Haploinsufficiency of phosphodiesterase 10A activates PI3K/AKT signaling independent of PTEN to induce an aggressive glioma phenotype

In this study, Nuechterlein et al. show that the deficiency of phosphodiesterase PDE10A activates PI3K/AKT/mTOR signaling independent of PTEN and promotes proneural-to-mesenchymal transition and gliomagenesis. Their work highlights that PDE10A may function as a potential tumor suppressor and that PDE10A-deficient glioblastomas may be therapeutically vulnerable to PI3K inhibition.

Abstract

Glioblastoma is universally fatal and characterized by frequent chromosomal copy number alterations harboring oncogenes and tumor suppressors. In this study, we analyzed exome-wide human glioblastoma copy number data and found that cytoband 6q27 is an independent poor prognostic marker in multiple data sets. We then combined CRISPR–Cas9 data, human spatial transcriptomic data, and human and mouse RNA sequencing data to nominate PDE10A as a potential haploinsufficient tumor suppressor in the 6q27 region. Mouse glioblastoma modeling using the RCAS/tv-a system confirmed that Pde10a suppression induced an aggressive glioma phenotype in vivo and resistance to temozolomide and radiation therapy in vitro. Cell culture analysis showed that decreased Pde10a expression led to increased PI3K/AKT signaling in a Pten-independent manner, a response blocked by selective PI3K inhibitors. Single-nucleus RNA sequencing from our mouse gliomas in vivo, in combination with cell culture validation, further showed that Pde10a suppression was associated with a proneural-to-mesenchymal transition that exhibited increased cell adhesion and decreased cell migration. Our results indicate that glioblastoma patients harboring PDE10A loss have worse outcomes and potentially increased sensitivity to PI3K inhibition.

IDH wild-type glioblastoma is the most common primary malignant brain tumor in adults (Ostrom et al. 2022). Patients diagnosed with this malignant solid tumor have grim outcomes and limited therapeutic options, which have improved only incrementally in the past few decades (Stupp et al. 2005, 2017). While the remarkable molecular heterogeneity of glioblastoma poses a challenge for drug development, it presents an opportunity to identify specific molecularly defined subsets of glioblastoma patients that will benefit from targeted therapy (Patel et al. 2014; Cimino et al. 2018b; Puchalski et al. 2018). Risk stratification alone has implications for clinical trial enrollment and design, and biological characterization of biomarkers associated with aggressive phenotypes may lead to the discovery of critical drug targets.

Somatic copy number alterations (SCNAs) largely characterize glioblastoma, with nearly all tumors harboring concurrent gain of whole chromosome 7 and loss of whole chromosome 10 (Cimino et al. 2017; Brat et al. 2018). These SCNA signatures have been recently incorporated into molecular grading criteria for glioblastoma (Brat et al. 2018; Louis et al. 2021). Additional SCNAs have been reported to be prognostic, such as the homozygous deletion of CDKN2A/B and EGFRvIII, amplification of CDK4 and MDM2, and gains of chromosomes 1, 19, and 20, none of which are universal in glioblastoma (Montano et al. 2011; Zheng et al. 2013; Geisenberger et al. 2015; Cimino et al. 2017, 2018b; Lu et al. 2020). Based on this premise, we hypothesized that additional SCNAs may be critical in IDH wild-type glioblastoma biology, at least for a subset of tumors, and that these genomic regions contain genes that drive gliomagenesis and present potential therapeutic vulnerabilities to treat this devastating disease. To address this, we sought to identify any novel genes and related pathways that may have a critical function in gliomagenesis.

Results

Distal chromosome 6q loss is an independent poor prognostic indicator in glioblastoma

We first performed an unbiased search for survival-associated, cytoband-level SCNA losses and gains in The Cancer Genome Atlas (TCGA) IDH wild-type glioblastomas (n = 358) and discovered that losses on the distal end of chromosome arm 6q were most prognostic (Supplemental Fig. S1). Loss of cytoband 6q27, the most distal cytoband on 6q, harbored the highest hazard ratio and was most significantly associated with survival (hazard ratio [HR] = 1.5, Cox's proportional hazard [CPH]) (Fig. 1A) and separated patient median survival by nearly 3 mo (Supplemental Fig. S2). Patient risk increased toward the distal end of 6q in the TCGA cohort and two other independent IDH wild-type glioblastoma cohorts from New York University (NYU; n = 68) and the Repository for Molecular Brain Neoplasia Data (REMBRANDT; n = 83) (Supplemental Fig. S3). In univariate analyses, loss of 6q27 was significantly prognostic in the NYU (P = 0.014, log rank) and REMBRANDT (P = 0.02, log rank) cohorts in addition to the TCGA cohort (false discover rate [FDR] = 0.16, log rank) (Fig. 1B). In multivariate analyses, loss of 6q27 was prognostic independent of age, KPS, sex, and MGMT promoter methylation status in the TCGA cohort (P = 0.04, CPH), prognostic independent of MGMT promoter methylation status and age in the NYU cohort (P = 0.01, CPH), and prognostic independent of age and sex in the REMBRANDT data set (P = 0.03, CPH) (Supplemental Figs. S4, S5). TCGA pan-cancer analysis of 6q27 loss confirmed its survival association in glioblastoma (Supplemental Fig. S6). Given the prognostic implications of 6q27 loss, we focused on identifying potential tumor suppressors in this genomic region.

Figure 1.

Nominating PDE10A as a potential glioblastoma tumor suppressor gene. (A) Unbiased cytoband-level log rank analysis identifies loss of cytoband 6q27 as the most prognostic in TCGA glioblastomas. (B) Kaplan–Meier curves show 6q27 loss is prognostic in the TCGA, NYU, and REMBRANDT cohorts. (C) Five genes in 6q27 had sgRNAs that caused cellular proliferation on a genome-wide CRISPR screen of glioblastoma stem-like cells (GSCs) and neural stem cells (NSCs). (D) PDE10A, AFDN, and PDCD2 have gene dosage effects on gene expression. (E) Only PDE10A showed lower expression in human glioblastoma compared with normal brain tissue. (F) In RCAS/tv-a mouse gliomas, only PDE10A showed significantly lower expression levels in mouse glioblastoma compared with normal mouse brain tissue. (HR) Hazard ratio, (OS) overall survival.

