Alterations in key cell signaling pathways in glioblastoma cancer stem cells in response to astrocytic differentiation, growth factor withdrawal and traditional 10% FBS media

CSCs derived from 8 newly diagnosed GBM patients, representing the main genomic drivers were selected for this study (Fig 1A). These CSC lines meet the requirement of the two defining stemness criteria: long term self-renewal and orthotopic tumor formation in immunocompromised mice [11,25,26]. FBS, a supplement supporting the culture of a variety of mammalian cells, contains hormones, growth factors, cell attachment factors and other nutrients [27]. Media supplemented with 10% FBS had been widely employed to propagate glioma cells in culture, prior to seminal data demonstrating genetic drift occurring from serial passaging under these conditions [28]. In contrast, 1–2% FBS added to defined media has been successfully used to differentiate human-induced pluripotent stem cells (hiPSC)-derived neural stem cells into astrocytic lineage [29,30]. Using this paradigm to model a differentiated phenotype in culture, CSCs growing in 3D cultures in defined neurosphere media, supplemented with, 20 ng/ml each epidermal (EGF) and basic fibroblast (bFGF) growth factors (NMGF), were differentiated into an astrocytic phenotype in NMGF supplemented with 2% FBS for 14 days. This differentiated CSC progeny is referred to here as “serum differentiated cells” (SDC). For comparison, we tested two additional growth conditions: growth factor-depleted neurosphere media (NM), or media supplemented with 10% FBS (10% FBS) (Fig 1B). After 2 weeks in culture, cell lysates were obtained and analyzed by Reverse Phase Protein Arrays (RPPA) in triplicates, as described [31,32]. The levels of 66 proteins or post-translational modifications (PTM) in the cell lysates were quantified (S1 Table). Pairwise comparisons between NMGF (CSC) culture and either NM or 2% FBS (SDC) for the 8 cell lines did not show a global shift in the levels of protein/PTMs levels (S1 Fig), but important patterns of alteration in specific signaling were observed. Phosphorylation (phospho) levels of serine (Ser)-residues in receptor-dependent r-Smad1/5/8, which are primarily activated by BMP ligand signaling, were notably upregulated in SDCs in relation to CSCs in all 8 GBM models (Fig 2A). Consistent with bone morphogenetic protein (BMP) ligands being present in FBS [33], increased levels of p-Smad1/5/8 were also observed in 10% FBS cultures for all models (S1 Table). The tumor suppressor phosphatase and tensin homolog (PTEN) function is frequently lost in GBMs through gene deletion and mutations, as seen in 7/8 GBM tumors represented in this study (Fig 1A), leading to activation of AKT [34]. AKT phosphorylation at both Y308 (by pyruvate dehydrogenase kinase 1 (PDK1)) and S473 (by mammalian target of rapamycin complex 2 (mTORC2)) was suppressed in the absence of growth factors (NM) for all lines, except for HF2927, which carries ligand independent epidermal growth factor receptor (EGFR) variant III (viii) amplification. In 2% FBS (SDCs), AKT phosphorylation was suppressed for HF2587, HF3035 and HF3077, and upregulated or unchanged in the remaining models (Fig 2A). In response to growth factor withdrawal, phosphorylation levels of S6 ribosomal proteins (RBS6), downstream of mTORC1, were decreased while phospho- extracellular signal-regulated kinases (ERK) increased for most CSC lines (Fig 2A). Phospho-EGFR (Y1173 and Y1068) levels were increased in response to growth factor withdrawal exclusively in the EGFRvIII line HF2927 (Fig 2A). Expression of MET proto-oncogene, receptor tyrosine kinase (cMET) was observed for all CSCs (S1 Table), but in the absence of MET ligand (hepatocyte growth factor) in the media, only HF2927 and HF3016 presented modest levels of MET activation, which was suppressed in 2% FBS (Fig 2A).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Glioblastoma patient-derived models and experimental design.

A) Clinical and molecular data associated with the GBM patient-derived models. GS, Gliosarcoma; M/F, male/female; RT, radiation therapy; TMZ, temozolomide; Rx1, first-line therapy; TTP, time to progression; OS, overall survival; U/M, unmethylated/methylated MGMT promoter. Red outline denotes extrachromosomal (ecDNA) amplification. MDM2, Mouse double minute 2 homolog; CDKN2A, cyclin-dependent kinase inhibitor 2A; NF1, Neurofibromin 1; MYC family, MYC proto-oncogene, bHLH transcription factor; ND, not determined; AMP, amplification; mut, mutation; hom del, homozygous deletion; WT, wild type; 0, diploid. B) Schematic depicting neurosphere culture in NMGF media for selection and amplification of cancer stem cells (CSCs) from surgical specimens, followed by 2 weeks incubation in altered conditions: Withdrawal of growth factors (NM), and addition of 2% or 10% FBS. For the purposes of this study, serum-differentiated cells (SDC) are CSCs progeny cultured in NMGF supplemented with 2% FBS. The effect of the different culture conditions on cell signaling were compared by RPPA. C) The sensitivity of CSC and SDC to single dose radiation (RT) or 5-day TMZ was measured in three select models. Transcriptional and epigenomic reprograming in SDC vs CSC, and in response to treatment were evaluated by bulk RNAseq and 450k DNA methylation array. Panels B and C were generated in BioRender (Created in BioRender. Berezovsky, A. (2024) https://BioRender.com/n49q520).

https://doi.org/10.1371/journal.pone.0315171.g001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Targeted proteomics reveals common and cell specific alterations in key signaling CSCs in response to differentiation and identifies interdependencies in activation of key signaling mediators.

