EGFRvIII expression initiates progressive gliomagenesis
To study the role of mutant EGFR in gliomagenesis, we generated double heterozygous mice carrying a conditional human EGFRvIII transgene [12] and cre under the control of the Nestin promoter [27] (nes-cre) which specifically activates EGFRvIII expression in the central nervous system, (Fig. 1a, b). EGFRvIII; nes-cre mice were born at expected frequencies with structurally normal brains. By 60 weeks of age, 100% of mice had succumbed to brain and/or spinal tumors (n = 48) (Fig. 1c).
Examination of brains prior to clinically overt disease (mice aged 12–28 weeks) revealed focal cellular expansions in the subventricular zone (SVZ) and small glioma precursors with proliferative activity, also described as “microneoplasias” [28, 29] (12/12 mice, Additional file 1, Fig. S1). Multiple lesions were detected bilaterally protruding into the lateral ventricles and from the brain surface (Fig. 1d). These microneoplasias expressed markers of neural stem cells and transit-amplifying cells, specifically Sox2, Nestin, PDGFRa, GFAP and Olig2, (Fig. 1g, Additional file 1: Fig. S2).
Mice later developed neurological signs due to one or multiple gliomas within the lateral ventricles and/or brain surface with evidence of subarachnoid involvement (40/48 mice had brain gliomas; mean survival 41.1 weeks); immunostaining for human EGFR and EGFRvIII confirmed EGFRvIII expression specifically and clonally in tumor cells of microneoplasias and gliomas (Additional file 1: Fig. S3, S4). These tumors had histological features comparable to those of human gliomas, Additional file 1: Fig. S5; a small proportion displayed necrosis and microvascular proliferation, characteristic of GBMs, Additional file 1: Fig. S6. The proportions of grade II, III, and IV tumors and their proliferative indices (Ki67) are shown in Fig. 1e, f and Additional file 1: Fig. S7. A histopathological difference between these mouse and human gliomas however is that EGFRvIII is largely found in human GBMs whereas the majority here are mouse LGGs, as observed in some other EGFRvIII mouse models [21].
Overall, these results show that EGFRvIII can initiate gliomagenesis in the brain, with the long latency reflecting the need for secondary mutations.
EGFRvIII drives spinal cord gliomas
In addition to brain tumors, EGFRvIII; nes-cre mice also developed multiple and widespread spinal tumors with 100% penetrance (48/48 mice, Additional file 2: Table S1), causing neurological deficits including limb weakness and ataxia. The tumors were located on the spinal cord surface and locally invaded surrounding soft tissue, nerve roots, and cranial nerve ganglia (Additional file 1: Fig. S8). At an advanced stage of tumor progression, tumor cells were seen invading the parenchyma, reminiscent of intramedullary spread of spinal astrocytomas in humans. They were present throughout the leptomeningeal space indicating leptomeningeal spread, a poor prognostic indicator in patients [30]. Tumors showed striking resemblance in histology and location with human leptomeningeal-disseminated spinal gliomas.
In 8/48 mice without established brain tumors (but with microneoplasia), there were widespread spinal tumors. Histology of spinal tumors universally classified them as grade II glioma even in the presence of grade IV intracranial gliomas, suggesting these are primary spinal gliomas, most likely arising independently (Fig. 1e). These spinal tumors expressed classical glioma markers, such as GFAP, Sox2, Olig2, and PDGFRa, and had a lower proliferative index than the corresponding brain tumors (Fig. 1f, g). These EGFRvIII-induced spinal tumors represent, to our knowledge, the first mouse model of spinal gliomas with leptomeningeal dissemination.