Nominating phosphodiesterase 10A (PDE10A) as a tumor suppressor candidate gene in cytoband 6q27

We next sought to determine whether any of the 40 genes located in 6q27 showed tumor-suppressive behavior. To do this, we turned to a previously published exome-wide functional CRISPR–Cas9 screen of human patient-derived glioblastoma stem-like cells (GSCs) and human neural stem cells (NSCs) (Toledo et al. 2015). Of the 40 genes located in cytoband 6q27, five (DLL1, PDE10A, C6orf118, AFDN, and PDCD2) had a single guide RNA (sgRNA) that caused increased cellular proliferation, suggesting potential tumor suppressor function (Fig. 1C). We tested whether any of these genes had a dosage response on gene expression, possibly indicating a haploinsuffiency phenotype. In the TCGA data set, this gene dosage effect was only observed in PDE10A (Padj < 0.001, Welch's t-test), AFDN (Padj < 0.001, Welch's t-test), and PDCD2 (Padj < 0.01, Welch's t-test) (Fig. 1D).

To further test which of these five genes may be tumor suppressors, we compared glioblastoma tissue with normal brain tissue to test whether any of these genes have decreased expression in tumor tissue compared with normal brain tissue. In human tissue, only PDE10A was expressed at lower levels in glioblastoma compared with normal brain tissue from the Genotype–Tissue Expression (GTEx) project (Padj < 1 × 10⁻⁵⁰, Welch's t-test) (Fig. 1E; Supplemental Fig. S7). Spatial RNA sequencing data of human glioblastoma tissue from the Ivy Glioblastoma Atlas Project also showed that regions of the tumor with higher tumor cell content have lower PDE10A expression than areas with lower tumor cell content (Puchalski et al. 2018). PDE10A expression was significantly lower in cellular tumor (CT), perinecrotic zone (PNZ), and infiltrating tumor (IT) compared with the tumor's leading edge (LE), confirming its lower expression in tumor tissue compared with nonneoplastic brain tissue (Supplemental Fig. S8). Immunohistochemistry using human glioblastoma tissue further confirmed decreased PDE10A expression in areas of highest cellular tumor content (Supplemental Fig. S8). As observed in human tissue, in a mouse glioma RNA sequencing data set (Ozawa et al. 2014), only Pde10a showed significantly lower expression levels in RCAS-induced gliomas compared with normal brain tissue (Fig. 1F; Supplemental Fig. S9). Thus, PDE10A was selected for further investigation as a haploinsufficient tumor suppressor candidate gene in glioblastoma based on the presented evidence: PDE10A was the only gene in cytoband 6q27 that showed cellular proliferation in a CRISPR–Cas9 knockout screen, decreased expression when 6q27 was deleted, and decreased expression in tumor tissue compared with normal brain tissue in both mice and humans. PDE10A is a member of one of the 11 phosphodiesterase families whose function is to regulate intracellular cyclic nucleotide levels by hydrolyzing adenosine cyclic monophosphate and guanosine cyclic monophosphate (Soderling et al. 1999). Furthermore, PDE10A is known to be highly expressed in the brain (Fujishige et al. 1999).

To examine PDE10A loss in the context of known poor prognostic indicators in glioblastoma, we investigated several clinical and pathological features and found that there was a strong association between MGMT promoter unmethylation status and PDE10A loss in the TCGA data set (Supplemental Figs. S10, S11). Conversely, gene body probes were hypermethylated in tumors with PDE10A loss relative to those with PDE10A intact. Furthermore, data from human glioblastoma cells in vitro supported PDE10A having a direct influence on MGMT expression and gene body methylation (Supplemental Fig. S12).

PDE10A expression is decreased in human and mouse central nervous system progenitor cells

We then determined whether PDE10A represents a haploinsufficient tumor suppressor, starting by examining PDE10A mutations and DNA methylation in TCGA, and found that PDE10A’s intact allele is rarely altered. Specifically, few homozygous deletions (1.7%) and inactivating mutations (0.3%) were observed on the intact allele, and methylation of the PDE10A promoter was not significantly greater in tumors with PDE10A loss than those without (P = 0.56, two-tailed t-test) (Supplemental Fig. S13). We next investigated whether decreased Pde10a gene expression (haploinsufficiency) is sufficient to induce a proliferative phenotype in the experimental RCAS/tv-a system.

To test the potential proliferative effects of Pde10a haploinsufficiency in the RCAS/tv-a mouse model, we first screened for effective short hairpin RNAs (shRNAs) against Pde10a mRNA and identified two that were effective in vitro (shPde10a#1 and shPde10a#4) (Supplemental Fig. S14). Next, primary mouse glioma precursor cell (GPC) neurospheres were transduced with either our RCAS-shPde10a constructs or RCAS-shScrambled. RCAS-positive cells were subsequently selected by using flow cytometry to enrich for RFP-expressing cells. We found that PDE10A suppression caused a significant increase in neurosphere formation compared with the scrambled control (P = 0.04, P = 0.02, two-way ANOVA) (Fig. 2A,B). Knockdown efficacy was confirmed by real-time qPCR, which showed that Pde10a expression was significantly lower, but nonzero, in each PDE10A knockdown compared with our control (P = 0.011, P = 0.016, one-way ANOVA) (Fig. 2C). We confirmed that both hairpins target the same Pde10a mRNA transcript variants; however, shPde10a#1 targets the open reading frame, while shPde10a#4 targets the 3′ untranslated region (UTR), which might possibly explain why shPde10a#1 reliably knocked down Pde10a mRNA levels more than shPde10a#4. This phenotype was rescued by the addition of an RCAS vector overexpressing PDE10A (Supplemental Fig. S15).

Figure 2.