A) Heatmap depicting log2 fold change (FC) relative to NMGF (CSC) for the top 5% most variable proteins/PTM. Each value represents mean log2FC for n = 3 RPPA measurements/group. Proteins/PTMs which were not detected under all three culture conditions were filtered out for each model (grey cells). MUT, mutation; WT, wild type; AMP, amplification; DEL, deletion B) Heatmap of Spearman correlation coefficient values comparing phosphorylation levels in key cell signaling mediators in 8 patient derived GBM models grown in 4 media conditions in triplicates.

https://doi.org/10.1371/journal.pone.0315171.g002

We used this unique targeted proteomics dataset, integrating various genomic backgrounds and culture conditions, to further investigate the signaling networks in glioblastoma. The Spearman correlation of key proteins and post-translational modifications levels in all 96 samples (8 GBM lines in 4 media conditions in triplicate) was calculated and displayed in the heatmap (Fig 2B). Glycogen synthase kinase-3 (GSK3)α/β is a multifunctional kinase frequently inhibited by phosphorylation downstream of PI3K/AKT, mTORC1 or MAPK signaling. Here, we observed that phosphorylation of GSK3α/β at S21/S9 was strongly correlated with AKT activation (Fig 2B). While mTORC1 is activated by AKT [35], in this dataset AKT activation did not correlate with markers of mTORC1 signaling, such as mTOR phosphorylation at the C-terminus S2448 residue [36], p70 ribosomal protein S6 kinase 1 (S6K) phosphorylation at T412, or phosphorylation of RBS6 [37] (Fig 2B). On the other hand, phosphorylation of AMP-activated protein kinase (AMPK) subunit β1 at S108 was highly correlated with mTORC1 activation (Fig 2B). AMPK αβɣ heterotrimeric complex is activated in response to low oxygen, low nutrients, and DNA damage, playing important roles in cancer [37]. AMPK is inactivated by AKT phosphorylation of AMPKα1 at S485 residue, while phosphorylation at AMPKβ1 S108 sensitizes AMPK to agonists [38], having a function in pro-survival pathways [39]. SRC proto-oncogene, non-receptor tyrosine kinase (Src) is a node of convergence for receptor mediated signaling pathways, activating substrates in the MAPK, JAK/Stat3, and PI3K/AKT pathways to promote tumor cell survival, proliferation and invasion in various cancers, including GBMs [40]. Src activity is inhibited by phosphorylation of Y527 by C-terminal Src kinase (CSK) and regulated by dephosphorylation by protein tyrosine phosphatases (PTPs), leading to conformational change and autophosphorylation of Y416 [41]. Levels of inhibitory Src phospho-Y527 were elevated in HF2354, HF2587, HF3016 and HF3077 relative to other CSC lines while the levels of Src phospho-Y416 were less variable among CSC lines and media conditions (S1 Table) and positively correlated with Src phospho-Y527 (r = 0.705, CI: 0.584 to 0.796), but not with phospho- signal transducer and activator of transcription 3 (Stat3 S727), a canonical downstream signal transducer (Spearman correlation, r = 0.302, CI: 0.1026 to 0.4788) (Fig 2B). This bi-phosphorylation of Src has been previously reported to reflect an intermediate state between open and closed kinase conformations with blockage of the SH3 domain leading to limited Stat3 phosphorylation activity, suggesting that additional stimulatory factors are required for downstream signaling [42]. Phospho-Stat3 was positively correlated with mTORK1 (phospho-mTOR, phospho-SK6) and MAPK (phospho-RSK) signaling (Fig 2B). Signaling through insulin like growth factor 1 receptor (IGF1R) and insulin receptor (IR) promotes cancer cell proliferation, survival, and treatment resistance in diverse malignancies. Upon ligand binding, IGF1R homodimer or IGF1R/IR heterodimer receptors recruit insulin receptor substrate (IRS) adaptor protein to transduce downstream signaling through binding of src-homology-2 (SH2) proteins, such as p85 to activate PI3K and growth factor receptor bound protein 2 (GRB2) to activate ERK [43]. All four media conditions tested here contain insulin at a final concentration of 50 μg/ml (N2 supplement), which has been shown to activate IGF1R, although at a lower rate than the canonical ligand IGF2 [44], while the FBS supplemented media contains additional insulin and IGF1/2 ligands [27]. Consistently, activation of IGF1R (phospho-Y1135/Y1136) /IR (phospho-Y1150/Y1151) presented little variation among GBM lines and culture conditions (S1 Table). mTORC1 phosphorylates IRS1 at S616 residue leading to the adaptor turnover [45], but p-IRS1 did not correlate with mTORC1 activation. Instead, a strong correlation of p-IRS1 with inactivating phosphorylation of forkhead box protein O1 (FOXO1/FKHR) on S256, which is attributed to AKT [46], was observed (Fig 2B). This targeted proteomics analysis using genomically diverse patient samples in different culture conditions reveled a complex pattern in the integration of signaling cascade nodes, likely reflecting deregulation of multiple oncogenic pathways observed in GBMs.