Whole-exome sequencing reveals the mutational landscape
To identify somatic mutations and copy number changes acquired after glioma initiation by EGFRvIII, we performed whole-exome sequencing (WES) on 17 tumors (9 brain and 8 spinal gliomas). To increase power for detection of recurrent mutations, WES analysis was performed on the pooled group of gliomas from all CNS compartments [31]. Across all tumors, we found 85 significant recurrently mutated genes with mutations in two or more tumors identified by MuSiC [32] (adapted for mouse data); most had single-nucleotide variants (SNVs) but some genes exhibited INDELS (Fig. 2a, Additional file 3: Table S2). The median number of exonic mutations per tumor was 29 of which missense mutations were the most common. Sub1, a transcriptional coactivator, was the most frequently mutated gene (6 mutations in 5/17 tumors, p = 1.1 × 10−16, FDR 2.27 × 10−12, likelihood ratio test, LRT) displaying INDELs and SNVs, all in splice sites suggesting loss of function. Trp53, a known tumor suppressor in human LGG and GBM [33], was the second most frequently mutated gene (5/17 tumors had a Trp53 missense mutation, all within Trp53’s DNA-binding domain; p = 1.13 × 10−12, FDR 7.75 × 10−9, LRT; Additional file 1: Fig. S9), validating the application of WES to identify relevant collaborative mutations. Similarly, Nf1, a known genetic driver of brain and spinal gliomas [34], was found to be mutated in two tumors (p = 0.0010, FDR 0.17, LRT). Other frequently mutated genes were Tead2, Nt5c2, Ces1c, Prex2, Uimc1, and Itga6. Tead2, a transcription factor in the Hippo pathway, had recurrent mutations across its TEA/ATTS (DNA-binding) domain (4 mutations in 3/17 tumors; p = 2.80 × 10−11, FDR 1.15 × 10−7, LRT), including splice site mutations and one frameshift mutation, suggesting loss of function. Uimc1 and Itga6 had three mutations each (p = 1.39 × 10−7 and FDR 1.9 × 10−4, p = 2.7 × 10−7 and FDR 3.2 × 10−4, LRT, respectively), all of which were INDELS and one of which caused a frameshift in Itga6 (Fig. 3i). These gliomas were all wild-type for Idh1, consistent with gliomas in humans in which IDH1 and EGFR mutations tend to be mutually exclusive.
In contrast to the relatively small number of recurrent mutations, EGFR-mutant tumors had complex genomes by DNA copy number analysis (Fig. 2b). Significant focal amplifications and deletions, identified by GISTIC2 [35], were evident in regions with known cancer genes, for example, significant focal Cdkn2a deletions (GISTIC q value = 1.39 × 10−5) were evident and EGFRvIII (in Col1a1 locus, GISTIC q value = 0.017) was recurrently amplified. Significantly recurrent focal deletions were present in a novel putative glioma driver Adgrl2 (GISTIC q value = 2.19 × 10−6, Additional file 4: Table S3). Several of the most significantly mutated genes were also in regions with frequent deletions, including Trp53, Tead2, and Uimc1, supporting putative tumor suppressive roles (Fig. 3i).
The significance and translational relevance of the most frequently mutated and/or focally deleted genes detected in mouse gliomas were assessed by comparison with human glioma datasets from The Cancer Genome Atlas (TCGA; n = 283 LGGs, 273 GBMs) [36, 37]. This revealed that TEAD2 is recurrently deleted in 48% of human LGGs in a mutually exclusive manner with TP53 (Bonferroni-adjusted p < 0.001, Fisher’s exact test, Additional file 1: Fig. S10). Recurrent deletions in previously unknown glioma genes NT5C2, ADGRL2, and UIMC1 were observed whilst SUB1, CES1, and ITGA6 were frequently methylated in human LGGs (Additional file 1: Fig. S10); frequent CNVs in these genes were also present in human GBMs. Subgroup analysis confirmed recurrent mutations/CNVs (> 2 tumors) in these genes specifically in EGFR-mutated/amplified human LGGs and GBMs. These data cross-validate the relevance of these novel putative drivers in humans.
Transcriptomic profiling defines glioma oncogenic pathways
To delineate the signaling pathways deregulated in tumors, we performed RNA-sequencing (RNA-seq) on 11 EGFRvIII-expressing mouse brain gliomas and 10 spinal gliomas.