PDE10A expression is developmentally regulated in human and mouse central nervous systems. (A,B) Primary mouse glioma precursor cells with PDE10A knockdowns showed increased neurosphere formation. (C) Real-time qPCR for neurospheres confirms that both hairpins for PDE10A reduce associated mRNA. (D) PDE10A knockdown led to increased SOX2 protein expression in mouse neurospheres. (E) SOX2 gene expression was negatively correlated with PDE10A expression in the TCGA and CGGA glioblastoma data sets. (F) PDE10A was more highly expressed in postmitotic differentiated neurons compared with progenitor and differentiated glial cells in mice. (G) PDE10A was also highly expressed in postmitotic neurons in humans.

We next investigated whether reduced PDE10A was associated with a neural progenitor phenotype. In mouse GPC neurospheres, PDE10A knockdown cells demonstrated increased SOX2 protein expression compared with control cells (Fig. 2D). Similarly, in human glioblastoma, SOX2 gene expression was negatively correlated with PDE10A expression (Fig. 2E). PDE10A expression was also inversely correlated with the progenitor marker CD44 and glial differentiation marker GFAP (Supplemental Fig. S16). PDE10A expression was positively correlated with the neuronal differentiation marker FOXBP3/NEUN (Supplemental Fig. S16). In addition, publicly available single-cell RNA sequencing data of the developing mouse and human brains gave further evidence of lower PDE10A expression in early central nervous system (CNS) cells. In mice, Pde10a mRNA was more highly expressed in postmitotic differentiated neurons compared with progenitor cells; in humans, PDE10A was also more highly expressed in mature postmitotic neurons (Fig. 2F,G). Together, these findings indicate that PDE10A is a developmentally regulated gene whose expression is low in stem cells and progenitor cells of the central nervous system. Overall, partial PDE10A loss was sufficient to drive a proliferative phenotype in mice and is further associated with a neural progenitor phenotype across human and mouse data sets.

PDE10A suppression induces an aggressive glioma phenotype in mice

Having established that PDE10A has a haploinsufficient effect on GPC neurospheres in vitro, we sought to assess whether PDE10A shows a tumor suppressor phenotype in vivo. PDGFA-driven gliomas were modeled using the RCAS/tv-a system, which included either of our two RCAS-shPde10a vectors or the RCAS-shScrambled control. Suppression of PDE10A led to decreased overall mouse survival (P = 0.001, log rank), higher CNS World Health Organization (WHO) histological grade (P = 0.04, Fisher's exact), and increased mitotic activity (P = 0.041, P = 0.045, two-tailed t-test) compared with the control (Fig. 3A–C).

Figure 3.

PDE10A suppression induces an aggressive phenotype in mouse glioma. (A–C) Mice harboring PDGF-driven gliomas with PDE10A knockdown suffered significantly worse survival, higher WHO histological grade, and higher mitotic activity in vivo compared with their control counterparts.

Pde10a suppression facilitates a proneural-to-mesenchymal transition in glioblastoma

We performed single-nucleus RNA sequencing with subsequent unsupervised clustering for two mice from each genotype (n = 6) (Supplemental Fig. S17). Cells in clusters expressing high levels of RCAS and hPDGFA were labeled tumor cells and separated from common nonneoplastic central nervous system cell types, including oligodendrocytes, astrocytes, neurons, and microglia, which were identified by standard gene markers (Fig. 4A; Supplemental Fig. S17). The remaining tumor cells separated into two clearly defined clusters, one of which (termed GBM2) was highly enriched for tumor cells from the PDE10A knockdown samples, as well as vascular endothelial growth factor A (VEGFA) expression (Fig. 4B,C). VEGFA expression is associated with hypoxia, which is reflected by histopathologic features palisading necrosis (PSN) and microvascular proliferation (MVP), which were more common in both knockdowns compared with the control (Fig. 4D).

Figure 4.

PDE10A suppression facilitates a proneural-to-mesenchymal transition in glioma. (A) UMAP plot of single-nucleus RNA sequencing of mouse glioma cells in vivo demonstrating cell type clustering. Cells in clusters expressing high levels of RCAS and hPDGFA were labeled tumor cells (GBM1 and GBM2); other common CNS cell types were identified by standard gene markers. (B) Tumor cluster GBM2 was highly enriched for tumor cells from both PDE10A knockdown gliomas compared with the scrambled control. (C) Tumor cluster GBM2 was also highly enriched for VEGFA expression (the top marker of this cluster). (D) PDE10A knockdown mouse gliomas were enriched for the VEGFA-associated histological features palisading necrosis (PSN) and microvascular proliferation (MVP). (E) The proportion of mesenchymal-like tumor cells in the knockdowns is higher in PDE10A knockdown tumor cells compared with the control in vivo. (F) The mesenchymal subtype is overrepresented in glioblastomas with PDE10A loss in the TCGA data set. (G,H) Western blotting showed unchanged total STAT3 levels with increased pSTAT3 levels in cells with PDE10A knockdown compared with the control cells in vitro.

Given that VEGFA expression is known to be increased in mesenchymal glioblastoma (Kim et al. 2021), we tested whether the PDE10A knockdown tumor cells were disproportionally mesenchymal-like compared with the scrambled control. To do this, we classified all tumor cells by published meta module-defined cell state (Neftel et al. 2019; Park et al. 2022). Because the mouse glioblastoma models are PDGFA-driven, most cells are inherently OPC-like. However, we did see an increased proportion of mesenchymal-like tumor cells in the PDE10A knockdown glioma cells compared with the control (Fig. 4E). We also observed that TCGA IDH wild-type glioblastomas with PDE10A loss are disproportionally classified as mesenchymal subtype when compared with those glioblastomas with PDE10A intact (P < 0.01, Fisher's exact) (Fig. 4F). We next checked for STAT3 activation, which has been implicated as a master regulator of the mesenchymal transcription signature in glioblastoma (Kim et al. 2021). In vitro, we observed that pSTAT3 was significantly increased in both PDE10A knockdowns compared with the control, with no significant difference in total STAT3 (P < 0.001, P = 0.036, two-tailed t-test) (Fig. 4G,H). Furthermore, we observed that expression of the mesenchymal markers CD44, YKL40, and VIM were inversely correlated with PDE10A expression in both the TCGA and CGGA glioblastoma data sets (Supplemental Figs. S16, S18). This indicates that tumors with reduced PDE10A expression undergo a proneural-to-mesenchymal-like phenotype transition, which has been linked to treatment resistance and increased tumor malignancy (Wang et al. 2016; Fedele et al. 2019). Interestingly, loss of 6q has been reported to correlate with the mesenchymal phenotype in glioblastoma (Wang et al. 2022).