Z-scores were calculated for each protein/PTM based on its distribution across all 8 models and 4 media conditions, and cumulative z-scores were determined for key signaling pathways and cellular functions (S1 Table). Interestingly, we observed a remarkable pattern of inhibition for key signaling pathway and functions in 10% FBS, in contrast with 2% FBS (SDC) cultures (Fig 3A), further corroborating the inadequacy of culturing GBM cells in 10% FBS [28,47]. Receptor tyrosine kinase activation (“RTK”) index, based on the levels of Tyr phosphorylation corresponding to activation of 11 RTKs (S1 Table), decreased in 2% FBS (SDCs) relative to CSCs in HF2927 and HF3016, the only 2 models harboring EGFR amplification on extrachromosomal DNA (ecDNA) [11], and increased in the remaining 6 GBM models (Fig 3A). The levels of EGFR were 10-fold higher for HF2927 CSCs, which also carries the EGFRvIII variant (Fig 1A), relative to HF3016, while the levels of phospho-EGFR at carboxy-terminal Y1068 and Y1173 residues were comparable for these 2 CSC lines (Fig 3A). Interestingly, EGF withdrawal for 2 weeks (NM) led to a 4-fold increase in the levels of EGFR activation in HF2927 (S1 Table, Fig 2A), suggesting an inhibitory effect of long-term exposure to EGF ligand on EGFRvIII. RTK activation correlated with downstream RTK signaling pathway (“TK_Signaling”), which includes both MAPK and PI3K/Akt/mTOR (“PI3K”) pathways (Fig 3A). As part of RTK signaling convergence, SHC Adaptor Protein 1 (SHC1) binds to phosphorylated EGFR Y1173, which is further phosphorylated on Y239 and Y317 resulting in recruitment of Grb2 and downstream activation of MAPK and PI3K/AKT pathways [46]. Phospho-Shc1 (Y317) was only observed for HF2927 and HF3016 CSCs, and in both cases upregulated upon growth factor withdrawal (NM) and downregulated in the presence of FBS, following the pattern of EGFR activation (S1 Table, Fig 2A). MAPK signaling activation and cytoskeletal function, consisting of regulation of actin and microtube dynamics (S1 Table), were maintained or increased in SDCs for all models except HF2927 and HF3016 (Fig 3A), analogous to what we observed for the RTK index. The PI3K/AKT/mTOR pathway activation score was low for HF2927 CSCs, relative to the other CSC lines, and remained low in NM and 2% FBS, despite genomic loss of PTEN, EGFRvIII amplification, and high level of EGFR activation (Fig 3A). PI3K/AKT/mTOR activation index was higher for the other 7 CSCs, maintained in SDCs, and decreased in the absence of growth factors (Fig 3A).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Comparative targeted proteomics of CSCs growing in 4 different conditions.

A) Heatmap depicting mean signaling pathway activation for triplicate samples in 4 culture conditions, for CSCs derived from 8 patients with diverse genomic landscape, transcriptional subclass and MGMT promoter methylation. Hierarchical clustering of rows was based on Spearman correlation analysis. B) Heatmap for protein/PTM z-scores involved in cell death and survival signaling for 3 models grown in NMGF (CSC), NM and 2% FBS (SDC), in triplicate (boxes in (A)).

https://doi.org/10.1371/journal.pone.0315171.g003

Since several of these signaling pathways are involved in cell survival and proliferation, we established a cell “viability index” measuring the net result from the sum of the level of protein/PTM associated with cell survival and proliferation minus the level of markers associated with programed cell death (S1 Table). We observed that the viability index varied among CSC lines and was not impacted by culture conditions in the same pattern, decreasing in SDC in half of the models and increasing in the other half (Fig 3A). Three models, HF2303, HF2927 and HF3016, representing each of the 3 transcriptional subclasses, diverse genomic features (Fig 1A), and a range of viability index values, were selected for further characterization. Levels of individual protein/PTM associated with cell death and survival are shown (Fig 3B). High PTEN expression was observed in HF2303, harboring pathogenic p.D92E mutation in the WPD loop [48], corresponding to a high level of phosphorylation at S380 in the carboxy terminal tail, a crucial post-translational modification for blocking proteasomal degradation [49]. High levels of AKT activation observed for HF2303 CSCs were further increased in SDCs, indicating that the p.D92E mutation impairs PTEN ability to inhibit AKT. Conversely, PTEN levels were very low in HF3016, which carries a pathogenic mutation in the P-loop, p.I122S [48], comparable to PTEN-null HF2927 (Fig 3B). Increased levels of phospho-ERK and phospho- c-Jun N-terminal kinase (JNK) were observed for HF2303 and HF2927 in response to growth factor withdrawal (NM), and in HF2303, the increase was also observed in SDCs, relative to CSCs. The levels of p53 protein were consistent with the mutation status for these lines (Fig 3B). Thus, our comparative analysis confirms that at the individual protein/PTM levels, differences in pathways regulating cell survival and death in untreated GBM cells are largely determined by the genomic landscape, with cell state playing a lesser role.

CSC presents increased sensitivity to temozolomide but not to radiation compared to their SDC progeny

MGMT epigenetic silencing is a clinically used biomarker to predict response to TMZ treatment. HF3016 and HF2303 CSCs were derived from tumors with unmethylated and HF2927 CSC from a tumor with methylated MGMT promoter (Fig 1A). Accordingly, HF2927 cells presented the lowest level of MGMT protein. Despite both having unmethylated promoter status, MGMT levels were 2.5-fold higher for HF3016 relative to HF2303 (Fig 4A, S1 Table). MGMT protein level in HF2303 was unchanged among all three culture conditions, while MGMT was significantly downregulated in SDCs relative to CSCs for HF2927 and HF3016 (Fig 4A). We confirmed that MGMT promoter was unmethylated for HF2303 and methylated for HF2927 cells and did not observe changes in methylation between CSC and SDCs for either of these models (S2 Fig). DNA methylation data is unavailable for HF3016 SDCs, preventing us from entirely ruling out differentiation-mediated alterations in MGMT promoter methylation in this model. We observed that MGMT protein expression pattern in the two models with unmethylated promoter, HF3016 and HF2303 (Fig 4A), was highly correlated (Spearman r = 0.823; 95% CI: 0.6213 to 0.9228) with the levels of activated NF-κB p65 (phospho-S536) (Fig 3B, S1 Table). Remarkably, the correlation between activated NF-κB and MGMT levels was significant for all models, regardless of promoter methylation status, but stronger for those presenting unmethylated MGMT (Fig 4B). To investigate to what extent the developmental state affected the sensitivity of GBM cells to DNA-damaging treatments, HF2303, HF2927 and HF3016 CSCs, along with their SDC progenies, were treated with TMZ and RT (Fig 1C). These models represent diverse molecular characteristics known to modulate response to DNA-damaging agents, such as TP53 status and MGMT promoter methylation. First, we compared cell proliferation rates between 3D CSC and 2D SDC cultures. SDCs proliferated at a higher rate than CSCs for HF3016 only, while no significant difference was observed for the other two lines (Fig 5A). Cells were then treated for 4 days with doses ranging from 0 (DMSO) to 400 mM TMZ, in quintuplicates. To compensate for possible binding of TMZ by proteins in FBS, even at the low concentration used, the serum-free neurosphere media was supplemented with bovine serum albumin (1%) [50]. Half maximal inhibitory concentration (IC50) and area above the curve (AAC) were calculated from dose-response curves generated by non-linear fitting. HF2927 cells were more sensitive to TMZ in comparison with the other two cell lines (Fig 5B), as expected based on MGMT promoter methylation and expression (Fig 4A). Differentiation of CSCs into SDCs significantly increased resistance to TMZ for all three lines (Fig 5B), despite downregulation of MGMT protein observed for HF2927 and HF3016 SDCs (Fig 4A).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. The impact of culture conditions on MGMT protein expression is cell line specific and correlates with NF-κB activation.