Compared with normal mouse brains (n = 6), EGFRvIII-brain gliomas show 2000 upregulated and 1784 downregulated genes (log2 fold change > ± 2 and Benjamini-Hochberg adjusted p value < 0.01, Additional file 5: Table S4). Gene ontology (GO) analysis of upregulated genes showed a significant enrichment for genes related to the cell cycle and mitosis, differentiation, and neurogenesis (FDR < 0.001). Downregulated genes showed enrichment for pathways such as neuron differentiation and migration (FDR < 0.001). Gene set enrichment analysis (GSEA) of differentially expressed genes in EGFRvIII-brain gliomas showed significantly enriched gene sets (p < 0.01) including p53, Wnt, MAPK, Jak-Stat, Rb pathways, and stemness, implicating these oncogenic pathways in driving gliomagenesis in cooperation with EGFRvIII (Fig. 2c, Additional file 1: Fig. S11).
The most “upregulated” gene was the EGFRvIII transgene, but as this human transgene is not present in normal tissue, fold change is not meaningful. The endogenous Egfr gene was also upregulated (mean log2 fold change = 3.71) in both brain and spinal tumors, suggesting both mutant EGFR and wild-type Egfr expression are advantageous to tumor growth (Fig. 2d), consistent with previous reports suggesting collaboration between the two as observed in human GBMs [38]. The majority of the top mutated genes are also expressed, including Sub1, Trp53, Tead2, Nt5c2, Prex2, Uimc1, and Itga6.
Hox (homeobox) genes have been implicated in escape from apoptosis, epithelial-mesenchymal transition, and angiogenesis in other cancers [39]. Nineteen of the 30 most strongly upregulated genes in the brain tumors were Hox genes (Benjamini-Hochberg adjusted p < 1 × 10−12, Fig. 2e), and these top genes associate with patient survival from human GBM TCGA data (Additional file 1: Fig. S12). Comparative analysis with large human GBM (Sun, Murat, and TCGA [37, 40, 41]) datasets revealed 14 of these most upregulated Hox genes in mice are also upregulated in human tumors, supporting a proposed role in oncogenesis [42, 43], Fig. 2f. In contrast, spinal tumors did not show such strong upregulation of Hox genes, although they did exhibit enrichment for the other oncogenic pathways (p < 0.01, Additional file 1: Fig. S13, S14, Additional file 6: Table S5).
Human gliomas may be classified according to gene expression profiles [44]. Comparison with human glioma subsets (Verhaak dataset) using GSEA revealed these mouse tumors showed strongest enrichment for the human mesenchymal GBM signature (q value < 0.01), although there was also weaker enrichment for the proneural and classical GBM signatures and negative enrichment for the neural GBM signature, Fig. 2g. Therefore, these tumors recapitulate key molecular features of a clinically relevant human GBM subset.
Transposon mutagenesis replaces genomic instability in glioma progression
Transposons have been successfully used for identifying cancer driver genes [45,46,47,48,49,50,51,52,53]. Mobilized piggyBac transposons randomly integrate in the genome and activate and/or disrupt gene expression [54]. Given large chromosomal aberrations or transcriptional changes make pinpointing driver genes difficult to identify, we performed a conditional piggyBac transposon mutagenesis screen in vivo to further identify genes that cooperate with mutant EGFR in gliomagenesis.
To limit transposition to the central nervous system, a conditional piggyBac transposase allele was activated by nes-cre (Fig. 3a, b). An experimental cohort of quadruple transgenic mice carrying conditional EGFRvIII, 20 copies of a piggyBac transposon (ATP1S2) [54], a conditional piggyBac transposase, and nes-cre were generated (EGFRvIII-PB, n = 72; Fig. 3b, see Methods). As controls, we established transgenic mice expressing EGFRvIII but lacking transposition (EGFRvIII; nes-cre = EGFRvIII-only, n = 48) and a set with transposition but lacking EGFRvIII (transposase; ATP1S2; nes-cre = PB-only, n = 20). Mean survival times between EGFRvIII-PB and EGFRvIII-only cohorts were similar (41.4 vs 41.1 weeks, p = 0.95, log-rank test), and both groups had similar incidences of brain and spinal gliomas (Fig. 3c, d, Additional file 1: Fig. S15, S16, S17). There was a trend towards increased GBMs in EGFRvIII-PB mice compared with EGFRvIII-only mice (13.9% vs 4.2% GBMs respectively; p = 0.082, two-sided chi-square test).