PDE10A suppression causes increased cell adhesion

We next analyzed these single-nucleus data in an unbiased manner by performing a gene ontology (GO) analysis on genes differentially expressed between tumor cells in the PDE10A knockdowns versus the control (Supplemental Table S1). Remarkably, nearly all top hits of this GO analysis were related to cell–cell adhesion or cell adhesion molecules (Fig. 5A). Among differentially expressed genes, several genes related to cell adhesion were up-regulated, some quite highly, including CADM1. Conversely, all differentially expressed genes involving focal adhesion complex (FAC) disassembly, such as DNM3, were down-regulated (Fig. 5B). These in vivo results were supported by Western blotting of GPCs in culture, where CADM1 levels were significantly higher in the knockdowns compared with the control (P = 0.004, P = 0.012, two-tailed t-test), and DNM3 levels were significantly lower in the knockdowns compared with the control (P = 0.014, P = 0.040, two-tailed t-test) (Fig. 5C,D). To test the functional consequences of adhesion gene dysregulation, we performed cell adhesion and migration assays in vitro. In these assays, we observed that there was a significant increase in cell adhesion in the PDE10A knockdown cells compared with the control (P < 0.0001, P < 0.0001, one-way ANOVA) (Fig. 5E). Conversely, there was a significant decrease in cell migration in the PDE10A knockdown cells compared with the control cells (P = 0.002, P = 0.003, one-way ANOVA) (Fig. 5F). Immunohistochemistry for matrix metalloproteinase 2 (MMP2) as a surrogate marker for tumor cell invasion in mouse tumor tissue showed a trend supportive of the direct relationship between PDE10A levels and glioma tumor cell invasion/migration (Supplemental Fig. S19). The in vitro functional assays confirmed the tumor cell phenotype from the gene expression and Western blot analyses that demonstrated up-regulation of cell adhesion molecule genes with concurrent down-regulation of FAC disassembly genes. Interestingly, dysregulation of cell adhesion genes has also been associated with a mesenchymal-like state in glioblastoma (Andl et al. 2010). Although the decreased cell migration is somewhat in opposition to the reported phenotype of mesenchymal glioma cells (Mikheeva et al. 2010; Piao et al. 2013), the cell adhesion phenotype, along with the other features discussed, is consistent with the proneural-to-mesenchymal transition phenotype. The lack of migration in these cells is likely due to the concomitant decreased FAC disassembly genes, leading to decreased focal adhesion turnover and the subsequent altered adhesion/migration phenotype (Nagano et al. 2012).

Figure 5.

PDE10A suppression causes increased cell adhesion and decreased single-cell migration. (A) In vivo single-nucleus RNA sequencing data showed that nearly all top hits of a gene ontology analysis comparing tumor cells from the PDE10A knockdowns with the controls were related to cell–cell adhesion or cell adhesion molecules. (B) Genes related to cell adhesion were up-regulated (e.g., CADM1), and genes involved in focal adhesion complex disassembly (e.g., DNM3) were down-regulated in the PDE10A knockdown tumor cells compared with the control. (C,D) Western blotting confirmed that CADM1 protein levels were higher, while DNM3 levels were lower, for PDE10A knockdown cells in vitro. (E,F) In vitro cell adhesion and cell migration assays showed a significant increase in cell adhesion with a concurrent significant decrease in cell migration in the PDE10A knockdown cells compared with the control.

PDE10A suppression activates PI3K/AKT/mTOR signaling

To explore potential mechanisms by which PDE10A suppression could induce an aggressive glioma phenotype in mice, along with a proneural-to-mesenchymal transition phenotype, we investigated the PI3K/AKT signaling axis because of its critical role in glioma, especially in the RCAS/tv-a system (Uhrbom et al. 2002; Rajasekhar et al. 2003; Hambardzumyan et al. 2008; Bleau et al. 2009). Furthermore, upstream RAS signaling was a top activated pathway in our single-nucleus RNA sequencing data (Supplemental Fig. S20). Immunohistochemistry (IHC) showed that there were increased pS6 levels, but not 4E-BP1 levels, in the mouse gliomas with PDE10A knockdowns compared with controls (Fig. 6A; Supplemental Fig. S21). To determine whether increased pS6 was due to either overactive PI3K signaling or loss of PTEN, we used mouse GPCs in vitro. Western blotting further showed that PTEN levels did not differ between the PDE10A knockdowns and control, implying that PTEN activity was not responsible for increased pAKT (Fig. 6B). Western blotting showed increased pPI3K and pAKT levels induced by PDE10A knockdown, while total PI3K and AKT levels were not altered, indicating PI3K/AKT pathway activation (Fig. 6C–F). Thus, we hypothesized that PDE10A may negatively regulate PI3K activity, a hypothesis supported by the effects of PI3K pharmacologic inhibition in vitro. Cells with PDE10A suppression had their proliferative phenotype silenced with increasing levels of either one of two PI3K inhibitors (alpelisib or paxalisib) in vitro (Fig. 6G,H). PDE10A suppression also conferred sensitivity to downstream mTOR inhibition with everolimus in vitro (Supplemental Fig. S22). Given that PDE10A loss partially mimics PTEN loss, we performed similar in vitro experiments with PTEN loss for comparison. As expected, PTEN loss caused decreased PI3K activation but increased AKT activation (Supplemental Fig. S23). In contrast to PDE10A loss, cells with PTEN loss were resistant to pharmacologic PI3K inhibition (Supplemental Fig. S24). However, cells with PTEN loss were similar to those with PDE10A loss in that they were sensitive to downstream mTOR inhibition by everolimus (Supplemental Fig. S24). Based on our results, we propose the model that PDE10A suppression leads to cell survival, growth, and proliferation by overactivation of the PI3K/AKT signaling pathway independent of PTEN activity (Fig. 6I).

Figure 6.