A) MGMT protein levels measured by RPPA for HF2303, HF2927 and HF3016, in biological triplicates (S1 Table). P-values for 1-way ANOVA and post-hoc TukeyHSD test for comparison of the effect of culture conditions on MGMT protein expression are shown. Pairwise comparison between culture conditions was only significant for HF2927 and HF3016. B) Correlation of the levels of MGMT and phospho-NFκB (S536), for all 8 GBM models and culture conditions (left panel) and for the 4 models presenting unmethylated MGMT promoter (right panel). Spearman r and 95% CI are shown.

https://doi.org/10.1371/journal.pone.0315171.g004

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. GBM CSCs are more sensitive to temozolomide than SDCs.

A) To compare cell proliferation rates, CSCs and SDCs for each line were plated in the respective media in 96-well assay plates (1,000 cells/well) and cell viability was measured over 11 days in culture (n = 5). Graphs represent mean (SE). The calculated doubling-times (DT, days) are shown, pairwise comparison between CSC and SDC values were significant for HF3016 only (*) p<0.01. B) Cells were plated in 96-well assay plates (2,000/plate) and treated for 5 days with the indicated doses of temozolomide, or equivalent DMSO control. Dose-response curves, the calculated IC50 concentrations with confidence interval (CI), and area above the curve (AAC) expressed as a fraction of total area with SE are shown. For each of the GBM patients, CSCs were significantly more sensitive to TMZ relative to SDCs, as shown by IC50 and AAC ratios, and p-values for dose-response curve comparison (F-test).

https://doi.org/10.1371/journal.pone.0315171.g005

We then compared the sensitivity of CSCs and SDCs to clinically relevant RT doses. Isogenic populations of CSCs and their SDC progenies were exposed to 2 gray (Gy) and 4 Gy RT doses and cultured for 5 days, surviving fractions (SF2 and SF4) relative to control treatment were determined (Fig 6A) as previously described [51]. Sensitivity to RT was variable among these models, as a high degree of resistance was observed for HF2303 cells relative to both HF3016 and HF2927; HF2927 was the most sensitive (Fig 6A). On the other hand, no significant difference in sensitivity to RT between CSCs and SDCs were observed for these three models (Fig 6A). To evaluate the long-term effect of radiation, we compared the tumorigenic potential of control and RT treated CSCs. Cells expressing firefly luciferase (fLuc) were implanted intracranially in nude mice immediately after radiation (4 Gy), or mock radiation (control). Tumor growth was monitored by weekly bioluminescence imaging and by daily monitoring of symptoms associated with tumor burden. The three irradiated CSC lines exhibited a delay in tumor growth relative to controls, assessed by symptom-free survival, with all mice eventually developing tumors, demonstrating that cells surviving 4 Gy RT dose were tumorigenic. HF2303 displayed the highest resistance to RT in vitro and exhibited the least significant difference in the survival curves. On the other hand, the HF2927 cells, which were the most sensitive to RT in vitro, displayed the lowest p-value in the curve comparison (Fig 6B and 6C). The degree of difference in survival curves (Log-rank p-values) ranked according to the in vitro sensitivity of each cell line to RT.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Response of glioblastoma CSCs and SDCs to ionizing radiation.

A) CSC and SDC received a single radiation dose (2Gy or 4Gy) or mock radiation (control), and cell viability was measured 5 days later (n = 3–4). No significant difference in sensitivity to RT between CSC and SDCs was observed for any of the lines (Holm-Sidak multiple comparisons test with adj p-value<0.05 threshold). The surviving fraction (SF) values among the 3 cell lines (CSC and SDC combined) were compared by one-way ANOVA, followed by Tukey’s multiple comparisons tests, adj p value<0.0001(***), <0.05 (*). B) Irradiated and control CSCs were implanted intracranially immediately after treatment, 3x10⁵ cells/mouse (n = 6-8/group). Symptom-free survival for the orthotopic PDX is shown in Kaplan-Meier curves, compared by log-rank test. C) Longitudinal live bioluminescense images for representative PDX in (B).