Genomic instability is a hallmark of cancer (including human gliomas) and a key driving force [55,56,57,58]. EGFRvIII has also been associated with genomic instability in vitro [59]. We hypothesized that the absence of reduced survival times of EGFRvIII-PB mice may reflect genomic instability providing secondary molecular alterations in EGFRvIII-only mice that is similar in consequence to transposon mutagenesis in EGFRvIII-PB mice. Supporting this, cytogenetic analysis revealed significantly more chromosomal aberrations in EGFRvIII-only compared to EGFRvIII-PB tumors (19 vs 6.4 mean number of chromosomal aberrations, p = 0.013, unpaired two-tailed t test; Fig. 3e, f). Whole-exome sequence of 20 brain and spinal gliomas from EGFRvIII-PB mice confirmed these had substantially less complex tumor-genomes with fewer copy number changes than EGFRvIII-only tumors (Fig. 3g). Nevertheless, whole chromosome 11 amplification was still common as well as focal amplifications of EGFRvIII (Col1a1 locus) and localized deletions in Cdkn2a and Adgrl2 in tumors arising from both cohorts. GISTIC2 analysis shows these alterations occur significantly more frequently than expected by chance (q value < 0.05; Additional file 1: Fig. S18, Additional file 4: Table S3), suggesting they provide a selective advantage for tumor progression.
Whole-exome sequence analysis revealed that while the median number of mutations was similar between the cohorts, their mutational profiles differed substantially. The top mutated genes identified in the EGFRvIII-PB tumors were Obscn, Hspg2, Rrbp1, Rpgrip1, and Atp5o which have unknown functions in cancer (Fig. 3h). Although the frequency of mutations in these genes was high (70–40%), Obscn and Hspg2 are particularly large genes (more likely to harbor mutations) and contained many synonymous changes, suggesting they were passengers. Nevertheless, in EGFRvIII-PB mice, there were low-frequency mutations in a subset of putative drivers we previously identified in EGFRvIII-only tumors, including frequent splice site mutations in Sub1 and Nt5c2, and mutations in Trp53, Tead2, Uimc1, and Itga6 (Fig. 3i).
We hypothesized genomic instability may be generated through oncogene-induced replicative stress [60]. We studied H2AX phosphorylation by immunostaining, which marks sites of DNA damage (focal nuclear staining) and replication stress (pan-nuclear staining) [61]. Mouse EGFRvIII-GBMs displayed large areas with a substantial fraction of cells showing intense pan-nuclear γ-H2AX and others with γ-H2AX foci (Additional file 1: Fig. S19a, b). Gene set enrichment analysis of RNA-seq data from these tumors revealed significant enrichment for upregulated gene sets involved in DNA repair, double strand break repair, base excision repair, and DNA damage checkpoints (Additional file 1: Fig. S19c). Specific DNA repair genes significantly upregulated in these tumors include Chek2, Xrcc2, Xrcc4, Ercc2, and Foxm1. These data are consistent with a model of oncogene-induced replication stress leading to genomic instability and activation of the DNA damage response (DDR), previously proposed for other oncogenes such as K-ras [62].
Together, these results suggest that piggyBac mutagenesis substitutes for genomic instability and highlight the relevance of transposon-mediated mutations for gliomagenesis. Replacing large chromosomal anomalies with precise genetic hits enables functional genomic interpretation.
