GliNS1, G179NS and G166NS GIC lines were grown in culture as previously described [6], [11]. Briefly, the cells were first grown as spheres in the first week before transferring to laminin-coated dishes, where they were grown as adherent monolayers in serum-free human Neurocult NS-A basal media (StemCell Technologies) supplemented with Glutamax, Hepes, N2, B27 (Invitrogen), EGF (10 ng/ml) and bFGF (10 ng/ml) (R&D Systems). GICs were grown to subconfluence, dissociated using TrypLExpress (Invitrogen), and then split 1∶2–1∶4. 2/3 of medium was replaced with fresh medium every 3–4 days. For differentiation, cells were cultured in DMEM/F12 media supplemented with 10% fetal bovine serum (FBS; Gibco, Life Technologies)) for two weeks

Novel human malignant glioblastoma initiating cell (GIC) cultures (U3NNN-MG series of GIC lines) used in this study are part of the Uppsala University Human Glioma Cell Culture (HGCC) collection, which comprises well-characterized GBM-derived cancer initiating cell cultures (Xie et al 2014, manuscript in preparation). This work was approved by the Uppsala ethical review board (2007/353). All GIC lines were used between passages 15 and 30.

GliNS1, G179NS and G166NS GIC lines, both undifferentiated and differentiated, were seeded on day 1 at 20% density onto laminin-coated 96 or 384 black well, flat bottom microplates (Corning). Compounds were added to the plates on day 2, followed by incubation for 48 hrs. FBS differentiated cells received serum-free media (50% human Neurocult NS-A basal media, 50% Neurobasal media (Gibco, Life Technologies), supplemented with Glutamax, Hepes, B27, ¼ N2, no growth factors) during chemical compound treatment. DMSO was used as negative control. Viability assay was performed using the CellTiterGlo assay (Promega, Madison, WI) according to the manufacturer's recommendations. Briefly, assay reaction buffer was added to the wells using an automated multipipette, followed by shaking the microplate for 30 seconds and 7 min incubation in the dark. Luciferase intensity reading was then taken using Victor2 (Perkin Elmer) with a setting of 1 s luminescence reading per well. Z-factor was calculated for each experiment. For every cell line (GliNS1, G179NS, G166NS and hf5205NS), at least 3 replicates were analyzed. Statistical calculations were performed with GraphPadPrism (GraphPad Software, Inc. San Diego, CA, USA). Dose-response data were processed using log-linear interpolation to obtain log IC₅₀ values.

Drug assays in novel GIC lines (U3NNN-MG series) were seeded in 384-well microplates (BD Falcon Optilux Cat. #353962) 24 hours prior to treatment using a Multidrop 384 liquid dispenser (ThermoScientific, Sweden). To ensure growth phase at end of the assay (∼70% confluency) cells were seeded at a density ranging between 2000–4000 cells/well. Drugs were transferred using the ECHO550 non-contact liquid dispenser (Labcyte, USA) to a 384 V- bottom polypropylene plate. The drugs were then diluted in medium and transferred using the MDT 384 head on a Janus automated workstation (PerkinElmer, USA) to the cell plates. Drugs were tested in 11-point dose dilution series and assayed for viability after 72 hours of treatment on an EnVision Multilabel reader (PerkinElmer, USA) using resazurin (R7017, SigmaAldrich, Stockholm, Sweden), at the excitation/emission wave- length 560/590 nm [20]. As a positive control, the drug doxorubicin was screened with the same dose-response curve setting, and wells containing negative DMSO controls at 4 different concentrations were assayed as well. The effect on viability of each drug dose was calculated as a viability ratio W =  Ytreated/Ycontrol, where Y represents the average fluorescence signal.

Two replicates were analyzed for the undifferentiated and differentiated cell lines GliNS1, G166NS and G179NS, while three replicates were analyzed for DMSO, A23187 and Thapsigargin treated GliNS1 and G166NS. Total RNA was extracted from cells grown to subconfluency using the RNeasy Mini Kit (Qiagen), following the manufacturer's instructions. Fluorometric quantitation of RNA concentration and quality was done using the Qubit RNA assay kit (Invitrogen). We used 300 ng of total RNA in the preparation of the TruSeq library, for which we used the Illumina Low-Throughput TruSeq RNA Sample Preparation Kit protocol resulting in barcoded cDNA. 50 ng of barcoded TruSeq products were used for Illumina RNA sequencing. Samples were sequenced on an Illumina HiSeq 2000 sequencer as single-end 51-nucleotide reads according to the manufactures protocol. Raw reads were mapped to the reference human genome and normalized data was generated for each genomic feature using STRT software [21]. Briefly, raw reads were aligned using Bowtie [22]. Mapped reads were normalized using reads per KB per million reads (RPKM) normalization method [23] whereas unmapped reads were removed. Differential gene expression analysis was done in R-studio using the DESeq package [24] and a script adopted from a previous paper [25]. Benjamini adjusted p-values were used for data analysis. Data analysis was done using Qlucore Omics Explorer 2.0 (Qlucore AB), PRISM 6 (GraphPad Software), DAVID (http://david.abcc.ncifcrf.gov) and GeneVenn (http://genevenn.sourceforge.net).