https://doi.org/10.1371/journal.pone.0315171.g006

Transcriptional reprogramming in response to CSC differentiation

To determine how differentiation state impacts transcriptional reprogramming in response to TMZ and RT, we selected the most resistant (HF2303) and the most sensitive (HF2927) models. We first determined the pattern of global transcriptome changes after differentiation of CSCs into SDCs from bulk RNAseq data. To identify differentially expressed genes between the two differentiation states, we employed the non-parametric NOISeq R package, and subsequently performed pathway enrichment analysis, as detailed in the Materials and Methods section. Numerous signaling pathways and cellular processes enriched in CSCs were cell line specific. It has been proposed that CSCs, from GBMs and other cancers, tend to rely more on oxidative phosphorylation (OXPHOS) than the more differentiated cancer cells [52]. Here, we observed that OXPHOS gene expression signature was enriched in HF2303 CSCs but not HF2927 CSCs (Fig 7A). Interestingly, reserve respiratory capacity was higher in untreated HF2303 relative to HF2927 cells, and higher in SDC compared to CSCs (S3 Fig). Replication-dependent histones were highly enriched in HF2303 CSCs, in the absence of enrichment in cell proliferation markers (Fig 7A, S2 Table), as previously observed for embryonic and induced pluripotent stem cells [53]. Several nuclear pseudogenes with sequence similarity to mitochondrially encoded 16S rRNA (e.g. MTRNR2L1, MTRNR2L3), enriched in HF2303 CSCs (A in S2 Table) are predicted to encode humanin-like peptides and to be involved in negative regulation of the execution phase of apoptosis [54] (Fig 7A). IL2-STAT5 signaling was specifically upregulated in HF2927 CSC in relation to SDC (Fig 7A), in agreement with levels of pSTAT5_Y694 being elevated in HF2927 CSC (4.697 ± 0.98), while undetectable in HF2927 SDCs and HF2303 cells (B in S1 Table). Patterns of transcriptome reprogramming in response to differentiation that were common to both HF2927 and HF2303 models included downregulation of genes involved in cholesterol biosynthesis and upregulation of genes associated with ECM components and organization, including collagen type IV alpha 6 chain (COL4A6), a target of SRY-box transcription factor 2 (SOX2) transcriptional regulation [55] (Fig 7B). The expression of several transcription factors associated with stemness and pluripotency, such as SOX2, spalt like transcription factor 2 (SALL2) and POU class 5 homeobox 1 (POU5F1 (Oct3/4)), was not altered upon differentiation of HF2303 and HF2927 CSCs into SDCs (A,B in S2 Table). On the other hand, other neural progenitor and stemness markers expressed in CSCs were down regulated upon differentiation: oligodendrocyte transcription factor 1 (OLIG1) in both lines; delta like canonical Notch ligand 3 (DLL3), integrin subunit alpha 6 (ITGA6), ETS variant transcription factor 4 (ETV4) in HF2927; fatty acid binding protein 7 (FABP7), oligodendrocyte transcription factor 2 (OLIG2) and ELOVL fatty acid elongase 2 (ELOVL2) [56] in HF2303 (Fig 7A and 7B).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Transcriptional reprograming in glioblastoma cells in response to CSC differentiation.

HF2927 and HF2303 CSC were grown in regular neurosphere media or in SDC cultures for 2 weeks. Total RNA from triplicate samples for each group was sequenced analyzed and DEG between CSC and SDC determined for each line. Total number of protein coding (PC) and non-coding (NC) genes, select genes and Metascape gene-set enrichment tests adjusted for multiple comparisons (q-value < 0.05) are shown for genes upregulated in CSCs (A) and in SDCs (C). The common DEGs for both lines along with enriched cellular processes are presented: 43 genes upregulated in CSCs (B) and 25 genes upregulated in SDCs (D). Figure generated using BioRender (Created in BioRender. Berezovsky, A. (2024) https://BioRender.com/n49q520).

https://doi.org/10.1371/journal.pone.0315171.g007

IFN and interleukins signaling were specifically upregulated in HF2303 SDCs (Fig 7C). Indeed 29 genes associated with IFN response pathways were upregulated in HF2303 SDCs (Table 1), including members of the 49-gene interferon-related DNA damage resistance signature (IRDS) [57]. Differentiation of CSCs into SDCs resulted in enrichment of the epithelial mesenchymal transition (EMT) signature for HF2303 but not for HF2927 (Fig 7C), consistent with the transcriptional subclasses affiliation, mesenchymal for HF2303 and classical for HF2927 (Fig 1A). In agreement with our observation that receptor dependent SMAD1/3/8 are activated in GBM SDCs (Fig 2A), inhibitor of DNA binding 1 (ID1), ID2 and ID3, BMP/retinoic acid inducible neural specific 1 (BRINP1), downstream of BMP signaling, were upregulated in HF2927 SDCs (Fig 7C). Several components of the TGFb superfamily signaling were upregulated in HF2303 SDCs, including transforming growth factor beta receptor 2 (TGFBR2), BMP1, latent transforming growth factor beta binding protein 1 (LTBP1), LTBP3, SMAD3, SMAD9 (Fig 7C and 7D; S2 Table). On the other hand, endogenous BMP4 mRNA was downregulated and BMP antagonists DAN family BMP antagonist (NBL1) and BMP and activin membrane bound inhibitor (BAMBI) were upregulated in HF2303 SDCs, suggesting a possible negative feedback loop, specific for this model. BAALC binder of MAP3K1 and KLF4 (BAALC), a regulator of developing neuroectoderm tissues overexpressed in acute leukemia and GBM [58], and a target of SOX2 in differentiated GBM cells [26], was upregulated in HF2303 SDCs (Fig 7C). Toll like receptor 4 (TLR4), which is expressed in both immune and neoplastic cells, was also upregulated in HF2303 SDCs (Fig 7C), in agreement with previous studies [59]. HIF1A, an IFN responsive gene (Table 1), along with downstream hypoxia markers were uniquely upregulated in HF2927 SDCs, under normal oxygen levels (Fig 7C)`. The 24 genes commonly upregulated in SDCs for both lines were enriched in ECM components (Fig 7D). Small nucleolar RNAs were also upregulated upon differentiation of both HF2303 and HF2927 (Fig 7D; S2 Table), including SNORD116 cluster, shown to have non-canonical functions in the differentiation, survival, and proliferation of neuronal cells [60].