PDE10A suppression activates the PI3K/AKT/pS6 pathway independent of PTEN. (A) Immunohistochemistry of mouse gliomas showed increased pS6 in the PDE10A knockdown tumors compared with the control. (B) Cellular PTEN levels were unaltered with PDE10A knockdown, as determined by Western blotting. (C,D) Western blotting showed unchanged total PI3K levels with increased pPI3K levels in cells with PDE10A knockdown compared with the control cells. (E,F) Western blotting showed unchanged total AKT levels with increased pAKT levels in cells with PDE10A knockdown compared with the control cells. (G,H) Mouse cells with PDE10A knockdown responded strongly to the PI3K inhibitors alpelisib and paxalisib in vitro with increasing dosage. (I) Proposed model of PDE10A as an inhibitor of the PI3K/AKT signaling pathway independent of PTEN activity.

Discussion

Although patients diagnosed with IDH wild-type glioblastoma have universally poor outcomes, some tumors are more aggressive than others and portend a shorter survival. Prognostic biomarkers capable of identifying patients predisposed to poor survival could have a significant impact on patient management. Furthermore, gathering an increased understanding of the biology associated with such biomarkers may lead to new therapies, which are urgently needed for glioblastoma patients for whom standard-of-care treatment has remained static for almost 20 years. Our study identified a novel poor prognostic biomarker (6q27 loss) in glioblastoma, identified PDE10A as a haploinsufficient tumor suppressor, interrogated the underlying biology of PDE10A loss, and proposed PI3K inhibition as a possible treatment for patients who harbor 6q27/PDE10A loss. These findings argue for incorporating 6q27/PDE10A loss into biomarker-driven approaches for early clinical trials and personalized medicine for patients with glioblastoma.

The magnitude of risk imparted by the loss of distal chromosome 6q in glioblastoma suggests that 6q loss may be a risk factor in other tumor types. In addition to glioblastoma, loss of 6q has been documented in posterior fossa ependymoma (Baroni et al. 2021), meningioma (Pérez-Magán et al. 2010; Maas et al. 2021), and WNT-activated medulloblastoma (Helgager et al. 2020), raising the possibility that our findings may be relevant to other brain tumors, in particular. More broadly, our TCGA pan-cancer analysis indicated that 6q27 loss was associated with poor survival in lower-grade glioma, colon adenocarcinoma, and stomach adenocarcinoma (Supplemental Fig. S2). Therefore, deeper investigation into the role of PDE10A in the development of these tumors is warranted.

PDE10A has been best studied in nonneoplastic central nervous system disorders (Wilson and Brandon 2015), although publications have reported PDE10A as having cancer type-specific roles in systemic solid cancers. However, in contrast to glioblastoma, PDE10A is overexpressed in other cancer types, including ovarian carcinoma (Borneman et al. 2022), colon carcinoma (Li et al. 2015), and non-small-cell lung carcinoma (NSCLC) (Fusco et al. 2018). This indicates a potential oncogenic role, rather than tumor suppressor role, in these cancer types. PDE10A overexpression in ovarian carcinoma (Borneman et al. 2022) and NSCLC (Fusco et al. 2018) has also been associated with worse patient survival. Anti-PDE10A IgGs causing paraneoplastic neurological effects have also been described for NSCLC, renal carcinoma, and pancreatic carcinoma, indicating that PDE10A may be overexpressed in these latter cancer types (Zekeridou et al. 2019). Rare PDE10A::BRAF fusions have been described in pediatric sarcomas; however, it is unclear whether the PDE10A fusion partner itself results in any altered cellular phenotype (Vairy et al. 2018; Hughes et al. 2021). Evidence of PDE10A's tumor-suppressive function has been documented in other cancers. Damaging mutations in PDE10A have been reported in metastatic gastric adenocarcinoma (Liu et al. 2016), which is consistent with the findings from our pan-cancer analysis of 6q27 loss. Additionally, inactivating PDE10A mutations have been implicated in the progression of prostate carcinoma (de Alexandre et al. 2015). Overall, the limited literature regarding the role of PDE10A in cancer is inconclusive, pointing to disparate functions that depend on specific cancer types under investigation.

To the best of our knowledge, PDE10A has not been well investigated in glioblastoma. One study reported a variable decrease in cell proliferation in vitro across three human glioblastoma cell lines (U87MG, A172, and T98G) in the presence of PDE inhibitors including papaverine, a reported PDE10A preferential inhibitor (Kopanitsa et al. 2021). The discrepancy in glioma cell proliferation between our current in vivo and in vitro findings and the reported effects of papaverine may be attributed to several reasons. First, in vitro pharmacologic treatment measures immediate or transient cell responses, whereas PDE10A loss in glioblastoma is a stable chronic genomic change in tumor cells. Second, papaverine is a preferential inhibitor for PDE10A but still shows biological activity targeting other PDEs (Ashrafi et al. 2023). In glioma cells, it is possible that other PDE subtypes might compensate for PDE10A loss, making them more available for papaverine inhibition. Third, papaverine, an alkaloid derivative, exhibits off-target activity in multiple organs and tissues (despite PDE10A being highly expressed only in the brain and testes) (Soderling et al. 1999; Ashrafi et al. 2023). Therefore, it is possible that papaverine has unknown protein targets outside of PDE family members. Fourth, while our in vivo studies focused on murine proneural-type PDGF-driven glioblastomas experimentally, the previously published human cell lines are more genetically and transcriptionally heterogeneous. Investigating the effects of PDE10A suppression related to other drivers of gliomagenesis in murine gliomas is needed to clarify this issue.