Transposon mutagenesis identifies EGFR-mutant glioma driver landscape
To identify the genetic driver landscape with piggyBac, common integration sites (CIS—genes into which the piggyBac transposon has recurrently inserted more frequently than expected by chance, p < 0.01) were identified by transposon-host PCR [63] and sequence analysis (quantitative insertion site sequencing, QI-seq). Gaussian kernel convolution was used to identify CIS from 46 brain and 50 spinal tumors [50]. Brain and spinal tumors from the same mice had different transposon integration sites, confirming these tumors arose independently. In total, 281 significant CIS genes were ranked according to the number of insertions across all tumors (Fig. 4a, Additional file 7: Table S6). Pathway analysis using Gene Ontology and Panther [64] revealed that CIS genes were enriched for oncogenic pathways including Ras-MAPK, Wnt, PI3K-AKT, and stem cell-related pathways (Additional file 1: Fig. S20). Analysis of CIS genes with STRING [65] showed PB mutagenesis significantly enriched for mutations that affect a functionally interacting network of proteins in gliomagenesis (Benjamini-Hochberg adjusted p = 4.9 × 10−13, hypergeometric test, Additional file 1: Fig. S21).
The highest-ranked CIS was Cdkn2a, followed by Nf1 (Fig. 4b). Loss-of-function mutations of CDKN2A and NF1 have been observed as drivers in a range of human gliomas including LGG and GBM [66, 67]. Interestingly, Spred1 (whose product also acts as negative regulator of the Ras pathway [68] and is a recently discovered melanoma tumor suppressor [69]) ranked within the top 10 CIS and exhibited a disruptive piggyBac insertional pattern, suggesting Spred1 acts as a novel tumor suppressor in glioma (Fig. 4c). Other MAPK signaling-related genes with recurrent mutations include Prkca, Pebp4, and Map3k1.
Genes involved in the PI3K-AKT oncogenic pathway were also identified including known tumor suppressor genes in gliomas such as Pten [70] and Pi3kr1 [71] as well as novel genes including Prex2, and the protein tyrosine phosphatases Ptpro and Ptprj, all with inactivating transposon insertional patterns. The glioma oncogene and PI3K-AKT activator, Pdgfrα [72], was also a CIS, with an insertional pattern consistent with gene activation (Additional file 1: Fig. S22). This supports the validity of our transposon screen in identifying both tumor suppressor genes and oncogenes. Other genes involved in the PI3K pathway with recurrent insertional mutations include Cbl and Pik3c3.
Several top CIS genes known from their function in nervous system development were not previously recognized as tumor suppressors. Sox6 and its paralog, Sox5, are expressed in a mutually exclusive pattern during brain development [73]—both were identified as CIS. Tcf12 and Tcf4, transcription factors implicated in neurogenesis [74], were also identified as CIS. Nav3, a gene belonging to the neuron navigator family predominantly expressed in the nervous system, had recurrent insertional mutations too. NAV3 silencing in breast cancer cells increased tumorigenicity in a xenograft model, supporting our data here for gliomas [75]. Inactivating transposon insertion patterns suggest tumor suppressor roles for these genes (Fig. 4c). Frequent insertional mutations were also observed in other genes with developmental roles: Qki, Zeb2, Dmd, Zfhx3, Zfhx4, and Exosc9.
To explore the evolutionary mechanisms underlying brain gliomas in our mouse model, we performed multi-region tumor sampling and QI-seq. This revealed intratumor heterogeneity, with clonal and subclonal piggyBac insertions, implying branching tumor evolution (Additional file 1: Fig. S23). With the exception of clonal Pdgfra and Nav3 insertions in one tumor, transposon insertions in MAPK/PI3K pathway and neurodevelopmental genes (including Nf1, Pten, Pik3r1, Ptprj, Sox6, Sox5, and Tcf4) were subclonal in these tumors, implying these were late evolutionary events. Altogether, piggyBac mutagenesis has comprehensively identified known and novel putative cancer genes and pathways driving EGFR-mutant gliomas.