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Differentially expressed genes associated with IFN response and signatures.

https://doi.org/10.1371/journal.pone.0315171.t001

Transcriptional response to sublethal treatment with temozolomide and radiation is determined by genomics and differentiation state

To compare the transcriptional response to temozolomide treatment at distinct differentiation states, GBM cells were treated in triplicate with the respective TMZ IC40 concentrations, calculated from dose-response curves in Fig 5B: 58 mM (HF2303 CSC), 25 mM (HF2927 CSC), 134 mM (HF2303 SDC), 70 mM (HF2927 SDC), or equivalent DMSO control for 4 days. Cells were then harvested and total RNA isolated and sequenced. Differentially expressed genes between treated and control samples for each model and culture type were determined as described for Fig 7 (C and D in S2 Table). Transcriptional response to 4-day TMZ treatment was largely regulated by wt p53 activation in HF2927 cells (Fig 8A). We observed that 18 out of the 36 genes commonly upregulated in HF2927 CSCs and SDCs in response to TMZ treatment (Table 2) are direct transcriptional targets of p53 [55]. However, the differentiation state of HF2927 cells also influenced p53 transcriptional specificity, as 23 and 26 p53 target genes were significantly upregulated in response to TMZ exclusively in CSCs or SDCs, respectively (Table 2). In contrast, interleukin-27 (IL-27) mediated signaling pathway was enriched in HF2303 SDC in response to TMZ (Fig 8A), based on the upregulation of Myxovirus resistance dynamin like GTPase 1 (MX1), 2’ 5’ oligoadenylate synthetase 1 (OAS1) and OAS2, all of which are members of the IRDS [57], providing further evidence for the proposed cooperation between IL-27 and IFN signaling in response to DNA damage [61]. To identify other components of IFN signaling in GBMs, we identified 100 genes whose expression presented the highest correlation to OAS1 (Spearman’s r > 0.55; q value < 5.0E-12) in The Cancer Genome Atlas (TCGA) GBM data set (CBIOPortal [62], accessed on 01/19/24). 62% of these genes are members of the hallmark IFN alpha or gamma response pathways, 24% are members of the IRDS [57], and this gene list also includes the IFN/STAT1 prognostic signature proposed for GBM proneural subtype [63] (S3 Table). We observed activation of IFN response in TMZ-treated HF2927 CSCs and SDCs, and in HF2303 SDCs (Fig 8A, Table 1). TMZ treatment induced expression of ECM components, and several neural developmental growth and survival factors, including brain derived neurotrophic factor (BDNF) and glial cell derived neurotrophic factor (GDNF), in HF2927 CSCs and SDCs (Fig 8A). Expression of other genes associated with neural development were down regulated in TMZ treated HF2927: platelet derived growth factor receptor alpha (PDGFRA), OLIG2, and prominin 1 (PROM1) in SDCs; neuropilin 1 (NRP1), GLI family zinc finger 1 (GLI1), and Kruppel like factor 9 (KLF9) in CSCs (Fig 8A). CDKN1A (p21), the main effector of p53-mediated downregulation of cell cycle genes, leading to cell cycle arrest, was upregulated in response to TMZ in both HF2927 CSCs and SDCs (Fig 8A and 8D). The more pronounced downregulation of cell cycle genes in HF2927 SDCs relative to CSCs (Fig 8A) is consistent with corresponding upregulation of Lin-37 DREAM MuvB core complex component (LIN37) in SDC but not in CSC (Fig 8A), as LIN37 is required for transcriptional repression by the DREAM complex, downstream of p53 signaling [64]. A possible explanation for these observations is that cell cycle-arrested TMZ-treated HF2927 CSCs may have undergone apoptosis to a greater extent at the time point analyzed, as they demonstrated increased sensitivity to TMZ treatment (Fig 5B). Furthermore, genes coding for proteins involved in EMT, including the master regulator twist family bHLH transcription factor 1 (TWIST1), were upregulated in response to TMZ exclusively in HF2927 SDCs, which could contribute to the increased resistance of the differentiated cells.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Transcriptional response to genotoxic treatment in glioblastoma cells.

Differentially expressed genes between treated and control samples were determined and enrichment performed as described for Fig 7. A) Transcriptional changes in CSCs and SDCs in response to 4-day TMZ treatment (IC40 concentrations). B) Transcriptional changes in CSCs and SDCs measured 5 days after one dose RT (4 Gy). C) RNA commonly downregulated in HF2303 CSCs in response to differentiation and RT treatment. D) mRNA expression (RPKM) of p53 transcriptional targets CDKN1A (p21), MDM2, and BBC3, and of cell cycle regulators in HF2303 and HF2927 control and treated cells, mean and SE of n = 3. Figure created with Biorender.com (Created in BioRender. Berezovsky, A. (2024) https://BioRender.com/n49q520).

https://doi.org/10.1371/journal.pone.0315171.g008

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Genes upregulated in HF2927 cells in response to temozolomide.