The Uppsala U3NNN-MG cell lines were not analyzed in replicates. For each cell line total RNA was extracted from cultured cells using the RNeasy Mini kit (Qiagen) and was labeled and hybridized on Affymetrix GeneChip Human Exon 1.0 ST arrays following the instructions of the manufacturer. The expression values were RMAnormalized using the Affymetrix Expression Console software.

Cells cultured attached on cover glass were fixed in 4% paraformaldehyde (PFA) for 15 min at room temperature (RT) followed by antibody incubation at 4°C overnight. The following primary antibodies were used: rabbit anti-glial fibrillary acidic protein (GFAP) (DAKO Cytomation) and mouse anti-beta III tubulin (Tuj1) (Millipore). The cells were then incubated with fluorescence-labeled secondary antibodies anti-rabbit Alexa Fluor 555 and Alexa Fluor 488 (Invitrogen) for 1 hr at RT. Nuclei were counterstained with DAPI and mounted using Immumount (DAKO). Images of stained cells were acquired with an Olympus FV1000 confocal microscope and processed with Adobe Photoshop CS5.1 software.

Cells were lysed in RIPA buffer and protein concentration was quantified using the Bicinchoninic Acid (BCA) Protein Assay Kit (SigmaAldrich; Pierce, Rockford, US). Equal amounts of protein samples were separated by electrophoresis using 10% Mini-PROTEAN TGX gels (Bio-Rad) or 4–12% Bis-Tris polyacrylamide gradient gel (NuPAGE, Invitrogen) under reducing conditions and transferred onto a PVDF membrane (Bio-Rad) or nitrocellulose membrane (Hybond-ECL, GE Healthcare, Sweden). Membranes were blocked in 5% milk for 1 hour and incubated overnight at 4°C with the following primary antibodies: rabbit anti-GRIA1 antibody (Abcam), rabbit anti-S100 alpha 6 antibody (Abcam), mouse anti-BLBP antibody (Millipore) and mouse anti-beta actin antibody (Abcam). The membranes were then incubated with the appropriate horseradish peroxidase-labeled secondary antibody (SigmaAldrich; GE Healthcare) before being revealed by chemiluminescence (Clarity Western ECL Substrate, Bio-Rad; SuperSignal West Pico Chemiluminescent Substrate (Pierce, Rockford, US)). The band intensities were quantified using ImageJ (NIH, USA).

To determine the relationship between human GIC lines and human fetal neural stem cells (NSC), we re-analyzed the microarray data from a previous study by Pollard et al [11]. While the first principal component segregated normal brain from NSCs and GICs (see [11]), the second principal component ranked GICs in relation to their similarity to NSCs, potentially reflecting aspects of stemness (Fig. 1A). The stemness-associated gene SOX2, the NSC marker BLBP (brain lipid-binding protein, encoded by the FABP7 gene) as well as the neuronal marker TUBB3 (Tubulin beta III), which may reflect high potency for concomitant neuronal differentiation, were expressed the highest in NSCs, and expression levels decreased in the order of the GliNS1 group > G179NS > G166NS (Fig. 1B). While the GliNS1 line was transcriptionally related to NSCs (NSC-proximal), the G166NS line, constituted the distal end of the ranking relative to NSCs (NSC-distal), expressing markers shared with microglia and reactive astrocytes and associated with inflammation, for example IL-6, CXCL2 (MIP-2), and CCL20 (Fig. 1B).

A de novo deep RNA sequencing of three of the GIC lines used in the Pollard et al study (GliNS1, G179NS and G166NS) and a NSC line (hf5205NS) showed that more genes were expressed in NSCs than GIC lines (Fig. 1C; S1), potentially reflecting their plasticity and ability to differentiate. Pairwise NSC-GIC gene enrichment and functional annotation (gene ontology) analysis unexpectedly showed that Ca²⁺ ion binding was the most significantly altered category in all three cell lines - a difference that increased in more NSC-distal GIC lines (Fig. 1D).