Comparative validation of CIS genes with human TCGA gliomas
To assess the clinical relevance of the putative glioma driver genes, we analyzed the frequency with which genetic alterations occur in our top CIS genes in 283 human brain LGGs and 273 GBMs from TCGA datasets [36, 37]. Aside from the known brain glioma tumor suppressors, CDKN2A, NF1, and PTEN, we found SPRED1 is deleted (heterozygous or homozygous) in 12% of LGGs and 27% of GBMs; and TCF12 deletions and/or truncating mutations are present in 15% of LGGs and 23% of GBMs—indeed SPRED1 and TCF12 are mostly co-deleted (p < 0.001, Fisher’s exact test) likely as part of a 15q deletion [76]. SOX6 is deleted with high frequency: 31% of LGGs and 18% of GBMs, Fig. 5a and Additional file 1: Fig. S23. Subgroup analysis confirmed these top CIS genes had recurrent mutations/CNVs (> 2 tumors) in EGFR-mutated/amplified human LGGs and GBMs.
QKI, UST, PPP1R14C, and MAP7, all mapping to chromosome 6q, are frequently co-deleted in both human LGGs and GBMs (Bonferroni-adjusted p < 0.001, Fisher’s exact test; Fig. 5a, Additional file 1: Fig. S24). In mice, all four genes had recurrent piggyBac insertions across their sequence (implying gene disruption), supporting the hypothesis that there are multiple putative tumor suppressors in this region [77]. Similarly, EXOSC9 and CLCN3 are frequently co-deleted on human chromosome 4q and both had disruptive transposon insertions in mice. These data illustrate the potential utility of piggyBac in pinpointing cancer drivers hidden within large copy number altered regions.
To understand the clinical relevance of top mutated novel genes, we analyzed the REMBRANDT [78] and TCGA GBM datasets for correlation of gene expression with patient survival (n = 329 and n = 348 tumor samples respectively): expression levels of SOX6, UST, QKI, PPP1R14C, TCF12, SPRED1, TEAD2, and NAV3 significantly correlated with patient survival in one or both of these independent datasets (p < 0.05, log-rank test, comparing patients with upper 30% vs lower 70% of expression levels, Fig. 5b–g). Moreover, deletions in these genes associate with correspondingly lower gene expression (Additional file 1: Figure, S25). Altogether, these results further support roles for these genes in gliomagenesis.
Validating the effects of transposon insertions from glioma transcriptomes
Transposition results in fusions with endogenous genes that can be detected by RNA-seq [79]. To produce direct evidence of piggyBac insertions affecting transcription of target CIS genes, we performed paired-end RNA-sequencing of 36 brain and spinal gliomas from EGFRvIII-PB mice and implemented IM-Fusion to detect gene-piggyBac fusion transcripts [80], Fig. 6a, b. Of the 281 CIS genes identified by QI-seq, 80 had supporting fusion transcripts from RNA-seq analysis (4.43 × 10−11, two-sided Fisher’s exact test, Fig. 6c). Moreover, 16 of the top 20 CIS genes had supporting fusion transcripts from at least one tumor, including Cdkn2a, Nf1, Pten, Sox6, Sox5, Spred1, and Tcf12 (all containing PB splice acceptor fusions, implying transcript termination; Fig. 6d). Other key genes with fusion transcripts suggesting disruption included Qki and Ust (Additional file 8: Table S7). Of the genes with the most fusion transcript sequencing reads containing PB splice donor (implying activating insertions, see Fig. 6b), Rad51b was also a CIS gene (Fig. 6e); its fusion transcripts found in two tumors imply a putative oncogenic role, supporting data demonstrating RAD51 inhibition radio-sensitizes gliomas by reducing DNA repair [81]. These transcriptomic signatures of piggyBac support the functional effects of the identified CIS genes on gliomas.