In summary, this study builds on our previous investigations of SCNAs in glioblastoma by using a systematic, unbiased, exome-wide analysis with revised WHO 2021 glioblastoma diagnostic criteria, including both histologically and molecularly defined tumors (Cimino et al. 2017; 2018a,b). We conducted an exome-wide analysis that identified cytoband 6q27 loss as the most prognostic cytoband SCNA that is found in a subset of TCGA IDH wild-type glioblastoma (molecularly defined and histologically defined), which we validated across multiple cohorts. Combining multiple genomic data sets, we then nominated PDE10A as a likely tumor suppressor located within cytoband 6q27. Using our RCAS/tv-a mouse model of glioma, we showed that PDE10A loss induces a haploinsufficiency tumor-suppressive phenotype, largely through biological processes associated with a proneural-to-mesenchymal transition. This PDE10A knockdown phenotype included susceptibility to PI3K and mTOR inhibitors. This preclinical phenotype suggests that PI3K and/or mTOR inhibitors may be a candidate therapy for further study in patients with glioblastoma harboring PDE10A loss, especially in those tumors with intact PTEN status.

Materials and methods

The Cancer Genome Atlas glioblastoma data set

Gene-level glioma somatic copy number alteration (SCNA) calls for The Cancer Genome Atlas (TCGA) IDH wild-type glioblastomas were downloaded from University of California, Santa Cruz (UCSC), Xena (https://xena.ucsc.edu) (Goldman et al. 2020). TCGA SCNA data were the thresholded output of the genomic identification of significant targets in cancer (GISTIC) 2.0 algorithm aligned to human genome assembly GRCh37 (hg19) (Mermel et al. 2011). Somatic mutation calls computed by the Multi-Center Mutation Calling in Multiple Cancers (MC3) Project (n = 358) and RNA sequencing RSEM counts (n = 208) for TCGA IDH wild-type glioblastoma were also downloaded (Goldman et al. 2020). Additionally, processed RNA-seq abundance values were downloaded from the recount2 (Collado-Torres et al. 2017) data set (n = 219) for TCGA primary IDH wild-type glioblastoma patients and converted to transcripts per million (TPM) units to be consistent with the Chinese Glioma Genome Atlas (CGGA) TPM data (Arora et al. 2020; Zhao et al. 2021). DNA methylation β-values from Illumina 450k arrays were also downloaded from UCSC Xena (n = 197). β-Values from multiple samples from the same patient were averaged. All idat files were processed through the DKFZ classifier (http://www.molecularneuropathology.org; classifier versions 11b4 and 12.3) and designated as RTK I/II/III or MES subtypes (Capper et al. 2018). Clinical characteristics including age, WHO grade, and overall survival were downloaded from UCSC Xena. TCGA IDH mutational status and 1p/19q codeletion data were ascertained as previously described (Nuechterlein et al. 2021). TCGA pan-cancer gene-level SCNA data (n = 10,845), RNA-seq RSEM data (n = 10,535), and clinical variables (n = 12,591) from TCGA patients diagnosed with one of 33 cancer types were downloaded from UCSC Xena.

Chinese Glioma Genome Atlas glioblastoma data set

Processed RNA-seq abundance values were downloaded from the recount2 data set for 270 patients from the Chinese Glioma Genome Atlas (CGGA) and converted to TPM units as previously described (Collado-Torres et al. 2017; Arora et al. 2020; Zhao et al. 2021). Only CGGA patients diagnosed as primary histological grade 4 IDH wild-type glioblastomas (N = 70) were analyzed.

New York University glioblastoma data set

Copy number segmentation files for IDH wild-type glioblastoma patients were obtained from an institutional data set at New York University (NYU) (n = 68). Thresholded gene-level copy number calls were computed using the GISTIC algorithm aligned to hg19. Age (n = 68) and MGMT promoter methylation status (n = 61) are available.

REMBRANDT glioblastoma data set

Binary CN4.cnchp files from Affymetrix human mapping 50K Hind240 (N = 240) and 50K Xba240 SNP arrays (N = 192) for 275 samples were downloaded from the REMBRANDT database (GEO data set GSE108475) (Gusev et al. 2018). Affymetrix power tools (https://github.com/rcallahan/affymetrix-power-tools) was used to convert the CN4.cnchp files into text files. Precomputed copy number and loss of heterozygosity analysis results from the CN4 algorithm were extracted from these files, and Hmm median log₂ ratio values were used to estimate the underlying DNA copy number variation using the Bioconductor package DNAcopy (https://bioconductor.org/packages/release/bioc/vignettes/DNAcopy/inst/doc/DNAcopy.pdf). GISTIC was then applied to calculate gene-level gains and losses. IDH wild-type status was determined as previously described (Nuechterlein et al. 2021).

Human neural stem cell and glioblastoma stem-like cell CRISPR–Cas9 screen data set

Previously published genome-wide CRISPR–Cas9 data derived from human neural stem cells and glioblastoma stem-like cells were obtained from Gene Expression Omnibus (GEO) GSE70038 (Toledo et al. 2015). Genes that mapped to chromosome 6q27 were further interrogated. Significantly overrepresented (log fold change >1 and false discovery rate <0.05) sgRNAs causing increased cell proliferation in any cell line were considered a positive hit for a candidate tumor suppressor.

Gene expression analyses comparing glioblastoma (TCGA and CGGA) and normal brain (GTEx) data sets

Raw RNA sequencing counts aligned to hg38 were downloaded from recount2 for 702 adult glioma samples from The Cancer Genome Atlas (TCGA) (The Cancer Genome Atlas Research Network et al. 2013), 270 adult glioma samples from the Chinese Glioma Genome Atlas (CGGA) (Zhao et al. 2021), and 1409 healthy normal brain samples from the Genotype–Tissue Expression Project (GTEx) (Carithers et al. 2015) across 12 GTEx-defined brain regions. Gene expression data from 19,142 protein-coding genes were transformed to get log₂ TPM-normalized counts as previously described (Arora et al. 2020). These counts were used to generate the Brain-UMAP. Differential expression analysis was performed using R/Bioconductor package DESeq2 (Love et al. 2014). Gene expression counts for IDH wild-type gliomas from both TCGA and CGGA were considered as one group and were compared with all samples from GTEx. A gene was considered differentially enriched if it had a fold change >1.5 and adjusted P-values (FDR) <0.05 using DESeq2.