https://doi.org/10.1371/journal.pone.0315171.t002

Ionizing radiation can damage molecules directly, exerting most of its toxicity to cancer cells through DNA double strand breaks (DSBs), and indirectly through the generation of free radicals. The complex cellular responses leading to cell cycle arrest, apoptosis, DNA repair and other processes occur over time [65]. The relative sensitivity of glioblastoma cells to RT varied according to tumor of origin but not according to differentiation status under the experimental parameters employed here (Fig 6A). Furthermore, pre-implantation treatment of CSCs with 4Gy led to delayed xenograft tumor growth for all models (Fig 6B). To integrate these observations with cellular responses, CSCs and SDCs were treated in triplicates with 4 Gy radiation dose or mock radiation control, followed by incubation in the respective growth media for 5 days prior to RNA isolation and sequencing. Under these conditions, few transcripts were altered in response to RT in HF2303 (Fig 8B), consistent with the resistance to 4 Gy dose observed for both SDCs and CSCs (Fig 6A). The only transcripts commonly upregulated in HF2303 CSCs and SDCs in response to RT were phospholipase C eta 2 (PLCH2) and SNORD3D (Fig 8B). Interestingly, 50 of the pseudogenes involved in negative regulation of the execution phase of apoptosis, which were downregulated in response to differentiation of CSCs into SDCs (Fig 7), were also downregulated in RT-treated HF2303 CSCs (Fig 8B and 8C). Compared to TMZ treatment, fewer IFN response genes were upregulated in RT treated cells, including MX1 which was upregulated in both CSC lines, and OAS2, HERC6, IFI44L and HLA-E in HF2927 CSCs (Table 1, Fig 8B). In contrast to the response to TMZ, where the cells were analyzed after 4 days of continuous treatment (Fig 8A), p53-mediated signaling in HF2927 was more attenuated 5 days after RT, as expected (Fig 8B). Few p53 transcriptional targets remained elevated in irradiated HF2927 cells, such as BCL2 Binding Component 3 (BBC3 / PUMA), a pro-apoptotic BCL-2 family member, while expression of several others, including MDM2, were not altered at the 5-day timepoint (Fig 8B and 8D). CDKN1A/p21, the main effector of p53 mediated cell cycle arrest significantly upregulated in response to TMZ (Fig 8A), was elevated to a lesser degree in response to RT, below the statistical threshold employed to determine DEGs (Fig 8D). However, sustained downregulation of cell cycle genes downstream of p53 activation was observed for RT-treated HF2927 CSCs and SDCs (Fig 8B). Indeed, 53 out of the 56 transcripts commonly downregulated in HF2927 CSC and SDCs in response to RT are transcriptional targets of the DREAM (dimerization partner, RB-like, E2F and multivulval class B) repressive complex [66] (Table 3). Among these targets, forkhead box M1 (FOXM1), a transcriptional activator of cell cycle genes, especially those involved in G2 and M phases [67], along with its transcriptional targets, including Polo-like kinase 1 (PLK1) which phosphorylates and activates FOXM1 in a positive feed-back loop [68], were downregulated exclusively in RT-treated cells at the analyzed time point (Fig 8B and 8D). Furthermore, DNA damage response and double strand break repair genes were downregulated in response to RT predominantly in HF2927 CSCs (Fig 8B), indicating that these cells would be sensitized to subsequent DNA damaging interventions.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Genes downregulated in HF2927 CSC and SDC in response to radiation.

https://doi.org/10.1371/journal.pone.0315171.t003

DNA methylation alteration pattern in differentiation of CSCs

To verify the contribution of epigenetic regulation to the transcriptional responses reported above, we analyzed the global DNA methylation profile of control and treated HF2303 and HF2927 CSCs and SDCs in triplicates, using Infinium Human Methylation 450K BeadChip. Differences in DNA methylation distinguished the samples according to the tumor of origin (HF2303 vs HF2927), as shown in the t-distributed stochastic neighbor embedding (t-SNE) plot (Fig 9A). Analysis of the differentially methylated positions (DMPs) and regions (DMRs) between CSC and SDC for each of these models identified 892 DMP and 12 DMRs for HF2303, and 27,431 DMP and 2 DMRs for HF2927 (S4 Table). Similar to the significant overlap in DEGs associated with CSC differentiation between the two cell lines (Figs 7B, 7D and 9B), we identified 384 DMPs common to both lines (Fig 9B, S4 Table), with a pattern of decreased b-values in SDCs relative to CSCs in HF2303, while HF2927 SDCs exhibited an increase in b-values. The intersection between genes mapped to the DMPs, at the transcription start sites or within CpG islands, and the genes differentially expressed between CSCs and SDCs were limited to 17 and 59 genes for HF2303 and HF2927, respectively. Among these genes, latent transforming growth factor beta binding protein 3 (LTBP3) was associated with DMPs and upregulation in SDC for both models (Fig 9B, S4 Table). Cholesterol homeostasis, hypoxia, growth factor binding, and ECM components were significantly represented in the 59 gene list from HF2927 (S4 Table). No significant DNA methylation changes were observed in the acute response to RT and TMZ treatment across both models and differentiation states (S4 Fig).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Genome-wide DNA methylation profiling groups samples by patient of origin and identifies DNA methylation alteration patterns between glioblastoma CSCs and their SDCs progeny.