Cytosolic Ca²⁺ signaling is balanced by various players, such as Ca²⁺ permeable ion channels that increase intracellular Ca²⁺ (e.g. voltage gated Ca²⁺ channels and glutamate receptors, here termed Ca²⁺ provokers) on one hand, and Ca²⁺ binders that reduce free intracellular Ca²⁺ on the other (here termed Ca²⁺ buffers) [26]. Direct pairwise comparisons between different GIC lines showed that the NSC-proximal GIC line (GliNS1) expressed a larger number of ion channel genes that include Ca²⁺ provokers compared to the NSC-distal GIC line (G166NS). These ranged from Ca²⁺ permeable ion channels (e.g. glutamate receptors: GRIA1, GRIA2, GRIK3 and regulatory subunits GRIN2B and GRIN2D), voltage-gated Ca²⁺ ion channels (CACNA1C/-E/-H) and Ca²⁺-activated potassium channels (e.g. KCNN3) (Fig. 1E). In contrast, the NSC-distal GIC line (G166NS) expressed higher levels of sensors of intracellular Ca²⁺ buffers (for example S100 family genes and CALB1).

To further delineate differences in Ca²⁺ gene expression between tested GIC lines and NSCs, expression of Ca²⁺ provokers and buffers was analyzed in more detail in all three GIC lines (S1 Fig.). As suggested in the previous pairwise comparisons, expression of glutamate receptors decreased in NSC-distal GIC lines (Fig. 2A). The AMPA receptor GRIA1 (Ca²⁺ permeable ionotropic glutamate receptor), which showed the highest expression among glutamate receptors, ranked the GIC lines identically to the order GliNS1> G179NS> G166NS seen in the PCA analysis (Fig. 1A) based on comparison with NSC gene expression. In contrast, expression of Ca²⁺ buffers/effectors increased with an inverse rank order of GliNS1 <G179NS <G166NS (Fig. 2B). This was particularly striking for the S100A6 gene that was most abundantly expressed in G166NS. Similar to the mRNA data, western blot analysis of GRIA1 revealed a 70% and more than 95% lower expression in G179NS and G166NS respectively when compared to the NSC-proximal GliNS1 line (Fig. 2B and 2E). The western blot analysis of S100A6 showed a 45% and 90% lower expression in G179NS and GliNS1, respectively, when compared to the NSC-distal G166NS line (Fig. 2C, 2D and 2E).

In addition Ca²⁺ provokers and buffers, plasma membrane localized Ca²⁺ transporters, such as sodium-calcium exchangers (NCX) belonging to the SLC8 family and the SERCA (sarco/endoplasmic reticulum Ca²⁺ ATPase) pump localized in the endoplasmic reticulum (ER) membrane, actively remove cytosolic Ca²⁺ to maintain homeostasis [27]–[30]. To explore potential functional implications of a differential expression of Ca²⁺ ion channels and Ca²⁺ binding genes, we next performed a drug-mediated challenge of Ca²⁺ homeostasis and signaling, in the GIC lines. Cells were exposed to either to the target independent cation (Ca²⁺) ionophore A23187 or the SERCA pump inhibitor Thapsigargin (Fig. 3), which increase cytosolic Ca²⁺ levels by two different mechanisms: A23187 by allowing Ca²⁺ to cross the normally impermeable cell membrane, and Thapsigargin by blocking import of Ca²⁺ into the ER. The GIC lines showed differences in sensitivity for both A23187 and Thapsigargin (Fig. 3A and 3B, respectively), remarkably with a rank order between the lines identical to that of the NSC-rooted transcriptome rank order, with the NSC-proximal GliNS1 being more sensitive than G179NS, while the NSC-distal G166NS was least sensitive to both drugs. Functional analyses thus show that NSC-proximal GICs with a higher expression of Ca²⁺ provokers, are more sensitive to disturbances in cytosolic Ca²⁺ regulation than GICs with a NSC-distal phenotype that express higher levels of Ca²⁺ buffers.

As sensitivity to Ca²⁺ drugs was associated with a NSC-like expression profile the question whether differentiation of GICs would affect Ca²⁺ sensitivity was investigated. To this end, three GIC lines were subjected to a differentiation protocol using fetal bovine serum (FBS). Validation of differentiation was done by transcriptome analysis of the GIC lines and their differentiated progeny using RNA sequencing (S1 Table). Principal component analysis of the global data set showed that changes in the transcriptome were distinct and segregated significantly between undifferentiated GICs and differentiated GICs (diffGICs) (Fig. 4A). Interestingly, GRIA1 expression that correlated with Ca²⁺ drug sensitivity, decreased in all GIC lines during differentiation (Fig. 4B), which suggested that differentiation status might affect Ca²⁺ sensitivity. Functional Ca²⁺ sensitivity was therefore assayed using A23187 (10 uM, 48 hours) in differentiated GICs and compared to undifferentiated GICs revealing a clearly reduced effect on cell viability in all GIC lines upon differentiation and with the strongest effect in the drug sensitive NSC-proximal GIC line (GliNS1)(Fig. 4C). These findings further support the data that Ca²⁺ sensitivity is associated with immature NSC-like GICs.

