrices using a minimal number of combinatorial experiments, thereby enabling more cost-effective, yet systematic combinatorial screens (191). Computationally, one needs to consider also higher-order combinations of more than two drugs to facilitate the current paradigm shift from the traditional ‘two drugs in combination’ to more complex ‘multi-drug cocktails’ (192). While there are a number of machine learning models for prediction of pairwise drug-dose combinations (193), accurate prediction of higher-order combination effects with more than two drugs remains an unsolved problem. A tensor learning model that enables the accurate prediction of pairwise dose-response combinations for each cancer cell line (194) has been recently developed. Higher-order tensor learning may enable generalising and learning from the currently still rather limited combination data to explore and score in an iterative manner so-far untested massive drug spaces of higher-order combinations (195), hence enabling adaptive learning of best combinations for each cell context or patient individually.

At the time of writing, a search for the terms ‘glioblastoma’ and ‘combination’ on the clinicaltrials.gov website returns 604 separate clinical trials; 29 not yet recruiting, 123 recruiting, 65 active but not recruiting, 102 suspended, terminated or withdrawn, 40 with status unknown and 242 completed. Despite the many successes of modern drug discovery strategies across several cancers, such approaches have yet to provide significant benefit for the majority of GBM patients (196). As detailed in this review, this failure can be attributed to several factors, including remarkable heterogeneity of GBM between and within individual patients, multiple redundant and adaptive pathway signalling mechanisms allowing GBM to escape from any substantial consequence of individual target perturbation, poor preclinical models, which fail to recapitulate the complex microenvironment and pathophysiology of clinical GBM. Below we outline some of the key technological advancements which are well placed to address past failures and maximize future opportunities of targeted therapies and their combinations in GBM.

DE conceived the review, wrote the manuscript and is the corresponding author. TJ is first author and wrote the manuscript. LM compiled the clinical trials tables and wrote the manuscript. JS compiled the clinical trials data. SE, GH and MF contributed to writing the manuscript and provided critical editing. TA and NC wrote the manuscript. All authors contributed to the article and approved the submitted version.