Mouse and human brain development single-cell RNA sequencing data sets

We downloaded and processed publicly available single-cell RNA sequencing data for both mouse and human brain development. For our mouse data set, the files “gene_count_cleaned.RDS,” “gene_annotation.csv,” and “cell_annotation.csv” were downloaded from https://oncoscape.v3.sttrcancer.org/atlas.gs.washington.edu.mouse.rna/downloads. These files contained processed normalized gene expression counts, gene model information, and information about sample, cell types, clusters, and trajectories, respectively. Gene count dot plots for the “neural tube and notochord trajectory” were made using Seurat for genes of interest (Satija et al. 2015; Butler et al. 2018; Stuart et al. 2019; Hao et al. 2021). For “cerebellum cells” in our human data set, the files for the raw gene count-sparse matrices (“gene_annotation.csv” and “cell_annotation.csv”) were downloaded from https://descartes.brotmanbaty.org/bbi/human-gene-expression-during-development. These files contained processed gene expression count, gene model information, and information about sample, cell types, clusters, and trajectories. Seurat (v 4.0) was used to normalize the data, and dot plots were made using Seurat for PDE10A (Satija et al. 2015; Butler et al. 2018; Stuart et al. 2019; Hao et al. 2021).

Generation of gliomas using the RCAS/tv-a mouse model system

All animal experiments were done in accordance with the Institutional Animal Care and Use Committees of the Fred Hutchinson Cancer Center and the National Institutes of Health. The RCAS/tv-a system used in this study has been previously described (Ozawa et al. 2014; Cimino et al. 2018a; Pattwell et al. 2020; Szulzewsky et al. 2020). Mice used for gliomagenesis were Nestin/tv-a;Cdkn2a(Ink4a–Arf)^−/−;Pten^fl/fl. Neonatal P0–P3 mice underwent intracranial injection of RCAS-containing Df-1 cells and were subsequently monitored and euthanized at first signs of neurological dysfunction. Mice that did not show neurological symptoms were euthanized at the study end point of 150 d. Brains were freshly harvested and processed accordingly for subsequent experiments.

RCAS vector production

Generation of RCAS shRNAs began by cloning shRNAs into the pENTR-RFP-H1 vector. The pENTR vector underwent predigestion by BglII and HindIII. Annealed shRNAs for PDE10A or scrambled control then underwent ligase reaction to the digested pENTR-RFP-H1 vector. The six shPDE10A sequences tested, along with the control scrambled sequences, are listed in Supplemental Table S2. For PDE10A overexpression, the mouse Pde10a coding sequence (Horizon Discovery 40048238) was amplified by PCR (primer set sequences in Supplemental Table S3), restriction-digested using SalI-HF and NotI-HF, and ligated into the pENTR-1A dual-selection vector. The ligated vectors were transformed into competent bacteria, selected for on kanamycin plates, purified, and Sanger-sequenced for confirmation of insertion and unaltered DNA sequences. Clonase reactions were performed using pENTR-RFP-H1 shRNA or pENTR-1A-PDE10A as the entry clone and RCAS destination vector (DV) as the destination vector. RCAS-RFP-H1 shRNAs and PDE10A overexpression were transformed into bacteria, grown, and selected on ampicillin plates. Purified RCAS vectors were Sanger-sequenced for confirmation. The RCAS vector containing human PDGFA (hPDGFA) has been previously described (Ozawa et al. 2014).

Primary mouse glioma precursor cell culture

Glioma precursor cells (GPCs) were isolated from P0 neonatal mouse brain tissue as previously described (Pattwell et al. 2020). Fresh media were prepared for each passage. For retroviral transduction of murine GPC neurospheres with RCAS-RFP-shPDE10A#1, RCAS-RFP-shPDE10A#4, or RCAS-RFP-shScrambled, RCAS virus was produced in DF-1 packaging cells maintained with serum-free neurosphere medium and was then diluted 1:1 with collection media (fresh murine neurosphere media). RCAS transduced cells were passaged twice, and retroviral incorporation was enriched for by gating RFP-positive cells using a MoFlo Astrios cell sorter (Beckman Coulter Life Sciences) at the Flow and Imaging Cytometry Core Facility of the National Institute of Neurological Disorders and Stroke. Flow-sorted RFP-positive cells were then grown again as neurospheres and passaged twice prior to each of the subsequent assays.

Quantitative real-time PCR

Total RNA was extracted from mouse neurosphere cells using RNeasy minikit (Qiagen). RNA was used to prepare cDNA using the SuperScript III first strand synthesis system according to the manufacturer's protocol (Invitrogen). SYBR Green real-time PCR was performed using primer sets, reagents, and protocols from Applied Biosystems in a 7900 HT Fast real-time PCR system. Each sample was analyzed in triplicate. Sequences for the primer sets used are in Supplemental Table S3.

Murine neurosphere formation assay

Cells were starved of growth factors 2 h prior to the assay and subsequently trypsinized into single-cell suspension. The cells were diluted serially and seeded at 1000–15 cells/well in biological triplicates. Five days after seeding, the number of spheres was quantified under a phase contrast microscope using a 0.5-cm size threshold.

In vitro drug treatment: alpelisib, paxalisib, and everolimus

Mouse neural stem cells were trypsinized into single-cell suspension, counted, and seeded at 5000 cells per well in a 96-well plate as biological triplicates in half growth factor media. DMSO was added to the lyophilized drugs to obtain stock concentrations of 100 mM alpelisib, 10 mM paxalisib, and 10 mM everolimus. The stock solutions were then serially diluted in DMSO to produce a 10-fold decrease in drug concentration (from 50 to –0.005 µM). To maintain a nontoxic level of DMSO, the drug solutions were subsequently diluted in half growth factor media. The plated cells were then treated with 50 μL of this drug–media solution and stored for 4 d at 37°C in a 5% CO₂ humidified incubator. On day four, an XTT cell proliferation assay kit (American Type Culture Collection 30-1011K) was used per the manufacturer's protocol.