Brain and spinal tumors share core genetic drivers
Of the 281 CIS genes, 206 (73%) were shared by both brain and spinal tumors, Fig. 7a. These include known tumor suppressors underlying multiple types of human gliomas, such as Cdkn2a, Nf1, and Pik3r1, as well as several putative tumor suppressors such as Sox6, Tcf12, and Spred1. However, the frequency of insertions in particular shared genes differed between brain and spinal tumors. For example, Pten had significantly more insertions in spinal than brain tumors (22 vs 8 insertions respectively, p = 0.008, Fisher’s exact test). Conversely, Sox6 has significantly more insertions in brain compared with spinal tumors (26 vs 3 insertions, respectively, p < 0.0001, Fisher’s exact test; Fig. 7b and Additional file 1: Fig. S26). Other CIS occurred uniquely in each tumor type, for example, Pdgfra had activating insertions in brain but not spinal tumors (4 and 0 insertions, respectively). Although CIS genes with lower frequency insertions require further characterization to confirm their tumor-type specificity, collectively, these results show there is a shared core set of driver genes for both brain and spinal gliomas.
Differential tumor suppressive effects of Pten in brain and spinal gliomas
PTEN loss is a common event and known to cooperate with EGFR in brain gliomas but its role is unclear in spinal tumors [83], with no previous mouse models (to our knowledge) showing whether Pten drives spinal gliomas. Pten was a CIS in both brain and spinal gliomas, Fig. 7c. To explore the role of Pten inactivation on brain compared with spinal gliomagenesis, we generated triple transgenic mice carrying the conditional allele of EGFRvIII, nes-cre and a conditional knockout Pten allele [82], PtenLoxp/+ (n = 11; Fig. 7d).
EGFRvIII; nes-cre; Pten+/− mice developed signs of spinal (focal paralysis) rather than brain disease (hydrocephalus or seizures) and showed a reduction in survival time compared with mice just carrying the EGFRvIII and nes-cre alleles (median age 13.0 vs 41.1 weeks, p < 0.001, log-rank test; Fig. 7e). Histological examination of EGFRvIII; nes-cre; Pten+/− mice identified extensive grade II gliomas surrounding the spinal cord at all levels with widespread leptomeningeal and nerve root invasion (9/9) (Fig. 7f, g). Of lesser clinical significance, microneoplasias in the SVZ and base of brain were observed. These data identify Pten as a novel spinal glioma tumor suppressor and suggest a stronger cooperative driving effect of Pten loss on spinal compared with brain tumors, highlighting context-dependent tumor suppressive effects (Fig. 7h).
Druggable targets in the glioma driver network
Knowledge of cancer driver landscapes presents opportunities for therapeutic strategies. Using canSAR [84], we have applied established chemogenomic technologies to pharmacologically annotate the glioma set of putative driver proteins identified here. This set was derived from all CIS genes (EGFRvIII-PB cohort) and from recurrent significantly mutated genes (EGFRvIII-only cohort); given loss-of-function (LOF) of several proteins identified directly lead to Akt activation (e.g., Pten, Ptpro, Pik3r1) [85,86,87], and Ras/Erk/Mek activation (e.g., Pdgfra, Nf1, Spred1) [88,89,90], these linked downstream oncoproteins were included as targets. The glioma set thus comprised 375 proteins. Each protein was assessed in multiple ways for “druggability” (probability of the protein being targeted by small molecule drugs). Comparative genomic analysis with human LGG and GBM data from TCGA confirmed all druggable genes in the set, except Ddx3y and Usp9y, are genetically altered in patients. CanSAR analysis revealed a highly druggable network of putative drivers, with 14 targets of approved drugs (for other indications), 3 targets of clinical investigational drugs, 34 targets under drug discovery or chemical biology investigation, and 96 proteins predicted to be druggable and thus of potential interest for future drug discovery efforts. In addition to targeted EGFR therapies, the network highlights targets being investigated clinically for glioma treatment, including not only PI3K, but also ESR1 and PDGFRA, Additional file 9: Table S8 and Additional file 10: Table S9.