A) t-distributed stochastic neighbor embedding (t-SNE) plot from β-values for all CpG sites comparing HF2303 and HF2927 CSCs and SDCs treated with control, RT or TMZ, in triplicates. B) Schematic showing the significant intersect for HF2303 and HF2927 of differentially expressed genes (DEGs) and differentially methylated positions (DMPs) between CSCs and SDCs. The intersect between DMPs mapped to genes (see Methods) and DEGs are shown. Significance of the intersections was determined using GeneOverlap, p-values and estimated odds ratio are shown; n.s., nonsignificant. C) Heatmap showing the β-values for the 387 DMPs common to HF2303 and HF2927. Fig 9B was created with Biorender.com (Created in BioRender. Berezovsky, A. (2024) https://BioRender.com/n49q520).

https://doi.org/10.1371/journal.pone.0315171.g009

Cell culture

Glioblastoma patient-derived cancer stem cell cultures were obtained from the live biobank collection at the Hermelin Brain Tumor Center, Henry Ford Hospital (Detroit, MI). These cells were derived from de-identified surgically resected glioblastoma specimens, collected with written informed consent, under a protocol approved by the HFH Institutional Review Board as previously described [90], in compliance with all relevant ethical regulations for research using human specimens. Cells were cultured in serum-free neurosphere media consisting of Dulbecco’s Modified Eagle Medium (DMEM)/F12 media (Invitrogen), N2 supplement (Gibco) and 0.5 mg/ml BSA, supplemented with growth factors 20 ng/ml EGF and 20 ng/ml bFGF (Peprotech) (NMGF), to select for CSCs [90]. CSC cultures were passaged in vitro and used prior to achieving passage 20. Identity of CSC lines was confirmed by comparing genotype with the patient germline using Short Tandem Repeat (STR) analysis. CSCs were dissociated and cultured for 2 weeks in three different conditions: NM lacking growth factors, and NMGF supplemented with 2% or 10% FBS (HyClone). Neural progenitor cells (NPCs) were derived from H9 human ESC, obtained from WiCell (Madison WI) as described [91] and differentiated in NMGF supplemented in 2% FBS [92].

Reverse Phase Protein Arrays analysis

Cell lysates were obtained in triplicate for each cell line and growth conditions described (Fig 1B). The levels of 66 proteins/post-translational modifications in the cell lysates were analyzed by Reverse Phase Protein Arrays (RPPA) at the Mason’s Center for Applied Proteomics and Molecular Medicine (George Mason University), essentially as described previously [31,32]. Statistical analyses of RPPA normalized values were conducted in the R statistical environment (4.2.1). Heatmaps were generated by ComplexHeatmap package in R (2.12.1) [93]. Z-scores for each protein and post-translational modification were calculated based on population mean and standard deviation ((x-μ)/σ). Fold-change was calculated using the gTools package (3.9.3) and computed as follows: num/denom if num>denom, and as -denom/num otherwise. Dendrograms were produced using Spearman correlation as the distance method.

Cell proliferation, radiation and temozolomide treatment

RNA sequencing

RNA quality assessment was performed using the Agilent RNA ScreenTape on the Agilent 2200 TapeStation. Library for RNA sequencing was prepared using TruSeq Stranded Total RNA LT (Illumina). 100 bps paired-end libraries were sequenced on the Illumina HiSeq 2500 following cBot clustering. Prior to alignment the sequencing adaptors were trimmed using FASTQ Toolkit v1.0 on Illumina’s BaseSpace portal. FastQC was used to generate quality metrics for assessment of FASTQ files [95]. RNA seq reads were aligned using TopHat v2.1.0 using default alignment parameters and GRCh38 reference genome [96,97]. After alignment, the mapped reads were further filtered by mappability score (MAPQ ≥ 10) and sorted by genomic position of the reads using samtools-0.1.19 [98]. PCR duplicate reads were processed and removed using rmdup function (options -S) in samtools. The cleaned BAM files were further analyzed using Integrative Genome Viewer (http://software.broadinstitute.org/software/igv/), processed with R function “FeatureCount” to quantify reads on various RNA types from GencodeV28 annotation. Differential gene expression analysis was performed using the non-parametric algorithm “NOISeq” in R [99], with q = 0.8 and fold change >2 or <0.5 thresholds on TMM normalized data for treated vs. control comparisons, and q = 0.8 on RPKM normalized data for long term transcriptome changes in untreated SDCs vs CSCs. Gene set enrichment tests for differentially expressed gene lists were performed using Metascape 3.5, and adjusted p-value (q) < 0.05 [100].

DNA methylation analysis

Genomic DNA was analyzed using Infinium Human Methylation 450K BeadChip system (Illumina), as described [91]. Raw methylation data was pre-processed for dye-bias normalization, detection p-value was calculated by comparing the signal intensity difference between array probes and a set of negative control probes on the array; probes with p-value greater than 0.01 filtered out. Beta values were calculated from pre-processed raw data using methylumi Bioconductor package in R [101]. Probes mapping to known SNPs and sex chromosomes were filtered out of dataset as recommended by Zhou et al. [102]. Beta values were corrected for batch effect using ComBat function within sva R package [103,104]. Probes were aligned to hg38 and annotated using Illumina methylation 450k manifest [105]. Differentially methylated positions between experimental groups (n = 3/group) were determined using champ.DMP function, and to determine differentially methylated regions we used the champ.DMR with Bumphunter method from the ChAMP package in ChAMPs R package [106] with raw p-values < 0.05, and adjusted p-value < 0.1. DMPs were annotated to genes using illuminaHg19 annotation employed by the ChAMPs package in R. Analyses to determine the intersect of DMPs, the genes mapped to DMPs and DEGs, within and between models, were carried out using GeneOverlap (version 1.34.0).

Cellular oxygen consumption rate

Cellular oxygen consumption rate (OCR) was measured using a Seahorse XF Cell Mito Stress Test in a XFe24 Bioanalyzer (Agilent). Cells were plated at a concentration of 30,000 (HF2303) or 40,000 (HF2927) per well on poly-ornithine hydrobromide (Sigma # 4957) coated plates in NMGF media (CSC) or 2%FBS media (SDC) and treated for 4 days with DMSO or TMZ IC40 concentrations, prior to basal oxygen consumption measurements, performed according to the manufacturer’s instructions.

Statistical analysis

Statistical analyses were performed using GraphPad Prism (v. 9.00) or R and are described in the results section for each experiment. Tumor growth in mouse PDX was evaluated by Kaplan-Meier survival curves, compared by the log-rank test with p<0.05 considered significant.