To explore potential additional genes correlating with Ca²⁺ sensitivity, transcriptome data from nine novel GIC lines (derived in an independent laboratory) was compared to Ca²⁺ sensitivity data from exposure to Thapsigargin (72 h exposure)(Fig. 5A and S2 Table). 7 out of the 9 lines have been shown to recapitulate the parent tumor (S3 Table). Analysis of correlation (Pearson) between NSC-markers and sensitivity to Thapsigargin (1 uM) revealed a significant correlation for nestin (NES) (Fig. 5B, outlier marked in red was excluded from the analysis) and brain lipid-bindig protein (BLBP/FABP7) (Fig. 5C) mRNA expression, while no correlation was found for SOX2 (data not shown). Western blot analysis further verified that calcium drug sensitive lines (U3017-MG and U3084-MG) expressed more BLBP protein than less sensitive lines (U3013-MG and U3035-MG) (Fig. 5D). The correlation analysis also confirmed a correlation between sensitivity to Thapsigargin and GRIA1 expression (Fig. 5E), which was corroborated by analysis of protein levels by western blot, as GRIA1 protein expression was only detected in the sensitive GICs (Fig. 5F). Further gene enrichment and gene ontology analyses implied genes involved in cell cycle regulation, oxygen, RNA and macromolecule metabolism, and not unexpectedly Ca²⁺-mediated signaling as correlating with Ca²⁺ drug sensitivity (Fig. 6A).

To identify genes in this data set that also associated with a NSC-proximal stemness signature in GICs, the set was further filtered for genes, which also (1) had a higher expression in GliNS1 compared to G166NS and (2) were downregulated upon differentiation (Fig. 6B). This retrieved a short-list of nine genes, two of which code for ion channels that may increase cytosolic Ca²⁺ (Ca²⁺ provokers), i.e. GRIA1 and the inward rectifier K⁺ channel KCNJ4, which may participate in maintaining a depolarized membrane potential required to activate voltage-gated Ca²⁺ channels and Ca²⁺ permeable glutamate receptors. In summary, the correlation between functional Ca²⁺ drug sensitivity and gene expression suggests participation towards sensitivity to drug-elicited Ca²⁺ overload, by a network of genes involved in maintaining Ca²⁺ homeostasis and membrane potential.

To identify intracellular responses to Ca²⁺ underlying the differential level of Ca²⁺ sensitivity in GICs, the NSC-proximal GliNS1 and NSC-distal G166NS were exposed to A23187 for 7 hours, followed by transcriptome analysis by RNA sequencing (“drug reactome mapping”)(S4 Table). In the most Ca²⁺ drug sensitive GIC line GliNS1, genes with significantly altered expression (Benjamini adjusted p-value <0.05) were analyzed by gene enrichment and gene ontology, which showed that cell cycle related genes were altered (Fig. 7B and S2 Fig.), suggesting cell cycle arrest prior to cell death. Not unexpectedly, genes involved in ER stress response were also enriched, as were genes in RNA metabolic processes. Interestingly, RNA metabolic process involved genes were also correlating with Thapsigargin sensitivity in the previous experiment (Fig. 6A).

Genes with altered expression after drug exposure were plotted against mean expression value (before drug exposure) to identify robustly altered genes with a potential biological significance. Strikingly, the GliNS1 line induced a clearly higher global transcriptome fold change than the less sensitive G166NS (Fig. 7A and 7C) suggesting a more potent onset of Ca²⁺ signaling in sensitive GICs. This may be the consequence by an inability to effectively reduce cytosolic Ca²⁺ levels. Interestingly, a very similar set of genes were altered in both the NSC-proximal (GliNS1) and the NSC-distal (G166NS) GIC lines, including Ca²⁺-binding genes acting as buffers (e.g. HSPA5 and HSP90B1) and Ca²⁺ related ER stress response (e.g. HERPUD1 and CALR). Also Ca²⁺-activated transcription factors (e.g NFATC2) were induced in both lines (Fig. 7A and 7C), suggesting that increased cytosolic Ca²⁺ could trigger a positive feedback mechanism. Interestingly, the G166NS line uniquely upregulated a set of anti-apoptotic genes, such as TP53 (coding for p53) and BCL2 genes (Fig. 7D), providing an explanation for lower sensitivity to e.g. Thapsigargin in addition to lower expression of Ca²⁺ permeable ion channels and higher expression of Ca²⁺ buffers.