Next, to validate the potential therapeutic effects of targeting these proteins with drugs, we analyzed large-scale drug sensitivity data from 21 human glioma cell lines (including EGFR-mutated and wild-type; 13 GBM, 8 LGG; GDSC [91]). Twenty-four drugs acting on our glioma network were tested by GDSC, of which 9 demonstrated significant growth inhibitory effects (IC50 Z score < − 2) and 15 showed partial inhibitory activity on at least one cell line, Additional file 11: Table S10. These results highlight potential efficacy of drugs targeting PI3K, AKT, MEK, ERK, EGFR, and PDGFRA, as well as APP, ESR1, SMARCA2, HDAC9, AURKC, and NAMPT in selected gliomas. Further testing is additionally required in genetically faithful models for drug sensitivity, but blood-brain barrier penetration is a challenge that will need to be overcome to realize the clinical potential of these observations. Nevertheless, such orthogonal demonstrations of functional genes and targets are essential for prioritizing potential therapeutics for preclinical and clinical trials.
EGFRvIII-glioma cells can serially engraft in recipient mice and are suppressed by afatinib
EGFRvIII-glioma cellular models are needed for pre-clinical studies. Cells from EGFRvIII-mouse GBMs were expanded ex vivo as gliomaspheres. We aimed to delineate the engrafting capacity of EGFRvIII-driven tumor cells as further evidence of their neoplastic nature by subcutaneous injection in the flanks of severe combined immunodeficient (NOD-SCID-γ) mice (Fig. 8a); subcutaneous rather intracranial injection was chosen as previous studies show the tumorigenic potential of mouse gliomas is equivalent for both methods [22] and tumor growth assessment was simplified. Tumors formed and were harvested within 20 days of injection in all mice (n = 6, Fig. 8c), and EGFRvIII was expressed in the vast majority of tumor cells as confirmed by immunostaining (Fig. 8b), demonstrating their transformed nature. Afatinib suppressed growth of these gliomaspheres with an IC50 of 0.1 μM (relative to equal volume vehicle treatment with DMSO; Additional file 1: Fig. S27a). Collectively, these data imply EGFRvIII is needed for initiation and maintenance of gliomagenesis in this model.
Novel genes drive EGFRvIII-tumor cell proliferation and drug sensitization
We next decided to explore the putative tumor suppressive effects of genes not previously linked to gliomagenesis, but strongly implicated by our mutational analysis and piggyBac experiments—Tead2, Nav3, and Spred1. We performed CRISPR-Cas9 knockout experiments using ex vivo EGFRvIII-gliomaspheres derived from mouse GBMs. Lentiviral transduction enabled Cas9 expression from these tumor cells, and subsequent targeted sgRNA transduction led to the production of frequent on-target indels in coding exons of Tead2, Nav3, and Spred1. A non-targeting sgRNA was transduced as a wild-type control. Tumor cells with these alterations were assessed for gliomasphere growth at 4 weeks post-sgRNA transduction—loss of each of these genes led to significantly increased gliomasphere proliferation (Tead2-loss—6.44x, Nav3-loss—5.04x, Spred1-loss—3.58x, relative cell viability compared with non-targeting sgRNA control (1x); p < 0.0001, < 0.0001, and 0.036 respectively, adjusted one-way ANOVA test, 3D, CellTiter-Glo 3D cell viability assay, Fig. 8d, e, Additional file 1: Fig. S27b, c, Additional file 12: Table S11). These results confirm that loss of these genes heightens tumor cell proliferation and gliomasphere growth and highlight the use of EGFRvIII-mouse gliomaspheres as a platform for functional genetic validation studies.
To demonstrate the utility of our model for pre-clinical drug testing, we conducted a proof-of-principle experiment comparing the sensitivity of EGFRvIII-gliomaspheres, with and without CRISPR-induced mutations in Nav3 and Spred1, to key small-molecule inhibitors. Although neither Spred1 nor Nav3 mutations affected tumor cell sensitivity to EGFR inhibition with afatinib (Additional file 1: Fig. S27a), loss of Spred1 or Nav3 increased sensitivity of EGFRvIII-tumor cells to MEK inhibitor treatment with trametinib (Fig. 8f, g). These preliminary results illustrate our model can be used for drug screening and the potential therapeutic relevance of EGFR-collaborative drivers.