Cell adhesion and migration assays

Cell adhesion was measured using xCELLigence real-time cell analysis (RTCA) technology with electronic microtiter plates (E-plates) according to the manufacturer's protocol. Each well in the E-plate was first filled with half growth factor media and allowed a 30-min equilibration period, and a baseline measurement was taken. Next, the single-cell suspension was seeded with an optimized cell density of 250,000 cells in each well, and cells were allowed to equilibrate and settle for 30 min. Cell migration was measured using xCELLigence RTCA technology with cell invasion and migration plates (CIM plates) according to the manufacturer's protocol. First, each well in the lower chamber of the CIM plates was filled with media + 10% FBS to serve as a chemoattractant. Then, the upper wells were promptly attached to the lower wells, a 1-h equilibration period was allowed, and a baseline measurement was taken. After baseline measurement, an optimized cell density of 150,000 cells in serum-free media was seeded into each upper well, and cells were allowed to equilibrate and settle for 30 min. For both assays, the cell index values were automatically obtained in real-time every 15 min for 48 h at 37°C and 5% CO₂.

Western blotting

Cell pellets were lysed with RIPA buffer (Sigma R0278) supplemented with a protease/phosphatase inhibitor cocktail (Cell Signaling Technology 5872). Equal concentrations of lysates were combined with NuPAGE LDS sample buffer and reducing agent (Novex NP0007 and NP0009), heated for 10 min at 85°C, and added to NuPAGE 4%–12% Bis-Tris gels (Invitrogen NP0321BOX). Gels were run, transferred to nitrocellulose membranes (Invitrogen IB23002), and blocked in 5% milk. Primary antibodies (listed in Supplemental Table S4) were incubated in milk overnight at 4°C in their optimized conditions. After thorough washing in PBST, antirabbit secondary antibodies (dilution 1:1000; KindleBiosciences R1004) were applied to the membranes for 1 h. Following additional wash steps, ECL substrate solution (KindleBiosciences R1004) was used as the detection method, and membranes were developed on an ImageQuant800 machine. Gel images were analyzed and quantified using ImageJ software version 1.54 (https://imagej.net/ij).

Immunohistochemistry

Formalin-fixed paraffin-embedded (FFPE) mouse brain tissue was cut into 5-μm-thick sectioned slides for immunohistochemistry. Immunostaining using primary antibodies (listed in Supplemental Table S4) was performed on an automated stainer (Leica Bond-Max). Antibody binding was visualized using a ready-to-use detection kit (Leica Biosystems) with chromogen-labeled 3,3′-diaminobenzidine (DAB). pS6-stained slides were digitized using a Leica Aperio Versa 200, and Leica's ImageScope software was used to visualize the scanned images. Two regions of interest (ROIs) per mouse tumor underwent color deconvolution (H DAB) and were analyzed using ImageJ software version 1.54 (https://imagej.net/ij).

Nucleus isolation for single-nucleus RNA sequencing for mouse gliomas

Mouse tumors were macro-dissected from normal-appearing brains and flash-frozen in liquid nitrogen. Nuclei were isolated from the frozen tissues using an adaptation of a previously established protocol (Matson et al. 2018). Briefly, each sample (two mice per condition were pooled for each sample) was placed on a clean Petri dish and washed with 800 mL of detergent–lysis buffer (0.1% Triton-X). Tissue was mechanically dissociated and homogenized. The crude nucleus homogenate was then passed through a 40-µm strainer and spun down at 3200g for 5 min at 4°C in a swing-bucket centrifuge. The pellet was resuspended in 1 mL of low-sucrose buffer and then gently layered above 4 mL of high-sucrose buffer without disrupting the density gradient. The gradient was then spun down at 3200g for 20 min at 4°C, and the ensuing pellet was resuspended in 25–150 mL of 0.04% BSA and 1 U/µL Sigma Protector RNase inhibitor (3335399001) in 1× PBS. An aliquot was taken to be stained with Acridine Orange dye (1:10 dilution) and counted/imaged using a Luna-FL dual-fluorescence cell counter (Logos Biosystems). Following the manufacturer's recommendations, the nucleus suspensions were processed via the single-cell gene expression assay (10X Genomics). Nuclei were loaded onto a Next GEM Chip G (10X Genomics) targeting a yield of 5000 nuclei.

Single-nucleus RNA sequencing, preprocessing, and analysis of mouse gliomas

For all single-nucleus experiments, the single-cell gene expression kit (10X Genomics) was used according to the manufacturer's protocol. Library preparation was performed according to the manufacturer's instructions. Libraries were pooled and sequenced on the NextSeq 500 system (Illumina) using high-output reagent kit v2.5 (150 cycles; Illumina). Sequenced reads from tumor samples were demultiplexed and aligned to the mm10 reference genome using the CellRanger (v6.1; 10X Genomics) mkfastq function with default settings, and subsequent counts were generated using the CellRanger count function (Schneider et al. 2017). Combined RCAS and hPDGFA expression was used to define murine neoplastic cells. To calculate RCAS expression, CellRanger version 7.1.0 was used to recover unmapped reads by converting BAM files processed by CellRanger back to fastq files (Zheng et al. 2017). The RCAS virus sequence was then added to the genome reference file, and CellRanger was used to realign the fastq files. The RCAS LTR reference sequence used is listed in Supplemental Table S5 (Alexander et al. 2020). The expression of hPDGFA (sequence in Supplemental Table S5) was computed in a similar manner. Seurat version 4.3.0 was used to run quality control on all scRNA-seq data, normalize gene counts, integrate the scRNA-seq data from each sample, conduct dimensionality reduction (PCA and UMAP), cluster cells, and perform differential gene expression (Hao et al. 2021). Gene ontology analysis and gene set enrichment analysis were performed using the R package clusterProfiler (R version 4.1.1) (Wu et al. 2021). Differential gene expression analysis was conducted using the R package DESeq2 (Love et al. 2014). Single-nucleus RNA sequencing data are available as GEO data set GSE243575 (https://www.ncbi.nlm.nih.gov/geo).

Statistical analysis

The log-rank test was used to test for univariate survival association. We used Fisher's exact test to assess statistical differences in binary data across two groups. Cox's proportional hazard (CPH) regression was used to compute hazard ratios (HRs) and for multivariate survival analysis. False discovery rate (FDR) was used for P-value correction when multiple hypotheses were tested. Statistical analyses were performed in Python version 3.8.6, R version 4.0.5, or GraphPad Prism version 9.3.1 unless otherwise specified.

Competing interest statement

The authors declare no competing interests.